paper_id
stringlengths 9
16
| version
stringclasses 26
values | yymm
stringclasses 311
values | created
timestamp[s] | title
stringlengths 6
335
| secondary_subfield
sequencelengths 1
8
| abstract
stringlengths 25
3.93k
| primary_subfield
stringclasses 124
values | field
stringclasses 20
values | fulltext
stringlengths 0
2.84M
|
|---|---|---|---|---|---|---|---|---|---|
1905.07153 | 1 | 1905 | 2019-05-17T07:56:21 | Embodied robots driven by self-organized environmental feedback | [
"q-bio.NC",
"nlin.AO"
] | Which kind of complex behavior may arise from self-organizing principles? We investigate this question for the case of snake-like robots composed of passively coupled segments, with every segment containing two wheels actuated separately by a single neuron. The robot is self organized both on the level of the individual wheels and with respect to inter-wheel coordination, which arises exclusively from the mechanical coupling of the individual wheels and segments. For the individual wheel, the generating principle proposed results in locomotive states that correspond to self-organized limit cycles of the sensorimotor loop.
Our robot interacts with the environment by monitoring the state of its actuators, that is via propriosensation. External sensors are absent. In a structured environment the robot shows complex emergent behavior that includes pushing movable blocks around, reversing direction when hitting a wall and turning when climbing a slope. On flat grounds the robot wiggles in a snake-like manner, when moving at higher velocities. We also investigate the emergence of motor primitives, viz the route to locomotion, which is characterized by a series of local and global bifurcations in terms of dynamical system theory. | q-bio.NC | q-bio | Embodied robots driven by self-organized
environmental feedback
Frederike Kubandt1, Michael Nowak1, Tim Koglin1,
Claudius Gros1, Bulcs´u S´andor2,1
1Institute for Theoretical Physics, Goethe University Frankfurt, Germany
2Department of Physics, Babes-Bolyai University, Cluj-Napoca, Romania
{kubandt,gros07}@itp.uni-frankfurt.de
http://itp.uni-frankfurt.de/~gros
Abstract. Which kind of complex behavior may arise from self-organizing
principles? We investigate this question for the case of snake-like robots
composed of passively coupled segments, with every segment contain-
ing two wheels actuated separately by a single neuron. The robot is self
organized both on the level of the individual wheels and with respect
to inter-wheel coordination, which arises exclusively from the mechani-
cal coupling of the individual wheels and segments. For the individual
wheel, the generating principle proposed results in locomotive states that
correspond to self-organized limit cycles of the sensorimotor loop.
Our robot interacts with the environment by monitoring the state of its
actuators, that is via propriosensation. External sensors are absent. In
a structured environment the robot shows complex emergent behavior
that includes pushing movable blocks around, reversing direction when
hitting a wall and turning when climbing a slope. On flat grounds the
robot wiggles in a snake-like manner, when moving at higher velocities.
We also investigate the emergence of motor primitives, viz the route
to locomotion, which is characterized by a series of local and global
bifurcations in terms of dynamical system theory.
1
Introduction
Wheeled snake-like robots [1] are a class of hypermobile robots [2] that are able to
navigate flexibly through rough terrains and restricted geometries. Movements
may be generated either via central pattern generators [3], or via top-down
commands [4], with the latter being a challenging task when a large number
of actuators is involved. An alternative is autonomous decentralized control,
which has been studied for the case of serpentine robots in terms of a chain of
locally coupled oscillators [5,6], and neurally-inspired generating schemes able of
sensorless pathfinding [7].
A key rationale for developing biologically inspired robots is the drive for
robust and highly adaptive designs [8,9]. Similarly, this is most of the time also
the motive for studying how adaptive locomotion can be realized [10], f.i. with
soft robots [11]. Abstracting from the direct engineering benefit, it is of particular
9
1
0
2
y
a
M
7
1
]
.
C
N
o
i
b
-
q
[
1
v
3
5
1
7
0
.
5
0
9
1
:
v
i
X
r
a
2
Train of cars
Fig. 1. Illustration of possible control schemes. Sensory information is processed, e.g.
by a neural network, and motor commands sent to the actuators. The actuators may
respond either rigidly (top), or elastically, viz compliant (bottom). Compliant actuators
may be realized, as illustrated here, via a direct feedback loop involving the state of the
actuator (propriosensation). In the limiting case of an embodied actuator, as considered
in this study, locomotion occurs also in the absence of a modulatory top-down signal.
interest to study generating mechanisms of locomotion in general, an approach
taken here. Our focus is on compliant locomotion generated by self-organizing
dynamical systems, which may take the form of either limit-cycle [12] or playful
behavior [13].
Compliance, which denotes the ability of a robot to react elastically to envi-
ronmental feedback [14], may be achieved in several distinct ways, which include
from the engineering perspective suitably designed actuators [15] and control
algorithms [16]. Compliant behavior can emerge on the other side also through
the reciprocal dynamical coupling of control, body and environment [17], the
sensorimotor loop. A particularly interesting limit is here, from the perspec-
tive of complex system theory, the limit of a fully reactive and hence embodied
controller. In this limit, the controller is inactive in absence of environmental
feedback, with the consequence that the sensorimotor feedback has not only a
modulating effect on locomotion, becoming instead essential. Locomotion then
arises via limit cycles and chaotic attractors that emerge within the sensorimotor
loop, the telltale sign of self-organized locomotion. It is hence important to ask,
as we will do in this study, how locomotion is generated in terms of a dynamical
systems bifurcation diagram.
Decomposing complex behavior into a series (or into a superposition) of basic
reusable building blocks, the motor primitives [18], is a well studied approach for
reducing the control problem of complex robots. Movement primitives may be
modeled by nonlinear dynamical systems [19] using, e.g., Gaussian mixture mod-
els [20], where the parameters of the dynamical system are either uniquely de-
controllersensorsactuatorlocomotionenvironmentembodied locomotionproprioceptionsensoryinputmotorcommandsmotorcommandscompliant actuator Self-organized embodied robots
3
fined or drawn from a suitable distribution [21,22]. Motor primitives can emerge
also from embodied dynamics in terms of chaotic itinerancy [23], or, alterna-
tively, as self-organized attracting states in the sensorimotor loop [24], that is
within the state space comprising the controller, the body of the robot and the
environment. Here we propose a new type of self-organized controller for wheeled
robots that leads to multiple fixpoint and limit-cycle attractor states and hence
to self-organized motor primitives in the sensorimotor loop. With the behavior of
the robot being self organized on the level of the individual wheels and with re-
spect to inter-wheel coordination, the resulting dynamics reflects its affordances
[25] when placed in simple but structured environments.
1.1 Control frameworks and the sensorimotor loop
Several in part non-exclusive routes for the generation of locomotion in robots
and animats do exist in generic terms. Standard top-down control, as illustrated
in Fig. 1, consists of a central processor generating motor commands either reac-
tively, in response to sensory inputs, or deliberately on its own [26]. The actuator
may be in turn stiff, as for industrial robots, or compliant, as for the muscles and
tendons of animals, reacting either passively or actively to external forces [15].
For the latter case, as sketched in Fig. 1, the actuator changes its stiffness upon
sensing its own state. Compliance arises then in response to propriosensation.
We are interested in locomotion that arises through the interaction of the
degrees of freedom of the robot, including both internal variables and the body,
with environmental feedback. The combined variables of the resulting sensori-
motor loop constitute then the phase space for dynamical attracting states, fixed
points, limit cycles and chaotic attractors, that correspond to self-organized be-
havioral primitives. The locomotion generated in this way is highly compliant in
the sense that the attracting states in the sensorimotor loop respond elastically
to additional top-down commands changing internal parameters.
2 Locomotive principles
Studies of real-world and simulated robots may focus either on performance,
and its improvement, or on the generative capabilities of locomotive principles.
The latter approach is gaining in importance in view of a recent study of the
neural coding of leg dynamics in flies, which showed that the dynamics of the
leg becomes dysfunction once the feedback loop between leg proprioception and
motor commands is cut [27]. These findings imply that self-organization plays
a commanding role in fly locomotion. Distributed computations has been found
to be of relevance for the nematode C. elegans [28]. Here we concentrate on
generative principles that are time reversal symmetric in the sense that a given
set of internal parameters allows the robot to move both forwards and backwards.
The direction selected by the robot then depends on the initial state, like a small
positive initial velocity or force.
4
Train of cars
Fig. 2. Illustration of a one-neuron controller simulating the transmission of classical
steam engines. The actual position x(a) = cos(ϕ) of the wheel drives, as described by
(1), the neural activity y(x) setting the target position x(t) = y. A simulated spring
with spring constant k between x(a) and x(t) generates subsequently the torque RFtan
acting on the wheel. Here Ftan = Fk sin(ϕ) denotes the tangential projection of the
spring force Fk = k(x(t) − x(a)).
2.1 Locomotion via time reversal symmetry breaking
Locomotion is parametrized typically by a velocity vector v = dr/dt that in-
corporates both the direction and the magnitude of the movement. Reversing
time t ↔ (−t) reverses then also the velocity vector. Here we are interested in
self-organized robots that break time reversal symmetry spontaneously, which
in our case implies that the attracting states in the sensorimotor loop come in
pairs that are related via time reversal symmetry. Whether the robot moves for-
or backwards depends then only on the initial conditions. For this purpose we
use the one-neuron controller illustrated in Fig. 2.
A wheel with a rotational angle ϕ is regulated individually via
τ x = cos ϕ − x,
y = tanh(ax) ,
(1)
where x is the membrane potential of the controlling neuron, y = tanh(ax)
the neural activity and τ the membrane time constant. The motor command is
proportional to the spring force
x(t) = y,
x(a) = cos(ϕ) ,
(2)
Fk = k(cid:0)x(t) − x(a)(cid:1),
where k is a spring constant and x(a) and x(t) respectively the actual and the
target position of the wheel in terms of a projection to the ground [29]. Note that
the angle ϕ, which enters the right-hand side of Eqs. (1) and (2) as cos(ϕ), is the
measured, the actual angle of the wheel. All forces, gravitational and mechanical,
impact the controller hence exclusively via their influence on the angle ϕ.
The controller simulates the transmission rod of a classical steam engine,
as sketched in Fig. 2, as it translates the bounded forth and back motion of
the neural activity y(t) into a rotational motion. Alternatively, instead of using
the angle ϕ as the determining variable, one could postulate a discrete map
ω(t) = tanh(aω(a)) between the actual and a target angular velocity [30], ω(a)
Self-organized embodied robots
5
and ω(a). This is not a problem for simulated robots, for which ω is a directly
accessible variable. To obtain a reliable estimate of the instantaneous angular
velocity for real-world robots working with duty cycles of the order of 20 Hz
would however be a challenge [29].
A controller enabling locomotive limit cycles to emerge in the sensorimotor
loop [12], as described here by (1) and (2), differs qualitatively from controlling
schemes employing local phase oscillators [31], for which a spontaneous reversal
of the direction of motion would not be possible.
2.2 Isolated wheel
The individual wheels of the simulated robots are controlled exclusively by (1)
and (2). There is no explicit inter-wheel coupling present. It is illustrative to
model, for comparison, an idealized isolated wheel with moment of inertia I, ra-
dius R, angle ϕ and angular velocity ω. The force Fk generated by the simulated
transmission rod then enters the equations of motion as a torque RFk sin(ϕ),
τ x = cos ϕ − x,
ϕ = ω,
I ω = R(Fk sin ϕ − f ω) ,
(3)
where f > 0 is a friction coefficient. Eq. (3) is manifestly invariant under ω ↔
(−ω), ϕ ↔ (−ϕ) and x ↔ x, which implies time-reversal symmetry in terms of an
invariance with respect to reversing the direction of motion. We will investigate
(3) further in Sect. 3.5, noting here that symmetry breaking may occur also in
embodied robots that incorporate forward world models [32].
3 Results
We used the LPZRobots physics simulation package [33] for the simulation of
robots composed of chains of 1-5 two-wheeled cars linked passively through hinge
joints, which are equipped with passively damped torsion springs. In the absence
of motor commands or external forces the equilibrium position of the hinge joints
induces a straight alignment of the connected body segments. During locomotion
the joints can store, on the other hand, potential energy when bent.
Shown in Fig. 3 is the trajectory of the train of cars climbing up a slope that is
intersected orthogonally by two other slopes. One observes wiggling and straight
locomotion together with direction reversal and large turns. In order to develop
an understanding we start by investigating the velocity profile of a robot on
an extended slope, concentrating on the dependence of the self-regulated steady
state velocity on the spring constant k of the actuator and on the inclination of
the slope. We note that the simulation cycle times of the LPZRobots simulation
package, which is based on the Open Dynamics Engine [34], are of the order of
50 ms.
3.1 Moving up and down an infinite slope
In Fig. 4 we present the velocity profile for a 5-segmented robot moving on
a slope parallel to the gradient, that is straight up and down. The downward
6
Train of cars
Fig. 3. Screenshots of the LPZRobots simulation environment. Left: A snake-like train
of cars composed of five passively coupled segments. Each segment contains two inde-
pendent wheels that influence each other exclusively through the mechanics of the body
and via the hinge joints connecting the individual segments. Right: A robot climbing
intersecting slopes on its own, with the silver line illustrating the ground trace of the
last segment. No explicit control signal has been given. The wiggling observed when the
robot moves fast on straight stretches disappears at lower velocities, as it is the case
when moving steeper up the slope. The train of cars reverses direction autonomously
when hitting the intersecting slope. The wiggling amplitude on the last leg increases
progressively while moving down, leading in the end to an upward curve (click for
movie).
velocity decreases in magnitude with decreasing slope and spring constant k, as
expected.
For the robot moving on a horizontal plane there exists a critical kc ≈ 0.54,
such that the limit-cycles corresponding to regular forward or backward move-
ment disappear for k < kc. For a spring constant of k = 0.2, which is below kc,
the robot moves therefore only when the slope has a finite downward inclination,
as shown in Fig. 4.
The torque exerted on the wheels is directly proportional to the spring con-
stant, being generated otherwise through the sensorimotor feedback. Moving
upward the slope the torque RFtan, and hence also the sensorimotor feedback,
needs to counter the gravitational downhill force FG. For an engine producing
a constant torque, the balancing of the tangential and the gravitational force
would lead to an uphill velocity vslope ∝ Ftan − FG that vanishes linearly and
hence continuously at a critical inclination. This is however not the case when
the motion is generated through sensorimotor feedback, as evident from the data
presented in Fig. 4. The sensorimotor feedback involves a self-consistency con-
dition that breaks down discontinuously, at finite values of the uphill velocity,
when the inclination of the slope becomes too large. It remains however the case
that larger spring constants k allow the robot to move up steeper slopes.
Self-organized embodied robots
7
Fig. 4. The velocity profile of a 5-segmented robot on a slope. The parameters are
a = 1 and k = 2.0/1.2/0.2 (yellow/red/green). Shown are the steady-state velocities
for moving directly upwards (positive vslope) and for moving down the slope (negative
vslope). A first order transition occurs for k = 2.0 (yellow dots) and k = 1.2 (red dots)
when increasing the inclination slowly, that is adiabatically, as indicated by the arrows.
The robot cannot move upwards for k = 0.2 (green dots), being subcritical.
3.2 Autonomous direction reversal
The trajectory of the robot presented in Fig. 3 hits twice an intersecting slope. A
robot equipped with actuators producing constant torques would move segment
by segment onto the intersecting slope, up to the point where the gravitational
downward pull of the increasing number of segments cancels with the locomotive
force. At this point the robot would remain in place.
The locomotive force of the snake-like robot presented here is however highly
compliant, being generated within the sensorimotor loop. The velocity profile
presented in Fig. 4 implies that an equilibrium position resulting from the bal-
ance of an upward locomotive force and the downhill gravitational pull is not
possible, namely that the limit-cycle attractors present in the sensorimotor loop
allow the robot to move only up- or downhill. We note that the stable fixpoint
attractors corresponding to a non-moving state, that exist in conjunction with
the limit-cycle locomotion for an isolated wheel, as discussed in Sect. 3.5, pos-
sesses only a vanishing small basin of attraction for the case of a train of cars. A
robot hitting a slope that is too steep is therefore likely to reverse direction, as
observed in Fig. 3, instead of being pulled into a fixpoint attractor and coming
to a stop.
3.3 Straight, meandering and chaotic modes
The last leg of the trajectory shown in Fig. 3 shows growing left- and rightward
swings. This is a typical behavior at larger velocities, here due to the slight
01234567%incline−2−101vslopek=2.00k=1.20k=0.20 8
Train of cars
Fig. 5. As a function of time, the measured angular velocity ω of one of the two wheels
of the first and of the last car. The membrane time constant is τ = 50 ms. Top: For
a = 0.5 and k = 0.5, a straight-moving mode Left and right wheels of a given car
are exactly synchronized. The small difference in amplitude in the ω(t) oscillations
between cars is due to small oscillations in the respective pitch angles. Note that ω(t)
is not exactly constant as resulting from (1). Mid: For a = 1.0 and k = 1.0, a regular
meandering mode Bottom: For a = 1.0 and k = 5.0, a chaotic mode resulting from
wide sideway swings of the tail (click for movies).
downhill direction, that results from a transversal mechanical instability of the
connected segments. The joints are elastic and therefore capable to store a certain
amount of energy, akin to what happens when a string starts to vibrate. The
final upturn of the robot may occur at an angle, interestingly, that puts the
robot below criticality. The robot will then stop moving and reverses direction.
In Fig. 5 we present the angular frequency ω(t) for one of the two wheels
of the first and the last car, respectively, of a five-segmented snake-like robot
moving on a flat ground. Shown are the timelines for two regular and for a
chaotic mode. In the first limit-cycle mode the train of cars moves straight. The
ten wheels are in this case synchronized in the sense that the small modulation
of the respective angular velocities, which occur because the controller (1) is
not rotationally invariant, appear all at exactly the same time. Their respective
28.028.228.4203040ω[Hz]firstlast50.050.250.450.650.851.0t[s]5075 Self-organized embodied robots
9
Fig. 6. A five-segmented robot bouncing off the wall, with the silver line tracing the
position of the darker segment. The only knowledge the robot disposes of the outside
world, and hence of the presence of a wall, is via propriosensation, that is via the
measurement of the angle of the wheels. The slight wiggling of the forward motion
causes the robot to be reflected at various angles, even though it bumps into the wall
perpendicularly (with respect to the average direction of locomotion). The direction
reversal results from the destruction of the forward limit cycle upon hitting the wall,
with the flow in phase space evolving subsequently towards the backward attractor.
amplitudes vary however from car to car, which implies that the pitch angles of
the individual cars oscillates, even though only slightly.
For larger average velocities, the limit-cycle straight mode tends to be become
unstable, making way to meandering and chaotic modes, as illustrated in Fig. 5.
We did not investigate the exact nature of the respective transition to chaos,
which may be due to a cascade of period-doubling within the space of meandering
modes [35]. We also note that our classification of the highly irregular mode as
chaotic relies here only on a visual inspection, as we did not apply a formal test
for the presence of deterministic chaos [36].
3.4 Interacting with a structured environment
The robots exhibit interesting behavioral patterns when situated in a struc-
tured environment. As a first example we show in Fig. 6 the interaction of a
5-segmented robot with a wall. Before hitting the wall the robot possesses, due
to time-reversal symmetry, both a forward- and a backward-moving limit cy-
cle. Approaching the wall the sensorimotor state of the robot is in the forward
limit cycle, which becomes however destroyed upon hitting the wall. The flow
in phase space, that is the evolution of the membrane potentials x of the indi-
vidual wheels, is then attracted by the remaining limit cycle, which is the one
corresponding to moving backward. The robot hence reverses direction.
The observed direction reversal occurs autonomously in the absence of top-
down control signals. As the robot possesses only sensors measuring the angles of
the wheels there are furthermore no external sensors present that would inform
the robot about the distance to the obstacle. One observes, as shown in Fig. 6,
that the angle at which the robot bounces varies considerably as a consequence
of the wiggling of the initial forward motion. In a slow-velocity non-wiggling
mode the bouncing occurs exactly perpendicularly.
10
Train of cars
Fig. 7. Superimposed screenshots from the LPZRobots simulation environment, with
a = 1 and k = 2. The silver line represents the ground trace of the darker of the two
end segments. The snake-like train of cars starts from the lower center (shadowed),
where it first hits a movable box (shadowed), bending and pushing the box around in
an inward spiral (as indicated by the shadowed box and robot in the middle). As the
angle of the spiral becomes steeper and the forward velocity smaller, the robot will
reach the point at which the forward limit cycle disappears, as seen also in Fig. 4.
The robot then stops for a short period during which the dynamics flows in phase
space autonomously towards the attractor corresponding to backward motion. Once
reached, the robot reverses direction, as one can see from the positioning of the dark
end segment (click for movie).
We also studied the interaction of the multi-segmented robot with movable
boxes, as illustrated in Fig. 7. Due to the mechanical feedback of the passive
but elastic hinge joints, the body of the train of cars bends, such that the robot
continues to push the box in ever smaller circles. During the push, the robot slows
down continuously, until a critical velocity is reached and the forward limit cycle
disappears. At this point the robot stops, reversing direction autonomously.
The robot disposes of a single controlling neuron per actuator, which ob-
tains in turn information about the external world only indirectly, namely via
the measured angle of the respective wheel. The behavior illustrated in Fig. 7,
such as pushing around boxes, is hence emergent and an example that embodied
controlling frameworks may lead robustly to novel behavioral patterns. Func-
Self-organized embodied robots
11
tionally, the ability of the snake-like robot to follow spiral-shaped trajectories,
when pushing a box, is a direct consequence of the highly elastic working regime
of the actuators, viz of compliance. By themselves, the wheels on the left-hand
and the right-hand side of the body would acquire, being subject to identical
controllers, identical angular velocities. The fact, however, that the turning speed
of the wheels is not determined by a top-down signal, but by the sensorimotor
feedback, allows the emerging limit cycles to adapt autonomously to environ-
mental forces. The wheels on opposite sides of the body react as a consequence
distinctly when the box exerts a non-symmetric resistance onto the robot. The
respective angular velocities then differ.
3.5 Theory for an isolated single wheel
The fixpoints of an isolated and non-moving single wheel, as described by Eq. (3),
are determined by ω = 0, x = cos ϕ and
sin ϕ = 0
or
y(x) = cos ϕ ,
compare Eqs. (3) and (2). The Jacobian J(ϕ) is in general
J(ϕ) =
−1/τ
0
ka(1 − y2) sin ϕ
− sin ϕ/τ 0
1
−f
0
A
,
(4)
(5)
where A = ky cos ϕ+k(sin2 ϕ−cos2 ϕ). We have set R = 1, measuring in addition
k and f relative to I.
-- For the first fixpoint sin ϕ = 0, that is for (x, ω, ϕ) = (1, 0, 0) and (−1, 0, π),
one has A(0, π) = k(±y(±1)− 1) = k(tanh(a)− 1), which is always negative.
The eigenvalues of the Jacobian are then
λ0(0, π) = −
1
τ
,
λ±(0, π) = −
f
2 ±
1
2
f 2 + A(0, π) ,
(6)
Since A(0, π) < 0, the fixpoints at ϕ = 0 and ϕ = π are always stable.
-- For the second set of fixpoints one has to solve the self-consistency condition
y(x) = tanh(ax) = x, with x = cos ϕ. The trivial solution x = 0, that is
(x, ω, ϕ) = (0, 0,±π/2) splits at ac = 1 via a pitchfork transition, allowing
for three coexisting fixpoints for a > ac. For small k the trivial fixpoint
x = 0 = cos ϕ = y with the Jacobian
(cid:112)
−1/τ ∓1/τ 0
0
±ka
0
1
k −f
J(ϕ = ±π/2) =
(7)
is a saddle for 0 < a ≤ 1, having two negative and one positive eigenvalues,
being stable however for a > 1. For larger values of the spring constant k,
the x = 0 solution undergoes a Hopf bifurcation leading to limit-cycle oscil-
lations.
12
Train of cars
Fig. 8. One-step heteroclinic route to locomotion for a < 1. Shown are stable limit
cycles (red) and stable, unstable manifolds (black, light/dark blue/green) and selected
sample trajectories (violet). The fixpoints at ϕ = 0, π are stable nodes/foci respectively
for small and larger spring constants k (top and upper middle panel), with the saddles
at ϕ = ±π/2 remaining unchanged in character for all k. A symmetric heteroclinic
connection between the saddles is created when increasing k further. For k ≈ 19.66
(lower bottom panel) a stable limit cycle (red) corresponding to limit-cycle locomotion
(bottom panel) is generated. The parameters are a = 0.95, τ = 0.2, I = 0.25 and
f = 0.5.
3.6 Routes to locomotion
It is interesting to study how limit-cycle locomotion arises from a configuration
of individual fixpoints upon increasing the force acting on the wheel, that is the
spring constant k.
In Fig. 8 we illustrate the case a < 1, for which the saddles at ϕ = ±π/2
do not undergo a pitchfork bifurcation yet. Locomotion arises in this case via
a one-step heteroclinic transition which allows for the generation of limit cycles
of finite amplitudes. A pair of stable and unstable limit cycles is eventually
produced when increasing the spring constant k, with the stable limit cycle
corresponding to a locomotive behavioral primitive. Note that the phase space
of (3) is three dimensional and that the flow shown in Fig. 8 corresponds to a
projection onto the (ϕ, ω)-plane. Trajectories may hence intersect.
−π−π/20π/2π-0.30.00.3−π−π/20π/2π-3.00.03.0−π−π/20π/2π-6.50.06.5−π−π/20π/2π-11.6-0.410.8k=0.80k=9.00k≈19.66k=21.00ω[Hz]ϕω[Hz]ϕω[Hz]ϕω[Hz]ϕ Self-organized embodied robots
13
Fig. 9. Multi-step heteroclinic route to locomotion for a > 1. The fixpoints at ϕ = 0, π
are always stable nodes, with the foci at ϕ = ±π/2 undergoing a Hopf bifurcation
upon increasing the spring constant k (top/upper middle panel). The resulting limit
cycle (red) corresponds to a periodic forth-and-back motion characterized by ω (cid:54)= 0
and a vanishing average ω = 0. The forth-and-back motion is destroyed by a sym-
metric heteroclinic transition (upper/lower middle panel), leading to an intermedi-
ate phase without locomotion. A second heteroclinic transition then generates stable
limit-cycle locomotion (lower middle/bottom panel). Parameters, besides a = 1.5, and
color-coding as for Fig. 8.
In Fig. 9 a multi-step route to locomotion for a > 1 is presented. One ob-
serves first a Hopf-bifurcation (HB) at ϕ = ±π/2, which leads to a first in-
termediate phase characterized by a closed limit-cycle in the (ϕ, ω)-plane. This
limit cycle, corresponding to small amplitude forth-and-back periodic motion,
is destroyed when hitting in a symmetric heteroclinic transition (SHE) the two
saddles present additionally for a > ac = 1. Motion ceases in the subsequent
second intermediate phase, for which the unstable trajectories emerging from
ϕ = ±π/2 lead to the fixpoints ϕ = 0, π. A stable limit-cycle emerges however
from a second heteroclinic transition (HE) when increasing the spring constant
k further, namely when the unstable manifold of one of the additional saddles
hits another saddle.
−π−π/20π/2π-0.80.00.8−π−π/20π/2π-1.70.01.7−π−π/20π/2π-2.90.02.9−π−π/20π/2π-3.90.03.9k=1.50k=2.40k=4.20k=5.60ω[Hz]ϕω[Hz]ϕω[Hz]ϕω[Hz]ϕ 14
Train of cars
For the parameters used for Fig. 9 there are hence for a > 1 two phases
without locomotion, viz for which the angular frequency ω decays to zero. The
average angular frequency ω vanishing for forth-and-back motion, but not for
limit-cycle locomotion:
ω → 0
HB
−−→ ω (cid:54)= 0, ω = 0
SHE
−−−→ ω → 0
HE
−−→ ω (cid:54)= 0, ω (cid:54)= 0
With the motor command being proportional to the spring constant k, it is some-
what intuitive that one needs a critical k for locomotion to emerge. Relatively
large spring constants have been used in Fig. 8 for illustrative purposes.
4 Conclusions
The gait of an animal corresponds to a coordinated pattern of limb movements
that repeats with a certain frequency. Gaits are generated typically by a cen-
tral pattern generator [3], that is by a central processing unit that produces
coordinated motor signals. We have examined here an alternative framework for
which the actuators of an animat are self active, with the dynamics of the indi-
vidual actuators resulting from the presence of limit-cycle attractors within the
sensorimotor loop. The actuators, in our case the wheels of a snake-like robot,
are coupled in this framework only via the mechanics of the body and by the
reaction of the environment. The gaits of the animat result in our framework
therefore from self-organizing principles. For a simulated wheeled snake-like an-
imat, we find that the robot interacts autonomously with the environment, e.g.
by turning on its own on a slope. The robot will also push a movable box around
for a while when colliding with one.
We have tested in addition that locomotion modes also arise for heteroge-
neous (not identical) body segments, e.g. when the controllers have different
spring constants ki, or different wheel sizes. The multi-segmented robot is capa-
ble of generating locomotion in particular when several actuators are subcritical,
i.e. with wheels which on their own would not maintain oscillatory dynamics.
Embodiment leads in our study robustly to emergent locomotion.
The here employed dynamical-system type approach to robotic locomotion
allows in addition to characterize the motion primitives in terms of self-organized
attractors formed in the extended phase space of the robot and environment. In-
corporating objects of the environment into the overarching dynamical system
allows in consequence dynamical system approaches also to classify computa-
tional models of affordance [37]. In this sense self-organizing attractors play an
important role in the generation of useful behavior for the discovery of dynamic
object affordances [38].
Acknowledgments
The support of the German Science Foundation (DFG) is acknowledged.
Self-organized embodied robots
15
References
1. Motoyasu Tanaka and Kazuo Tanaka. Control of a snake robot for ascending and
descending steps. IEEE Transactions on Robotics, 31(2):511 -- 520, 2015.
2. Grzegorz Granosik. Hypermobile robots -- the survey. Journal of Intelligent &
Robotic Systems, 75(1):147 -- 169, 2014.
3. James K Hopkins, Brent W Spranklin, and Satyandra K Gupta. A survey of
snake-inspired robot designs. Bioinspiration & biomimetics, 4(2):021001, 2009.
4. Lars Pfotzer, Sebastian Klemm, Arne Ronnau, Johann Marius Zollner, and Rudiger
Dillmann. Autonomous navigation for reconfigurable snake-like robots in challeng-
ing, unknown environments. Robotics and Autonomous Systems, 89:123 -- 135, 2017.
5. Takahide Sato, Takeshi Kano, and Akio Ishiguro. On the applicability of the
decentralized control mechanism extracted from the true slime mold: a robotic
case study with a serpentine robot. Bioinspiration & biomimetics, 6(2):026006,
2011.
6. Takahide Sato, Takeshi Kano, and Akio Ishiguro. A decentralized control scheme
for an effective coordination of phasic and tonic control in a snake-like robot.
Bioinspiration & biomimetics, 7(1):016005, 2011.
7. Jordan H Boyle, Sam Johnson, and Abbas A Dehghani-Sanij. Adaptive undu-
latory locomotion of a c. elegans inspired robot. IEEE/ASME Transactions on
Mechatronics, 18(2):439 -- 448, 2013.
8. Shigeo Hirose and Makoto Mori. Biologically inspired snake-like robots. In Robotics
and Biomimetics, 2004. ROBIO 2004. IEEE International Conference on, pages
1 -- 7. IEEE, 2004.
9. Pal Liljeback, Kristin Ytterstad Pettersen, Øyvind Stavdahl, and Jan Tommy
Gravdahl. A review on modelling, implementation, and control of snake robots.
Robotics and Autonomous Systems, 60(1):29 -- 40, 2012.
10. Shinya Aoi, Poramate Manoonpong, Yuichi Ambe, Fumitoshi Matsuno, and Flo-
rentin Worgotter. Adaptive control strategies for interlimb coordination in legged
robots: a review. Frontiers in neurorobotics, 11:39, 2017.
11. M Calisti, G Picardi, and C Laschi. Fundamentals of soft robot locomotion. Journal
of The Royal Society Interface, 14(130):20170101, 2017.
12. Laura Martin, Bulcs´u S´andor, and Claudius Gros. Closed-loop robots driven by
short-term synaptic plasticity: Emergent explorative vs. limit-cycle locomotion.
Frontiers in Neurorobotics, 10:12, 2016.
13. Ralf Der, Frank Hesse, and Georg Martius. Rocking stamper and jumping snakes
from a dynamical systems approach to artificial life. Adaptive Behavior, 14(2):105 --
115, 2006.
14. Alexander Sprowitz, Alexandre Tuleu, Massimo Vespignani, Mostafa Ajallooeian,
Emilie Badri, and Auke Jan Ijspeert. Towards dynamic trot gait locomotion:
Design, control, and experiments with cheetah-cub, a compliant quadruped robot.
The International Journal of Robotics Research, 32(8):932 -- 950, 2013.
15. Ronald Van Ham, Thomas G Sugar, Bram Vanderborght, Kevin W Hollander,
IEEE Robotics & Automation
and Dirk Lefeber. Compliant actuator designs.
Magazine, 16(3), 2009.
16. Andrea Calanca, Riccardo Muradore, and Paolo Fiorini. A review of algorithms for
compliant control of stiff and fixed-compliance robots. IEEE/ASME Transactions
on Mechatronics, 21(2):613 -- 624, 2016.
17. Rolf Pfeifer, Max Lungarella, and Fumiya Iida. The challenges ahead for bio-
inspired'soft'robotics. Communications of the ACM, 55(11):76 -- 87, 2012.
16
Train of cars
18. Tamar Flash and Binyamin Hochner. Motor primitives in vertebrates and inverte-
brates. Current opinion in neurobiology, 15(6):660 -- 666, 2005.
19. Auke Jan Ijspeert, Jun Nakanishi, and Stefan Schaal. Movement imitation with
nonlinear dynamical systems in humanoid robots. In Robotics and Automation,
2002. Proceedings. ICRA'02. IEEE International Conference on, volume 2, pages
1398 -- 1403. IEEE, 2002.
20. S Mohammad Khansari-Zadeh and Aude Billard. Learning stable nonlinear dy-
namical systems with gaussian mixture models. IEEE Transactions on Robotics,
27(5):943 -- 957, 2011.
21. Alexandros Paraschos, Christian Daniel, Jan R Peters, and Gerhard Neumann.
Probabilistic movement primitives. In Advances in neural information processing
systems, pages 2616 -- 2624, 2013.
22. Heni Ben Amor, Gerhard Neumann, Sanket Kamthe, Oliver Kroemer, and Jan
Peters. Interaction primitives for human-robot cooperation tasks. In Robotics and
Automation (ICRA), 2014 IEEE International Conference on, pages 2831 -- 2837.
IEEE, 2014.
23. Jihoon Park, Hiroki Mori, Yuji Okuyama, and Minoru Asada. Chaotic itinerancy
within the coupled dynamics between a physical body and neural oscillator net-
works. PloS one, 12(8):e0182518, 2017.
24. Bulcs´u S´andor, Tim Jahn, Laura Martin, and Claudius Gros. The sensorimotor
loop as a dynamical system: How regular motion primitives may emerge from self-
organized limit cycles. Frontiers in Robotics and AI, 2:31, 2015.
25. Anthony Chemero and Michael T Turvey. Gibsonian Affordances for Roboticists.
Adaptive Behavior, 15(4):473 -- 480, 2007.
26. Danial Nakhaeinia, SH Tang, SB Mohd Noor, and O Motlagh. A review of control
architectures for autonomous navigation of mobile robots. International Journal
of Physical Sciences, 6(2):169 -- 174, 2011.
27. Akira Mamiya, Pralaksha Gurung, and John C Tuthill. Neural coding of leg pro-
prioception in drosophila. Neuron, 100(3):636 -- 650, 2018.
28. Harris S Kaplan, Annika LA Nichols, and Manuel Zimmer. Sensorimotor inte-
gration in caenorhabditis elegans: a reappraisal towards dynamic and distributed
computations. Philosophical Transactions of the Royal Society B: Biological Sci-
ences, 373(1758):20170371, 2018.
29. Bulcs´u S´andor, Michael Nowak, Tim Koglin, Laura Martin, and Claudius Gros.
Kick control: using the attracting states arising within the sensorimotor loop of
self-organized robots as motor primitives. Frontiers in neurorobotics, 12, 2018.
30. Ralf Der, Frank Guttler, and Nihat Ay. Predictive information and emergent
cooperativity in a chain of mobile robots. In ALIFE, pages 166 -- 172, 2008.
31. Yuichi Ambe, Shinya Aoi, Timo Nachstedt, Poramate Manoonpong, Florentin
Worgotter, and Fumitoshi Matsuno. Simple analytical model reveals the func-
tional role of embodied sensorimotor interaction in hexapod gaits. PloS one,
13(2):e0192469, 2018.
32. Ralf Der and Georg Martius. Behavior as broken symmetry in embodied self-
In Artificial Life Conference Proceedings 13, pages 601 -- 608.
organizing robots.
MIT Press, 2013.
33. Ralf Der and Georg Martius. The Playful Machine: Theoretical Foundation and
Practical Realization of Self-Organizing Robots, volume 15. Springer Science &
Business Media, 2012.
34. Russell Smith. Open dynamics engine, 2006.
35. Claudius Gros. Complex and adaptive dynamical systems: A primer. Springer,
2015.
Self-organized embodied robots
17
36. Hendrik Wernecke, Bulcs´u S´andor, and Claudius Gros. How to test for partially
predictable chaos. Scientific reports, 7(1):1087, 2017.
37. Philipp Zech, Simon Haller, Safoura Rezapour Lakani, Barry Ridge, Emre Ugur,
and Justus Piater. Computational models of affordance in robotics: a taxonomy
and systematic classification. Adaptive Behavior, 25(5):235 -- 271, 2017.
38. Ralf Der and Georg Martius. Self-organized behavior generation for musculoskele-
tal robots. Frontiers in Neurorobotics, 11:8, 2017.
|
1709.02443 | 1 | 1709 | 2017-09-07T20:35:38 | Fractional cable equation for general geometry, a model of axons with swellings and anomalous diffusion | [
"q-bio.NC"
] | Different experimental studies have reported anomalous diffusion in brain tissues and notably this anomalous diffusion is expressed through fractional derivatives. Axons are important to understand neurodegenerative diseases such as multiple sclerosis, Alzheimer's disease and Parkinson's disease. Indeed, abnormal accumulation of proteins and organelles in axons is a hallmark feature of these diseases. The diffusion in the axons can become to anomalous as a result from this abnormality. In this case the voltage propagation in axons is affected. Another hallmark feature of different neurodegenerative diseases is given by discrete swellings along the axon. In order to model the voltage propagation in axons with anomalous diffusion and swellings, in this paper we propose a fractional cable equation for general geometry. This generalized equation depends on fractional parameters and geometric quantities such as the curvature and torsion of the cable. For a cable with a constant radius we show that the voltage decreases when the fractional effect increases. In cables with swellings we find that when the fractional effect or the swelling radius increase, the voltage decreases. A similar behavior is obtained when the number of swellings and the fractional effect increase. Moreover, we find that when the radius swelling (or the number of swellings) and the fractional effect increase at the same time, the voltage dramatically decreases. | q-bio.NC | q-bio |
Fractional cable equation for general geometry, a model of
axons with swellings and anomalous diffusion
Erick J. L´opez-S´anchez (1)∗, Juan M. Romero (2) †, Huitzilin Y´epez-Mart´ınez(3) ‡
(1)Posgrado en Ciencias Naturales e Ingenier´ıa,
Universidad Aut´onoma Metropolitana, Cuajimalpa.
Vasco de Quiroga 4871, Santa Fe Cuajimalpa, D. F. 05300 M´exico.
(2)Departamento de Matem´aticas Aplicadas y Sistemas,
Universidad Aut´onoma Metropolitana-Cuajimalpa,
M´exico, D.F 05300, M´exico
(3)Universidad Aut´onoma de la Ciudad de M´exico,
Prolongaci´on San Isidro 151, San Lorenzo Tezonco, Iztapalapa,
Ciudad de M´exico 09790, M´exico.
Abstract
Different experimental studies have reported anomalous diffusion in brain tissues and
notably this anomalous diffusion is expressed through fractional derivatives. Axons are im-
portant to understand neurodegenerative diseases such as multiple sclerosis, Alzheimer's
disease and Parkinson's disease. Indeed, abnormal accumulation of proteins and organelles
in axons is a hallmark feature of these diseases. The diffusion in the axons can become
to anomalous as a result from this abnormality.
In this case the voltage propagation
in axons is affected. Another hallmark feature of different neurodegenerative diseases is
given by discrete swellings along the axon. In order to model the voltage propagation
in axons with anomalous diffusion and swellings, in this paper we propose a fractional
cable equation for general geometry. This generalized equation depends on fractional pa-
rameters and geometric quantities such as the curvature and torsion of the cable. For a
cable with a constant radius we show that the voltage decreases when the fractional effect
increases. In cables with swellings we find that when the fractional effect or the swelling
radius increase, the voltage decreases. A similar behavior is obtained when the number
of swellings and the fractional effect increase. Moreover, we find that when the radius
swelling (or the number of swellings) and the fractional effect increase at the same time,
the voltage dramatically decreases.
1
Introduction
In biological organisms there are hydrogen atoms in abundance, particularly in water and fat.
These atoms allow to study biological organisms with non invasive techniques, such as Magnetic
Resonance Imaging (MRI). Indeed, using a magnetic field and the Zeeman effect, the diffusion
process of water molecules in biological tissues can be mapped [1, 2]. Thus, using these tech-
niques it is possible to study some physiological properties of biological tissues, for example
water molecules diffusion patterns can reveal microscopic details about tissue architecture and
can reflect interactions with many obstacles, such as macromolecules, fibers, and membranes
∗[email protected]
†[email protected]
‡[email protected]
1
[1, 2, 3, 4]. Notably, in neuroscience water diffusion can provide information about white matter
integrity, fiber density, uniformity of nerve fiber direction, axonal membranes and cytoskele-
ton properties, etc [3, 4].
In fact, water diffusion has been used to detect and characterize
different neurodegenerative diseases [5]. For instance, using Diffusion Tensor Imaging (DTI),
altered diffusion has been detected in white matter of subjects with multiple sclerosis [6, 7, 8].
Also, using DTI, white matter alterations were found in the corpus callosum of subjects with
Huntinton's disease [9]. Moreover, using DTI, relevant white matter abnormalities were found
in Alzheimer's disease [10, 11].
In addition, using MRI, hippocampal atrophy has been de-
tected in Parkinson's disease [12]. Even more, different experimental studies have shown that
water diffusion in some tissues can not be described by a Gaussian model, but as an anomalous
diffusion expressed through fractional calculus [13, 14, 15, 16, 17, 18], in particular in brain
tumours [19, 20, 21] and dendrites [22]. Furthermore, some authors have suggested that the
anomalous diffusion parameters might play a role in the diagnosis of brain diseases [23].
It
is worth mentioning that in different neurodegenerative diseases are reported abnormal accu-
mulations of proteins and organelles which generate disruption axonal transport [24, 25, 26].
Additionally, different theoretical studies support the claim that anomalous diffusion appears
in a heterogeneous medium [27, 28, 29]. For these reasons, diffusion and anomalous diffusion in
brain tissues are relevant in the study of the brain physiology. Other studies about anomalous
diffusion in cellular system can be seen in [30, 31, 32, 33].
Another essential aspect to understand how the brain works is given by its electrical activity.
Thus, in order to obtain a reasonable model of the brain, it is important to know how the voltage
propagates in brain tissues with anomalous diffusion. In particular, due that axons are crucial
in neuron-to-neuron communications and those can be described as cables, we should know
how the voltage propagates in a cable with anomalous diffusion. In this respect, to study the
voltage V (x, t) in a straight cylindrical cable with a circular cross-section of constant diameter
d0 and anomalous diffusion, recently some authors have proposed a fractional cable equation
as follows [34, 35]
cM
∂V (x, t)
∂t
= βνDνt
∂2V (x, t)
∂x2 − iion
,
(1)
(cid:32) d0
4rL
(cid:33)
where cM denotes the specific membrane capacitance, rL denotes the longitudinal resistance
and iion is the ionic current per unit area into and out of the cable,
0 ≤ ν ≤ 1,
(2)
is the Riemann-Liouville fractional operator [36] and βν is a constant with (time)1−ν dimen-
sions. The passive cable case, namely when iion = V /rM (where rM is the specific membrane
resistance) was used to study electrodiffusion of ions in nerve cells [34, 35].
∂1−ν
∂t1−ν ,
Dνt =
ν = constant,
Additionally, there are different physiological phenomena where the geometry is important.
For instance, axons with non trivial geometry are important to understand some neurodegen-
erative diseases, indeed discrete swellings along the axons appear in neurodegenerative diseases
such as Alzheimer's disease, Parkinson's disease, HIV-associated dementia and multiple scle-
rosis. In fact, axons with a diameter of approximately 1µm with a swelling with diameter of
approximately 5µm, are reported in Parkinson's disease [37]. In addition, axons with a diame-
ter of approximately 4µm with a swelling with a diameter of approximately 60µm are reported
in multiple sclerosis [38]; axons with a diameter of approximately 1.5µm with a swelling train,
where the swelling diameter varies between 4µm and 10µm, are reported in Alzheimer's disease
[41, 39, 40]; axons with a diameter of approximately 6µm and swellings with a diameter of
approximately 43µm are reported in HIV-associated dementia [42, 43, 44, 45, 46]. Other sizes
of the axonal swellings can be see in [47, 48]. Some theoretical studies on cables with non
cylindrical geometry can be seen in [49, 50, 51, 52, 53, 54, 55, 56]. Because the equation (1)
only describes axons with cylindrical geometry, therefore, in order to study the voltage propa-
gation in a cable with non trivial geometry and anomalous diffusion, this equation should be
2
generalized.
In this paper, we study the voltage propagation in a cable with anomalous diffusion and
non trivial geometry. For this purpose, we introduce a fractional cable equation for a general
geometry. This generalized equation depends on fractional parameters βν and ν and geometric
quantities such as the curvature and torsion of the cable. For different cable geometries, we
show that with regard to model a system where the voltage decreases, we should suppose that
βν increases when ν decreases. For a straight cylinder with a constant radius we show that the
voltage depends on neither the curvature nor the torsion of the cable and it decreases when βν
increases and ν decreases. In addition, cables with swellings are studied. In these last cable
geometries we find that when the swelling radius increases or βν increases and ν decreases, the
voltage decreases dramatically.
This paper is organized as follows: in the section 2 we propose a fractional cable equation
with a general geometry; in the section 3 we analyse some general properties of the generalized
fractional cable equation; in the section 4 we consider the cylindrical cable with a constant ra-
dius; in the section 5 we study cable with swellings. Finally, in the section 6 a summary is given.
2 Fractional cable equation in a general geometry
It is well known that to study the geometric properties of a three dimensional curve (cid:126)γ the arc
length parameter
(cid:115)
(cid:90) x
s =
0
d(cid:126)γ(ζ)
dζ
· d(cid:126)γ(ζ)
dζ
dζ,
(3)
is a friendly parameter. Indeed, using the arc length parameter (3) we can construct the vectors
of the Frenet-Serret frame [57]
d(cid:126)γ(s)
ds
= T ,
N =
d T
ds(cid:12)(cid:12)(cid:12) d T
ds
(cid:12)(cid:12)(cid:12),
B = T × N ,
(4)
where T is the unit vector tangent, N is the normal unit vector and B is the binormal unit vector
to the curve. Furthermore, using the arc length and the Frenet-Serret frame, the Frenet-Serret
formulas can be obtained as follow [57]
d T
ds
= κ N ,
d N
ds
= −κ T + τ B,
d B
ds
= −τ N ,
(5)
where κ, τ are the curvature and torsion of the curve (cid:126)γ, respectively.
We can employ the Frenet-Serret frame to construct a cable model. Actually, we can propose
a general cable as the region bounded by the following surface
(cid:126)Σ(θ, s) = (cid:126)γ(s) + f1(θ, s) N (s) + f2(θ, s) B(s),
(6)
where θ is an angular variable. Notice that employing the angular coordinate θ, the functions
f1(θ, s), f2(θ, s) and the vectors N (s), B(s) we are constructing the cable over the curve (cid:126)γ(s).
In Fig. 1 we can see a representation of the surface (6). For instance, a cable with a deformed
circular cross-section, where the radius R depends on the angle θ, can be described by the
surface (6) where
f1(θ, s) = R(θ, s) cos θ,
f2(θ, s) = R(θ, s) sin θ.
Notice that in this case the cross-section area is given by
(cid:90) 2π
0
a(s) =
1
2
R2(θ, s)dθ.
3
(7)
(8)
Some geometric quantities as the area of a surface can be written in terms of the first
fundamental form, which is constructed with the inner product on the tangent space of a
surface as follows [57]
g =
where
(cid:32)
(cid:33)
,
E F
F G
(9)
(10)
(11)
(12)
,
,
.
E =
G =
F =
∂(cid:126)Σ(θ, s)
∂s
∂(cid:126)Σ(θ, s)
∂θ
∂(cid:126)Σ(θ, s)
∂s
∂s
· ∂(cid:126)Σ(θ, s)
· ∂(cid:126)Σ(θ, s)
· ∂(cid:126)Σ(s, θ)
∂θ
∂θ
Now, let us remember that the curvature κ(s) at a point P of the curve (cid:126)γ(s) is defined as
the inverse of the radius of the osculating circle at P, see Ref. [58]. Then, if the radius of the
osculating circle is small, the surface (6) describes a cable with a big curvature. In addition,
notice that when the radius of the osculating circle is smaller than the radius of the cable, the
cable surface touches itself. In the literature there are not reported axons with big curvature,
then in this paper we suppose that at each point of the curve (cid:126)γ(s) the radius of the osculating
circle is larger than the radius of the cable. Namely, at each point of the curve (cid:126)γ(s), we suppose
that the curvature κ(s) is smaller than R−1(s) and the following inequality
κ(s)R(s) < 1
(13)
is satisfied.
The axon geometry is important for diverse physiological processes, such as the voltage
propagation. In this respect, according to Ref. [56], when the cable geometry is given by (6)
the cable equation is
∂V (s, t)
∂t
=
rLcM
(cid:82) 2π
0 dθ
det g(θ, s)
∂
∂s
a(s)
∂V (s, t)
∂s
− iion
cM
,
(cid:32)
(cid:33)
where a(s) is the cable cross-section and
(cid:113)
(cid:34)
det g(θ, s) =
R2(θ, s)
− τ
∂R(θ, s)
(cid:33)2
R2(θ, s) +
∂θ
(cid:32)∂R(θ, s)
∂θ
(cid:33)2(cid:35) 1
2
.
+ (1 − κ(s)R(θ, s) cos θ)2
1
(cid:113)
(cid:32) ∂R(θ, s)
∂s
(14)
(15)
Notice that the equation (14) depends on geometric quantities as the curvature κ and torsion
τ of the cable.
Then, in order to study the voltage propagation in a cable with general geometry and anoma-
lous diffusion, we can employ the equations (1) and (14) to propose a generalized fractional
cable equation as follows
(cid:34)
(cid:82) 2π
(cid:113)
1
rLcM
0 dθ
(cid:32)
(cid:33)
(cid:35)
∂
∂s
det g(θ, s)
a(s)
∂V (s, t)
∂s
− iion
cM
,
(16)
∂V (s, t)
∂t
= βνDνt
where
Dνt =
∂1−ν
∂t1−ν ,
ν = constant,
0 ≤ ν ≤ 1,
(17)
4
is the Riemann-Liouville fractional operator [36] and βν is a constant with (time)1−ν dimen-
sions. The voltage in an infinite cable has to satisfy the Dirichlet boundary condition and a
finite cable has to satisfy the Neumann boundary condition [59, 60], then the solutions of the
equation (16) should satisfy these boundary conditions.
In the general case, iion depends on the voltage and the equation (14) is a non linear
differential equation. However, in the passive cable model we can take
iion =
V (s, t)
rM
.
(18)
Therefore, the fractional cable equation for the passive cable model with the geometry given
by (6) is
(cid:34)
∂V (s, t)
∂t
= βνDνt
(cid:82) 2π
(cid:113)
1
rLcM
0 dθ
det g(θ, s)
(cid:32)
(cid:33)
(cid:35)
∂
∂s
a(s)
∂V (s, t)
∂s
− V (s, t)
rM cM
.
(19)
In the next sections we will study some solutions of this equation.
3 A qualitative analysis
For a non trivial geometry, to find solutions of the equation (19) is a difficult task. However,
let us provide a qualitative analysis of this equation. In this respect, we propose the following
voltage
In this case the equation (19) implies the following equations
(cid:82) 2π
(cid:113)
1
rLcM
0 dθ
det g(θ, s)
V (s, t) = T (t)X(s).
(cid:32)
(cid:33)
∂X(s)
∂
∂s
a(s)
∂s
− X(s)
rM cM
∂T (t)
∂t
(20)
(21)
(22)
(23)
(24)
(25)
= −λX(s),
= −λβνDνtT (t),
(cid:16) da(s)
(cid:17)2
ds
2a2(s)
.
− d2a(s)
ds2
a(s)
where λ is a constant.
Moreover, if we take
X(s) =
ψ(s)(cid:113)
a(s)
,
the spacial equation (21) can be written as
where
rLcM
U (s) =
−∂2ψ(s)
∂s2 + U (s)ψ(s) = 0,
(cid:19)
(cid:18)
λ − 1
rM cM
− 1
2
det g(θ, s)
(cid:113)
(cid:82) 2π
0 dθ
a(s)
Observe that in this last equation the parameters ν and βν do not appear. In fact, this spacial
equation is the same spacial equation which appears in the non fractional case [56]. As well,
observe that the parameter λ does not depend on ν neither on βν.
5
Furthermore, the Laplace transform of the temporal equation (22) implies
Tν(ζ) = T (0)
ζ ν−1
(ζ ν + βνλ)
,
(26)
where Tν(ζ) is the Laplace transform of the function Tν(t). The inverse Laplace transform of
the function (26) is given by [36]
where
Tν(t) = T (0)Eν,1 (−βνλtν) ,
(cid:88)
n≥0
zn
Γ(νn + Λ)
Eν,Λ(z) =
is the Mittag-Leffler type function [36]. Notice that if ν = 1 we obtain the usual solution
T1(t) = T (0)e−λt.
In addition, if ν = 0 we obtain
T0(t) = T (0)
(cid:88)
n≥0
(−λβ0)n,
which is a constant. We can observe that the equation (30) only makes sense if
in this case we get
β0λ < 1,
T0(t) =
T (0)
1 + β0λ
.
(27)
(28)
(29)
(30)
(31)
(32)
In the section IV, we can see that, for realistic parameter for cylindrical cable, λ is the order
of 10−3(sec)−1. Then, the inequality (31) is satisfied if β0 is smaller than 103sec.
Notice that the strongest fractional effect is obtained when ν is close to zero and in this
case the function Tν(t) is close to the constant (32), which decreases when β0 increases. Then,
in order to model a system where the voltage decreases, we should take β0 bigger than one, but
satisfying the inequality (31). If we take β0 close to zero, the voltage does not decrease.
Now, remember that the usual case is obtained with ν = 1 and β1 = 1. Additionally, when ν
is close to zero, βν should reach its maximum value, in fact β0 should be bigger than one. Thus,
for a system where the voltage decreases, we can suppose that βν increases when ν decreases.
Fig. 2 shows the function (27) for different βν and ν values. In this we can see that the
function (27) decreases when βν increases and ν decreases.
Then, using the functions (27) and (23), we found the following voltage
V (s, t) =
Eν,1 (−βνλtν) ψ(s),
(33)
T (0)(cid:113)
a(s)
where the function ψ(s) satisfies the equation (24). From the equation (33) we can see that
when the cable cross-sectional area a(s) increases the voltage decreases. As well, when βν in-
creases and ν decreases the voltage decreases.
Therefore, this qualitative analysis suggests that in an axon with transport or geometrical
defects the voltage decreases.
6
4 Cylindrical cable with constant radius
When the sectional area is a constant, that is R(s) = R0 =constant, the equation (15) does not
depend on the curvature neither on the torsion of the cable. Indeed, in this case the equation
(19) becomes
(cid:32) R0
(cid:33)
∂V (s, t)
∂t
= βνDνt
∂2V (s, t)
∂s2 − V (s, t)
rM cM
2cM rL
,
(34)
which is equivalent to the fractional cable equation for a straight cylindrical cable (1). How-
ever, observe that the equation (34) depends on the arc length parameter (3) instead of the lab
frame coordinate. This shows that the natural variables for the voltage are given by geometric
quantities of the cable.
For these geometries, in the finite cable case, the solution of the equation (34) is given by
(cid:88)
n≥0
(cid:20)
−βν
(cid:18) nπR0
2cM rLl
V (s, t) =
bnEν,1
(cid:19)
(cid:21)
+
1
rM cM
tν
cos nπ
s
l
,
(35)
where l is the length of the cable and Eν,Λ(z) is the Mittag-Leffler function (28).
Fig. 3 shows the fractional cable equation solution for a cylindrical cable for different ν and
βν values and the initial condition
V (s, 0) = A
(cid:18)
(cid:18) sπ
(cid:19)(cid:19)
l
1 + cos
,
(36)
where A = 0.05mV · cm 1
2 and l = 0.13cm. In this figure, we can see that the voltage decreases
when βν increases and ν decreases. Then, anomalous diffusion implies that voltage decreases.
To obtain an exact solution of the equation (34) is a difficult task for an arbitrary initial
condition. Then, in order to study this case we employ a numerical method. For integer
derivatives, spacial and temporal, we use a second order finite differences method. Moreover,
for the temporal fractional derivatives we use a second order scheme taken from the Fractional
Integration Toolbox [61]. The mesh size was chosen as follows: first, we begin with 1024 points
along the s-axis and 100 points at the time. These numbers were increasing until the difference
between two successive solutions was almost null. The number of spatial and temporal points
used in the simulations are shown in Table 1. The system is solved using the Gauss-Seidel
iterative method, with a tolerance of 10−10. In addition, we impose the Neumann boundary
conditions [59]
∂V (s0, t)
∂s
=
∂V (sns, t)
∂s
= 0.
(37)
Fig. 4 shows the numerical solution of the equation (34) for a cylindrical cable for different
ν and βν values and the initial condition (36). This figure shows that the numerical solution is
close to the analytical solution.
Table 1: Number of points, ns and nt, and number of times of the refinement.
Refinement
First time
Second time
Third time
ns
1024
2048
4096
nt
100
500
2000
7
A more realistic initial condition is given by the function
V (s, 0) =
A√
2πσ
− s2
e
2σ2 ,
(38)
where A = 0.00128mV · cm 1
2 and σ = 0.004cm. Fig. 5 shows the numerical solution of equation
(34) for different ν and βν values and the initial condition (38). The numerical values of ν and
βν used in this simulation are shown in Table 2.
Notice that both the exact and the numerical solutions of the equation (34) provide a volt-
age which decreases when βν increases and ν decreases. Then, anomalous diffusion implies
that voltage decreases. It is worth mentioning that some experimental studies show that in
brain tumours there is anomalous diffusion [19, 20, 21]. As well, in different neurodegenerative
diseases disruption axonal transport are reported [24, 25, 26]. In these cases, the fractional
cable equation implies that the voltage decreases.
In the next section we will study the equation (19) with non constant radius.
5 Cables with swellings
It can be shown that when a cable has vanished curvature and the cable radius is given by
or
R(s) = R0 (1 + α1 sin α2s)
(cid:16)
R(s) = R0
1 + α1 sin2 α2s
(cid:17)
,
(39)
(40)
the spatial equation (24) is similar to the spatial equation for the straight cylindrical cable.
For this reason, the voltage in a cable with radius (39) or (40) is similar to the voltage in
the straight cylindrical cable [56]. Hence, in the fractional case, when a cable has vanished
curvature and the radius is given by (39) or (40) the voltage will be similar to voltage (35).
5.1 Circular cross-section
The cable equation (19) is hard to solve for a general cable geometry. However, for some
cases this equation can be simplified. For instance, when the cable has a circular cross-section,
namely when R(θ, s) = R(s), the equation (19) does not depend on the torsion of the cable τ
and becomes
rLcM R(s)(cid:82) 2π
0 dθ
π ∂
∂s
(cid:114)
(cid:17)
(cid:16)
(1 − κ(s)R(s) cos θ)2 +
R2(s) ∂V (s,t)
∂s
(cid:16) dR(s)
(cid:17)2
ds
− V (s, t)
rM cM
.
∂V (s, t)
∂t
= βνDνt
(41)
In the following subsections we will study cables which model axons with swellings.
Table 2: Numerical values of ν and βν.
ν
1
0.9
0.7
0.5
0.3
βν
1
1.5 (sec)0.1
(sec)0.3
4
16 (sec)0.5
37 (sec)0.7
8
5.2 Cable with Gaussian swelling
In different neurodegenerative diseases, focal axonal swellings are found such as the Fig. 6(a),
see for example [38]. This geometry can be modelled with a cable with radius
(cid:16)
1 + α1e−α2(s−α3)2(cid:17)
R(s) = R0
.
(42)
For this cable geometry, if the initial condition is given by (38), the numerical solution of the
equation (41) can be seen in the Fig. 6(b), (c) and (d). In these figures we can observe that
the voltage decreases faster than the voltage of the cylindrical cable. In addition, notice that
for this cable geometry when βν increases and ν decreases the voltage of the cable decreases
faster than when β1 = 1 and ν = 1. Then, in an axon with a swelling and anomalous diffusion
the voltage strongly decreases.
5.3 Cable with Gaussian swellings
In this section we study the numerical solution for the equation (41) for a cable with the radius
(cid:16)
1 + α1e−α2s2 + α1e−α2(s−α3)2 + α1e−α2(s−2α3)2 + α1e−α2(s−3α3)2(cid:17)
R(s) = R0
(43)
and with the initial condition (38).
A cable with this geometry can be seen in the figure 7(a). It is worth mentioning that axons
with this geometry have been reported in different studies [48]. The numerical solution of the
equation (41) for this cable radius can be observed in Fig. 7(b), (c) and (d). In this case the
voltage decreases faster than the voltage of the cylindrical cable. Moreover, when βν increases
and ν decreases, the voltage decreases faster than non fractional voltage (β1 = ν = 1).
In
addition, we can see that the voltage in a cable with this geometry decreases faster than the
voltage in a cable with the geometry (42).
5.4 Amorphous swelling
A more realistic model for an axon with swelling is given by Fig. 8(a). This geometry can be
described by a cable with an amorphous swelling with radius
R(θ, s) = R0
1 + α1e−α2(s−α3)2 + α4 sin θ cos α5s
.
(44)
(cid:16)
(cid:17)
The numerical solution of the voltage for this cable can be seen in Fig. 8(b), (c) and (d).
This voltage is different from the voltage for the cable with radius (42). Then, geometric
inhomogeneities affects the voltage propagation in a cable.
6 Summary
In different neurodegenerative diseases such as multiple sclerosis, Alzheimer's disease and
Parkinson's disease are reported abnormal accumulations of proteins and organelles which gen-
erate disruption axonal transport.
In addition, recently different experimental studies have
found anomalous diffusion in brain tissues. Notably this diffusion is expressed through frac-
tional calculus. Another hallmark feature of some neurodegenerative diseases is given by axonal
discrete swellings along the axons.
In order to study the voltage propagation in a cable with both of these hallmark features of
the neurodegenerative diseases, we proposed a fractional cable equation with non trivial geom-
etry. This equation depends on geometric quantities such as the curvature and torsion of the
cable, as well as the fractional parameters βν and ν. It is worth mentioning that the parameter
βν depends on ν. In this respect, we showed that with regard to model a system where the
voltage decreases, we should suppose that βν increases when ν decreases. Furthermore, in this
9
new cable equation the strongest fractional effect is obtained when ν is close to zero.
For a straight cylinder with a constant radius we showed that the voltage decreases when
the fractional effect increases. Notice that in this case the cable does not have swellings. Then,
if there is an abnormal accumulation of proteins and organelles in an axon, the diffusion can
be hindered and become to anomalous diffusion. In this case our results suggest that the volt-
age decreases in axons. Indeed, when the fractional effect is strong, the voltage may be blocked.
In addition, cables with swellings and anomalous diffusion were studied. Regarding this, we
studied cable geometries similar to some axons reported in the literature. For all these cable
geometries, we found that when the fractional effect increases, the voltage decreases. Further-
more, we found that the voltage dramatically decreases when the cable has a big swelling or has
many swellings. Then, these results also suggested that in axons with swellings and abnormal
accumulations of proteins and organelles the voltage decreases and may be blocked.
How axonal transport defects and deformed geometry of axons are related each other is an
important problem. In order to study this problem, in our model the geometry quantities of
the cable should be related with the parameters ν and βν. In a future work the above problem
and the active case, where non linear interactions play an important role, will be studied.
Acknowledgments
This work was supported in part by conacyt-sep 47510318 (J.M.R.). We are grateful to
referees for providing valuable comments.
References
[1] I. L. Pykett, NMR Imaging in Medicine, Scientific American 246, 78 (1986).
[2] I. L. Pikett, J. H Newhouse, F. S. Buonanno, T. J. Brady, M. R. Goldman, J. P. Kistler,
and G. M. Pohost, Principles of nuclear magnetic resonance imaging, Radiology 143, 157
(1981).
[3] J.-D. Tournier, S. Mori, and A. Leemans, Diffusion tensor imaging and beyond, Magnetic
Resonance in Medicine 65, 1532 (2011).
[4] C. Beaulieu, The basis of anisotropic water diffusion in the nervous system - a technical
review, NMR Biomed., 15, 435 (2002).
[5] J. Goveas, L. O' Dwyer, M. Mascalchi, M. Cosottini, S. Diciotti, S. De Santis, L. Passa-
monti, C.Tessa, N. Toschi, and M. Giannelli, Diffusion-MRI in neurodegenerative disor-
ders, Magnetic Resonance Imaging 33, 853 (2015).
[6] S. K. Song, S. W. Sun, M. J. Ramsbottom, C. Chang, J. Russell, and A. H. Cross, Dys-
myelination revealed through MRI as increased radial (but unchanged axial) diffusion of
water, NeuroImage 17, 1429 (2002).
[7] D. J. Werring, C. A. Clark, G. J. Barker, A. J. Thompson, and D. H. Miller, Diffusion ten-
sor imaging of lesions and normal-appearing white matter in multiple sclerosis, Neurology
52 1626 (1999).
[8] M. Filippi, M. Cercignani, M. Inglese, M. A. Horsfield, and G. Comi, Diffusion tensor
magnetic resonance imaging in multiple sclerosis, Neurology 56 304 (2001).
[9] H. D. Rosas, D. S. Tuch, N. D. Hevelone, A. K. Zaleta, M. Vangel, S. M. Hersch, and D. H.
Salat, Diffusion tensor imaging in presymptomatic and early Huntington's disease: Selec-
tive white matter pathology and its relationship to clinical measures, Movement Disorders
21, 1317 (2006).
10
[10] S. J. Choi, K. O. Lim, I. Monteiro, and B. Reisberg, Diffusion tensor imaging of frontal
white matter microstructure in early Alzheimers disease: a preliminary study, Journal of
Geriatric Psychiatry and Neurology 18, 12 (2005).
[11] N. H. Stricker, B. C. Schweinsburg, L. Delano-Woodd, C. E. Wierengad, K. J. Bangen, K.
Y.Haaland, L. R. Frank, D.P.Salmon, and M.W.Bondi, Decreased white matter integrity in
late-myelinating fiber pathways in Alzheimer's disease supports retrogenesis, NeuroImage
45 10 (2009).
[12] M. P. Laakso, K. Partanen,P. Riekkinen, M. Lehtovirta, E. L. Helkala, M. Hallikainen,
T. Hanninen, P. Vainio, and H. Soininen, Hippocampal volumes in Alzheimer's disease,
Parkinson's disease with and without dementia, and in vascular dementia: an MRI study,
Neurology 46, 678 (1996).
[13] Y. Bai, Y. Lin, J. Tian, D. Shi, J. Cheng, E. M. Haacke, X. Hong, B. Ma, J. Zhou,
and M. Wang, Grading of gliomas by using monoexponential, biexponential, and stretched
exponential diffusion-weighted MR imaging and diffusion kurtosis MR imaging, Radiology
278, 496 (2016).
[14] M. M. Karaman, Y. Sui, H. Wang, R. L. Magin, Y. Li, and X. J. Zhou, Differentiating
low- and high-grade pediatric brain tumors using a continuous-time random-walk diffusion
model at high b-values, Magnetic Resonance in Medecine 76 4, 1149 (2016).
[15] B. Xu, L. Su, Z. Wang, Y. Fan, G. Gong, W. Zhu, P. Gao, and J.-H. Gao, Anomalous
diffusion in cerebral glioma assessed using a fractional motion model, Magnetic Resonance
in Medecine (2017). DOI: 10.1002/mrm.26581.
[16] R. L. Magin, O. Abdullah, D. Baleanu, and X. J. Zhoua, Anomalous diffusion expressed
through fractional order differential operators in the Bloch-Torrey equation, Journal of
Magnetic Resonance 190, 255 (2008).
[17] K. M. Bennett, K. M. Schmainda, R. T. Bennett, D. B. Rowe, H. Lu, and J. S. Hyde,
Characterization of continuously distributed cortical water diffusion rates with a stretched-
exponential model, Magnetic Resonance in Medicine 50, 727 (2003).
[18] K. M. Bennett, J. S. Hyde, and K. M. Schmainda, Water diffusion heterogeneity index
in the human brain is insensitive to the orientation of applied magnetic field gradients,
Magnetic Resonance in Medicine 56, 235 (2006).
[19] B. Xu, L. Su, Z. Wang, Y. Fan, G. Gong, W. Zhu, P. Gao, and J.-H. Gao, Anomalous
diffusion in cerebral glioma assessed using a fractional motion model, Magnetic Resonance
in Medicine, (2017) doi:10.1002/mrm.26581.
[20] M. M. Karaman, H. Wang, Y. Sui, H. H. Engelhard, Y. Li, and X. J. Zhou, A fractional
motion diffusion model for grading pediatric brain tumors, NeuroImagen: Clinical 12, 707
(2016).
[21] X.-J. Zhou, Q. Gao, O. Abdullah, and R. L. Magin, Studies of anomalous diffusion in
the human brain using fractional order calculus, Magnetic Resonance in Medicine 63, 562
(2010).
[22] F. Santamaria, S. Wils, E. De Schutter, and G. J. Augustine, Anomalous diffusion in
Purkinje cell dendrites caused by spines, Neuron 52, 365 (2006).
[23] S. Qin, F. Liu, I. W. Turner, Q. Yu, Q. Yang, and V. Vegh, Characterization of anomalous
2 -weighted magnetic
relaxation using the time-fractional Bloch equation and multiple echo T ∗
resonance imaging at 7 T , Magnetic Resonance in Medicine 77, 1485 (2017).
[24] E. Chevalier-Larsen and E. L. F. Holzbaur, Axonal transport and neurodegenerative disease,
Biochimica et Biophysica Acta - Molecular Basis of Disease 1762, 1094 (2006).
11
[25] K. J. De Vos, A. J. Grierson, S. Ackerley, and C. C. J. Miller, Role of axonal transport in
neurodegenerative diseases, Annu. Rev. Neurosci. 31, 151 (2008).
[26] S. Millecamps and J.-P. Julien, Axonal transport deficits and neurodegenerative diseases,
Nature Reviews Neuroscience 14, 161 (2013).
[27] R. Metzler and J. Klafter, The random walk's guide to anomalous diffusion: a fractional
dynamics approach, Phys. Rep. 339, 1 (2000).
[28] D. S. Novikov, E. Fieremans, J. H. Jensen, and J. A. Helpern, Random walks with barriers,
Nature Physics 7, 508 (2011).
[29] M. Palombo, A. Gabrielli, V. D. P. Servedio, G. Ruocco, and S. Capuani, Structural
disorder and anomalous diffusion in random packing of spheres, Scientific Reports 3, 2631
(2013).
[30] J. M. Haugh, Analysis of reaction-diffusion systems with anomalous subdiffusion, Biophys-
ical Journal 97, 435 (2009).
[31] M. J. Saxon, A biological interpretation of transient anomalous subdiffusion. I. Qualitative
model, Biophysical Journal 92, 1178 (2007).
[32] M. Weiss, M. Elsner, F. Kartberg, and T. Nilsson, Anomalous subdiffusion is a measure
for cytoplasmic crowding in living cells, Biophysical Journal 87, 3518 (2004).
[33] M. J. Saxon, Anomalous subdiffusion in fluorescence photobleaching recovery: a Monte
Carlo study, Biophysical Journal 81, 2226 (2001).
[34] B. I. Henry, T. A. M. Langlands, and S. L. Wearne, Fractional cable models for spiny
neuronal dendrites, Phys. Rev. Lett. 100, 128103 (2008).
[35] T. A. M. Langlands, B. I. Henry, and S. L. Wearne, Fractional cable equation models for
anomalous electrodiffusion in nerve cells: infinite domain solutions, Journal of Mathemat-
ical Biology 59, 761 (2009).
[36] L. Debmath and D. Bhatta, Integral Transforms ans Their Applications 2nd ed. (Chapman
and Hall/CRC, NewYork, 2007).
[37] J. E. Galvin, K. Uryu, V. M.-Y. Lee, and J. Q. Trojanowski, Axon pathology in Parkinson'
s disease and Lewy body dementia hippocampus contains α-, β-, and γ-synuclein, Proc.
Natl. Acad. Sci. USA 96, 13450 (1999).
[38] B. D. Trapp, J. Peterson, R. M. Ransohoff, R. Rudick, S. Mork, and L. Bo, Axonal
transection in the lesions of multiple sclerosis, New Engl. J. Med. 338, 278 (1998).
[39] V. E. Johnson, W. Stewart, and D. H. Smith, Axonal pathology in traumatic brain injury,
Exp. Neurol. 246, 35 (2013).
[40] D. Krstic and I. Knuesel, Deciphering the mechanism underlying late-onset Alzheimer
disease, Nat. Rev. Neurosci. 9, 25 (2013).
[41] H. Xie, J. Guan, L. A. Borrelli, J. Xu, A. Serrano-Pozo, and B. J. Bacskai, Mitochondrial
alterations near amyloid plaques in an Alzheimer's disease mouse model, J. Neurosci. 33,
17042 (2013).
[42] H. Budka, G. Costanzi, S. Cristina, A. Leehi, C. Parravicini, R. Trabattoni, and L. Vago,
Brain pathology induced by infection with the human immunodeficiency virus (HIV), Acta
Neuropathologica 75, 185 (1987).
[43] F. Gray, F. Chr´etien, A. V. Vallat-Decouvelaere, and F. Scaravilli, The Changing Pattern
of HIV Neuropathology in the HAART Era, J. Neuropathol. Exp. Neurol. 62, 429 (2003).
12
[44] M. Kaul, G. A. Garden, and S. A. Lipton, Pathways to neuronal injury and apoptosis in
HIV-associated dementia, Nature 410, 988 (2001).
[45] F. Raja, F. E. Sherriff, C. S. Morris, L. R. Bridges, and M. M. Esiri, Cerebral white matter
damage in HIV infection demonstrated using β-amyloid precursor protein immunoreactiv-
ity, Acta Neuropathol. 93, 184 (1997).
[46] R. Ellis, D. Langford, and E. Masliah, HIV and antiretroviral therapy in therain: neuronal
injury and repair, Nat. Rev. Neurosci. 8, 33 (2007).
[47] M. H. Magdesian, F. S. Sanchez, M. Lopez, P. Thostrup, N. Durisic, W. Belkaid, D. Lia-
zoghli, P. Grutter, and D.R. Colman, Atomic force microscopy reveals important differences
in axonal resistance to injury, Biophys. J. 103, 405 (2012).
[48] G. M. G. Shepherd and K. M. Harris, Three-dimensional structure and composition of
CA3→CA1 axons in rat hippocampal slices: Implications for presynaptic connectivity and
compartmentalization, J. Neurosci. 18, 8300 (1998).
[49] H. Anwar, C. J. Roome, H. Nedelescu, W. Chen, B. Kuhn, and E. De Schutter, Den-
dritic diameters affect the spatial variability of intracellular calcium dynamics in computer
models, Front. Cell. Neurosci. 8, 168 (2014).
[50] P. Vetter, A. Roth, and M. Hausser, Propagation of action potentials in dendrites depends
on dendritic morphology, Neurophysiol 85, 926 (2001).
[51] A. D. Bird and H. Cuntz, Optimal current transfer in dendrites, PLoS Comput Biol. 12,
e1004897 (2016).
[52] E. R. Kandel, J. H. Schwartz, and T. N. Jessell, Principles of Neural Science (McGraw-Hill,
New York, 2000).
[53] J. C. Fiala and K. M. Harris, Dendrite structure, in Dendrites, edited by G. Stuart, N.
Spruston, and M. Hausser (Oxford University Press, Oxford, 1999), p.1.
[54] R. R. Poznanski, Modelling the electrotonic structure of starburst amacrine cells in the
rabbit retina: A functional interpretation of dendritic morphology, Bull. Math. Biol. 54,
905 (1992).
[55] P. D. Maia, M.A. Hemphill, B. Zehnder, C. Zhang, K. K. Parker, and J. N. Kutz, Diagnos-
tic tools for evaluating the impact of Focal Axonal Swellings arising in neurodegenerative
diseases and/or traumatic brain injury, Journal of Neuroscience Methods 253, 233 (2015).
[56] E. L´opez-S´anchez and J. M. Romero, Cable equation for general geometry, Phys. Rev. E.
95, 022403 (2017).
[57] M. P. Do Carmo, Differential Geometry of Curves and Surfaces (Dover, NewYork, 2016).
[58] M. Spivak, A Comprehensive Introduction to Differential Geometry Vol. II (Publish or
Perish, Houston 1975).
[59] G. B. Ermentrout and D. H. Terman, Mathematical Foundation of Neuroscience (Springer,
London, 2010).
[60] P. C. Bressloff, Waves in Neural Media: From Single Neurons to Neural Fields (Springer,
London, 2014).
[61] T. M. Marinov, N. Ramirez, and F. Santamaria, Fractional Integration Toolbox, Fractional
Calculus and Applied Analysis, 16, 670 (2013).
13
Figure 1: Cable with general geometry. The vectors T , N , B are shown in two different points
on the curve (cid:126)γ.
Figure 2: Mittag-Leffler with different values of βν and ν.
(a) Voltage for the cylindrical cable, with ν = 0.7 and β0.7 = 15 (sec)0.3. (b) Voltage
Figure 3:
vs t in s = 0 with different values of βν and ν. (c) Voltage vs s at time t = 12 sec for different
values of βν and ν. Parameter values used for simulations correspond to realistic dendritic
parameters as in [54]: cM = 1 mF/cm2, rM = 3000 Ω cm2, rL = 100 Ω cm, R0 = 10−4 cm. The
initial condition is given by (36).
14
(a) Voltage for the cylindrical cable, with ν = 0.7 and β0.7 = 15 (sec)0.3. (b) Voltage
Figure 4:
vs t in s = 0 with different values of βν and ν. (c) Voltage vs s at time t = 12 sec for different
values of βν and ν. Parameter values used for simulations correspond to realistic dendritic
parameters as in [54]: cM = 1 mF/cm2, rM = 3000 Ω cm2, rL = 100 Ω cm, R0 = 10−4 cm. The
initial condition is given by (36).
(a) Voltage for the cylindrical cable, with ν = 0.5 and β0.5 = 16 (sec)0.5. (b) Voltage
Figure 5:
vs t in s = 0 with different values of βν and ν. (c) Voltage vs s at time t = 7 sec for different
values of βν and ν. Parameter values used for simulations correspond to realistic dendritic
parameters as in [54]: cM = 1 mF/cm2, rM = 3000 Ω cm2, rL = 100 Ω cm, R0 = 10−4 cm. The
initial condition is given by (38).
15
Figure 6:
(a) Cable with geometry (42). (b) Voltage for the cable with the radius (42) and
ν = 0.5, β0.5 = 16 (sec)0.5. (c) Voltage vs t in s = 0 with different values of βν and ν. (d)
Voltage vs s at time t = 7 sec for different values of βν and ν. Parameter values used for
simulations correspond to realistic dendritic parameters as in [54]: cM = 1 mF/cm2, rM =
3000 Ω cm2, rL = 100 Ω cm, R0 = 10−4 cm, α1 = 10, α2 = 0.11 cm−2, α3 = 0 cm. The initial
condition is given by (38).
16
(a) Cable with geometry (43).
Figure 7:
(b) Voltage for the cable with a gaussian train
swellings, radius (43) and ν = 0.5, β0.5 = 16 (sec)0.5. (c) Voltage vs t in s = 0 with dif-
ferent values of βν and ν. (d) Voltage vs s at time t = 7 sec for different values of βν and
ν. Parameter values used for simulations correspond to realistic dendritic parameters as in
[54]: cM = 1mF/cm2, rM = 3000 Ω cm2, rL = 100 Ω cm, R0 = 10−4 cm, α1 = 10, α2 =
0.11 cm−2, α3 = 0.06 cm. The initial condition is given by (38).
17
Figure 8:
(a) Cable with geometry (44). (b) Voltage for the cable with the radius (44) and
ν = 0.5, β0.5 = 16 (sec)0.5. (c) Voltage vs t in s = 0 with different values of βν and ν. (d) Voltage
vs s at time t = 7 sec for different values of βν and ν. Parameter values used for simulations
correspond to realistic dendritic parameters as in [54]: cM = 1 mF/cm2, rM = 3000 Ω cm2, rL =
100 Ω cm, R0 = 10−4 cm, α1 = 10, α2 = 0.11 cm−2, α3 = 0 cm, α4 = 10, α5 = 0.11 cm−1. The
initial condition is given by (38).
18
|
1303.0402 | 1 | 1303 | 2013-03-02T16:50:14 | Organizational properties of second-order visual filters sensitive to the orientation modulations | [
"q-bio.NC"
] | Here we show psychophysical study of second-order visual mechanisms sensitive to the orientation modulations. Selectivity to orientation, phase and spatial frequency of modulation is measured. Bandwidths for phase (0,5{\pi}) and orientation (33,75 deg) are defined, but there is no evidence for spatial frequency selectivity. | q-bio.NC | q-bio | Proceedings of the 10th International Conference "Applied Optics - 2012", St.
Petersburg. P. 331-334. 2012
ORGANIZATIONAL PROPERTIES OF SECOND-ORDER VISUAL
FILTERS SENSITIVE TO THE ORIENTATION MODULATIONS
Babenko V.V., Miftakhova M.M., Yavna D.V.
[email protected]
Here we show psychophysical study of second-order visual mechanisms sensitive to
the orientation modulations. Selectivity to orientation, phase and spatial frequency of
modulation is measured. Bandwidths for phase (±0,5π) and orientation (±33,75 deg) are
defined, but there is no evidence for spatial frequency selectivity.
Studying preattentive mechanisms of spatial grouping of local information functioning
at early stages of visual processing and known as second-order mechanisms is necessary for
understanding how local features filtered on the level of primary visual cortex are united into
cognitive blocks.
Generally to describe principles of second-order stimuli processing filter-rectify-filter
model is used ([1] for review). Firstly this model was non-specific to the modulated parameter
(contrast, orientation or spatial frequency (SF)).
Thus, if contrast of texture elements lying in the periphery of the second-order
receptive field decreases, then second-order filter ’s response increases. But we come to the
similar results if orientation or SF of elements varies. In other words second-order filters pass
any modulation. Kingdom et al. put this non-specificity into question for the first time in 2003
[2]. His research carried out in the paradigm of subthreshold summation and the number of
latter studies using masking [3] and adaptation paradigms consider the existence of
independent channels, selectively sensitive to the type of second-order information.
Previously we have shown properties of the second-order visual mechanisms sensitive
to the contrast [4] and SF [5] modulations. This research is devoted to the measuring tunings
of second-order filters sensitive to the orientation modulations.
Procedure
3 observers with normal or corrected-to-normal vision participated in the research.
Experiment was carried out using masking, two-alternative forced-choice procedure,
and staircase method.
There were 3 experimental series: measuring selectivity to orientation, SF and phase of
modulation. Textures composed of gabor micropatterns were used as stimuli. Amplitude of
modulation threshold in the presence of masking stimuli was measured in all experiments.
In each experimental series 5 masks were presented. In the first experiment axis tilt
was varied in a range of 0-90 dergees with 22.5 increment from the test stimulus. In the
second experiment masks with phase shift from 0 to 1π were used, 0.25π increment. And in
the third experimental series SF of the modulation was changed from -2 to 2 octaves with 1
octave increment. In the control series non-modulated texture was used as a mask.
Selectivity to the orientation of modulation
Studying selectivity of second-order mechanism to the orientation of modulation
revealed significant decrease of threshold amplitude with increasing of the axis tilt of mask’s
modulation (Fig. 1)
Proceedings of the 10th International Conference "Applied Optics - 2012", St.
Petersburg. P. 331-334. 2012
Fig. 1. Threshold amplitude dependence on the orientation of modulation. Summary data for 3
observers. Abscissa shows modulation axis tilt in mask in deg, ordinate shows measured
threshold in arbitrary units. Dash-line corresponds to the threshold level when masking by
non-modulated mask.
Curves for all observers were similar. Maximum of the masking effect was observed
when axis tilts of modulation in mask and test where the same. With the increasing of the
difference between test and mask significant threshold decrease was measured. Threshold for
masking by 67.5 and 90 degrees axis tilts is about the same and equals 7.55 and 7.74 arbitrary
units. Bandwidth at the half of the amplitude is estimated at ±33.75 degrees.
It should be noted that bandwidth for orientation of second-order mechanisms
sensitive to the orientation modulations is narrower than bandwidth for second-order
mechanisms sensitive to the contrast modulations [4].
Selectivity to the phase of modulation
When masking by stimuli with shifted phase of modulation the following change in
threshold was found (Fig.2):
Fig.2. Threshold amplitude dependence on the phase shift of modulation. Summary data for 3
observers. Abscissa shows phase shift of masks’s modulation function (parts of π), ordinate
shows threshold in arbitrary units. Dash-line corresponds to the threshold level when masking
by non-modulated mask.
It was observed that modulation threshold amplitude decreases with the increase of
phase shift of mask’s modulation. Therefore we consider second-order mechanisms sensitive
to the orientation modulations are sensitive to the phase shift of modulation.
In this experimental series 2 observers have shown bandpass tuning, but one didn’t
show this tuning. One explanation is subjective difficulty of performing task for this observer,
another explanation is that mechanisms that lie higher than second-order mechanism influence
on the task performing.
Tuning of filters to the phase of the envelope is estimated as 0.5π at the half of
amplitude, and it is consistent with analogous bandwidth of mechanisms sensitive to the
Proceedings of the 10th International Conference "Applied Optics - 2012", St.
Petersburg. P. 331-334. 2012
contrast modulations.
Selectivity to the spatial frequency of modulation
Studying selectivity to SF of modulation have shown the increase of threshold
amlitude with increasing of mask’s SF from -2 to +2 octaves (Fig.3).
Fig.3. Threshold amplitude dependence on the SF of modulation. Summary data for 3
observers. Abscissa shows SF of mask in octaves, ordinate shows threshold in arbitrary units.
Dash-line corresponds to the threshold level when masking by non-modulated mask.
When SF of test and mask are the same threshold amplitude equals 10.5. With
decreasing of mask’s SF perception of test stimulus improves; with increasing of SF threshold
slightly rises.
Results for two observers have shown that detection threshold is significantly lower
for low mask’s frequencies versus thresholds for high mask’s frequencies.
However our data doesn’t lead to the conclusion about second-order SF sensitivity of
mechanisms sensitive to the orientation modulations.
Discussion
Our data allows us to estimate tunings of filters sensitive to the orientation
modulations to orientation (±33,75 deg) and phase (±0,5 π) of modulation.
In our previous studies we investigated second-order mechanisms sensitive to the
contrast [4] and spatial frequency [5] modulations. According to the data on selectivity we
suggested possible forms of receptive fields of underlying second-order vision mechanisms.
Receptive fields of mechanisms sensitive to the contrast modulations are likely to have
elongated shape and inhibitory subfields. Second-order mechanisms sensitive to the spatial
frequency modulations are likely to have concentric receptive fields.
Our current data doesn’t allow us to assess SF tunings of second-order mechanisms
sensitive to the orientation modulations, because existence and properties of spatial frequency
selectivity are not clear. So lack of spatial frequency selectivity can be explained by
organizational properties of second-order filters sensitive to orientation modulations.
Generally our data is consistent with conclusions of Hallum et al. [6] about the existence of
differences in mechanisms processing second-order orientations in the processing of high and
low spatial frequencies.
References
1. Graham, N.V. Beyond multiple pattern analyzers modeled as linear filters (as classical V1
simple cells): Useful additions of the last 25 years // Vision Res. Vol. 51, N. 13. P. 1397–1430.
2011.
2. Kingdom, F.A.A., Prins, N., and Hayes, A. Mechanism independence for texture-
Proceedings of the 10th International Conference "Applied Optics - 2012", St.
Petersburg. P. 331-334. 2012
modulation detection is consistent with a filter-rectify-filter mechanism // Vis. Neurosci. Vol.
20, N.1. P. 65–76. 2003.
3. Babenko V.V., Yavna D.V. Specificity of the visual second-order mechanisms // Perception
37 Abstract Supplement.
P. 78. 2008.
4. Babenko V.V., Yavna D.V., Solov’ev A.A., Miftakhova M.B. Spatial selectivity of visual
∗∗
mechanisms sensitive to contrast modulation
This article is a development of the report
“Adjusting second-order visual filters to the orientation of the envelope,” presented at the
conference Applied Optics–2010 // Journal of Optical Technology, Vol. 78, Issue 12, pp. 771-
776. 2011.
5. Miftakhova M.B., Babenko V.V. Features of the second-order visual filters sensitive to the
spatial frequency modulations // Proceedings XVI International Conference on
Neurocybernetics. Rostov-on-Don: SFU-Press. Vol.1 P. 432-434. 2012
6. Hallum L.E., Landy M.S., Heeger D.J. Human primary visual cortex (V1) is selective for
second-order spatial frequency // J. Neurophysiol.
Vol. 105, N 5. P. 2121–2131. 2011.
|
1303.6175 | 1 | 1303 | 2013-03-25T15:43:48 | Compression as a universal principle of animal behavior | [
"q-bio.NC",
"cs.CL",
"cs.IT",
"cs.IT",
"physics.data-an",
"q-bio.QM"
] | A key aim in biology and psychology is to identify fundamental principles underpinning the behavior of animals, including humans. Analyses of human language and the behavior of a range of non-human animal species have provided evidence for a common pattern underlying diverse behavioral phenomena: words follow Zipf's law of brevity (the tendency of more frequently used words to be shorter), and conformity to this general pattern has been seen in the behavior of a number of other animals. It has been argued that the presence of this law is a sign of efficient coding in the information theoretic sense. However, no strong direct connection has been demonstrated between the law and compression, the information theoretic principle of minimizing the expected length of a code. Here we show that minimizing the expected code length implies that the length of a word cannot increase as its frequency increases. Furthermore, we show that the mean code length or duration is significantly small in human language, and also in the behavior of other species in all cases where agreement with the law of brevity has been found. We argue that compression is a general principle of animal behavior, that reflects selection for efficiency of coding. | q-bio.NC | q-bio | Compression as a universal principle of animal
behavior
Short title: Compression in animal behavior
Ramon Ferrer‐i‐Cancho1,*, Antoni Hernández‐Fernández1,2, David Lusseau3,4, Govindasamy
Agoramoorthy5, & Minna J. Hsu6, Stuart Semple7
(1) Complexity and Quantitative Linguistics Lab. Departament de Llenguatges i Sistemes
Informàtics. TALP Research Center. Universitat Politècnica de Catalunya, Barcelona (Catalonia),
Spain.
(2) Departament de Lingüística General. Universitat de Barcelona, Barcelona (Catalonia), Spain.
(3) Institute of Biological and Environmental Sciences, University of Aberdeen, Aberdeen, UK.
(4) Marine Alliance Science and Technology for Scotland, University of Aberdeen, Aberdeen,
UK.
(5) College of Environmental and Health Sciences, Tajen University, Yanpu, Taiwan, ROC.
(6) Department of Biological Sciences, National Sun Yat‐sen University, Kaohsiung, Taiwan,
ROC.
(7) Centre for Research in Evolutionary and Environmental Anthropology, University of
Roehampton, London, UK.
*Author for correspondence ([email protected]).
Keywords: law of brevity, compression, animal communication, language, animal behavior
1
ABSTRACT
A key aim in biology and psychology is to identify fundamental principles underpinning the
behavior of animals, including humans. Analyses of human language and the behavior of a
range of non‐human animal species have provided evidence for a common pattern underlying
diverse behavioral phenomena: words follow Zipf’s law of brevity (the tendency of more
frequently used words to be shorter), and conformity to this general pattern has been seen in
the behavior of a number of other animals. It has been argued that the presence of this law is
a sign of efficient coding in the information theoretic sense. However, no strong direct
connection has been demonstrated between the law and compression, the information
theoretic principle of minimizing the expected length of a code. Here we show that minimizing
the expected code length implies that the length of a word cannot increase as its frequency
increases. Furthermore, we show that the mean code length or duration is significantly small in
human language, and also in the behavior of other species in all cases where agreement with
the law of brevity has been found. We argue that compression is a general principle of animal
behavior, that reflects selection for efficiency of coding.
1. INTRODUCTION
Understanding the fundamental principles underpinning behavior is a key goal in biology and
psychology (Grafen, 2007; Gintis, 2007). From an evolutionary perspective, similar behavioral
patterns seen across different animals (including humans) may have evolved from a common
ancestral trait, or may reflect convergent evolution; looking for shared quantitative properties
of behavior across diverse animal taxa may thus allow identification of the elementary
processes constraining or shaping behavioral evolution (e.g., Hailman et al., 1985; McCowan et
al. 2002; Sumpter, 2006). Recent evidence of consistent patterns linking human language with
vocal communication and other behavior in a range of animal species has pointed to the
2
existence of at least one such general principle (Ferrer‐i‐Cancho & Lusseau, 2009; Semple et
al., 2010). Here, we detail the nature of this evidence, before describing and mathematically
exploring the principle in question – compression – that we propose underlies the consistent
patterns discovered.
Words follow Zipf’s law of brevity, i.e. the tendency of more frequently used words to be
shorter (Zipf, 1935, Strauss et al., 2007), which can be generalized as the tendency of more
frequent elements to be shorter or smaller (Ferrer‐i‐Cancho & Hernández‐Fernández, 2013).
Statistical laws of language have been studied outside human language (see Ferrer‐i‐Cancho &
Hernández‐Fernández, 2013 for an overview), and to our knowledge, the first report of
conformity to this generalized brevity law outside human language is found in the pioneering
quantitative research by Hailman et al. (1985) on chick‐a‐dee calls. More recently, accordance
with the generalized law of brevity has been found in dolphin surface behavioral patterns
(Ferrer‐i‐Cancho & Lusseau, 2009), the vocalizations of Formosan macaques (Semple et al.,
2010) and a subset of the vocal repertoire of common marmosets (Ferrer‐i‐Cancho &
Hernández‐Fernández, 2013). A lack of conformity to the law has been found in the
vocalizations of golden‐backed uakaris (Bezerra et al., 2011) and ravens (Ferrer‐i‐Cancho &
Hernández‐Fernández, 2013).
Statistical patterns of language, and the law of brevity in particular, offer a unique chance to
reframe research on language universals (Evans & Levinson, 2009) in three ground‐breaking
directions, according to the current state of the art. Firstly, in extending typology of linguistic
universals beyond human language and considering the possibility of universals of behavior
across species. Secondly, in considering that some language universals may, in fact, not be
specific to language or communication but an instance of universals of animal behavior.
Thirdly, in changing the stress of the quest from universal properties of language to universal
principles of language (or behavior) along the lines of modern quantitative linguistics (Zipf,
3
1949; Köhler, 1987; Köhler, 2005). More important than expanding the collection of universal
properties or delimiting what is universal or not, is (a) a deep understanding of the principles
explaining the recurrence of these regularities regardless of the context and (b) the role of
those principles, perhaps hidden, in exceptions to widespread regularities. If these regularities
are not inevitable (Ferrer‐i‐Cancho et al., 2013), a parsimonious hypothesis would be a minimal
set of principles that are independent from the context and thus universal. An important
research goal is defining such a set, if it exists. Furthermore, this third point could contribute to
reconciling the wide diversity of languages with the need for unifying approaches, outside of
the realm of a “language faculty” or innate specializations for language (Evans & Levinson,
2009).
The discovery of conformity to statistical laws of language outside human language raises a
very important research question: are the findings simply a coincidence, or are there general
principles responsible for the appearance of the same statistical pattern across species, or
across many levels of life? As the arguments against the simplicity of language laws within and
outside human language are falling (e.g., Ferrer‐i‐Cancho & Elvevåg, 2010; Ferrer‐i‐Cancho &
McCowan, 2012; Hernández‐Fernández et al., 2011; Ferrer‐i‐Cancho et al., 2013), here we
propose the principle of compression ‐ the information theoretic principle of minimizing the
expected length of a code ‐ as an explanation for the conformity to the law of brevity across
species. Hereafter, we consider the term “brevity”, the short length of an element, as an effect
of “compression”.
In his pioneering research, G. K. Zipf argued that the law of brevity was a consequence of a
general principle of economy (Zipf, 1949). He used the metaphor of an “...artisan who will be
obliged to survive by performing jobs as economically as possible with his tools” (p. 57). The
artisan is a metaphor for a speaker (or a community of speakers) while tools is a metaphor for
words. The artisan should decrease the size or the mass of the tools that he uses more
4
frequently to reduce the amount of work (Zipf, 1949, pp. 60‐61). Along similar lines, it is well‐
known that S. Morse and A. Vail optimized the Morse code simply by choosing the length of
each character approximately inversely proportionally to its frequency of occurrence in English
(Gleick, 2011). Although it has been claimed that conformity to the law of brevity is a sign of
efficient coding in both human language (Zipf, 1949) and animal behavior (Ferrer‐i‐Cancho &
Lusseau, 2009; Semple et al., 2010), no strong direct connection with standard information
theory, and with coding theory in particular, has been shown.
Imagine that
n is defined as the number of elements of the repertoire or vocabulary.
pi is defined as the probability of producing the i‐th most likely element.
ei is defined as the energetic cost of that element. ei could be the number of letters or
syllables of the i‐th most likely word or the mean duration of the i‐th most likely
vocalization or behavior of a non‐human animal.
Then, the mean energetic cost of a repertoire or vocabulary can be defined as
(cid:3041)
(cid:1831) (cid:3404) (cid:3533) (cid:1868)(cid:3036) (cid:1857)(cid:3036) .
(cid:3036)(cid:2880)(cid:2869)
Indeed, a particular case of the definition of E in Eq. 1 is ECL, the mean code length, which is the
(1)
cost function that is considered in standard information theory for the problem of
compression, being ei the length of the code used to encode the i‐th element of the set of
symbols (Cover & Thomas, 2006, pp. 110). Data compression consists of finding the code
lengths that minimize ECL given the probabilities p1,...pi,...,pn. ECL provides an objective measure
of coding efficiency. In his pioneering research, G. K. Zipf called his own version of Eq. 1 the
“minimum equation” (in the sense that it is the function to minimize in order to reduce effort),
with pi being frequency of use of a tool, and ei being work, i.e. the product of the “mass” of the
tool and the “distance” of the tool to the artisan (Zipf, 1949, p. 59). Therefore, G. K. Zipf’s view
5
constitutes a precursor of coding theory. However, he never tested if ECL is significantly small in
human languages, using a rigorous statistical approach. Hereafter, the definition of ECL from
information theory is relaxed so that ei can be not only the bits used to code the i‐th element
but also its size, length or duration. ei > 0 is assumed here as we focus on ‘elements’ that can
be observed in a real system, although some models of behavior produce elements of length
zero, and then ei = 0 for such empty elements (see Ferrer‐i‐Cancho & Gavaldà 2009 for a
review of these models). Pressure for minimizing ECL in animal behavior may arise from a
number of different sources.
Firstly, there may be a need to minimize the direct energetic costs of producing a behavior;
evidence from a range of vertebrate and invertebrate species indicates, for example, that
energy availability may limit the duration of calling behavior (Gillooly & Ophir, 2012; Thomas,
2002; Klump, 2005; Bennet‐Clark, 1998; Fletcher, 1997). Furthermore, Waters & Jones (1995)
showed theoretically that energy
is directly proportional to duration
in acoustic
communication by integrating the energy flux density, as it is well‐known in fluid mechanics,
according to which the sound energy flux (a) is given by the time integral of the squared sound
pressure (Landau & Lifshitz, 1987) and (b) can be roughly approximated by A ∙ t, where t is the
pulse duration and A is its amplitude, for relatively short stimuli, as Baszczyk (2003) explains.
According to general acoustics, the energy of a sound wave is ξ = P ∙ t, where P is its sound
power and t is its duration and P = I ∙ S, where I is the sound intensity and S is the area of the
propagation surface (Kinsler et al., 2000). For a sound wave of amplitude A, the sound energy
becomes (Fahy, 2001, Kinsler et al., 2000):
tP
·
tSI
··
tA
·
2
,
(2)
namely, the energy of a signal depends linearly on time and the squared amplitude. This
suggests a priori it is not only duration that matters but also amplitude. However, amplitude
6
determines the energy of a sound wave at a given time and therefore reducing the amplitude
reduces the reach of the signal due to the natural degradation of the signal with distance, and
interference caused by other sounds (Bennet‐Clark, 1998). Amplitude fluctuations during the
propagation of sound have different effects on receivers, depending on wave frequency, and
interfere with their ability to detect directionality (Wiley & Richards, 1978). Minimizing energy
by amplitude modulation puts the success of communication at risk. Furthermore, amplitude is
highly determined by body size (Bennet‐Clark, 1998; Fletcher, 2004) as Gillooly & Ophir (2010)
demonstrate, because there is a dependency between sound power and amplitude. Therefore,
our energy function ECL is capturing the contribution to sound energy that can be more easily
controlled, namely that of duration.
Secondly, shorter signals may have advantages independent of energetic production costs. In
predation contexts, it may be beneficial for calling prey to decrease conspicuousness to
predators (Ryan et al., 1982; Endler, 1993; Hauser, 1996) or for calling predators to decrease
conspicuousness to prey (Deecke et al., 2005). In addition, shorter signals suffer less from
problems linked to reverberation, an important phenomenon that degrades signals in
environments containing solid structures (Waser & Brown, 1986). For the vocalizations of
rainforest primates, for example, there appears to be an upper call duration limit of 200‐
300ms, to minimize such interference (Brown & Sinott, 2006, p. 191). Interestingly in relation
to this point, the law of brevity has been found in one subset of the repertoire of common
marmosets where all but one call type (the exception being the submissive squeal) are below
the 300ms limit; the other subset, where the law is not found, (a) consists of calls that are
above 300ms in duration and (b) contains all long‐distance communication calls (Ferrer‐i‐
Cancho & Hernández‐Fernández, 2013). If sound pressure, P, falls inversely proportional to the
distance from the sound source (according to the so‐called distance law of sound attenuation),
and sound intensity (and then energy, recall ξ = I ∙ S) falls inversely proportional to the square
of the distance (Landau & Lifshitz, 1987; Kinsler et al 2000) then long‐distance calls must
7
generally show an increase either in sound intensity or duration, or both. Therefore, ECL does
not measure all the costs of a repertoire, as it does not include intensity and sometimes it may
be advantageous to increase duration rather than reducing it. By focusing on ECL rather than on
E, we hope to shed light on the importance of compression in animal behavior.
Here we aim to provide support for the principle of compression (minimization of ECL) as a
general principle of animal behavior. We do not propose that compression is the only principle
of animal behavior, or the only principle by which behavior is optimized. The design of
language is a multiple constraint engineering problem (Evans & Levinson, 2009; Köhler, 2005)
and the same applies to the communication systems of other species (Endler, 1993). The
appearance of design in communication systems can exist without a designer or intentional
engineering (Cornish, 2010, Kirby et al., 2008).
Our notion of principle should not be confused with explanation. Principles are the ingredients
of explanations. Using the physical force of gravity as an example helps to illustrate our notion
of principle. The universal force of gravity explains why objects fall, but when a rocket flies
upwards its movement does not constitute an exception to the force. The force is still acting
and is involved in explaining, for instance, the amount of fuel that is needed to fly in the
opposite direction of the force. As the falling of an object is a manifestation of the force of
gravity, we propose the law of brevity in animal behavior is a manifestation of a principle of
compression. Just as the force of gravity is still acting on a rocket moving away from the Earth,
so we should not conclude prematurely that compression is not a relevant principle of
behavior when the law of brevity is not detected.
Although the Earth and Venus are very different planets, the force of gravity is valid in both
(indeed universal) and physicists only care about the difference in its magnitude. Similarly, we
do not assume that human language and animal behavior cannot have compression in
common, no matter how large the biological, social, cognitive or other differences are. Science
8
is founded on parsimony (Occam’s razor), and from an evolutionary perspective hypothesizing
that the principle of compression is common to humans and other animal species is a priori
simpler than hypothesizing that compression is not shared. We are adopting the perspective of
standard statistical hypothesis testing, where the null hypothesis is that there is no difference
between two populations e.g. two different species (Sokal and Rohlf, 1995). The research
presented here strongly suggests that there is no need to adopt, a priori, the more demanding
alternative hypothesis of an intrinsic difference between humans and other species’ linguistic
and non‐linguistic behavior for the particular case of the dependency between the size or
length of the units and their frequency. However, according to modern model selection
theory, a good model of reality has to comply with a trade‐off between its parsimony and the
quality of its fit to reality (Burnham & Anderson, 2004). A key and perhaps surprising result
that will be presented in this article is that ECL is never found to be significantly high, in spite of
apparently clear advantages in certain situations of increasing signal/behavior length (e.g. for
long distance communication). It could interpreted that even when the direct effect of
compression is not observed, compression still has a role, just as a rocket heading to space is
still being attracted by the Earth’s gravitational field. It is still too early to conclude that
compression has no role even when there is no evidence that ECL is being minimized; it is
important to note that there are serious statistical limits for detecting efficient coding,
especially in small repertoires (Ferrer‐i‐Cancho & Hernández‐Fernández, 2013).
The remainder of the article is organized as follows. Section 2 shows that the energetic cost of
an element (e.g. the length of a word) cannot increase as frequency increases in a system that
minimizes ECL. Thus, the law of brevity is an epiphenomenon of compression. Section 3 shows
direct evidence of compression in animal behavior, in particular by demonstrating that ECL is
significantly small in all the cases where conformity to the law of brevity has previously been
found. Section 4 discusses these findings.
9
2. THE MATHEMATICAL RELATIONSHIP BETWEEN COMPRESSION AND THE LAW OF BREVITY
By definition of pi, one has
p1 ≥ p2 ≥ p3 ≥ ... pn.
If Zipf’s law of brevity was agreed with fully, one should also have
e1 ≤ e2 ≤ e3 ≤ ... en.
(3)
(4)
Here a mathematical argument for Zipf’s law of brevity as a requirement of minimum cost
communication is presented.
Imagine that the law of brevity is not agreed with fully, namely that there exist some i and j
such that i < j, ei > ej., and thus Eq. 4 does not hold perfectly. One could swap the values of ei
and ej. That would have two important consequences. Firstly, the agreement with the law of
brevity (Eq. 4) would increase. Secondly, ECL would decrease as is shown next. To see that ECL is
(cid:1831)(cid:3004)(cid:3013)(cid:3046) (cid:3046) (cid:3404) (cid:1831)(cid:3004)(cid:3013) (cid:3398) (cid:1868)(cid:3036) (cid:1857)(cid:3036)(cid:3398)(cid:1868)(cid:3037) (cid:1857)(cid:3037) (cid:3397) (cid:1868)(cid:3036) (cid:1857)(cid:3037)(cid:3397)(cid:1868)(cid:3037) (cid:1857)(cid:3036) .
reduced by swapping ei and ej, we calculate Es
CL, the mean cost after swapping,
After rearranging Eq. 5, it is obtained that the increment of cost, = Es
CL ‐ ECL is
(5)
= (pi ‐ pj)(ej ‐ ei).
(6)
< 0 means that the cost has reduced i.e. compression has increased. To see that ≤ 0 with
equality if and only if pi = pj, notice that pi ‐ pj ≥ 0 (by definition of pi and pj and i < j) and ej ‐ ei <
0 (recall we are considering the case ei > ej and thus ei = ej is not possible). In sum, a system
that minimizes ECL needs to follow Zipf’s law of brevity (Eq. 4), otherwise there is a pair of
element costs (ei and ej) that can be swapped to reduce ECL. Finally, note that the idea of
swapping of costs was introduced to demonstrate the necessity of the law of brevity (Eq. 4)
10
from the minimization of ECL, not to argue that swapping is the evolutionary mechanism
through which durations or length are optimized according to frequency.
3. DIRECT EVIDENCE OF COMPRESSION
3.1. Methods.
ECL is estimated taking pi as the relative frequency of the i‐th most frequent element of a
sample. If ECL is significantly small in animal behavior that would be a sign of compression to
some degree. Whether ECL is significantly small or not can be determined by means of a
defined as
randomization test (Sokal & Rohlf, 1995, pp. 803‐819). To this end, ECL’, a control ECL, is
(cid:3041)
(cid:1831)(cid:3004)(cid:3013) ′ (cid:3404) (cid:3533) (cid:1868)(cid:3036) (cid:1857)(cid:3095)(cid:4666)(cid:3036)(cid:4667) .
(cid:3036)(cid:2880)(cid:2869)
where π(i) is a permutation function (i.e. a one‐to‐one mapping from and to integers 1,2,...,n).
(7)
The left p‐value of the test is defined as the proportion of permutations where ECL’≤ ECL and the
right p‐value is defined as the proportion of permutations where ECL’ ≥ ECL. The left p‐value is
estimated by TL/T, where T is the total number of random permutations used and TL is the
number of such permutations where ECL’≤ECL. Similarly, the right p‐value is estimated by TR/T,
where TR is the number of random permutations where ECL’ ≥ ECL. T=105 uniformly random
permutations were used.
3.2. Data
We adopt the term type for referring not only to word types, but also to types of vocalization
and behavioral patterns. We used a dataset that comprises type frequency and type
size/length/duration from the following species: dolphins (Tursiops truncatus), humans (Homo
sapiens sapiens), Formosan macaques (Macaca cyclopis), common marmosets (Callithrix
jacchus), golden‐backed uakaris (Cacajao melanocephalus) and common ravens (Corvus corax).
11
For humans, seven languages are included: Croatian, Greek, Indonesian, US English, Russian,
Spanish and Swedish. The data for dolphins come from Ferrer‐i‐Cancho and Lusseau (2009), for
human languages from Ferrer‐i‐Cancho and Hernández‐Fernández (2013), for Formosan
macaques from Semple et al. (2010), for common ravens from Conner (1985) and finally, those
for common marmosets and golden backed uakaris come from Bezerra et al. (2011).
The mean duration of a call type (e in our notation) is defined as e = D/f where f is the number
of times that the call has been produced and D is total duration (D is the sum of all the
durations of that call). For common marmosets, D defines two partitions of the repertoire: the
low D partition, where the law of brevity is found, and the high D partition, where the law is
not found (further details on this partitioning are given in Ferrer‐i‐Cancho & Hernández‐
Fernández, 2013). For dolphins, the definition of the size of a behavioral pattern in terms of
elementary behavioral units ‐ see Ferrer‐i‐Cancho & Lusseau, (2009) for a description of these
elementary units ‐ is subject to some degree of arbitrariness. For this reason, two variants of
behavioral pattern size are considered: one where the elementary behavioral unit “two” is not
used and another where elementary unit “stationary” is not used. A summary of the features
of the dataset is provided in Table 1.
Table 1 summarizes the results of the randomization test in all the species where the law of
brevity has been studied so far with the exception of chick‐a‐dee calls (Hailman et al., 1985),
for which data are not available for reanalysis. The conclusions reached by Hailman et al.
(1985) on chick‐a‐dees, namely that the law of brevity holds in bouts of calls but does not (at
least not as clearly) in individual calls must be interpreted carefully. Hailman et al.’s (1985)
analysis relies on visual inspection of data, which is highly subjective: visual conformity with
the law of brevity is clear for bouts of calls but disagreement with the law in individual calls is
not obvious. Hailman et al. (1985) did not perform a correlation test to test for conformity to
the law of brevity; by contrast, we are determining if the law holds by means of a correlation
12
test between frequency and length/duration. Even raw plots of frequency versus length in
human language show substantial dispersion, as shown for example in Fig. 1(a) of Ferrer‐i‐
Cancho & Lusseau (2009).
13
Table 1. Summary of the features of the dataset and the results of the compression test. Units:
s stands for seconds; char stands for characters; e.b.u. stands for elementary behavioral units.
ECL was rounded to leave two significant digits. When not bounded, estimated left and right p‐
values were rounded to leave only one significant digit. Conformity to the law of brevity means
that the correlation between frequency and size/duration is negative and significant.
Species
Behavior
Golden‐backed
uakaris
Common
marmosets
Common
ravens
Dolphins
Formosan
macaques
Humans
Vocalizations
Vocalizations
Vocalizations
(low D cluster)
Vocalizations
(high D cluster)
Vocalizations
Surface
behavioral
patterns
Id. without
“stationary”
Id. without “two”
Vocalizations
Greek
Russian
Croatian
Swedish
US English
Spanish
Indonesian
Conformity to
law of brevity
No (Bezerra et
al., 2011)
No (Bezerra et
al., 2011)
Yes (Ferrer‐i‐
Cancho &
Hernández,
2013)
No (Ferrer‐i‐
Cancho &
Hernández,
2013)
No (Ferrer‐i‐
Cancho &
Hernández,
2013)
Yes (Ferrer‐i‐
Cancho &
Lusseau, 2009)
Yes (Semple et
al., 2010)
Yes
(Ferrer‐i‐Cancho
& Hernández‐
Fernández,
2012)
n
7
12
6
6
18
31
31
31
35
4203
7908
15381
19164
24101
27478
30461
ECL
0.14 s
0.46 s
left p‐
value
0.09
right p‐
value
0.9
0.5
0.5
0.048 s
0.006
1
0.59 s
0.1
0.9
0.25 s
0.3
0.6
1.3 e.b.u.
0.0002
1
1.2 e.b.u.
0.001
1.3
e.b.u.
0.18 s
3.9
char
4.6 char
3.9 char
3.2 char
3.8 char
3.9 char
4.2 char
0.002
0.008
<10‐5
<10‐5
<10‐5
<10‐5
<10‐5
<10‐5
<10‐5
1
1
1
>1‐10‐5
>1‐10‐5
>1‐10‐5
>1‐10‐5
>1‐10‐5
>1‐10‐5
>1‐10‐5
14
4. DISCUSSION
Table 1 presents a number of interesting results. First, ECL was significantly small in all cases
where conformity to the law of brevity (a significant negative correlation between frequency
and size/length/duration) had been reported. This provides further support for the intimate
relationship between the
law of brevity and the
information theoretic principle of
compression. Second, there is no evidence of redundancy maximization, the opposite of
compression, in the species that we have analyzed: ECL was never significantly high. This result
was unexpected for two reasons: pressure for compression is not the only factor shaping
repertoires (Bezerra et al., 2011; Ferrer‐i‐Cancho & Hernández‐Fernández, 2013; Endler, 1993)
and pressure for compression can cause a signal to be more sensitive to noise in the
environment. Coding theory indicates that redundancy must be added in a controlled fashion
to combat transmission errors (Cover & Thomas, 2006, pp. 184). In cases where signals can be
mistaken for each other, for example due to noise in the environment, forming words by
stringing together subunits (e.g. combining 'phonemes' into 'words') allows a system to
increase its capacity to communicate (Plotkin & Nowak, 2000). Therefore, redundancy
maximization is a conceivable alternative to compression (redundancy minimization).
A range of phenomena or situations may select against compression in signals. Firstly, as
mentioned above, environmental noise may drive signals in the opposite direction to that of
code minimization. Elongation of signals is one of a number of adaptations to noise, and this is
known as the Lombard effect in human language and the communication systems of other
species (Zöllinger & Brumm, 2011; Brumm & Zöllinger, 2011); another major strategy to
combat noise is to increase redundancy, in the temporal or spatial organization of signals
(Richards & Wiley, 1980; Ay et al., 2007). Secondly, capacity for compression may be
constrained in some species more than others, due to biological features of the species
(Gillooly & Ophir, 2010) or its environment (Wiley & Richards, 1978). For example, the ability
15
to use low‐frequency signals to reduce attenuation and thus maximize transmission success is
not always possible for smaller animals, due to allometric constraints that limit the frequency
(i.e. pitch) of a signal as a function of an animal’s body mass (Fletcher, 2004). The same applies
to maximizing transmission success by increasing amplitude (Brumm & Slabbekoorn, 2005): a
species with a small body cannot produce a sound with high amplitude (Bennet‐Clark, 1998;
Fletcher, 2004) because amplitude is limited by body mass (Gillooly & Ophir, 2010). Thirdly,
reverberation is one of the fundamental problems posed by the forest environment, and one
way to overcome this is to package the overall signal into brief pulses that end prior to the
delay time expected for the main first reflection of the pulse (Brown & Sinott, 2006, p.191). A
continuous signal of duration t could be converted into a package of duration t + u, where u is
the total duration of the inter‐pulse silences. Finally, in relation to human language, elongation
occurs in the context of adults’ child‐directed speech, which tends to be slower, with syllable
lengthening, longer pauses, etc. (Saxton, 2010, p. 81 and references therein).
When considering these processes that may drive an increase in signal size, it is important to
note that if elongation concerned all the types of elements equally (by a constant
proportionality factor), then the law of brevity would remain. This suggests that the law of
brevity can only be violated if elongation concentrates on a subset of the repertoire.
Consistent with this idea, common marmoset vocalizations do not conform to the law at the
level of the whole repertoire, but do within a subset of the repertoire where none of the long‐
distance calls ‐ calls highly sensitive to noise in the environment and other signal distorting
factors ‐ is found (Ferrer‐i‐Cancho & Hernández‐Fernández, 2013). This finding suggests that
analysis of logically selected subsets of the vocal repertoire of ravens and golden‐backed
uakaris, two species where conformity to the law has not been found, is worthwhile.
If mean length maximization plays a role in the species we have examined, it is not strong
enough to surface in a statistical test. It is possible that these species introduce redundancy at
16
other levels: below the type of element level, i.e. within the shape or structure of the element,
or above the element level, i.e. in the way sequences of elements are constructed. It is also
possible that individuals can achieve a similar goal to redundancy by varying instead call
amplitude (e.g., Brumm & Slabbekoorn, 2005). The trade‐off between compression (mean
length minimization) and the need for successful signal transmission, where new element type
formation (e.g. words) by combining elements (e.g. ‘phonemes’) is a well‐known strategy
(Plotkin & Nowak, 2000), may explain why a clear manifestation of length maximization is
difficult to find at our level of analysis.
The existence of true universals in the sense of exceptionless properties is a matter of current
debate in human language (Evans & Levinson, 2009) and animal behavior (Bezerra et al., 2011;
Ferrer‐i‐Cancho & Hernández‐Fernández, 2013). We believe it is important to investigate the
real scope of statistical patterns of language such as the law of brevity in world languages and
animal behavior. The analysis of exceptions to the law of brevity is full of subtle statistical and
biological details (Ferrer‐i‐Cancho & Hernández‐Fernández, 2013; Bezerra et al., 2011).
However, it is of even greater theoretical importance to investigate which are the universal
principles of behavior. They are the arena where unification (universality) and the complex
nature of reality, including exceptions to language patterns or peculiar situations where certain
language patterns emerge, may reconcile. We are not claiming that the law of brevity is a
hallmark of language (as opposed to simpler forms of communication), but an example of
constrained or convergent evolution by an abstract principle of compression acting upon
communicative and non‐communicative behavior and reflecting selection for efficiency of
coding. This principle provides an avenue for the optimization of the behavior of a species. The
exact nature of the solutions adopted may depend on the chances and needs the species
encounters during its evolutionary history (Monod, 1972).
ACKNOWLEDGEMENTS
17
We are grateful to S. E. Fisher, S. Kirby, Rick Dale and three anonymous reviewers for helpful
remarks. This work was supported by the grant Iniciació i reincorporació a la recerca from the
Universitat Politècnica de Catalunya and the grants BASMATI (TIN2011‐27479‐C04‐03) and
OpenMT‐2 (TIN2009‐14675‐C03) from the Spanish Ministry of Science and Innovation.
REFERENCES
Ay, N., Flack, J. & Krakauer, D.C. (2007). Robustness and complexity co‐constructed in
multimodal signalling networks. Philosophical Transactions of the Royal Society of London B
362 (1479), 441–447.
Bennet‐Clark, H.C. (1998). Size and scale effects as constraints in insect sound communication.
Philosophical Transactions of the Royal Society of London B, 353, 407‐419.
Baszczyk, J.W. (2003). Startle response to short acoustic stimuli in rats. Acta Neurobiologica
Experimentalis, 63: 25‐30.
Bezerra, B. M., Souto, A. S., Radford, A.N. & Jones, G. (2011). Brevity is not always a virtue in
primate communication. Biology Letters, 7, 23‐25.
Brown, C. H. & Sinnot, J. M. (2006). Cross‐species comparisons of vocal perception. In: S.
Greenberg & W. A. Ainsworth (Eds.), Listening to speech: an auditory perspective (pp. 183‐
201). London: Routledge.
Brumm, H. & Slabbekoorn, H. (2005). Acoustic communication in noise. Advances in the Study
of Behavior 35, 151‐ 209.
Brumm, H. & Zöllinger, S. A. (2011). The evolution of the Lombard effect: 100 years of
psychoacoustic research. Behaviour 148, 1173‐1198.
Burnham, K. P., & Anderson, D. R. (2002). Model selection and multimodel inference: a
practical information‐theoretic approach, 2nd Ed. New York: Springer‐Verlag.
18
Conner, R. N. (1985). Vocalizations of common ravens in Virginia. Condor, 87, 379‐388.
Cornish, H. (2010). Investigating how cultural transmission leads to the appearance of design
without a designer in human communication systems. Interaction Studies, 11(1), 112‐137.
Cover & T. M. & Thomas, J. A. (2006). Elements of information theory (2nd edition). Hoboken,
NJ: Wiley.
Deecke, V. B., Ford, J. K. B. & Slater, P. J. B. (2005). The vocal behaviour of mammal‐eating
killer whales (Orcinus orca): communicating with costly calls. Animal Behaviour, 69, 395‐405.
Endler, J. A. (1993). Some general comments on the evolution and design of animal
communication systems. Philosophical Transactions of the Royal Society of London B, 340, 215‐
225.
Evans, N., & Levinson, S. C. (2009). The myth of language universals: Language diversity and its
importance for cognitive science. Behavioral and Brain Sciences, 32, 429‐492.
Fahy, F.(2001). Foundations of Engineering Acoustics. Oxford: Elsevier.
Ferrer‐i‐Cancho, R. & Elvevåg, B. (2010). Random texts do not exhibit the real Zipf's law‐like
rank distribution. PLoS ONE, 5 (3), e9411.
Ferrer‐i‐Cancho, R., Forns, N., Hernández‐Fernández, A., Bel‐Enguix, G. & Baixeries, J. (2013).
The challenges of statistical patterns of language: the case of Menzerath's law in genomes.
Complexity, in press. doi: 10.1002/cplx.21429.
Ferrer‐i‐Cancho, R. & Gavaldà, R. (2009). The frequency spectrum of finite samples from the
intermittent silence process. Journal of the American Society for Information Science and
Technology 60 (4), 837‐843.
19
Ferrer‐i‐Cancho, R. & Hernández‐Fernández, A. (2013). The failure of the law of brevity in two
New World primates. Statistical caveats. Glottotheory, 4(1), in press.
Ferrer‐i‐Cancho, R. & Lusseau, D. (2009). Efficient coding in dolphin surface behavioral
patterns. Complexity, 14, 23‐25.
Ferrer‐i‐Cancho, R. & McCowan, B. (2012). The span of correlations in dolphin whistle
sequences. Journal of Statistical Mechanics, P06002.
Fletcher, N.H. (1997). Sound in the Animal World, Acoustics Australia, 25, 2‐69.
Fletcher, N. H. (2004). A simple frequency‐scaling rule for animal communication, Journal of
the Acoustical Society of America, 115, 2334–2338.
Gillooly, J . F. & Ophir, A. G. (2010). The energetic basis of acoustic communication.
Proceedings of the Royal Society B, 277, 1325‐1331.
Gintis, H. (2007). A framework for the unification of the behavioral sciences. Behavioural and
Brain Sciences, 30, 1‐16.
Gleick, J. (2011). The Information: A History, a Theory, a Flood. New York: Pantheon Books,
Random House.
Grafen, A. (2007). The formal Darwinism project: a mid‐term report. Journal of Evolutionary
Biology, 20, 1243‐1254.
Hailman, J. P., Ficken, M. S. & Ficken, R. W. (1985). The ‘chick‐a‐dee’ calls of Parus atricapillus:
a recombinant system of animal communication compared with written English. Semiotica 56,
191–224.
Hauser, M. D. (1996). The Evolution of Communication. Cambridge, Massachusetts: MIT Press.
20
Hernández‐Fernández, A., Baixeries, J., Forns, N. & Ferrer‐i‐Cancho, R. (2011). Size of the whole
versus number of parts in genomes. Entropy, 13, 1465‐1480.
Kinsler, L.E., Frey, A.R., Coppens, A.B., & Sanders.,J.V.(2000). Fundamentals of Acoustics.
Hoboken, NJ: John Wiley & Sons, Inc. Fourth Edition.
Kirby, S., Cornish, H., & Smith, K. (2008). Cumulative cultural evolution in the laboratory: An
experimental approach to the origins of structure in human language. Proceedings of the
National Academy of Sciences, 105(31), 10681‐10686.
Klump, G. (2005). Evolutionary adaptations for auditory communication. In: J. Blauert (Ed.),
Communication Acoustics (pp. 27‐46) Berlin: Springer‐Verlag.
Köhler, R. (1987). Systems theoretical linguistics. Theoretical Linguistics, 14, 241‐257.
Köhler, R. (2005). Synergetic Linguistics. In: R. Köhler, G. Altmann & R. G. Piotrowski (Eds.),
Quantitative Linguistik. Ein
internationales Handbuch. Quantitative Linguistics: An
International Handbook (pp. 760‐775) Berlin, New York: Walter de Gruyter.
Landau, L.D. & Lifshitz, E.M. (1987) Fluid Mechanics. Oxford: Pergamon Press. Second Edition.
McCowan, B., Doyle, L. R. and Hanser, S. F. (2002). Using information theory to assess the
diversity, complexity and development of communicative repertoires. Journal of Comparative
Psychology 116, 166‐172.
Monod, J. (1972). Chance and Necessity. Knopf: London.
Plotkin, J.B. & Nowak, M.A. (2000) Language evolution and information theory. Journal of
Theoretical Biology, 205, 147‐159.
Richards, D. G., and R. H. Wiley (1980). Reverberations and amplitude fluctuations in the
propagation of sound
in a forest:
implications for animal communication. American
Naturalist, 115, 381‐399.
21
Ryan, M. J., Tuttle, M. D. & Rand, A. S. (1982). Bat predation and sexual advertisement in a
Neotropical frog. American Naturalist, 119, 136–139.
Saxton, M. (2010). Child language. Acquisition and development. Los Angeles, US: Sage.
Semple, S., Hsu, M. J. & Agoramoorthy, G. (2010). Efficiency of coding in macaque vocal
communication. Biology Letters, 6, 469‐471.
Sokal, R. R. & Rohlf, F. J. (1995). Biometry: the principles and practice of statistics in biological
research. New York: Freeman.
Strauss, U., Grzybek, P. & Altmann, G. (2007). Word length and word frequency. In P. Grzybek
(Ed.), Contributions to the science of text and language (pp. 277–294) Dordrecht: Springer.
Sumpter, D. J. T. (2006) The principles of collective animal behaviour. Philosophical
Transactions of the Royal Society of London B: Biological Sciences, 361, 5–22.
Thomas, R. J. 2002. The costs of singing in nightingales. Animal Behaviour, 63, 959‐966.
Waser, P. M. & Brown, C. H. (1986). Habitat acoustics and primate communication. American
Journal of Primatology, 10, 135‐154.
Waters D. A., & Jones G. (1995). Echolocation call structure and intensity in five species of
insectivorous bats. The Journal of Experimental Biology, 198, 475–489.
Wiley R.H. & Richards, D. G. (1978). Physical constraints on acoustic communication in the
atmosphere: implications for the evolution of animal vocalizations. Behavioral Ecology and
Sociobiology, 3, 69‐94.
Zipf, G. K. (1935). The psycho‐biology of language: an introduction to dynamic philology.
Cambridge, MA: MIT Press.
22
Zipf, G.K. (1949). Human behaviour and the principle of least effort. Cambridge, MA: Addison‐
Wesley.
Zöllinger, S. A. & Brumm, H. (2011). The Lombard effect. Current Biology, 21, R614‐615.
23
|
1208.2100 | 1 | 1208 | 2012-08-10T06:54:09 | Input Statistics and Hebbian Crosstalk Effects | [
"q-bio.NC",
"math.DS"
] | As an extension of prior work, we study inspecific Hebbian learning using the classical Oja model. We use a combination of analytical tools and numerical simulations to investigate how the effects of inspecificity (or synaptic "cross-talk") depend on the input statistics. We investigated a variety of patterns that appear in dimensions higher than 2 (and classified them based on covariance type and input bias). The effects of inspecificity on the learning outcome were found to depend very strongly on the nature of the input, and in some cases were very dramatic, making unlikely the existence of a generic neural algorithm to correct learning inaccuracy due to cross-talk. We discuss the possibility that sophisticated learning, such as presumably occurs in the neocortex, is enabled as much by special proofreading machinery for enhancing specificity, as by special algorithms. | q-bio.NC | q-bio |
Input Statistics and Hebbian Crosstalk Effects
Anca Radulescu1
May 21, 2018
1Department of Mathematics, 395 UCB, University of Colorado, Boulder, 80309-0395
phone: (303)492-6617; fax: (303)492-7707; email: [email protected]
Abstract
As an extension of prior work, we study inspecific Hebbian learning using the classical Oja model. We use
a combination of analytical tools and numerical simulations to investigate how the effects of inspecificity
(or synaptic "cross-talk") depend on the input statistics. We investigated a variety of patterns that
appear in dimensions higher than 2 (and classified them based on covariance type and input bias). The
effects of inspecificity on the learning outcome were found to depend very strongly on the nature of the
input, and in some cases were very dramatic, making unlikely the existence of a generic neural algorithm
to correct learning inaccuracy due to cross-talk. We discuss the possibility that sophisticated learning,
such as presumably occurs in the neocortex, is enabled as much by special proofreading machinery for
enhancing specificity, as by special algorithms.
Keywords: Hebbian learning, cross-talk, biased input statistics, negative correlation, spectrum, n-dimensional
dynamics, bifurcation.
1
Introduction
In this paper, we revisit some fundamental questions in computational neuroscience, related to unsupervised
learning in cortical networks. We use a simple model of learning (previously studied by the author) to
study how learning occurs when the model incorporates transmission inspecificity ("synaptic errors"). We
focus in particular on a few crucial questions: To what extent and under which circumstances can synaptic
inspecificity facilitate or prevent learning? Are certain input distributions more easily learned than others,
or more affected by inspecificity? Can a small level of cross-talk induce significant changes (bifurcations)
in the system's asymptotic dynamics?
The paper is organized as follows.
In the introduction, we present the model (which we will call
throughout the paper the "inspecific Oja model") and we overview the basics of its dynamic behavior
(Section 1.1). In Section 2 we start by investigating numerically how a 3-dimensional Oja inspecific network
processes different classes of input distributions, preserving some of the dynamical aspects found in the
2-dimensional phase plane [12], but also introducing new features, specific to higher dimensions. In Section
3, we study analytically, in an n-dimensional example, the behavior observed numerically in the previous
section. The Section 4 interprets the numerical and analytical results in the biological context of a learning
cortical network.
1
1.1 The inspecific Oja model
Oja [11] showed that a simple neuronal model could perform unsupervised learning based on Hebbian
synaptic weight updates incorporating an implicit "multiplicative" weight normalization, to prevent un-
limited weight growth [10].
Our focus is on studying a single-output network, learning an input distribution according to Oja's
rule [11]. More precisely, the output neuron receives, through a set of n input neurons, n signals x =
(x1, ..., xn)T drawn from an input distribution P(x), x ∈ Rn, transmitted via synaptic connections of
strengths ω = (ω1, ..., ωn)T . The resulting scalar output y is generated as the weighted sum of the inputs
y = xT ω. The synaptic weights ωi are modified by implementing first a Hebb-like strengthening pro-
portional to the product of xi and y , followed by an approximate "normalization" step, maintaining the
Euclidean norm of the weight vector close to one. The input covariance matrix C = xT x can be used as an
appropriate long-term characterization of the inputs, to study the expected long-term convergence of the
weight vector (i.e., learning), by approximating it with the asymptotic behavior of w(t) = (cid:104)ω(t + 1)ω(t)(cid:105)
in the equation [13]:
= γ(cid:2)Cw −(cid:0)wT Cw(cid:1) w(cid:3)
dw
dt
Oja [11] showed that this simple model acts, when applicable, as a principal component analyzer for the
input distribution. Finding principal components could be very useful in the brain for data compression
and transmission, since for Gaussian data such representations have statistically optimal properties, and
often neural signals are approximately Gaussian.
Recent data suggest [9, 4, 5] that weight updates may be affected by each other, for example due to
unavoidable residual second messenger diffusion between closely spaced synapses. In our recent work we
examined how such "crosstalk" would affect the Oja model [13]. We formalized learning inspecificity via
an error matrix E ∈ Mn(R) that has positive entries, is symmetric and equal to the identity matrix
I ∈ Mn(R) in case the error is zero. Consistently with our previous studies in both two [12] and higher
dimensions [13], we consider the error matrix of the form
···
···
. . .
···
q
...
q
E =
(1)
q
...
1
n
where 0 < <
(n − 1) = 1. The inspecific learning equations become:
is the "transmission error" and
1
n
< q < 1 is the "transmission quality," satisfying q +
= γ[ECw − (wT Cw)w]
(2)
dw
dt
We have noted previously that an equilibrium for this system is any vector w = (w1...wn)T such that
ECw = (wT Cw)w, i.e., an eigenvector of EC (with corresponding eigenvalue λw), normalized, with
respect to the norm (cid:107)·(cid:107)C =(cid:112)(cid:104)·,·(cid:105)C (defined as (cid:104)v, u(cid:105)C = vT Cu, for all u, v ∈ Rn), so that (cid:107)w(cid:107)C = λw.
Generically, EC has a strictly positive, unique maximal eigenvalue, and the corresponding eigendirection
is orthogonal in (cid:104)(cid:104)·,·(cid:105)(cid:105)C to all other eigenvectors of EC.
We have also shown that the eigenvalues of the Jacobian matrix at an equilibrium w are given by
−2γλw and −γ[λw − λvj], where λw and λvj,∀j = 1, n − 1 are the n eigenvalues of EC (noting first that
Bw = {w, v1, ...vn−1}, the completion of w to a basis of eigenvectors of EC, orthogonal with respect to
the dot product (cid:104)·,·(cid:105)C, also forms an eigenvector basis for the Jacobian). We concluded that, if EC has a
2
v + δ1
±c
...
±c
C =
±c
v + δ2
±c
···
···
. . .
···
±c
±c
...
v + δn
unique largest eigenvalue, then a normalized eigenvector w is a local hyperbolic attracting equilibrium for
(2) iff it corresponds to the maximal eigenvalue of EC. If EC has a multiple largest eigenvalue, the system
will exhibit a set of nonisolated, neutrally attracting equilibria (all normalized eigenvectors spanning the
principal eigenspace, in this case of dimension ≥ 2). Some of the computational details are summarized
in Appendix 1 (e.g., a description of the attraction basins, supporting the absence of cycles in the phase
space) and further expanded in our previous work [13, 12].
Since the nature and position of the equilibria depends on the spectral properties of EC, the next
task is, naturally, to study the spectral changes of EC when perturbing the system by increasing the
transmission inspecificity. In our previous work on the model, we found that the effects of perturbations on
the system's dynamics can depend very strongly on the characteristics of the input distribution (correlation
sign, degree of bias). In our first study we only considered learning of positively correlated n-dimensional
input distributions, and found a smooth degrading of the learning outcome with increasing error, but no
sudden changes in dynamics [13]. In our second study, we discovered that negatively correlated inputs can
induce a bifurcation (stability swap of equilibria, through a critical stage) when increasing the error, even
in as simple as a two-dimensional system; this bifurcation only occurred, however, in the case of unbiased
inputs [12].
Here, we want to extend this work and investigate the effects of cross-talk in higher dimensional
networks, when learning a variety of classes of input distributions, both biased and unbiased. More
precisely, we will consider as potential covariance matrices all combinations of the form:
(3)
where we can assume without loss of generality that δ1 ≥ δ2 ≥ . . . ≥ δn ≥ 0. For any k ≤ n, we will say
that the input has bias loss of order k if δ1 = . . . = δk. We hypothesize that, even though the background
covariance ±c is taken for simplicity to be uniform in absolute value, the inspecific learning rule will lead
to interesting dynamics, in particular when the inputs exhibit some degree of cross-correlation.
Since our analysis will focus on symmetric matrices C with possible off-diagonal elements, we have to
first ask whether/when such a matrix can constitute the covariance matrix of a distribution of n-dimensional
vectors. While establishing equivalent conditions may be difficult even for small dimensions [14], a simple
sufficient criterion valid for any dimension is diagonal dominance. It is known that a symmetric diago-
nally dominant matrix with real, non-negative diagonal entries is positive semi-definite, hence implicitly a
covariance matrix, from the finite-dimensional case of the spectral theorem1. If we are willing to impose
v + δn > (n − 1)c as a (biologically plausible) upper bound on how large the input cross-correlations c
can be with respect to the auto-correlations v, diagonal dominance clearly follows, and C is thus automat-
ically guaranteed to be a covariance matrix. An interesting direction would be to interpret biologically the
significance of an n-dimensional input distribution with negative correlations [7]; this question is, however,
beyond the scope of this paper.
2 Classes of inputs and bias effects on 3-dimensional dynamics
We study here how input patterns affect the effects of inspecificity in driving the dynamics of a 3-
dimensional network – the lowest dimension for which the question applies, and which captures the essence
of this behavior even in higher-dimensional systems.
In this section, we will inspect all combinatorial
possibilities of cross-correlation sign and auto-correlation bias, and determine the effect of increasing error
1If X is an n × 1 column vector-valued random variable whose covariance matrix is the n × n identity matrix. Then
√
CX) =
cov (
√
C = C.
C cov(X)
√
3
on the dynamics in each case. In the next section, we will support with some rigorous proofs the main
results obtained here through numerical simulations (we used the Matlab software, version 7.2.1).
2.1 Input covariance patterns
We studied separately all combinatorial possibilities for the input statistics, with uniform absolute value
covariance; in other words, we considered covariance matrices of the form:
v + δ1
±c
±c
C =
±c
±c
v + δ2 ±c
±c
v
(4)
where, as before, v > 2c, and δ1 ≥ δ2 ≥ 0 (i.e., allowing bias of any order).
Let's first note that, based on the number of negative upper-diagonal entries of C, we distinguish 4
combinatorial classes: (A) all positive covariance (one configuration), (B) one negative entry (3 configura-
tions), (C) two negative entries (3 configurations), (D) and all negative entries (one configuration). We will
study the spectra of the inspecific matrices EC, and the differences that occur in these when considering
different classes of C, as well as different degrees of bias: from fully biased (δ1 > δ2 > 0) to partly biased
(δ1 = δ2 > 0) to fully unbiased (δ1 = δ2 = 0). In this section, we illustrate the behavior of the eigenvalues
of EC as the quality q is changing in the interval (1/3, 1] (representing quality higher than error).
For fully biased inputs (δ1 > δ2 > 0), the behavior is indistinguishable between classes2: the largest
eigenvalue remains separated from the second largest for the whole range of q (as shown for one example
in Figure 1a), determining the eigenvector to gradually drift from the direction of the principal component
of C (blue curve in Figures 1a). For any value of q, the system has two hyperbolically attracting equilibria
(the normalized principal eigenvectors of EC, whose basins are separated by an invariant plane). In Figure
2 we show the evolution of a set of trajectories, to illustrate convergence to the two attractors in the phase
space, as well the dynamics within the separating plane. (We will encounter similar behavior for other
classes of inputs as well, for which we will refer to the same Figure, since the same phase space evolution
remains a qualitatively accurate depiction.)
For loss of bias of order one (δ1 = δ2 > 0), we distinguish three possible types of behavior.
Separated leading eigenvalues (behavior "typical" to class A, as illustrated in Figure 1b). This corre-
sponds to a slow depreciation of the learned vector as q decreases (blue curve in Figure 1b). The phase
space behavior resembles qualitatively that for unbiased inputs, illustrated in Figure 2.
Crossing of leading eigenvalues (behavior "typical" to class D, as illustrated in Figure 1c and fur-
ther discussed in Section 3), which produces a sudden swap of the attractors from one eigendirection to
another, orthogonal, one (phenomenon we have described previously in a 2-dimensional model [12]). This
corresponds to a crash in learning at a critical value of the quality q∗ (which depends on parameters as
q∗ = v+δ+c
v+δ−c ). Low inspecificity (q > q∗) has in fact no effect on learning in this case: although the leading
eigenvalue changes, the principal direction remains the same, so the system will converge generically to the
same outcome as in the absence of error. This may seem like a very desirable input distribution to learn
in the presence of inpecificity; however, one has to keep in mind that, if the cross-correlations are small in
absolute value c, then q∗ will be very close to 1. Such perfect learning will therefore only happen when
inspecificity is almost insignificantly small. The more disturbing this becomes, when we recall that at the
2Since the spectra depend qualitatively on all parameter values, we present here the results of a numerical investigation,
rather than an rigorous analytical study, which would be extremely cumbersome. In contrast, we will later prefer an analytical
approach to the classification in fully unbiased case, where the computations become more tractable.
4
Figure 1: Spectral changes induced by increasing inspecificity, for various inputs schemes. In all panels
we show, with respect to the quality q = 1 − 2: the evolution of the eigenvalues, with color-coding black for largest
eigenvalue, red for the second largest and green for the lowest (top subplot); the angle between the inspecific stable
vector and the correct attracting direction(s) (bottom subplot). In all panels, v = 1, c = 0.2. The classification is as
follows: A. For fully biased inputs (δ1 = 2, δ2 = 1), the three eigenvalues remain separated. For partly-biased inputs
(δ1 = δ2 = 1), there are three cases, depending on the number of negative cross-correlations and on their placement:
the leading eigenvalues can remain separated (B), they can cross at a critical values of q = q∗ (C), or they can
approach significantly for some value of q, but "avoid" crossing (D). For fully unbiased inputs, we found four cases,
classified simply by the number of negative off-diagonal cross-correlations (and not by their geometry): all positive
cross-correlations - leading eigenvalues remain separated (E); one negative cross-correlation - leading eigenvalues
only coincide at q = 1, and immediately separate (F); two negative cross-correlations - leading eigenvalues may
approach each other, in an avoided crossing of magnitude depending on parameters, but remain separated (G); all
negative cross-correlations - leading eigenvalues coincide on a whole interval, as quality depreciates from q = 1 to a
critical value. In this case, the system has a curve of half-neutral attractors, which persists until q reaches the critical
value, when a different, orthogonal, eigenvector takes over as stable direction.
5
Figure 2: Phase space trajectories for fully biased inputs. A. In the absence of error, the system converges
generically to the two normalized vectors in the principal direction wC of the covariance matrix C. The attraction
basins are separated by the subspace (cid:104)w, wC(cid:105) = 0 (the shaded plane). B. For error = 0.2, the system converges
generically to the two normalized vectors in the principal direction wEC of the modified covariance matrix EC. The
attraction basins are separated by the subspace (cid:104)w, wEC(cid:105) = 0 (the shaded plane). Parameters used: v = 1, c = 0.2,
δ1 = 2, δ2 = 1. Color coding: trajectories evolve in time from darker towards lighter shades.
end of the "good" interval lies the bifurcation, crashing the equilibria to a completely irrelevant direction;
so any fault of the system in the direction of slightly miscalculating the limits for the permissible error,
would have dire consequences. If the network does not have an additional, good estimator of its degree
of inspecificity, it may not only learn an irrelevant outcome, but also have no knowledge of it. In Figure
3, we represent three phase space plots: before, at and after the bifurcation point. While Figures 3a and
3c illustrate the typical phase space with two hyperbolically stable equilibria (one representing accurate,
error-free learning, the other – inaccurate learning for a post-critical error), the bifurcation phase space
is qualitatively different: the system has no hyperbolic attractors, but rather a closed curve (ellipse) of
half-stable equilibria (neutral along the direction of the curve). Clearly, the outcome of learning is in this
case extremely dependent on the initial conditions (although, as commented in previous work [12], the
stochastic version of the system will rather have noise-driven stationary solutions that drift around this
attracting ellipse).
If this phase-plane dynamics were specific only to this bifurcation, one may find it
justified to overlook its occurrence in the context of generic dynamics. However, this is not the case; as
shown below, there are classes of inputs for which such a attracting-ellipse slice represents the natural state
of the system, and persists for an entire inspecificity range.
"Avoided crossing" of the eigenvalues (a hybrid behavior observed in mixed cases from classes B and
C, in which the eigenvalues approach, without actually crossing, at a value q = q∗, which depends on all
parameter values). While the principal eigenvectors never swap in this case, the learning has a significantly
rapid depreciation around q∗ (see blue curve in Figure 1d).
For bias loss of order two, the computations are much simplified by the absence of bias, so we can carry
out analytically a complete classification. The result is presented concisely in Theorem 2.1, explained in
more detail in the proof of the theorem, then further interpreted for the remainder of the section.
Theorem 2.1. For order two input bias δ1 = δ2 = 0, the dynamic behavior of the system is classified by
the classification of the input covariance sign, A-D.
Proof. For order two input bias, all classes A-D can be generated from three symbolic structures:
6
Figure 3: Bifurcation in attractor dynamics for partly biased inputs, all negative cross-correlations.
A. For small error, the attractors (the two normalized principal eigenvectors of EC) don't differ too much from
the correct attractors (the two normalized principal eigenvectors of EC). The attraction basins are separated by the
subspace (cid:104)w, wC(cid:105) = 0 (the shaded plane). B. For critical error = −c
v+δ−c , the system exhibits an ellipse of neutrally-
stable equilibria (yellow curve contained in the shaded plane). C. For error past the critical value, the attractors have
moved significantly far from the correct positions. Parameters used: v = 1, c = 0.2, δ = δ1 = δ2 = 1. Color coding:
trajectories evolve in time from darker towards lighter shades.
v
c
c v
c
c
c
c v
, C2 =
v −c
−c
c
v
c
c
c
v
C1 =
and C3 =
v
c −c
c
v
v
c
c
−c
Class A represents Structure C1 with c > 0, and Class D, represents Structure C1 with c < 0. Class B
can be obtained from Structures C2 and C3 with c > 0, while Class C can be obtained from Structures
C2 and C3 with c < 0.
Computing directly the spectrum for C1, we get one simple eigenvalue ξ1 = v + 2c (whose eigenvector
is also error-independent) and one double eigenvalue ξ2 = (1 − 3e)(v − c). If c > 0 (Class A), ξ1 always
dominates (Figure 1e). If c < 0 (Class D), the double eigenvalue ξ2 = (1 − 3e)(v − c) takes over for error
smaller than the critical value <
(Figure 1h).
−c
v − c
Also by direct computation, one notices that bf C1 and C2 have the same spectral decomposition. One
eigenvalue is given by ξ1 = (1 − 3e)(v + c), while the other two ξ2 ≥ ξ3 are the roots of the quadratic
polynomial P (X) = X 2 + (c− 2v − 5ec + 3ev)X + (6ec2 − cv − 3ev2 − 2c2 + v2 + 3ecv). It is easy to see that
P (ξ1) = −8ec(1− 3e)(v + c). If c > 0 (Class B), then P (ξ1) < 0, hence ξ2 ≥ ξ1, with equality at = 0, and
ξ1 ≥ ξ3, with equality when = 1/3 (Figure 1g). If c < 0 (Class C), then P (ξ1) > 0 and ξ1 < (ξ2 + ξ3)/2,
7
hence ξ1 ≤ ξ2 < ξ3, with equality when = 0 and = 1/3 (Figure 1f).
2
The theorem allows us some immediate class-specific interpretations in the context of phase space dynamics
and learning.
Figure 4: Bifurcation in attractor dynamics for partly biased inputs, all negative cross-correlations.
A. For small error, the system has an ellipse of neutrally-stable equilibria (yellow curve). This ellipse is stable, in
the sense that it persists for a whole interval of errors, from = 0 until = −c
v−c . B. For error past the critical value,
the ellipse is destroyed, but the new attractors are significantly far from the plane of the ellipse. Parameters used:
v = 1, c = 0.2, δ1 = 2 = δ2 = 0. Color coding: trajectories evolve in time from darker towards lighter shades.
Class A. The leading eigenvalue is constant, and always separated from the second (double) eigenvalue.
Moreover, the principal component of EC does not change when the error increases, so in this case the
learning is fully accurate for any degree of inspecificity (Figure 1e). This is a class of input statistics which
is completely error-proof.
Class B. This falls within the typical case of separated leading eigenvalues, where the system learns, for
any error value, the leading eigendirection of EC (which degrades smoothly from the principal component
of C; see Figure 1g, and Figure 2). Depending on parameters, the eigenvalue curves with respect to q
may exhibit a significant point of minimal separation (see "avoided crossing"), where the learning outcome
(leading eigenvector of EC) deteriorates very fast.
Class C. In the error-free case, the matrix C has a double leading eigenvalue, and the system has a
whole closed curve of neutrally attracting equilibria (in the eigenplane spanned by the corresponding
eigenvectors). When error is introduced, the two leading eigenvalues segregate, and one of the eigenvectors
takes over, which determines an immediate complete switch in the learning outcome. In this case, even
the smallest degree of inspecificity leads to favoring one specific direction, slightly detaching off the plane
where the "real" equilibria are contained (notice that the cosine of the accuracy angle, represented by the
blue curve in Figure 1f, does not fall too far off the perfect value cos(θ) = 1). We may interpret this as the
error helping the system "make up its mind" in the presence of too much ambiguity in the input statistics.
Class D. In the error-free case, the matrix C has a double leading eigenvalue, and the system has again
a whole curve of neutral equilibria, contained in the corresponding eigenplane. When subject to errors up
to a critical value q∗ = v+c
v−c , the leading eigenvalues change, but remain equal; furthermore, the subspace
spanned by the two corresponding eigenvectors remains unchanged, hence the learning process retains the
original ambiguity. Past the critical error value, the eigenvalues swap, and the eigendirection of the new
leading eigenvalue (of multiplicity one) is orthogonal to the previous plane (Figure 1h). In other words,
8
past the critical error value, the system will learn, but with such low accuracy, that the result of learning
is useless.
3 An analytical application in higher dimensions
We will work out an analytical computation which suggests that the cases described in Section 2 can
be extended to classify the behavior of the higher dimensional inspecific Oja system. For simplicity, we
consider only one application: for negatively cross-correlated inputs:
−c
−c
...
v + δ1
−c
v + δ2
C =
(5)
···
···
. . .
···
−c
v + δn
−c
...
−c
Although this is perhaps the least biologically sound case, we feel that it is mathematically the most inter-
esting, and describes the opposite scenario from the case of all positively cross-correlated inputs (which is
mathematically the least interesting). As suggested by the numerical computations in 3 dimensions, covari-
ance matrices that exhibit other positive/negative patterns of cross-correlations are expected to produce
hybrid dynamics between these two extreme ends. These dynamics will depend not only on the number
of negative correlations, but also on their distribution within the covariance matrix. A random matrix
analysis may be able to classify behavior for all input patterns, but this is not within the scope of this
study. In this section, we only present the main analytical results we obtained for our application; proofs
of the statements and additional comments can be found in Appendix 2.
Fully biased case. We first consider the covariance biases δj's to be distinct: δ1 > δ2 > . . . > δn−1 >
δn = 0. We want to study the eigenvalues and eigenspaces of the modified covariance matrix EC. The
characteristic polynomial of EC can be expressed as (see details in Appendix 2):
where for all j = 1, n, we called fj = (v + δj − c) + c and Xj(λ) = q(v + δj − c) + c − λ.
We consider λj = (q − )(v + δj − c); clearly: λ1 > λ2 > ... > λn.
In Appendix 2, we show how
these values can be used to partition the real line and separate the roots of ∆. In particular, we prove the
following:
Proposition 3.1. In the biased case δ1 > δ2 > . . . > δn, the matrix EC has n real distinct eigenvalues
ξ1 > ξ2 > . . . > ξn.
We can define, as in the 2 and 3-dimensional applications, the "critical" error values, for which fj(∗
j ) =
0, ∀j ∈ 1, n:
j . As increases from 0 to 1/n, it traverses the values = ∗
v + δj − c
(6)
n (since δ1 > δ2 > . . . > δn). Clearly, for all j ∈ 1, n, we have fj > 0 iff
j . When is in the intervals between two
so that 0 < ∗
> ∗
1 < ∗
2 < . . . < ∗
−c
∗
j =
9
∆(λ) = det(EC − λI) =
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
X1(λ)
f2
X2(λ)
f1
...
f1
f2
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
···
···
. . .
··· Xn(λ)
fn
fn
...
Figure 5: A simple example of how the characteristic polynomial ∆ of EC and its roots change as the
quality q decreases, for dimension n = 3 and fixed parameters v = 1, c = −0.2, δj = j/3, so that ∗
1 ∼ 0.091,
2 ∼ 0.107, ∗
3 ∼ 0.130, for j ∈ 1, 3. Each different color represents a different values of q: q = 0.98 (red), q = 0.805
∗
(blue), q = 0.76 (green), q = 0.6 (pink). The continuous curves correspond to the graph of the polynomial for
different q's, and the bullets represent (along the x-axis) the points λj = (q − )(v + δj − c), for j = 1, 3. The figure
shows how the order of the position of the roots of ∆ changes with respect to the points of the partition λ3 < λ2 < λ1
(which in turn travel down the axis as q decreases). For q = 0.98 (i.e., = 0.01 < ∗
1), λ1 > ξ1 > λ2 > ξ2 > λ3 > ξ3.
For q = 0.805 (where = 0.0975 ∈ [∗
3]),
ξ1 > λ1 > ξ2 > λ2 > λ3 > ξ3. For q = 0.6 (where = 0.2 > ∗
2]), ξ1 > λ1 > λ2 > ξ2 > λ3 > ξ3. For q = 0.76 (where = 0.12 ∈ [∗
1, ∗
3), ξ1 > λ1 > ξ2 > λ2 > ξ3 > λ3.
2, ∗
consecutive critical values ∗
j , each two consecutive roots of ∆ are separated by at least one λj. When
reaches each critical value j, the root ξj crosses from one interval to another through the stage ξj = λj.
Losing the bias. Suppose now that, for j ∈ 1, n, δj = δj+1 + ζj, and allow some of the ζj → 0; in the
limit, this results in a loss of bias in the covariance matrix C (v + δj = v + δj+1 for some index j). In
consequence, λj − λj+1 → 0. It follows that in the limit of ζ = 0 and ξ = λ1 = λ2, so that the maximal
eigenvalue of EC preserves its multiplicity =1. This situation changes if we introduce an order two bias
loss δ1 = δ2 = δ3 (i.e. if we make both ζ1 and ζ2 approach zero simultaneously). Then λ1 − λ2 → 0 and
λ2 − λ3 → 0, so that, the two leading roots collide into a double root λ3 = ξ2 = λ2 = ξ1 = λ1. This justifies
the following proposition:
Proposition 3.2. Suppose < ∗
results in a leading eigenvalue of multiplicity k − 1 for the modified covariance matrix EC.
1. An order k bias loss of the covariance matrix C of the type δ1 = . . . = δk
4 Discussion
In this study, we considered a learning network based on the classical unsupervised learning model of
Oja, extended it to allow synaptic cross-talk (encoded either as inspecificity , or as transmission quality
q = 1 − (n − 1)) and we showed how different input patters can exacerbate, or at the contrary, efface the
effects of this cross-talk on the asymptotic outcome of learning.
We made a few simplifying assumptions: we considered uniform magnitude of input cross-correlations
(i.e, uniform absolute value c of the off-diagonal elements of C), and uniform error (the Hebbian adjust-
ment of any weight was equally affected by error, and did not depend either on the strength of that weight
or its identity). Such "isotropicity" seems like a reasonable basic assumption, and has been further moti-
vated and discussed in our previous work [13, 12]. Furthermore, it allowed us to identify other features of
the input distribution, crucially consequential on the learning dynamics and outcome: the cross-correlation
signs, and the input bias.
10
When observing the (qualitative and quantitative) effects that the presence of cross-talk can have on
the system's asymptotic behavior, we noted that these can vary substantially, depending on the input
second-order statistics. We found that, in specific highly unbiased cases, the cross-talk has no effect on
the presence and position of the asymptotic attractors (Figure 1e). In other cases, the depreciation of the
asymptotic outcome with error is so slow, that small errors have virtually no effect on learning (Figure
1a,b; also see Figure 2 for a phase space illustration).
Other significant classes of inputs, however, exhibited a sudden change of the attracting direction from
an almost perfect input principal component estimator to a direction almost orthogonal to the original. This
occurred either in the form of an eigenvalue swapping bifurcation in dynamics (producing the instantaneous
loss of learning accuracy at a critical error value; see Figure 1c, and also 3 for an illustration of phase space
transitions), or in the milder form of an eigenvalue "avoided crossing," (inducing a smooth, yet very steep
depreciation of the learned direction at a specific error, see Figure 1d,g). As discussed in our previous
work, these two latter effects can be practically undistinguishable: learning works reasonably well for small
enough errors; for errors past the crash value, the outcome becomes irrelevant to the input statistics, and
the system is essentially encoding information on the cross-talk pattern itself.
Finally, we found that in instances of highly unbiased inputs, learning may lead to an ambiguous
outcome (double leading eigenvalue), even in the absence of cross-talk (Figures 1f,h and 4). This is an
occurrence we have not encountered in our previous, more restrictive, versions of the model, since it
requires inputs with concomitant negative cross-correlations and loss of bias of order > 2). Our current
analysis shows that the fashion in which the cross-talk handles input ambiguity (i.e., nonisolated, neutrally-
attracting equilibria) depends quite significantly on the number and (in this case also) geometry of the
negative correlations within the input. Depending on these, we distinguished two cases. One in which even
the smallest degree of cross-talk helps the system make an asymptotic selection for one particular direction
in the eigenspace spanned by the multiple eigenvalue. The other, in which no small degree of inspecificity
can perturb this "stable ambiguity". The level of critical cross-talk that can finally destroy the curve of
neutrally-stable equilibria also pushes the system to learn an orthogonal direction, hence irrelevant to the
main features of the original input statistics.
While this extension still only considers a very simple model of learning, it helps us re-iterate an impor-
tant idea, which we have formulated before [13, 6, 8, 12]). A central problem for biological learning seems
to be that the activity-dependent processes that lead to connection strength adjustments cannot be com-
pletely synapse specific [6, 1]. This raises the possibility that sophisticated learning, such as presumably
occurs in the neocortex, is enabled as much by special machinery for enhancing specificity, as by special
algorithms [3]. It seems therefore possible that a key biological factor in learning problems is not just in
finding good architectures, techniques and estimative algorithms, but also in perfecting the relevant plas-
ticity apparatus. We have suggested that learning plasticity errors are analogous to mutations, and that
cortical circuitry might reduce such errors in the same manner as "proofreading" reduces DNA copying
mistakes. This further suggests that problems of survival and reproduction are so diverse that no single
algorithm can solve them all, so that no "universal" or "canonical" cortical circuit should be expected.
However, if every specialized algorithm relies on extraordinarily specific synaptic weight adjustment, then
finding machinery that allows such specificity would indeed be equivalent to discovering new neurobio-
logical general principles. We have speculated that an important part of such machinery, at least in the
neocortex, might lie outside the synapse itself, in the form of complex circuitry performing a proofreading
operation analogous to that procuring accuracy for polynucleotide copying [1, 3, 2]. Let us note that such
machinery would be less necessary if update inaccuracy merely degraded learning, rather than preventing
it (possibility which our model does not exclude). Even so, when temporarily unfavorable input statistics
lead to imperfect learning because of Hebbian inspecificity, the degraded weights might still be a useful
starting point for better learning when input statistics improve.
11
Acknowledgements
To my mom and dad, who took care of my world while I have been working, and to Alex, who allowed
them to take over and rule the world.
References
[1] P. Adams and K. Cox. A new interpretation of thalamocortical circuitry. Philosophical Transactions
of the Royal Society of London. Series B: Biological Sciences, 357(1428):1767, 2002.
[2] P. Adams and K. Cox. From life to mind: two prozaic miracles. In Integral Biomathics, volume 67,
page 02. 2012.
[3] PR Adams and KJA Cox. A neurobiological perspective on building intelligent devices. Neuromorphic
Eng, 3:2–8, 2006.
[4] G.Q. Bi. Spatiotemporal specificity of synaptic plasticity: cellular rules and mechanisms. Biological
Cybernetics, 87(5):319–332, 2002.
[5] T. Bonhoeffer, V. Staiger, and A. Aertsen. Synaptic plasticity in rat hippocampal slice cultures: local "
hebbian" conjunction of pre-and postsynaptic stimulation leads to distributed synaptic enhancement.
Proceedings of the National Academy of Sciences, 86(20):8113, 1989.
[6] K.J.A. Cox and P.R. Adams. Hebbian crosstalk prevents nonlinear unsupervised learning. Frontiers
in computational neuroscience, 3, 2009.
[7] S. Durrant. Negative correlation in neural systems. Doctoral thesis, 2010.
[8] T. Elliott. Cross-talk induces bifurcations in nonlinear models of synaptic plasticity. Neural Compu-
tation, pages 1–68, 2012.
[9] C.D. Harvey and K. Svoboda. Locally dynamic synaptic learning rules in pyramidal neuron dendrites.
Nature, 450(7173):1195–1200, 2007.
[10] C. Malsburg. Self-organization of orientation sensitive cells in the striate cortex. Biological Cybernetics,
14(2):85–100, 1973.
[11] E. Oja. Simplified neuron model as a principal component analyzer. Journal of mathematical biology,
15(3):267–273, 1982.
[12] A. Radulescu and P. Adams. Hebbian crosstalk and input segregation. preprint, 2010.
[13] A. Radulescu, K. Cox, and P. Adams. Hebbian errors in learning: an analysis using the oja model.
Journal of theoretical biology, 258(4):489–501, 2009.
[14] R. Vasudeva. How negative can the product-moment correlation coefficient be? Resonance, 3(5):73–75,
1998.
12
Appendix 1
The symmetric, positive definite matrix C ∈ Mn(R) defines a dot product in Rn as:
(cid:104)v, w(cid:105)C = vT Cw
Although both C and E are symmetric, the product EC is not symmetric in the Euclidean metric.
However, in a new metric defined by the dot product (cid:104)·,·(cid:105)C, EC is symmetric. (Indeed, for any pair of
vectors u, v ∈ Rn, we have
(cid:104)ECu, v(cid:105)C = (ECu)tCv = utCtEtCv = utCECv = (cid:104)u, ECv(cid:105)C
In consequence, EC has a basis of eigenvectors, orthogonal with respect to the dot product (cid:104)·,·(cid:105)C.
The following theorem, describing the equilibria of the system (2), is immediate.
Theorem 1. An equilibrium for the system is any vector w = (w1...wn)T such that ECw = (wT Cw)w,
i.e., an eigenvector of EC (with corresponding eigenvalue λw), normalized, w.r.t. the norm (cid:107)·(cid:107)C = (cid:104)·,·(cid:105)C,
so that (cid:107)w(cid:107)C = λw.
ECw = λww,
(cid:107)w(cid:107)C = λw
If we additionally assume (generically) that EC has strictly positive maximal eigenvalue of multiplicity
one, then the corresponding eigendirection is orthogonal in (cid:104)·,·(cid:105)C to all other eigenvectors of EC.
Take then w to be an equilibrium of the system (2), i.e. an eigenvector of EC, with eigenvalue λw =
(wT Cw)w > 0. To establish stability, we calculate the Jacobian matrix at w to be
w = γ(cid:2)EC − 2w(Cw)T − (wT Cw)I(cid:3)
Df E
Then we get the following:
Theorem 2. Suppose EC has a multiplicity one largest eigenvalue. An equilibrium w (i.e., by theorem
(), an eigenvector of EC with eigenvalue λw, normalized so that (cid:107)w(cid:107)C = λw) is a local hyperbolic attractor
for (2) iff it is an eigenvector corresponding to the maximal eigenvalue of EC.
Proof. Fix an eigenvector w of EC, with ECw = λww. Then:
Df E
ww = −2γλww
Recall that the vector w can be completed to a basis B of eigenvectors, orthogonal with respect to the dot
product (cid:104)·,·(cid:105)C. Let v ∈ B, v (cid:54)= w, be any other arbitrary vector in this basis, so that ECv = λvv, and
(cid:104)w, v(cid:105)C = wtCv = 0. We calculate:
wv = −γ[λw − λv]v
Df E
So B is also a basis of eigenvectors for Df E
w. The corresponding eigenvalues are −2γλw (for the eigenvector
w) and −γ[λw − λv] (for any other eigenvector v ∈ B, , v (cid:54)= w). An equivalent condition for w to be
a hyperbolic attractor for the system (2) is that all the eigenvalues of Df E
w are < 0. Since γ, λw > 0,
this condition is further equivalent to having −γ(λw − λv) < 0 , for all v ∈ B , v (cid:54)= w. In conclusion, an
for all v (cid:54)= w (i.e. λw is the maximal
equilibrium w is a hyperbolic attractor if and only if λw > λv ,
eigenvalue, or in other words if w is in the direction of the principal eigenvector of EC).
13
Such attractors always exist provided that the condition of Theorem 2 is met (i.e., EC has a maximal
eigenvalue of multiplicity one). Then the network learns, depending on its initial state, one of the two
stable equilibria, which are the two (opposite) maximal eigenvectors of the modified input distribution,
normalized so that (cid:107)w(cid:107)C = λw. Next, we aim to show that these two attractors are the system's only
hyperbolic attractors.
2
Theorem 3. Suppose the the modified covariance matrix EC has a unique maximal eigenvalue λ1. Then
the two eigenvectors ±wEC corresponding to λ1, normalized such that (cid:107)w(cid:107)C = λ1, are the only two at-
tractors of the system. More precisely, the phase space is divided into two basins of attraction, of wEC and
−wEC respectively, separated by the subspace (cid:104)w, wEC(cid:105) = 0.
Proof. We make the change of variable u =
Cw. The system then becomes:
√
u = Au − (utu)u
√
√
C symmetric matrix, having the same eigenvalues as EC. More precisely, w is an
CE
Cv is an eigenvector of A with eigenvalue µ; hence any two
where A =
eigenvector of EC with eigenvalue µ iff
distinct eigenvectors of A are orthogonal in the regular Euclidean dot product.
√
(7)
Consider then v to be the leading eigenvector of A, and let u = u(t) be a trajectory of the system (7). We
want to observe the evolution in time of the angle between the variable vector u and the fixed vector v,
measured as:
We differentiate and obtain:
cos θ =
(cid:104)v, u(cid:105)
(cid:107)v(cid:107) · u(cid:107)
− (cid:107)v(cid:107) sin(θ) θ =
(vt u)(cid:107)u(cid:107)2 − (vtu)(utu)
(cid:107)u(cid:107)3
The numerator of this expression is
h(u) = (utu)(vtAu) − (vtu)(utAu)
(8)
We are interested in the sign of h(u); to make our computations simpler, we can diagonalize A in a basis of
orthogonal eigenvectors A = PtDP, where D is the diagonal matrix of eigenvalues and P is an orthogonal
matrix whose columns are the eigenvectors. Then:
h(u) = (ztz)(ytDz) − (ytz)(ztDz)
where y = Pv and z = Pu, so that Dy = DPv = λ1y (where λ1 is the largest eigenvalue of EC, assumed
to have multiplicity one). Then:
n(cid:88)
h(u) = (ytz)
(λ1 − λj)z2
j
Hence, if ytz > 0, then h(u) > 0. In other words: if vtu > 0 then −(cid:107)v(cid:107) sin(θ) θ > 0, hence that θ < 0. For
our original system, this means that any trajectory starting at a w with (cid:104)w, wEC(cid:105) > 0 converges in time
towards the principal eigenvector wEC of the matrix EC.
2
j=2
14
Appendix 2
(cid:88)
(cid:88)
(cid:88)
We want a concise description of the modified input matrix EC. To begin with, we can express the matrices
E and C individually as: E = M + (q − )I and C = cM + (v − c)I +
δjAj, where I is the n × n
identity matrix, M is the n × n matrix with uniform unit entries, and, for any j = 1, n, Aj is the matrix
with zero entries except Aj(j, j) = 1. Note, for future computations, that M2 = nM and that MAj is
the matrix with the only nonzero entries being ones along the j-th column. Unless otherwise specified, the
summations are for j = 1, n. The product EC will then be
EC = [(v − c) + c(q − ) + cn]M + (q − )(v − c)I +
δjMAj + (q − )
djAj
In matrix form, this translates as
EC =
X1(λ)
f2
f1
...
f1
X2(λ)
f2
···
···
. . .
··· Xn(λ)
fn
fn
...
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
where, ∀j = 1, n, we called fj = (v + δj − c) + c and Xj(λ) = q(v + δj − c) + c − λ.
Fully biased case
We first consider the covariance biases δj's to be distinct: δ1 > δ2 > . . . > δn−1 > δn = 0. We will prove
that the polynomial ∆ has n real roots ξ1 ≥ ξ2 ≥ . . . ≥ ξn, and we will find approximating bounds for
their positions on the real line.
Remark first that the end behavior of ∆(λ) is given by:
Consider λj = (q − )(v + δj − c); clearly: λ1 > λ2 > ... > λn. We will use these values to partition the
real line and separate the roots of ∆. To begin, we calculate, for all i, j = 1, n:
lim
λ→−∞ ∆(λ) = ∞ and lim
λ→+∞ ∆(λ) = (−1)n∞
(9)
In particular: Xj(λj) = fj, ∀j = 1, n. By raw and column manipulations, it can be shown that, ∀j = 1, n
Xi(λj) = fi + (q − )(δi − δj)
∆(λj) = fj(q − )n−1(cid:89)
i(cid:54)=j
(δi − δj)
(10)
In consequence: sign(∆(λj)) = sign(fj)(−1)n−j.
Recall that fj = (v + δj − c) + c, hence f1 > f2 > . . . > fn. To continue our discussion and establish the
signs of ∆ at all partition points λj, we need to establish the index j for which the values fj switch sign.
For each j ∈ 1, n, consider the "critical" error values, for which fj(∗
j ) = 0, ∀j ∈ 1, n:
so that 0 < ∗
1 < ∗
2 < . . . < ∗
(11)
n (since δ1 > δ2 > . . . > δn). Clearly, for all j ∈ 1, n, we have fj > 0 iff > ∗
j .
v + δj − c
−c
∗
j =
15
Remark. A safe assumption that would allow us to study all cases that may appear is to consider
v > (n − 1)c, which guarantees ∗
j < 1/n, ∀j ∈ 1, n. This insures a complete discussion, since then
∈ [0, 1/n] is allowed to reach and cross over all the critical values ∗
j , creating a possible swap in the order
of the eigenvalues of EC, as we will show later. The proof for the other cases will be omitted, since it is
just a simplification of the present argument. In fact, the only crossover of true interest to us is = ∗
1,
where the eigenvalue swap involves the two largest eigenvalues and thus affects the position of the system's
attracting equilibria, corresponding to the normalized eigenvectors of the maximal eigenvalue; the other
j , for j ≥ 2, only affect the stable/unstable spaces of the saddle equilibria. In this light,
critical values = ∗
the condition on the entries of the covariance matrix can be loosened to v > (n − 1)c − δ1.
We distinguish the following cases:
(I.) For 0 ≤ < ∗
1. This implies fj < 0, ∀j ∈ 1, n. Then
sign(∆(λj)) = sign(fj)(−1)n−j = (−1)(−1)n−j = (−1)n−j+1
(12)
From (4),(4) and (12), we obtain the following sign table:
λ
sign(∆(λ))
−∞ λn
(−)
(+)
λn−1
. . . λ2
(+)
. . .
(−1)n−1
λ1
(−1)n
+∞
(−1)n
From the Intermediate Value Theorem and the Fundamental Theorem of Algebra, it follows that the
polynomial ∆(λ) has n real roots ξ1 > ξ2 > . . . > ξn, such that:
− ∞ < ξn < λn < ξn−1 < λn−1 < . . . < λ2 < ξ1 < λ1 < ∞
(13)
(II.) For ∗
p < < ∗
p+1. Then f1, . . . , fp > 0 and fp+1, . . . , fn < 0. Similarly as in (I.), we have:
λ
sign(∆(λ))
−∞ λn
(−)
(+)
λn−1
. . . λp+1
(+)
. . .
(−1)n−p
λp
(−1)n−p
. . . λ1
. . .
(−1)n−1
+∞
(−1)n
hence the polynomial ∆(λ)) has roots ξ1 > ξ2 > . . . > ξn, such that:
− ∞ < ξn < λn < ξn−1 < λn−1 < . . . < ξp+1 < λp+1 < λp < ξp < . . . < λ1 < ξ1 < ∞
n < < 1/n. Then f1, . . . , fn > 0 and we have
(III.) For ∗
(14)
λ
−∞ λn
sign(∆(λ))
(+)
(+)
λn−1
(−)
. . . λ2
. . .
(−1)n
λ1
(−1)n−1
+∞
(−1)n
and the polynomial ∆(λ)) has roots ξ1 > ξ2 > . . . > ξn, such that:
− ∞ < λn < ξn < λn−1 < ξn−1 < . . . < λ1 < ξ1 < ∞
(15)
In particular, we have proved the following lemma in the main text:
Proposition 3.1. In the biased case δ1 > δ2 > . . . > δn, the matrix EC has n real distinct eigenvalues
ξ1 > ξ2 > . . . > ξn.
16
Losing the bias
Suppose now that, for j ∈ 1, n, δj = δj+1 + ζj, and allow some of the ζj → 0; in the limit, this results in a
loss of bias in the covariance matrix C (v +δj = v +δj +1 for some index j). In consequence, λj −λj+1 → 0.
Let's study the changes of the maximal root ξ1 as ζ1 → 0 (i.e., we eliminate the bias between the
two most correlated components of the matrix C. Suppose ∈ [0, ∗
1]. This calculation can be extended
similarly to the other intervals for ; however, we will only discuss here the case ∈ [0, ∗
1], since it is the
only one that relates directly to the position and multiplicity of the leading root of ∆. It also agrees with
our goal to study the behavior of the system for small enough transmission errors. According to (13), we
have
−∞ < ξn < λn < ξn−1 < λn−1 < . . . < λ2 < ξ1 < λ1 < ∞
Since λ1 → λ2, it follows that in the limit of ζ = 0 and ξ = λ1 = λ2, so that the maximal eigenvalue of
EC preserves its multiplicity =1. This situation changes if we introduce an order two bias loss δ1 = δ2 = δ3
if we make both ζ1 and ζ2 approach zero simultaneously). Then λ1 − λ2 → 0 and λ2 − λ3 → 0, so
(i.e.
that, the two leading roots collide into a double root λ3 = ξ2 = λ2 = ξ1 = λ1. This justifies the following
proposition:
Proposition 3.2. Suppose < ∗
results in a leading eigenvalue of multiplicity k − 1 for the modified covariance matrix EC.
1. An order k bias loss of the covariance matrix C of the type δ1 = . . . = δk
This proposition can be generalized to encompass bias loss anywhere in the inputs, and any interval
for the error . Below, we give a more general statement, which follows by repeating the argument for the
case we already analyzed, but could also be proved more directly.
Theorem. S uppose that the matrix C is allowed to exhibit bias loss in all possible ways, so that it can
be written in block form as (5), where there exist k1, k2, . . . , kN ∈ 1, n, with
kj = n and such that
N(cid:88)
j=1
δ1 = . . . = δk1 = ν1
δk1+1 = . . . = δk2 = ν2
...
δkN−1+1 = . . . = δkN = νN
ν1 > ν2 > . . . > νN
with
Then the characteristic polynomial ∆ of EC has all real eigenvalues. More precisely, these eigenvalues
are λj = (q − )(v + δj − c) with multiplicity kj − 1, for all j ∈ 1, N , and N additional eigenvalues
ξ1 ≥ ξ2 ≥ . . . ≥ ξN .
Remark. The order of these eigenvalues, depending on the the error value with respect to the critical
error values ν∗
, is the same as described in the cases (I.)-(III.) above.
−c
j =
v + νj − c
17
|
1711.00928 | 1 | 1711 | 2017-11-02T20:45:07 | Physiological Tremor Increases when Skeletal Muscle is Shortened: Implications for Fusimotor Control | [
"q-bio.NC"
] | The involuntary force fluctuations associated with physiological (as distinct from pathological) tremor are an unavoidable component of human motor control. While the origins of the physiological tremor are known to depend on muscle afferentation, it is possible that the mechanical properties of muscle-tendon systems also affect its generation, amplification and maintenance. In this paper, we investigated the dependence of physiological tremor during tonic, isometric plantarflextion torque at 30% of maximum at three ankle angles. The amplitude of physiological tremor increased as calf muscles shortened in contrast to the stretch reflex whose amplitude decreases as muscle shortens. We used a closed-loop simulation model of afferented muscle to explore the mechanisms responsible for this behavior. We demonstrate that changing muscle lengths does not suffice to explain our experimental findings. Rather, the model consistently required the modulation of gamme-static fusimotor drive to produce increases in physiological tremor with muscle shortening--while successfully replicating the concomitant reduction in stretch reflex amplitude. This need to control gamma-static fusimotor drive explicitly as a function of muscle length has important implication. First, it permits the amplitudes of physiological tremor and stretch reflex to be decoupled. Second, it postulates neuromechanical interaction that require length-dependent gamma drive modulation to be independent from alpha drive to the parent muscle. Lastly, it suggests that physiological tremor can be used as a simple, non-invasive measure of the afferent mechanisms underlying healthy motor function, and their disruption in neurological conditions. | q-bio.NC | q-bio |
Physiological Tremor Increases when Skeletal Muscle is Shortened:
Kian Jalaleddini 1,§, Akira Nagamori 1,§, Christopher M. Laine 1, Mahsa A. Golkar 2, Robert
Implications for Fusimotor Control
E. Kearney 2, Francisco J Valero-Cuevas 1,3,‡
1Division of Biokinesiology and Physical Therapy, University of Southern California, USA
2Department of Biomedical Engineering, McGill University, Canada
3Department of Biomedical Engineering, University of Southern California, USA
§ Equal Contribution
‡ Corresponding Author
1 Key Point Summary
• In tonic, isometric, plantarflexion contractions, physiological tremor increases as the ankle joint becomes
plantarflexed.
• Modulation of physiological tremor as a function of muscle stretch differs from that of the stretch reflex
amplitude.
• Amplitude of physiological tremor may be altered as a function of reflex pathway gains.
• Healthy humans likely increase their γ-static fusimotor drive when muscles shorten.
• Quantification of physiological tremor by manipulation of joint angle may be a useful experimental probe
of afferent gains and/or the integrity of automatic fusimotor control.
2 Abstract
The involuntary force fluctuations associated with physiological (as distinct from pathological) tremor are an
unavoidable component of human motor control. While the origins of physiological tremor are known to de-
pend on muscle afferentation, it is possible that the mechanical properties of muscle-tendon systems also affect
its generation, amplification and maintenance. In this paper, we investigated the dependence of physiological
tremor on muscle length in healthy individuals. We measured physiological tremor during tonic, isometric plan-
tarflexion torque at 30% of maximum at three ankle angles. The amplitude of physiological tremor increased
as calf muscles shortened in contrast to the stretch reflex whose amplitude decreases as muscle shortens. We
used a published closed-loop simulation model of afferented muscle to explore the mechanisms responsible for
this behavior. We demonstrate that changing muscle lengths does not suffice to explain our experimental find-
ings. Rather, the model consistently required the modulation of γ-static fusimotor drive to produce increases
in physiological tremor with muscle shortening-while successfully replicating the concomitant reduction in
stretch reflex amplitude. This need to control γ-static fusimotor drive explicitly as a function of muscle length
has important implications. First, it permits the amplitudes of physiological tremor and stretch reflex to be
decoupled. Second, it postulates neuromechanical interactions that require length-dependent γ drive modula-
tion to be independent from α drive to the parent muscle. Lastly, it suggests that physiological tremor can be
used as a simple, non-invasive measure of the afferent mechanisms underlying healthy motor function, and their
disruption in neurological conditions.
1
3 Introduction
Physiological tremor is the unavoidable tendency of muscles to generate involuntary, rhythmic oscillations in
the frequency range of 5-12 Hz (Lippold, 1971; Sowman and Turker, 2005). There is an emerging consensus
that muscle afferentation and the integrity of the spinal cord stretch reflex circuitry lie at the heart of physiologi-
cal tremor (Lippold, 1971; Young and Hagbarth, 1980; Christakos et al., 2006; Laine et al., 2016; Cresswell and
Loscher, 2000). Given the cardinal role that disrupted muscle afferentation plays in the clinical presentation of
neurological conditions, it is critical to develop noninvasive means to assess reflex function through measure-
ments of physiological tremor amplitude. One possibility is to leverage the known increase in stretch reflex
amplitudes with muscle stretch, potentially due to increased firing of muscle spindles (Mirbagheri et al., 2000;
Mileusnic et al., 2006). Dependence of physiological tremor on muscle stretch has not been characterized. Do-
ing so might enable noninvasive methods to assess the integrity of spinal circuitry without the need for electrical
or mechanical perturbations.
Spindle primary afferents are sufficiently sensitive to respond to the changes in muscle length induced by
physiological tremor in contracting muscles Hagbarth and Young (1979). Since stretch reflex is often used
to assess the sensitivity of spindle afferents, we hypothesized, therefore, that the amplitude of physiological
tremor would increase with muscle stretch provided muscle activation was held constant. The human ankle
plantarflexor muscle is a well established model to study muscle and joint neuromechanics (Mugge et al., 2010;
De Vlugt et al., 2010), stretch reflex responses (Kearney et al., 1997; Alibiglou et al., 2008; Toft et al., 1991),
afferent gains (e.g., H reflex) (Capaday and Stein, 1986), and physiological tremor (Laine et al., 2014). Conse-
quently, we investigated physiological tremor in healthy adults as they produced tonic isometric plantarflexion
ankle torque at different joint angles. To elucidate the mechanical and neural factors that could explain our
results, we also simulated a closed-loop model of afferented muscle as in (Laine et al., 2016), scaled to the
architecture of the gastrocnemius .
Our findings show that the amplitude of physiological tremor is related to (and can be manipulated through)
changes in muscle length. Further, our simulations establish that it is not sufficient to simply alter muscle
mechanics, but adjustment of spindle sensitivity is necessary to replicate the changes in the physiological tremor
associated with muscle stretch. Given that a physiologically-plausible fusimotor mechanism could explain our
experimental results, we discuss the implications of this for the development of simple and powerful techniques
for assessing the integrity of reflex pathways in health and disease.
4 Methods
4.1 Experimental Methods
Seven subjects with no history of neuromuscular disease were recruited (Age of 32 ± 2.2, 2 female). Subjects
gave their informed consent prior to participation, and all procedures were approved by the McGill University
Institutional Review Board.
4.1.1 Experimental Setup
The subjects lay supine while the left foot was placed inside a custom-made fiberglass boot rigidly attached to
a hydraulic actuator (Figure 1). The boot constrained the whole limb movement except the ankle movement in
the sagittal plane. The actuator was programmed to maintain a particular ankle angle.
EMG signals were recorded from the main plantarflexor and dorsiflexor muscles using standard, active,
single differential surface electrodes supplied with the Bagnoli System (Delsys, Inc., Boston). EMGs were
recorded from the medial and lateral heads of the gastrocnemius, soleus and tibialis anterior muscles. The
reference electrode was a DermaSport electrode (American Imex, Irvine) placed on the left knee joint. The
EMG signals were amplified 1000 times and high-pass filtered at 20 Hz prior to acquisition.
The ankle angle was measured using a 6273 BI Technologies potentiometer (TT Electronics, Woking, UK).
The ankle torque was measured using a 2110-5K Lebow transducer (Honeywell, Columbus). All signals were
2
Figure 1: Subjects lay supine with the knee extended and were instructed to produce a constant
isometric plantarflexing ankle torque of 30% of their maximum voluntary contraction, aided by an
overhead monitor. Three joint angles were tested: 20◦ plantarFlexed (PF), 5.7◦ PF from neutral
(MID), and 8.6◦ dorsiflexed (DF) from neutral where the neutral angle was defined as the 90 ◦ angle
between the shank and foot. Ankle torque and EMG from the main plantarflexors and dorsiflexor
were recorded.
3
AnkleTorqueSurfaceEMGIsometric PlantarexingTorqueDF: 8.6 DorsiFlexed PostureTorqueVisual FeedbackMID: 5.7 PF PosturePF: 20 PlantarFlexedPostureTarget Torque Level !"# low-pass filtered at 486.3 Hz and digitized at 1 kHz with 24-bit resolution using a set of NI-4472 A/D cards
(National Instruments, Austin).
4.1.2 Experimental Trials
The experiment consisted of three types of trials: (I) Maximum Voluntary Contraction (MVC) trials. Subjects
were instructed to perform five maximum voluntary contractions separated by 10 s rest intervals. The MVC was
taken as the maximum torque in this trial. (II) Passive torque trials of 10 s. At each joint angle, subjects were
instructed to relax and the passive torque was taken as the average of the torque. (III) Tonic Contraction trials
of 30 s. Each subject was instructed to maintain a tonic voluntary contraction of 30% of MVC torque aided by
a visual feedback of their filtered torque in real-time with an eighth-order, low-pass, Bessel filter with 0.7 Hz
cutoff.
The hydraulic actuator moved the ankle to the angle randomly selected from: 20◦ PlantarFlexed (PF), 5.7◦
plantarflexed from neutral (MID), and 8.6◦ DorsiFlexed (DF) from neutral where the neutral angle was defined
as the 90 ◦ angle between the shank and foot. At each angle, the subjects performed one MVC trial, one passive-
torque trial and three tonic contraction trials successively. One minute of rest was imposed between trials to
prevent muscle fatigue.
4.1.3 Data Analysis
To describe the tremor in torque during each trial, we conducted a spectral analysis. First, the tonic contraction
torque was normalized by subtracting the passive torque and dividing by the MVC torque at each angle. Second,
the contribution of slow-varying torque was removed using a high-pass, sixth order, butterworth filter with 5
Hz cut-off frequency. Third, the power spectral density of the torque was calculated using 1 s Gaussian-tapered
windows. The power of the physiological tremor was calculated as the average of the power spectral density in
the 5-12 Hz band. Lastly, the power of the physiological tremor was normalized to that of the DF posture for
analysis of the group results.
We also analyzed the coherence between the EMG signals of the plantarflexor muscles and torque. This
helps to verify a relationship between torque tremor and tremor-band oscillations in muscle activation. First, the
EMG signals were full-wave rectified, and the magnitude-squared coherence was estimated with 1 s Gaussian-
tapered windows. For statistical reasons, the magnitude-squared coherence was normalized using Fisher's
r-to-z transform zFisher(f ) = atanh((cid:112)c(f )) where c(f ) is the magnitude-squared coherence at a given
frequency (f ) and atanh is the inverse hyperbolic tangent. Next, the transformed values were converted to
standard zstandard(f ) scores. A bias inherent in this method was estimated (and then removed) by averaging
zstandard(f ) over the frequency range of 100 to 300 Hz and excluding the harmonics of 60 Hz Baker et al.
(2001, 2003). Fig. 2 illustrates these analyses on representative data.
To estimate the overall α drive to the calf muscles and test for possible agonist-antagonist co-activation, we
analyzed the amplitude of the EMG signals of the plantarflexor and dorsiflexor muscles. We calculated the Root
Mean Squared (RMS) values of EMG signals in each trial and normalized these values to the RMS of the EMG
signals recorded during the maximum contractions.
4.2 Simulation Methods
4.2.1 Closed-loop Simulations of Afferented Muscle
We used a computational model of afferented musculotendon to investigate neuromechanical factors that might
contribute to modulation of physiological tremor with muscle stretch. This model was an extension of our
previously published model of (Laine et al., 2016) by incorporating the more realistic muscle model of (Brown
et al., 1996; Cheng et al., 2000; Song et al., 2008b,a), and an offset on the muscle spindle Ia firing rate to
account for presynaptic inhibition mechanisms (Raphael et al., 2010; Tsianos et al., 2014). Figure 3 presents
a conceptual description of this model that comprises previously published models of a musculotendon unit
(Brown et al., 1996; Cheng et al., 2000; Song et al., 2008b,a), muscle spindle (Mileusnic et al., 2006), and Golgi
Tendon Organ (GTO) (Elias et al., 2014), and a tracking controller (Laine et al., 2016). Detailed descriptions of
4
Figure 2: experimental methods: (A) Representative 1 s snapshot of torques recorded at three joint
angles (torques were offset for the purpose of visual comparison); (B) torque high-pass filtered with
a sixth order Butterworth filter with 5 Hz cut-off frequency; (C) power spectral density of torque cal-
culated using a 1 s Gaussian window; (D) the EMG-torque coherence estimated using 1 s Gaussian
window. The shaded area represents the physiological tremor band (5-12 Hz).
5
01Time (s)Raw Torque01Time (s)0.5 Nm1 HzHigh-Pass FilterFiltered Torque0102001234x10-8(A)(B)Filtered TorqueSpectrum(C)Frequency (Hz)(D)Frequency (Hz)EMG-TorqueCoherence01020024681012Power Spectral Density (%MVC)/HzStandard Z ScoreSigni(cid:31)cance ThresholdDFMID PF2 each model and model parameters can be found in the corresponding references and only brief descriptions are
given here.
The muscle model describes the known physiological properties of muscle and tendon that relate to mus-
cle force production, including the non-linear passive behavior of muscle and tendon, force-length and force-
velocity relationships of muscle, and length- and velocity-dependent activation-force dynamics (Brown et al.,
1999; Brown and Loeb, 2000). In this study, we specifically modeled the medial head of the gastrocnemius
muscle using previously published anatomical data (Table 1) (Arnold et al., 2010; Elias et al., 2014; Wickiewicz
et al., 1983). We adjusted the parameters that relate to passive muscle (Lr1 = 1.05) and tendon (cT = 100) stiff-
nesses to account for known changes in the intrinsic stiffness of muscle as a function of joint angle (Mirbagheri
et al., 2000). We set the muscle and tendon lengths to have optimal lengths at the PF posture based on previ-
ously published passive stiffness data (Weiss et al., 1986). To simulate the experimental conditions, the muscle
fiber and tendon lengths were varied as a function of joint angle based on previously published experimental
measurements (Iwanuma et al., 2011; Kawakami et al., 1998). We obtained muscle fiber length changes of 0.6
and 1.0 cm, and tendon length changes of 0.2 and 0.4 cm from PF to MID and from PF to DF joint angles,
respectively.
The feedback control system that generates neural drive to the muscle (see Fig 3) consisted of a tracking
controller and proprioceptive feedback pathways. The tracking controller generates a command signal based
on the error between the actual force output and reference force (Laine et al., 2016). This controller was not
designed to represent specific neural pathways, but it ensures successful force tracking (Laine et al., 2016).
Proprioceptive feedback pathways consisted of Ia and II afferent feedback arising from muscle spindles and
Ib afferent feedback from GTO. The contribution of each proprioceptive feedback pathway was controlled by
"presynaptic inhibition control" input (Raphael et al., 2010; Tsianos et al., 2014). Each feedback pathway was
appropriately delayed based on the simulated distance of 0.8 m between the spinal cord and muscle as well
as the conduction velocities of the respective afferent and efferent fibers (Elias et al., 2014) (Table 1). Signal
Dependent Noise (SDN) was added to the neural drive such that the force variability of the model matched
approximately that recorded experimentally (coefficient of variation of ≈3 %). The SDN was modeled as
zero-mean signal with variance of 0.3 generated by low-pass filtering white noise with a 4th-order Butterworth
filter at 100 Hz. The neural drive was first passed through an activation filter, which accounts for low-pass
dynamics of calcium uptake in skeletal muscle (Song et al., 2008a). The resulting activation signal induces
muscle contraction and associated dynamics within the musculotendon unit.
4.2.2 Simulation Protocols
We simulated MVC and passive force trials at each joint angle similar to those acquired experimentally. To do
so, we used a ramp-and-hold reference force input consisting of a 1 s zero activation, 1 s ramp-up, and 13 s
maximum activation phases. We removed all the feedback components (proprioceptive feedback and tracking
controller). We calculated the passive and MVC forces as the average force during the last 0.5 s of the zero
activation and maximum activation phases, respectively.
Next, we simulated tonic, isometric contraction trials for 100 s at 30% MVC (Figure 4). We analyzed the
last 90 s of the simulated force data to exclude the transient response. The simulated force data were analyzed
in the same way as the experimental data.
To explore contributions of mechanical and neural factors to modulation of physiological tremor as a func-
tion of joint angle, we performed a sensitivity analysis. First, we performed a Monte-Carlo analysis of 20 trials
to obtain the statistics of the estimates by simulating tonic, isometric contractions. Second, given the experi-
mental reports that Ia afferent feedback is important for the modulation of physiological tremor (Lippold, 1971),
we manipulated the gain of the Ia afferent feedback as a function of joint angle, either through presynaptic con-
trol of Ia afferent or fusimotor drive to the muscle spindle. We increased presynaptic inhibition by 20% (i.e.,
decreased presynaptic control input levels by 20%) at the DF angle compared to the other two joint angles to
mimic the changes in H-reflex amplitudes observed experimentally (Hwang, 2002). We also studied the effect
of fusimotor drive by independently varying γ-dynamic and γ-static drive from 10 to 250 pps with increment
of 20 pps at each joint angle. In this paper, we present the "optimal" set of γ-dynamic and γ-static drives that
produced behavior most similar to that observed experimentally.
6
Figure 3: Schematic representation of the afferented muscle model used in simulation study. The
muscle model receives neural drive from multiple sources: tracking controller, muscle spindle feed-
back and Golgi Tendon Organ (GTO) feedback. Muscle spindle receive adjustable fusimotor (γ)
drive and make mono- and disynaptic excitatory connections to the α-motoneuron in the spinal cord
through primary (Ia) and secondary (II) afferents, respectively. GTO makes disynaptic inhibitory
connection to α-motoneuron. Each proprioceptive feedback pathway receives presynaptic inhibition
control. The tracking controller provides a control signal based on the error between the reference
and output forces. The resulting muscle neural drive is passed through an activation filter that ac-
counts for the Ca + dynamics and produces muscle contraction. Afferent and efferent pathways have
realistic time delays.
7
γ-motoneuronα-motoneuronII a(cid:29)erentIb a(cid:29)erentTracking ControllerSpinal CordReference ForceForceGolgi Tendon Organ(GTO)Muscle SpindleIa a(cid:29)erentIbIIPresynapticControl Table 1: Model parameters of the medial head of the gastrocnemius muscle
Mass (g)
Optimal fiber length (cm)
Tendon length (cm)
Pennation angle (deg)
75
10.1
23.5
5
Ia afferent conduction velocity (m/s)
64.5
II afferent conduction velocity (m/s)
32.5
Ib afferent conduction velocity (m/s)
59
α-motoneuron conduction velocity (m/s)
56
Synaptic delay (ms)
2
Quantification of Muscle Stiffness
We quantified changes in the mechanical stiffness of muscle as a function of joint angle to explore potential
mechanical explanations for modulation of physiological tremor. The afferented muscle model was first deaffer-
ented by removing all feedback components. Next, muscle length perturbations of four amplitudes (0.05, 0.1,
0.15, and 0.2 cm) were applied directly to the musculotendon unit during sustained contraction at 30% MVC,
each 20 times. For each perturbation, the musculotendon unit was shortened for 2 s and then lengthened to the
original length over 50 ms (Figure 4 C top panel). The difference in the mean muscle force (Fdiff) before and
after the perturbation was calculated in a time-window of 1 s (shown in pink shaded areas in Figure 4 C bottom
panel).
Stiffness at each joint angle was quantified as the slope of the first-order regression line fitted between
perturbation amplitudes and Fdiff (on 20× 4 = 80 points). We used first-order regression based on the HC4
method (Wilcox, 2012), which allows for heteroscedasticity (unequal variance among groups) and therefore
provides an accurate estimate of the confidence interval of the slope (i.e. stiffness).
Quantification of Stretch Reflex
We simulated a stretch reflex experiment to demonstrate the consistency of the model with previous experimental
observations. Here, 20 direct ramp-and-hold perturbations (amplitude of 0.1 cm and velocity of 2 cm/s, shown
in Figure 4 D top panel) were applied to the afferented muscle model at each simulated muscle length. The
reflex amplitude (phasic response) was calculated as the difference between the mean force level in a 50-ms
window before the perturbation and peak muscle force within a 100 - 300 ms window after the perturbation
onset (shown in pink shaded areas in Figure 4 D bottom panel).
4.3 Statistical Analyses
Statistical analyses were performed to identify the dependence of outcome variables (e.g., total power, physio-
logical tremor power, and stretch reflex amplitude) on joint angle (independent variable). For experimental data,
we used the paired, Wilcoxon signed rank test (built-in function in R) on the normalized data.
8
Figure 4: Simulation methods: (A) Representative 1 s snapshot of the simulated muscle force; (B)
representative power spectral density of muscle force. The shaded area represents physiological
tremor range; (C) simulated muscle stiffness experiment. The difference in muscle force between
the shortened and lengthened phases (shaded areas) was obtained to compute muscle stiffness (see
Methods); (D) simulated stretch reflex experiment. Reflex amplitude was obtained as the peak force
value in the 100-300 ms window after the perturbation onset (shaded area).
9
For simulation results, we used an unpaired version of a heteroscedastic one-way ANOVA for means
(Wilcox, 2012). If a significant interaction was found between an outcome variable and joint angles, correspond-
ing post-hoc analysis that accounts for multiple comparisons was performed to identify significant differences
between pairs (Welch's method, lincon function in "WRS2" package) (Wilcox, 2012).
We performed the statistical analyses in the R environment for statistical computing (The R Foundation for
Statistical Computing, Vienna, Austria). The significance level was set to p < 0.05.
5 Results
To investigate potential changes in the level of neural drive to muscle as a function of joint angle, we calculated
the average RMS EMG values across the calf muscles, i.e. the three ankle plantarflexor muscles. We found no
significant difference in the RMS values across the three joint angles (p >= 0.12, Figure 5). Next, we computed
the RMS values of the tibialis anterior muscle to examine potential changes in the level of co-contraction as a
function of joint angle. Once again, we found no differences in the level of muscle activation across joint angles
(p >= 0.12). We also analyzed the individual calf muscles using the same technique and we found no significant
differences in the level of muscle activation across joint angles (p >= 0.12). These results show that there was
no angle-dependent modulation of plantarflexor and dorsiflexor muscle activation.
Next, we tested for potential contribution of co-activation of the ankle dorsiflexor in modulation of physio-
logical tremor. To this end, we normalized the EMG RMS values to those of the DF posture and computed the
difference between the calf (average across the three plantarflexors) and tibialis anterior muscle RMS EMG as
a function of ankle angle. We correlated changes in physiological tremor with the difference in the EMG RMS
values. Figure 6 demonstrates that this correlation was not significant. In fact, the slope of the fit line was not
significantly different than zero (p > 0.44). This result reinforces the lack of angle dependent changes in tibialis
anterior/calf muscles coactivation.
To ensure that physiological tremor was related to the plantarflexor's activities, we computed the coherence
between the plantarflexor EMGs and torque. The EMGs showed significant coherence (larger than significance
threshold shown by the dotted line) with plantarflexion torque in the physiological tremor range in all subjects
(Figure 2 (D) shows a typical coherence plot). This validates the use of the plantarflexor muscle in our sim-
ulations for investigating physiological tremor. Since on average, the medial head of gastrocnemius muscle
showed higher coherence compared to the other two muscles, we opted to use the architectural parameters of
the gastrocnemius medialis in the simulation of the afferented muscle model.
5.1 Physiological Tremor Increased with Muscle Shortening.
Figure 7(A) demonstrates that there was no consistent pattern of change in total torque power as a function of
joint angle (p = 1, 0.125, and 1, for MID-DF, PF-DF, and PF-MID, respectively). However, the amplitude of
the physiological tremor (5-12 Hz) showed a consistent pattern across all subjects in Figure 7 (B). Specifically,
physiological tremor was larger in the PF than MID, and DF postures (p = 0.031 and p = 0.016, respectively)
and in the MID than DF posture (p = 0.016). These results demonstrate that modulation of torque fluctuations
was frequency-specific, and that power in the physiological tremor range increased as the muscle shortened.
5.2 Simulations of Afferented Muscle to Evaluate Potential Neuromechan-
ical Factors to Explain Experimental Results
To examine potential origins of physiological tremor modulation with respect to joint angle, we simulated tonic,
isometric contraction trials at the muscle and tendon lengths corresponding to the three ankle angles tested
experimentally (Figure 4(A)). Simulations of afferented muscle can replicate physiological tremor, similar to
the experimental findings (Figure 4(B)).
We used a systematic process of elimination to identify the mechanical and neural factors responsible for
the angle dependent modulation of physiological tremor (Figure 8(A-1)). The considered factors included
musculotendon mechanics, presynaptic gain of Ia afferent feedback, and fusimotor drives.
10
Figure 5: RMS level of the EMG signals of the calf muscles (average of gastrocnemius medialis,
gastrocnemius lateralis, and soleus) does not change significantly as a function of joint posture. Each
point represents the EMG RMS level of a subject.
11
DFMIDPFAnkle Posture00.20.4EMG-RMSp = 0.12p = 0.45p = 1DFMID PF Figure 6: Changes in physiological tremor power were not correlated with possible co-activation of
the tibialis anterior and calf muscles: Panels show changes in tremor power as a function of changes
in the difference of the tibialis anterior and calf muscles RMS EMG values: (A) between the MID
and DF postures; (B) between the PF and DF postures; (C) between the PF and MID DF postures.
5.2.1 Effect of Mechanical Factors
Figure 8(A-2) shows the changes in physiological tremor power as a function of joint angle when all parameters
related to afferent feedback were kept constant across muscle lengths. Contrary to the experimental results, we
found that mechanical factors by themselves resulted in a decrease in physiological tremor when the muscle
shortened. This indicates that our experimental observations cannot be explained by neuromechanical factors
related to changes in muscle length, such as mechanical properties of the muscle-tendon system and changes in
the background activity of afferents.
5.2.2 Effect of Presynaptic Inhibition
To explore effects of afferent feedback modulation on physiological tremor amplitude, we ran simulations with
different afferent feedback gains. A previous study demonstrated that the level of presynaptic inhibition of Ia
afferents can vary with joint angle (Hwang, 2002). We tested this by increasing the presynaptic inhibition by
20% at DF (from presynaptic control level of -0.1 to -0.12) compared to the other two joint angles (Hwang,
2002). This modulation had little impact on physiological tremor amplitude (Figure 8 (C)), suggesting that
modulation of presynaptic inhibition of Ia afferent feedback does not suffice to reproduce our experimental
results.
5.2.3 Effect of Fusimotor Drive
Next, we tested the influence of fusimotor drive modulation. We first increased γ-dynamic as the muscle was
shortened. As shown in Figure 8(A-4), this was not sufficient to replicate experimental observations either.
From all the tested combinations of γ-dynamic and γ-static (see methods), we found that only increases in
γ-static resulted in greater physiological tremor in the plantarflexed posture. Figure 8(A-5) demonstrates that
increasing γ-static from 10 pps to 70 pps and to 190 pps in the DF, MID and PF angles with a fixed γ-dynamic
of 110 pps results in a consistent monotonic increase in physiological tremor amplitude similar to the pattern
we identified experimentally. Further, physiological tremor amplitude was significantly larger at the PF angle
compared to the other two angles (p < 0.01 for both comparisons) and in the MID compared to the DF angle (p
< 0.01). These results suggest that the modulation of physiological tremor observed experimentally could stem
from modulation of γ-static fusimotor drive with muscle/tendon stretch.
12
MID-DFSlope p-value : 0.4400.51EMG TA - EMG Calf-5051015Tremor PowerPF-DFSlope p-value : 0.50-1012EMG TA - EMG CalfPF-MIDSlope p-value : 0.85-1012EMG TA - EMG Calf(A)(B)(C) Figure 7: Torque power as a function of joint angle for individual subjects. (A) torque total power
shows no consistent pattern of change as a function of joint angle; (B) torque power in the physiolog-
ical tremor range (5-12 Hz) increases from dorsiflexed to plantarflexed angles.
13
(A)0123Total Power(Normalized to DF)(B)048121234567Subject NumberPhysiological Tremor Power(Normalized to DF)DFMID PF 5.2.4 Model Validation: Reproducing Muscle Stiffness
To confirm the validity of the chosen model parameters, we tested our model to see if it exhibited the known
angle dependent changes in muscle stiffness (Mirbagheri et al., 2000). Figure 8(B-2) shows that the model's
muscle stiffness increased from the PF to MID and to DF. This is consistent with a previous experimental
report that showed that ankle stiffness increased monotonically from 27◦ plantarflexion toward full dorsiflexion
as shown in Figure 8(B-1) (Mirbagheri et al., 2000). This further supports our conclusion that modulation of
physiological tremor as a function of joint angle cannot be explained by alterations in mechanical properties of
the musculotendon unit.
5.2.5 Model Validation: Reproducing Stretch Reflex Response
We further tested if the known effect of joint angles on stretch reflex amplitude (Mirbagheri et al., 2000) emerges
from our simulation when fusimotor parameters are set to replicate the experimental behavior of physiological
tremor. Figure 8(C-2 and C-5) shows that the stretch reflex response increased from PF to MID and to DF,
which is again consistent with previously reported observations Figure 8(C-1) (Mirbagheri et al., 2000). This
result confirms that our model produces physiologically realistic modulation of both physiological tremor and
stretch reflex amplitude.
5.2.6 Experimental Study: Validation of Stiffness and Stretch Reflex Response
We recruited one subject to further validate, simultaneously, the angle-dependence of both stretch reflex ampli-
tudes and muscle stiffness. We placed the ankle at the PF, MID and DF postures and applied pulse perturbations
to the ankle. The peak amplitude of the pulses was 1.15◦. The duration of each pulse was 1s and time interval
the between pulses was a random uniform variable with minimum of 2s and maximum 4s.
We used the technique for the simulation data to identify stiffness and reflex responses. Stiffness was 1.88
± 0.24 (Nm/◦) at the DF, 1.44 ± 0.16 (Nm/◦) at the MID, and 0.97 ± 0.16 (Nm/◦) at the PF postures. Stretch
reflex response was 0.94 ± 0.60 (Nm) at the DF, 0.37 ± 0.20 (Nm) at the MID, and 0.15 ± 0.05 (Nm) at the PF
postures. The direction and magnitude of changes in stiffness and stretch reflex response are very comparable
to those of the simulation model and the previous works (see Figure 8).
6 Discussion
We studied the dependence of physiological tremor on the length of the muscle-tendon complex. We hypothe-
sized that tremor amplitude would be proportional to the gain of the stretch reflex loop and therefore increase
when the muscle was lengthened. Surprisingly, we found the opposite: tremor amplitude increased when mus-
cle was shortened. We explored potential mechanical and neural mechanisms underlying this observation by
simulating a model of an afferented muscle-tendon system controlled in closed-loop. Using this simulation,
we were able to exclude basic mechanical properties of muscle responsible for our experimental observations.
Importantly, we found that replicating experimental results required increasing the activation of γ-static fusimo-
tor drive as the muscle was shortened. This previously-undescribed relationship between fusimotor drive and
muscle length (independent of movement, changes in α-drive, or antagonist muscle activity) may begin to fill
important gaps in our knowledge of fusimotor activity during voluntary actions. It may also provide novel
avenues for experimental characterization of fusimotor activity, or its dysfunction in the context of injury or
disease.
6.1 Limitations of the Study
As with any computational model, our results help to understand experimental observations, but are not direct
mechanistic proofs. No simulation is complete (Valero-Cuevas et al., 2009), and we intentionally used a simple
implementation which lacks biophysical models of spiking neurons, Renshaw cells, or supraspinal circuitry. Our
controller, though reminiscent of a cortical long latency reflex loop in conception, is not a neurophysiological
14
Figure 8: Simulation results showing that only an increase in γ-static is sufficient to replicate the
experimentally observed muscle length dependence of physiological tremor. We used an afferented
muscle model and systematically tested different factors that can contribute to physiological tremor.
By process of elimination we identified the sufficient conditions under which the model replicates ex-
perimental observations. The first column evaluates ankle-angle dependent changes in physiological
tremor in healthy adults(A-1); and in simulation when (A-2) only muscle length was varied and the
neural gains were held constant, when (A-3) presynaptic inhibition was increased at the dorsiflexed
posture, when (A-4) γ-dynamic was increased at the plantarflexed posture, and when (A-5) γ-static
was increased at the plantarflexed posture. The second column shows changes in muscle stiffness
across joint angles (i.e, muscle lengths) in healthy adults (B-1) and in our model (B-2). The third
column shows changes in stretch reflex responses in healthy adults(C-1) as well under the simulated
conditions described above (C-2 to C-5).
15
ExperimentalResultsMuscleLengthPresynapticInhibitionGammaDynamicGammaStaticPhysiological TremorSti(cid:31)nessStretch Re(cid:30)exResponse✔✗✗✗●Adapted from Mirbagheri et al. (2000)Adapted from Mirbagheri et al. (2000)(A-1)(B-1)(C-1)(A-2)(A-3)(A-4)(A-5)(C-2)(C-3)(C-4)(C-5)Do changes in these paramters su(cid:29)ce to reproduceexperimental results?Not ApplicableDFMIDPFDFMIDPFDFMIDPF(B-2) model of descending control, and we only consider a single isolated muscle rather than a coordinated group of
agonists/antagonists, as is almost always the case in reality.
The surface EMG measurement suffers from several issues including: cross-talk especially in ankle muscles
Winter et al. (1994), dispersion of action potentials Phinyomark et al. (2012), changes in muscle geometry, etc.
We should emphasize that while we used surface EMG measurement in this study, the precise correspondence
between EMG and torque was not the primary focus. Rather, we used EMG to demonstrate that (i) tremor had
neural origin (versus mechanical) by showing the torque variability in the tremor frequency range is coherent
with EMG, (ii) the average activation level in calf muscles is held constant as a function of joint angle, and (iii)
antagonistic co-activation was minimal and did not have a posture-dependent component. While we expect these
issues to have small effects in our final conclusions, it is within interest to use intramuscular EMG recording to
fully characterize the EMG-torque coherence.
Even with these limitations, our prediction of neural (rather than purely mechanical) responses to altered
muscle length remains a valid and particularly a straightforward explanation for our experimental observations.
In this sense, the simplicity of our simulation is a strength, since basic pathway gain adjustments can explain
many different aspects of tremor modulation.
In future, it is within our goals to include more anatomically and physiologically faithful structures of the
tendon-driven limbs. More specifically, we are interested to include multiple synergistic and antagonistic mus-
cles actuating the joint. We will use muscle models scaled to the three calf, and tibialis anterior muscles. We
will include more spinally mediated neural pathways connecting these muscles. This includes classical auto-
genic and heterogenic neural (heteronomous spindle pathways shared between the calf muscles and reciprocal
inhibition between the calf and tibialis anterior muscles) Nichols et al. (2016), Renshaw cells, and propriospinal
interneurons Raphael et al. (2010). Inclusion of β motoneurons may also expand our understanding of neural
control of mechanoreceptors. These neurons innervate both intrafusal and extrafusal muscle fibers Emonet-
D´enand et al. (1992); Jami et al. (1982); Manuel and Zytnicki (2011); Burke and Tsairis (1977). However, our
current understanding of these neurons is very limited and further research into this matter is required to pave
the way.
Moreover, as the next logical step, we will study physiological tremor across different muscle activation and
co-activation levels during tonic and time-varying contraction profiles. Based on the current knowledge of neu-
romuscular control, we can hypothesize that physiological tremor increases with muscle activation due to α− γ
co-activation and facilitation of recruitment of larger motoneurons. We also expect during co-activation of the
ankle plantarflexor/dorsiflexor muscles, reciprocal inhibition plays an important role in modulating physiologi-
cal tremor as a function of joint angle. Thus, when dorsiflexor is stretched (PF angle) or more activated, it has a
larger spindle afferent activity which in turn inhibits the motoneurons of the calf muscles. The degree to which
this inhibition can alter the physiological tremor is subject to future study and validation against experimental
data. This will provide critical insight into the coordination of muscles in force production tasks.
6.2 Muscle Length Dependence Modulation of Fusimotor Drive
Direct recording of fusimotoneurons in humans is not yet possible. Nevertheless, recording from afferent nerves
using microneurography has indirectly revealed that many factors can influence fusimotor activity in humans, in-
dependently of the physical state of the muscle (Dimitriou, 2016; Ribot-Ciscar et al., 2000; Vallbo and Hulliger,
1981; Prochazka et al., 1985).
Although a direct dependence of fusimotor drive on muscle length alone has not been reported previously, it
has been shown that activity of fusimotoneurons (specifically the γ-static motorneurons) co-varies with the ankle
joint angle during locomotion in decerebrate cat (Taylor et al., 2000, 2006; Ellaway et al., 2015). The favored
interpretation of this is the well-known hypothesis of simultaneous α− γ co-activation, Loeb (1984). Our study
suggests that the length of the muscle itself (independently of changes in α-drive) may contribute significantly
to the decoupled modulation of fusimotor activity (and more specifically the static fusimotor activity).
Our findings of independent modulation of γ-static with muscle length would also explain the common
interpretation that α − γ co-activation as a function of joint angle is a means to compensate for shortening of
intrafusal fibers. This has been demonstrated by (1) (Lan and He, 2012) that have shown the necessity of γ-static
control during arm locomotion with controlled spindle sensitivity using a simulation study ; and (2) (Grandjean
16
and Maier, 2014, 2017) who have shown a similar γ-static mechanism during wrist movement. Importantly, the
results of these modeling studies have almost exclusively been interpreted in the context of α − γ co-activation.
Deviations from strict α− γ co-activation have already been reported, e.g. in Dimitriou (2016) during active
naturalistic movements of the wrist joint and in (Ribot-Ciscar et al., 2009) during passive movement of the ankle
joint. More recently, it has been demonstrated that adjustment of spindle sensitivity via motorneuron efferent
control is not only proportional to the α drive of the parent muscle but also inversely proportional to the drive
of the antagonistic muscle thanks to the reciprocal inhibition pathways. Thus, a more general form of the α − γ
co-activation reflects the balance of activity between antagonistic muscles Dimitriou (2014). Nevertheless, our
findings of length-dependent, differential modulation of fusimotor drive (static only) that does not comply with
α − γ co-activation but provides a mechanistic approach to explain some of such deviations.
6.3 Alternative Mechanisms
Previous experimental studies have revealed that ankle joint stiffness, on average, decreases with plantarflexion
until 27◦ plantarflexion (Sobhani Tehrani et al., 2013; Jalaleddini et al., 2015). Thus, a reduction of muscle
stiffness with muscle shortening during plantarflexion might explain our results (Davidson and Charles, 2016).
Our results come from the systematic exploration of the capabilities of our model under different conditions.
For example, if we only change muscle length (Figure 8(row 2)) we see its consequence to physiological tremor,
intrinsic muscle stiffness and stretch reflex response. This case demonstrates that our model is capable of
replicating the intrinsic stiffness of muscles in this task (B-2), yet did not replicate the observed physiological
tremor modulation when γ-static was constant (A-2).
Rows 3 and 4 in Figure 8 demonstrate that how changing the gain of the afferent feedback (while also
including changes in muscle length) did not replicate experimental observations of physiological tremor either
(A-3, presynaptic inhibition, and A-4, γ-dynamic).
Consequently, the mechanical properties (muscle stiffness), presynaptic inhibition, and γ-dynamic fusimo-
tor drive -by themselves- do not explain all experimental observations of physiological tremor, muscle stiff-
ness and stretch reflex response. These results strengthen our conclusion that fusimotor drive decoupled from α
drive (i.e., changes in γ-static with changes in muscle length) is a potentially important mechanism.
6.4 Stretch Reflex and Physiological Tremor Amplitudes Can Be Decou-
pled
Physiological tremor is characterized by synchronized firing of motoneurons in the range of 6-12 Hz, which is
believed to emerge from self-sustaining oscillations of excitation in the stretch reflex loop (Hagbarth and Young,
1979; Lippold, 1970; McAuley and Marsden, 2000; Takanokura and Sakamoto, 2005). Previous studies have
shown that the amplitude of physiological tremor increased whenever afferent gain increased, for example by
cognitive/perceptual aspects of force control tasks (Laine et al., 2015b), or by direct manipulation of afferent
nerves (Christakos et al., 2006).
Accordingly, it is not unreasonable to expect that physiological tremor amplitude should increase when
muscle was lengthened, as muscle lengthening is known to increase stretch reflex amplitudes (Mirbagheri et al.,
2000)-yet this relationship has not been conclusively established as obligatory.
In contrast to this line of
reasoning, our data demonstrate decoupling of physiological tremor and stretch reflex amplitudes.
Decoupling of measures of afferent sensitivity has been previously reported (Maluf et al., 2007). We have
even recently documented that increases in γ-static fusimotor drive can actually decrease the stretch reflex
amplitude (Jalaleddini et al., 2017). Our results here now provide further evidence to suggest that the amplitudes
of the stretch reflex and physiological tremor are separable-likely because they are driven by distinct, but
complementary and sometimes overlapping, afferent information. This is likely because stretch reflex amplitude
depends on sensitivity/gain along different pathways, each of which operates on a slightly different time-scale,
and responds to different stimuli (length, velocity, force, etc).
We propose that assessment of muscle force variability might be an important avenue for future investigation
of reflex modulation in health and disease. While we have identified a potential fusimotor strategy that suffices
17
to replicate changes in physiological tremor, muscle stiffness, and stretch reflex response, it will require further
research to fully characterize the relationship between fusimotor drive and physiological tremor.
6.5 Clinical Implications
Rhythmic oscillation in execution of motor tasks is clinically significant in a variety of disorders including
Parkinson's disease (Ko et al., 2015; Vaillancourt et al., 2001), Dystonia (Xia and Bush, 2007; Chu and Sanger,
2009), bruxism (Laine et al., 2015a), essential tremor (H´eroux et al., 2010; Gallego et al., 2017), etc. Moreover,
abnormal gains and thresholds in the spinal loop circuitry have been documented in pathologies such as spinal
cord injury, stroke, and multiple sclerosis (O'Sullivan et al., 1998; Jobin and Levin, 2000; Kamper and Rymer,
2000; Tuzson et al., 2003).
The results of this study suggest that the variability in isometric ankle torque production at different dorsi-
plantarflexion angles could serve as a simple, non-invasive, quick and clinically-practical tool to characterize
the integrity of afferent and reflex circuitry. Thus, it can serve as a means to explore neuromechanical coupling
and reflex pathway gains during natural behavior-and without the need for mechanical perturbation or nerve
stimulation. Because the amplitude of physiological tremor is negligible for most motor control tasks, it has been
often considered as noise. Our results show it might be a rich source of physiologically relevant information.
The alternative of using direct central recordings in humans to assess reflex pathways is severely limited by
practical/ethical considerations (McAuley and Marsden, 2000).
Our modeling environment makes it feasible to systematically vary all physiological parameters of the sys-
tem (e.g.
fusimotor drive, presynaptic inhibition, muscle and load mechanical parameters, etc) to replicate
experimental observations. As our capabilities to simulate neuromechanical interactions and reflex modulation
becomes more complete, it may be possible to guide the design of experiments and methods to extract important
information about the state of injury or dysfunction.
7 Acknowledgment
Research reported in this publication was supported by the National Institute of Arthritis and Musculoskeletal
and Skin Diseases of the National Institutes of Health under Awards Number R01 AR-050520 and R01 AR-
052345 to FVC. The contents of this endeavor is solely the responsibility of the authors and does not necessarily
represent the official views of the National Institutes of Health. This work was also supported by the Canadian
Institutes of Health Research to REK, and Fonds Qu´eb´ecois de la Recherche sur la Nature et les Technologies
to KJ an MG.
References
Alibiglou, L., W. Z. Rymer, R. L. Harvey, and M. M. Mirbagheri
2008. The relation between ashworth scores and neuromechanical measurements of spasticity following
stroke. Journal of neuroengineering and rehabilitation, 5(1):18.
Arnold, E. M., S. R. Ward, R. L. Lieber, and S. L. Delp
2010. A model of the lower limb for analysis of human movement. Annals of biomedical engineering,
38(2):269–279.
Baker, S., R. Spinks, A. Jackson, and R. Lemon
2001. Synchronization in monkey motor cortex during a precision grip task. i. task-dependent modulation in
single-unit synchrony. Journal of Neurophysiology, 85(2):869–885.
Baker, S. N., E. M. Pinches, and R. N. Lemon
2003. Synchronization in monkey motor cortex during a precision grip task. ii. effect of oscillatory activity
on corticospinal output. Journal of Neurophysiology, 89(4):1941–1953.
18
Brown, I. E., E. J. Cheng, and G. E. Loeb
1999. Measured and modeled properties of mammalian skeletal muscle. ii. the effectsof stimulus frequency
on force-length and force-velocity relationships. Journal of Muscle Research & Cell Motility, 20(7):627–643.
Brown, I. E. and G. E. Loeb
2000. Measured and modeled properties of mammalian skeletal muscle: Iv. dynamics of activation and
deactivation. Journal of muscle research and cell motility, 21(1):33–47.
Brown, I. E., S. H. Scott, and G. E. Loeb
1996. Mechanics of feline soleus: Ii design and validation of a mathematical model. Journal of Muscle
Research & Cell Motility, 17(2):221–233.
Burke, R. E. and P. Tsairis
1977. Histochemical and physiological profile of a skeletofusimotor (β) unit in cat soleus muscle. Brain
research, 129(2):341–345.
Capaday, C. and R. Stein
1986. Amplitude modulation of the soleus h-reflex in the human during walking and standing. Journal of
Neuroscience, 6(5):1308–1313.
Cheng, E. J., I. E. Brown, and G. E. Loeb
2000. Virtual muscle: a computational approach to understanding the effects of muscle properties on motor
control. Journal of neuroscience methods, 101(2):117–130.
Christakos, C. N., N. A. Papadimitriou, and S. Erimaki
2006. Parallel neuronal mechanisms underlying physiological force tremor in steady muscle contractions of
humans. Journal of neurophysiology, 95(1):53–66.
Chu, W. T. V. and T. D. Sanger
2009. Force variability during isometric biceps contraction in children with secondary dystonia due to cere-
bral palsy. Movement Disorders, 24(9):1299–1305.
Cresswell, A. and W. Loscher
2000. Significance of peripheral afferent input to the α-motoneurone pool for enhancement of tremor during
an isometric fatiguing contraction. European journal of applied physiology, 82(1-2):129–136.
Davidson, A. D. and S. K. Charles
2016. Fundamental principles of tremor propagation in the upper limb. Annals of Biomedical Engineering,
Pp. 1–15.
De Vlugt, E., J. H. de Groot, K. E. Schenkeveld, J. Arendzen, F. C. van der Helm, and C. G. Meskers
2010. The relation between neuromechanical parameters and ashworth score in stroke patients. Journal of
neuroengineering and rehabilitation, 7(1):35.
Dimitriou, M.
2014. Human muscle spindle sensitivity reflects the balance of activity between antagonistic muscles. Journal
of Neuroscience, 34(41):13644–13655.
Dimitriou, M.
2016. Enhanced muscle afferent signals during motor learning in humans. Current Biology, 26(8):1062–
1068.
Elias, L. A., R. N. Watanabe, and A. F. Kohn
Spinal mechanisms may provide a combination of intermittent and continuous control of hu-
2014.
man posture: predictions from a biologically based neuromusculoskeletal model. PLoS Comput Biol,
10(11):e1003944.
19
Ellaway, P. H., A. Taylor, and R. Durbaba
2015. Muscle spindle and fusimotor activity in locomotion. Journal of anatomy, 227(2):157–166.
Emonet-D´enand, F., J. Petit, and Y. Laporte
1992. Comparison of skeleto-fusimotor innervation in cat peroneus brevis and peroneus tertius muscles. The
Journal of physiology, 458(1):519–525.
Gallego, J. A., J. L. Dideriksen, A. Holobar, E. Rocon, J. L. Pons, and D. Farina
2017. Neural control of muscles in tremor patients. In Converging Clinical and Engineering Research on
Neurorehabilitation II, Pp. 129–134. Springer.
Grandjean, B. and M. A. Maier
2014. Model-based prediction of fusimotor activity and its effect on muscle spindle activity during voluntary
wrist movements. Journal of computational neuroscience, 37(1):49–63.
Grandjean, B. and M. A. Maier
2017. Emergence of gamma motor activity in an artificial neural network model of the corticospinal system.
Journal of computational neuroscience, 42(1):53–70.
Hagbarth, K.-E. and R. R. Young
1979. Participation of the stretch reflex in human physiological tremor. Brain: a journal of neurology,
102(3):509–526.
H´eroux, M., G. Pari, and K. Norman
2010. The effect of contraction intensity on force fluctuations and motor unit entrainment in individuals with
essential tremor. Clinical Neurophysiology, 121(2):233–239.
Hwang, I.-S.
2002. Assessment of soleus motoneuronal excitability using the joint angle dependent h reflex in humans.
Journal of Electromyography and Kinesiology, 12(5):361–366.
Iwanuma, S., R. Akagi, S. Hashizume, H. Kanehisa, T. Yanai, and Y. Kawakami
2011. Triceps surae muscle–tendon unit length changes as a function of ankle joint angles and contraction
levels: the effect of foot arch deformation. Journal of biomechanics, 44(14):2579–2583.
Jalaleddini, K., M. A. Golkar, D. L. Guarin, E. S. Tehrani, and R. E. Kearney
2015. Parametric methods for identification of time-invariant and time-varying joint stiffness models. IFAC-
PapersOnLine, 48(28):1375–1380.
Jalaleddini, K., C. M. Niu, S. C. Raja, W. J. Sohn, G. E. Loeb, T. D. Sanger, and F. J. Valero-Cuevas
2017. Neuromorphic meets neuromechanics part II: The role of fusimotor drive. Journal of Neural Engineer-
ing.
Jami, L., K. Murthy, and J. Petit
1982. A quantitative study of skeletofusimotor innervation in the cat peroneus tertius muscle. The Journal of
Physiology, 325(1):125–144.
Jobin, A. and M. F. Levin
2000. Regulation of stretch reflex threshold in elbow flexors in children with cerebral palsy: a new measure
of spasticity. Developmental Medicine & Child Neurology, 42(8):531–540.
Kamper, D. G. and W. Z. Rymer
2000. Quantitative features of the stretch response of extrinsic finger muscles in hemiparetic stroke. Muscle
& nerve, 23(6):954–961.
Kawakami, Y., Y. Ichinose, and T. Fukunaga
1998. Architectural and functional features of human triceps surae muscles during contraction. Journal of
Applied Physiology, 85(2):398–404.
20
Kearney, R. E., R. B. Stein, and L. Parameswaran
1997. Identification of intrinsic and reflex contributions to human ankle stiffness dynamics. IEEE Transac-
tions on Biomedical Engineering, 44(6):493–504.
Ko, N.-h., C. M. Laine, B. E. Fisher, and F. J. Valero-Cuevas
2015. Force variability during dexterous manipulation in individuals with mild to moderate parkinsons dis-
ease. Frontiers in aging neuroscience, 7.
Laine, C., S¸. Yavuz, J. DAmico, M. Gorassini, K. Turker, and D. Farina
2015a. Jaw tremor as a physiological biomarker of bruxism. Clinical Neurophysiology, 126(9):1746–1753.
Laine, C., S¸. Yavuz, and D. Farina
2014. Task-related changes in sensorimotor integration influence the common synaptic input to motor neu-
rones. Acta Physiologica, 211(1):229–239.
Laine, C. M., E. Martinez-Valdes, D. Falla, F. Mayer, and D. Farina
2015b. Motor neuron pools of synergistic thigh muscles share most of their synaptic input. Journal of
Neuroscience, 35(35):12207–12216.
Laine, C. M., A. Nagamori, and F. J. Valero-Cuevas
2016. The dynamics of voluntary force production in afferented muscle influence involuntary tremor. Fron-
tiers in Computational Neuroscience, 10.
Lan, N. and X. He
2012. Fusimotor control of spindle sensitivity regulates central and peripheral coding of joint angles. Fron-
tiers in computational neuroscience, 6:66.
Lippold, O.
1970. Oscillation in the stretch reflex arc and the origin of the rhythmical, 8-12 c/s component of physiological
tremor. The Journal of Physiology, 206(2):359.
Lippold, O.
1971. Physiological tremor. Scientific American, 224(3):65–73.
Loeb, G. E.
1984. The control and responses of mammalian muscle spindles during normally executed motor tasks.
Exercise and sport sciences reviews, 12(1):157–204.
Maluf, K. S., B. K. Barry, Z. A. Riley, and R. M. Enoka
2007. Reflex responsiveness of a human hand muscle when controlling isometric force and joint position.
Clinical neurophysiology, 118(9):2063–2071.
Manuel, M. and D. Zytnicki
2011. Alpha, beta and gamma motoneurons: functional diversity in the motor system's final pathway. Journal
of integrative neuroscience, 10(03):243–276.
McAuley, J. and C. Marsden
2000. Physiological and pathological tremors and rhythmic central motor control. Brain, 123(8):1545–1567.
Mileusnic, M. P., I. E. Brown, N. Lan, and G. E. Loeb
2006. Mathematical models of proprioceptors. i. control and transduction in the muscle spindle. Journal of
neurophysiology, 96(4):1772–1788.
Mirbagheri, M., H. Barbeau, and R. Kearney
2000. Intrinsic and reflex contributions to human ankle stiffness: variation with activation level and position.
Experimental Brain Research, 135(4):423–436.
21
Mugge, W., D. A. Abbink, A. C. Schouten, J. P. Dewald, and F. C. van der Helm
2010. A rigorous model of reflex function indicates that position and force feedback are flexibly tuned to
position and force tasks. Experimental brain research, 200(3-4):325–340.
Nichols, T. R., N. E. Bunderson, and M. A. Lyle
2016. Neural regulation of limb mechanics: insights from the organization of proprioceptive circuits. In
Neuromechanical modeling of posture and locomotion, Pp. 69–102. Springer.
O'Sullivan, M., S. Miller, V. Ramesh, E. Conway, K. Gilfillan, S. McDonough, and J. Eyre
1998. Abnormal development of biceps brachii phasic stretch reflex and persistence of short latency heterony-
mous reflexes from biceps to triceps brachii in spastic cerebral palsy. Brain, 121(12):2381–2395.
Phinyomark, A., S. Thongpanja, H. Hu, P. Phukpattaranont, and C. Limsakul
2012. The usefulness of mean and median frequencies in electromyography analysis. In Computational intel-
ligence in electromyography analysis-A perspective on current applications and future challenges. InTech.
Prochazka, A., M. Hulliger, P. Zangger, and K. Appenteng
1985. fusimotor set: new evidence for α-independent control of γ-motoneurones during movement in the
awake cat. Brain research, 339(1):136–140.
Raphael, G., G. A. Tsianos, and G. E. Loeb
2010. Spinal-like regulator facilitates control of a two-degree-of-freedom wrist. Journal of Neuroscience,
30(28):9431–9444.
Ribot-Ciscar, E., V. Hospod, J.-P. Roll, and J.-M. Aimonetti
2009. Fusimotor drive may adjust muscle spindle feedback to task requirements in humans. Journal of
neurophysiology, 101(2):633–640.
Ribot-Ciscar, E., C. Rossi-Durand, and J.-P. Roll
2000. Increased muscle spindle sensitivity to movement during reinforcement manoeuvres in relaxed human
subjects. The Journal of Physiology, 523(1):271–282.
Sobhani Tehrani, E., K. Jalaleddini, and R. E. Kearney
2013. Linear parameter varying identification of ankle joint intrinsic stiffness during imposed walking move-
In 2013 35th Annual International Conference of the IEEE Engineering in Medicine and Biology
ments.
Society (EMBC), Pp. 4923–4927. IEEE.
Song, D., N. Lan, G. Loeb, and J. Gordon
2008a. Model-based sensorimotor integration for multi-joint control: development of a virtual arm model.
Annals of biomedical engineering, 36(6):1033–1048.
Song, D., G. Raphael, N. Lan, and G. Loeb
2008b. Computationally efficient models of neuromuscular recruitment and mechanics. Journal of Neural
Engineering, 5(2):175.
Sowman, P. F. and K. S. Turker
2005. Methods of time and frequency domain examination of physiological tremor in the human jaw. Human
movement science, 24(5):657–666.
Takanokura, M. and K. Sakamoto
2005. Neuromuscular control of physiological tremor during elastic load. Medical science monitor,
11(4):CR143–CR152.
Taylor, A., R. Durbaba, P. Ellaway, and S. Rawlinson
2006. Static and dynamic γ-motor output to ankle flexor muscles during locomotion in the decerebrate cat.
The Journal of physiology, 571(3):711–723.
22
Taylor, A., P. Ellaway, R. Durbaba, and S. Rawlinson
2000. Distinctive patterns of static and dynamic gamma motor activity during locomotion in the decerebrate
cat. The Journal of physiology, 529(3):825–836.
Toft, E., T. Sinkjaer, S. Andreassen, and K. Larsen
1991. Mechanical and electromyographic responses to stretch of the human ankle extensors. Journal of
neurophysiology, 65(6):1402–1410.
Tsianos, G. A., J. Goodner, and G. E. Loeb
2014. Useful properties of spinal circuits for learning and performing planar reaches. Journal of neural
engineering, 11(5):056006.
Tuzson, A. E., K. P. Granata, and M. F. Abel
2003. Spastic velocity threshold constrains functional performance in cerebral palsy. Archives of physical
medicine and rehabilitation, 84(9):1363–1368.
Vaillancourt, D. E., A. B. Slifkin, and K. M. Newell
2001. Regularity of force tremor in parkinson's disease. Clinical Neurophysiology, 112(9):1594–1603.
Valero-Cuevas, F. J., H. Hoffmann, M. U. Kurse, J. J. Kutch, and E. A. Theodorou
2009. Computational models for neuromuscular function. IEEE reviews in biomedical engineering, 2:110–
135.
Vallbo, A. and M. Hulliger
1981. Independence of skeletomotor and fusimotor activity in man? Brain research, 223(1):176–180.
Weiss, P., R. Kearney, and I. Hunter
1986. Position dependence of ankle joint dynamicsi. passive mechanics. Journal of biomechanics, 19(9):727–
735.
Wickiewicz, T. L., R. R. Roy, P. L. Powell, and V. R. Edgerton
1983. Muscle architecture of the human lower limb. Clinical orthopaedics and related research, 179:275–
283.
Wilcox, R. R.
2012. Introduction to robust estimation and hypothesis testing. Academic Press.
Winter, D., A. Fuglevand, and S. Archer
1994. Crosstalk in surface electromyography: theoretical and practical estimates. Journal of Electromyogra-
phy and Kinesiology, 4(1):15–26.
Xia, R. and B. M. Bush
2007. Modulation of reflex responses in hand muscles during rhythmical finger tasks in a subject with writers
cramp. Experimental brain research, 177(4):573–578.
Young, R. R. and K.-E. Hagbarth
1980. Physiological tremor enhanced by manoeuvres affecting the segmental stretch reflex. Journal of
Neurology, Neurosurgery & Psychiatry, 43(3):248–256.
23
|
1902.01257 | 2 | 1902 | 2019-06-13T09:16:27 | Cortical thickness and functional networks modules by cortical lobes | [
"q-bio.NC"
] | This study aims to investigate topological organization of cortical thickness and functional networks by cortical lobes. First, I demonstrated modular organization of these networks by the cortical surface frontal, temporal, parietal and occipital divisions. Secondly, I mapped the overlapping edges of cortical thickness and functional networks for positive and negative correlations. Finally, I showed that overlapping positive edges map onto within-lobe cortical interactions and negative onto between-lobes interactions. | q-bio.NC | q-bio |
Cortical thickness and functional networks modules by
cortical lobes
Vesna Vuksanovi´c
Aberdeen Biomedical Imaging Centre
University of Aberdeen
[email protected]
June 14, 2019
1 Abstract
This study aims to investigate topological organization of cortical thickness and func-
tional networks by cortical lobes. First, I demonstrated modular organization of these
networks by the cortical surface frontal, temporal, parietal and occipital divisions. Sec-
ondly, I mapped the overlapping edges of cortical thickness and functional networks for
positive and negative correlations. Finally, I showed that overlapping positive edges map
onto within-lobe cortical interactions and negative onto between-lobes interactions.
2 Introduction
Classical, localized characterization of human brain surface morphology has moved in
recent years towards maps of inter-regional co-variations of gray matter volume or thick-
ness measurements in structural magnetic resonance imaging data (sMRI) [1]. To make
such cortical maps or brain networks (graphs) the edges (links) between the regions
(nodes) are defined by the strength of correlation between regional volume or cortical
thickness. The physiological mechanisms underlying the thickness correlation among
cortical areas remain unclear. Neuroimaging findings propose strong correlation of the
cortical thickness measurements between regions that are directly connected via axonal
connections [2, 3]. However, only a fraction of variations in regional thickness correla-
tions across the cortex has been explained by direct axonal links [3]. In this context
it is interesting to ask whether the regional thickness (morphological) co-variations also
reflect functional associations (or parts of them). Recent studies suggest that functional
network properties, such as the strengths or the number of correlations, can impact
structural properties. The synchronization of neuronal activity in response to specific
1
functional demands might induce synchronized plastic changes among related regions
during brain development [4] or degenerative diseases [5]. More pertinent to this study,
such interplay between regional structural and functional properties may contribute to
strong thickness correlations observed within the visual areas of the occipital lobe [6].
Here it is also worth noting that, while it has been observed that functional modules
mirror the local brain anatomy a number of studies show deviations between these two
suggesting many-to-one function-structure mapping [7].
The natural divisions of the cortical surface -- frontal, temporal, parietal and occipital
-- have been defined by regional morphological characteristics, which are also known to
support different functions [8, 9]. In this context, questions arise about the relationship
between cortical surface inter-regional correlations, i.e., how do variations in functional
and thickness correlations map onto each other? To what extent these cortical struc-
tural and functional correlations share similar topological organization? If topological
measures of functional network organization are structurally relevant we might expect
them to be impacted at the lobe level of the cortical surface organization.
To address these questions I examined thickness and functional correlational net-
works based on two different neuroimaging modalities -- structural and functional MRI.
The networks were represented by 68 × 68 correlation matrices, where functional and
thickness network maps inter-regional correlations between brain activity and thickness
measurements respectively. The main focus of the study is to shed light on intrinsic
topological properties of the frontal, temporal, parietal and occipital divisions, which
may impact structural and functional interactions across the cortical surface. To this
end I studied how modular organization of the cortical surface divisions affect overlap
of the cortical thickness and functional correlations across cortical regions.
3 Methods
The MRI data sets considered in this short communication are from a public database
(http : //f con1000.projects.nitrc.org/) provided by the Max Planck Institute (MPI)
Leipzig. A study group included 37 (16M/21F) participants between 20 and 42 years.
MRI data was acquired on 3 Tesla Magnetom Tim Trio scanner (Siemens, Erlangen,
Germany) using a 32-channel head coil. T1-weighted images were acquired using a
MPRAGE sequence (TR = 1.3 s; TE=3.46 ms; flip angle=10deg; FOV=256× 240mm2;
176 sagittal slices; voxel size = 1× 1× 1.5mm). Functional MRI/EPI data were acquired
on a 3T MRI scanner (Siemens Tim Trio) using TR=2.3 sec, TE=30ms, 3 × 3 in-plane
resolution, 3 mm slice thickness, 1 mm gap between slices. Each scanning session was a
task-absent ("resting state") scan lasting 7.6 minutes during which subjects were asked
to fixate a fixation cross.
Processing the structural MRI data was done using FreeSurfer v5.3.0 (https://
surfer.nmr.mgh.harvard.edu/) pipeline according to the procedure described in more
details in [6]. Cortical thickness was measured on N = 68 cortical surface regions
segmented according to the Desikan-Keliany Atlas (DKA) [10]. FSL v5.0.11 (https:
//fsl.fmrib.ox.ac.uk/fsl/fslwiki/) was used to process fMRI data and extract
2
Figure 1: Workflow diagram of the cortical networks' construction on processed struc-
tural and functional magnetic resonance images.
(Upper panel) Structural magnetic
resonance images (sMRI) are processed using FreeSurfer v5.3.0 pipeline: cortical surface
was reconstructed and regional cortical thickness measures were averaged over N = 68
structures of the Desikan-Keliany Atlas (DKA). Structural correlation network was con-
structed on partial correlations between regional thickness across all participants (S =
37) while controlling for age, sex and mean CT. (Lower panel) Functional MRI (fMRI)
were parcellated and time series were averaged over the same 68 DKA regions. Individual
functional correlations (FC) matrix was constructed on pair-wise correlations between
the fMRI time series, Fisher z-transformed, averaged over S subjects and transformed
back to linear Pearson coefficient. Group-wise CT and FC matrices were reordered ac-
cording to node affiliation with cortical lobes (frontal, temporal, parietal and occipital)
and their coupling topology was assessed to capture positive and negative correlations
maps.
BOLD fMRI time-series on the same 68 cortical regions of the DKA parcellations. The
fMRI data were processed using the FSL toolbox FEAT fMRI analysis, similarly to our
previous work [11] and global signal regression was applied in order to reduce motion
artifacts [12].
3.1 Brain Graph Construction
Brain graphs considered in this study were constructed on correlations between either
regional cortical thickness or BOLD activity measured by structural or functional MRI.
In these correlational networks, a brain region represent a network node and a pair-wise
correlation between the regional measurements represents network's edge/link or con-
nection between the nodes. See Fig. 1 that summarizes analysis pipeline and approach.
3
3.1.1 Functional Correlations
To obtain the FC matrices, time series of the 68 DKA regions were calculated by av-
eraging the respective BOLD signals over all regional voxels. The correlation matrices
were obtained for each subject in the study. Since voxels are in MNI space and a given
voxel has approximately the same anatomical position in all subjects, the individual cor-
relation matrices can be averaged across subjects. In detail, each matrix is first Fisher
z-transformed and averaged across subjects and then transformed back into correlation
coefficients again as described in previous studies [11, 13]. The resulting correlation
matrix fij, (i, j = 1, . . . , N ), was used to create a group-wise functional correlation ma-
trix between the regions of interest. See Fig. 1 (lower panel) for analysis pipeline and
approach.
3.1.2 Cortical Thickness Correlations
A group-wise structural network was constructed on inter-regional cortical thickness
correlations across all study participants. The correlations were calculated between each
pair of 68 regional thickness measurements while controlling for age, gender and mean
CT, similarly to previously described methodology [6, 3]. See Fig. 1 (upper panel) for
analysis pipeline and approach.
3.2 Brain Graph Analysis
3.2.1 Network Density
Network topological properties depend on the network thresholding. Threshold affects
network density (κ), i.e., number of links relative to the total number of all possible links
in the network [14]. The threshold considered here was set to yield a fully connected
network, i.e., one network component. In addition, the threshold was applied to ensure
that both networks CT and FC have the same density, i.e., to ensure that the networks
are analyzed and compared across the same number of links.
3.2.2 Modularity by Lobes
Each network was assessed at the scale of its lobar organization -- the frontal, tempo-
ral, parietal and occipital divisions of the cortical surface. Modularity index (Q), was
calculated to determine whether cortical lobes as conventionally defined correspond to
modules of the CT/FC network. This can be done by calculating the Q according to lobe
(by employing vector of nodal affiliation with the particular lobe as initial community af-
filiation vector). The modularity index quantifies the observed fraction of within-module
degree values relative to those expected if connections were randomly distributed across
the network. In the context of this short communication, cortical lobes have been consid-
ered and analyzed as network modules. Since the constructed CT/FC network contains
both positive and negative edge strengths, I used the asymmetric generalization of the
modularity quality function introduced in Rubinov and Sporns [15]. Since the algorithm
4
returns varying results on different runs, to control for the variance in the analysis, mod-
ularity index Q was calculated 100 times for each network, i.e., from running the same
analysis 100 times on CT/FC network. In other words, index Q was estimated from
100 runs on CT/FT network and on 100 randomized (CT/FC each) versions of these
networks. (See section 4.2) for details.) For networks of similar size, it is accepted that a
Q value above 0.3 is a good indicator of the existence of significant modules in a network
[16].
4 Results
4.1 Network Density
As described in the Methods, a single value for network density was obtained by con-
sidering percolation threshold for each network. This yielded a network density of κ
= 0.16, which includes both positive and negative edges. This choice of the threshold
insures that each network (either CT or FC) has the same (N = 362) total number of
positive and negative links.
4.2 Modularity by Lobes
Modularity index of the CT and FC networks by lobes were 0.446 and 0.442 (mean
values on 100 runs for each network) respectively, thus indicating modular organization of
CT/FC networks by cortical lobes. To get the confidence interval for network modularity
distribution by lobar (frontal, temporal, parietal and occipital) divisions, 200 randomized
networks were generated on CT/FC networks (100 each), i.e., each null network model
was tested to obtain distribution of the modularity indexes (Qs) on randomly distributed
correlations of the CT/FC network. Fig. 2(B) shows how the distributions of 100 Q
values on random CT/FC networks differ from those calculated on real networks (i.e.,
on 100 calculations of the modularity index for each network).
4.3 Convergent Cortical Thickness and Functional Networks Maps
Overlapping thickness and functional networks were calculated by element-wise multi-
plication of the two matrices. Fig. 2(D) shows products of this multiplication for the
coupling of positive and negative correlations/networks. The survived network edges
of positive correlations map intra-lobar and inter-hemispheric homologous (off-diagonal
matrix elements). In contrast, the coupled negative networks map exclusively onto inter-
lobar connections.
In particular, the frontal lobe regions -- rostral anterior cingulate,
rostral middle frontal and superior frontal -- are correlated with regions of the occipital
lobe. The mean distance between the coupled networks nodes (Nagree) was (72± 30)mm
for positive and (81 ± 30)mm for negative network nodes (p > 0.05), Fig. 2(C).
5
Figure 2: Cortical thickness and functional correlations networks. (A) Group-wise cor-
tical thickness (CT) and functional correlations (FC) networks arranged (from top to
bottom) according to nodal affiliations with frontal, temporal, parietal and occipital
lobes. (B) Histograms of modularity index (Q) by lobe calculated on 100 randomized
CT (magenta) and FC (cyan) networks and compared against the modularity index
calculated on group-wise CT (magenta) and FC (cyan) networks (values of 0.446 and
0.442, respectively). (C) Histogram of the cortical distances for the convergent positive
(red) and negative (blue) networks. Vertical, color-coded lines represent mean value of
the corresponding cortical distances (i.e., distances between regions with the overlapping
positive or negative correlations). (D) Overlapping edges of positive and negative CT
and FC networks; red -- positive correlations; blue -- negative correlations.
6
5 Discussion
This study demonstrates convergence of the cortical thickness and functional correlations
networks, with reference to modular organization of the conventional lobar divisions.
Noticeably different patterns of the overlapping correlations were observed for positive
and negative networks. In contrast to the positive networks' overlaps, which map al-
most exclusively within-lobe interactions, negative networks show only between-lobes
interactions.
Both cortical networks -- inter-regional thickness and functional correlations -- demon-
strated modular organizations that mirror conventional frontal, temporal, parietal and
occipital divisions of the cortical surface. Thus allowing for inferences about overlapping
interactions of these networks by cortical lobes. While several previous studies reported
modular topology of the cortex by cortical morphological (thickness and volume) [17, 18]
or functional [19] measurements, no data are available on functional modules as an in-
stinct property of the cortical surface lobar divisions. Motivated by the lobe-specific
patterns of atrophy progression in two types of dementia, we have recently reported
similar findings on cortical thickness and surface area networks in healthy elderly sub-
jects [6]. This study represents first effort to map patterns of positive and negative
correlations in cortical thickness and functional networks in healthy adults.
Several interesting results observed when mapping overlapping links between the cor-
tical thickness and functional networks by lobar divisions need further attention. First,
the approach adopted here was to compare the networks at a single value of network
density. The advantage of this approach is that is maps only the most pronounced cor-
relations within the both networks. This means that the correlations statistically not
different from zero were not considered in the analysis. However, the obtained topologies
may include different numbers of correlations for a larger study-group size. Thus, the
calculation of the maps of overlapping correlations across a range of networks densities
(based also on a larger study group) may reveal some of the cortical interactions that did
not pass statistical significance test in this study. Second, the definition of conventional
lobar divisions used in this study was based on the Desikan-Kelliany cortical parcellation.
Although well established, there exist alternative approaches for the cortical segmenta-
tion by using different brain templates [20]. Since adoption of different brain templates
may have influence on the patterns of network correlations, further studies could validate
these results using different anatomical (or functional) classifications. Finally, a study
combining structural, functional and diffusion-weighted MRI data could test to what ex-
tent variations and strengths of positive and negative correlations can be inferred from
underlying direct axonal links or synchronous/asynchronous functional links. These fu-
ture studies could also resolve problems of estimating networks individually for each
subject.
Acknowledgements
The author would like to acknowledge the support of the Maxwell compute cluster funded
by the University of Aberdeen.
7
References
[1] Evans, A. C. Networks of anatomical covariance. Neuroimage 80, 489 -- 504 (2013).
[2] Lerch, J. P. et al. Mapping anatomical correlations across cerebral cortex (macacc)
using cortical thickness from mri. Neuroimage 31, 993 -- 1003 (2006).
[3] Gong, G., He, Y., Chen, Z. J. & Evans, A. C. Convergence and divergence of
thickness correlations with diffusion connections across the human cerebral cortex.
Neuroimage 59, 1239 -- 1248 (2012).
[4] Hyde, K. L. et al. The effects of musical training on structural brain development:
a longitudinal study. Annals of the New York Academy of Sciences 1169, 182 -- 186
(2009).
[5] Witiuk, K. et al. Cognitive deterioration and functional compensation in als mea-
sured with fmri using an inhibitory task. Journal of Neuroscience 34, 14260 -- 14271
(2014).
[6] Vuksnaovi´c, V., Staff, R., Ahearn, T., Murray, A. & Claude, W. Cortical thick-
ness and surface area networks in healthy aging, alzheimer's disease and behavioral
variant fronto-temporal dementia. Int J Neur Sys (in press).
[7] Park, H.-J. & Friston, K. Structural and functional brain networks: from connec-
tions to cognition. Science 342, 1238411 (2013).
[8] Kandel, E. R. et al. Principles of neural science, vol. 4 (McGraw-hill New York,
2000).
[9] Mesulam, M.-M. From sensation to cognition. Brain: a journal of neurology 121,
1013 -- 1052 (1998).
[10] Desikan, R. S. et al. An automated labeling system for subdividing the human
cerebral cortex on mri scans into gyral based regions of interest. Neuroimage 31,
968 -- 980 (2006).
[11] Vuksanovi´c, V. & Hovel, P. Functional connectivity of distant cortical regions:
role of remote synchronization and symmetry in interactions. NeuroImage 97, 1 -- 8
(2014).
[12] Ciric, R. et al. Benchmarking of participant-level confound regression strategies
for the control of motion artifact in studies of functional connectivity. Neuroimage
154, 174 -- 187 (2017).
[13] Vuksanovi´c, V. & Hovel, P. Dynamic changes in network synchrony reveal resting-
state functional networks. Chaos: An Interdisciplinary Journal of Nonlinear Science
25, 023116 (2015).
8
[14] van Wijk, B. C., Stam, C. J. & Daffertshofer, A. Comparing brain networks of
different size and connectivity density using graph theory. PLoS One 5, e13701
(2010).
[15] Rubinov, M. & Sporns, O. Complex network measures of brain connectivity: uses
and interpretations. NeuroImage 52, 1059 -- 1069 (2010).
[16] Clauset, A., Newman, M. E. & Moore, C. Finding community structure in very
large networks. Physical review E 70, 066111 (2004).
[17] Chen, Z. J., He, Y., Rosa-Neto, P., Germann, J. & Evans, A. C. Revealing modular
architecture of human brain structural networks by using cortical thickness from
mri. Cerebral cortex 18, 2374 -- 2381 (2008).
[18] Bassett, D. S. et al. Hierarchical organization of human cortical networks in health
and schizophrenia. Journal of Neuroscience 28, 9239 -- 9248 (2008).
[19] Meunier, D., Achard, S., Morcom, A. & Bullmore, E. Age-related changes in mod-
ular organization of human brain functional networks. Neuroimage 44, 715 -- 723
(2009).
[20] Tzourio-Mazoyer, B. et al. Automated anatomical labeling of activations in SPM
using a macroscopic anatomical parcellation of the MNI MRI single-subject brain.
NeuroImage 15, 273 -- 289 (2002).
9
|
1110.0452 | 1 | 1110 | 2011-10-03T19:41:26 | Onset of negative interspike interval correlations in adapting neurons | [
"q-bio.NC"
] | Negative serial correlations in single spike trains are an effective method to reduce the variability of spike counts. One of the factors contributing to the development of negative correlations between successive interspike intervals is the presence of adaptation currents. In this work, based on a hidden Markov model and a proper statistical description of conditional responses, we obtain analytically these correlations in an adequate dynamical neuron model resembling adaptation. We derive the serial correlation coefficients for arbitrary lags, under a small adaptation scenario. In this case, the behavior of correlations is universal and depends on the first-order statistical description of an exponentially driven time-inhomogeneous stochastic process. | q-bio.NC | q-bio |
Onset of negative interspike interval correlations in adapting neurons
Eugenio Urdapilleta1, ∗
1Divisi´on de F´ısica Estad´ıstica e Interdisciplinaria & Instituto Balseiro,
Centro At´omico Bariloche, Avenida E. Bustillo Km 9.500,
S. C. de Bariloche (8400), R´ıo Negro, Argentina
Negative serial correlations in single spike trains are an effective method to reduce the variability
of spike counts. One of the factors contributing to the development of negative correlations between
successive interspike intervals is the presence of adaptation currents. In this work, based on a hidden
Markov model and a proper statistical description of conditional responses, we obtain analytically
these correlations in an adequate dynamical neuron model resembling adaptation. We derive the
serial correlation coefficients for arbitrary lags, under a small adaptation scenario.
In this case,
the behavior of correlations is universal and depends on the first-order statistical description of an
exponentially driven time-inhomogeneous stochastic process.
I.
INTRODUCTION
Spike-frequency adaptation (SFA) is one of the main
adaptation mechanisms in neural systems [1, 2]. As its
name implies, the effect defining SFA is the observation of
a scaling in the input-output relationship between the in-
jected current (or a stimulus property) and the firing rate
of a spiking neuron, from an initial to a stationary map-
ping [3 -- 6]. Several mechanisms can contribute to SFA
(for example, depressing synapses [7]); however, the most
prominent mechanism accounting for SFA is the presence
of (probably simultaneous) spike-related currents, which
produce a negative feedback to the neuron in time scales
ranging from tens to thousands of milliseconds [4, 8 -- 11].
Even when the full impact of these currents on neural
coding is not completely understood, it is known that
they contribute to the processing of static as well as tem-
poral signals. For the processing of temporal signals, it is
worth pointing out the high-pass filtering characteristics
due to SFA [4, 12, 13] and its related sensitivity to input
fluctuations [5], the forward masking effect [14, 15], the
selectivity to complex stimuli [16, 17], and the enhanced
reliability of temporal coding [6].
For the case of static signals, given the absence of a
temporal structure in the input, a neural system makes
use of a rate code to represent them. A rate code is
defined as the number of spikes in a certain tempo-
ral window and it is completely described by the spike
count statistics [18]. Spike-related adaptation currents
have a twofold impact on this code. First, they modify
the tuning curve between signals and responses or the
input-output relationship mentioned above, which helps
to match dynamic ranges [2 -- 6]. Second, they introduce
negative correlations between successive interspike inter-
vals (ISIs) in stationary neural spike trains [6, 13 -- 15, 19].
While the first effect reflects the modification of the mean
spike count, the second effect implies a strong change in
its variance (and higher-order properties). This change
∗[email protected]
arises from the fact that the presence of negative cor-
relations in a point process reduces the long-term vari-
ability in the counting process that defines the rate code
[20]. Taken together, both effects deeply affect the en-
coding reliability [19, 21 -- 24]. The presence of negative
correlations also affects the coding capabilities of other
related schemes; for example, coding of slowly varying
signals through a mechanism called noise shaping [25 --
27] (demonstrated for negative correlations arising from
a history-dependent threshold, but also valid for spike-
related adaptation currents), or adaptation-based inde-
pendent codes [28].
Correlated single spike trains constitute a nonrenewal
point process. In neural systems, this kind of process is
relatively ubiquitous [29 -- 33] (for reviews, mostly based
on negative correlations, see also [19, 23]).
In gen-
eral, there is a coexistence of processes that evoke op-
posite effects on the ISI correlations, and therefore on
the counting statistics in single neurons (adaptation cur-
rents and other regularizing processes such as synaptic
depression and negative feedback versus filtering and in-
put correlations, among others). Such interesting scenar-
ios have attracted relative attention within the theoret-
ical neuroscience community, and several studies focus
on these statistics or related properties in different situ-
ations [23, 24, 26, 28, 33 -- 37]. Related with our present
work, we should note three approaches that have been de-
rived recently: a population-based scheme for adapting
ensembles [38] (see also [23, 24, 28]), a directed discrete
representation for counting events with a general internal
structure [39] (see also [40, 41]), and a general framework
for nonrenewal processes as a hidden Markov model [42].
In particular, our work can be framed within the general
approach obtained by van Vreeswijk [42].
In this work, based on the results we have found in a
previous study about the first-passage-time (FPT) prob-
lem in an exponentially driven Wiener process [43], we
derive how the resulting negative correlations arise in the
spike train of a dynamical, although simple, process re-
sembling the basic features of a spike-related adaptation
current added to a spiking neuron in the presence of
fast additive fluctuations. The expressions we find are
strictly valid in a slight adaptation regime, where the
complete FPT density is not necessary and perturbation
techniques can be applied, so they remain valid for other
dynamical models (with the spike-related adaptation cur-
rent considered here, or similar). In this way, the onset of
correlations due to adaptation is general across different
models, provided the additive noise is fast. In the final
part of the work, we use the statistics of the FPT problem
and the emerging correlations due to adaptation to assess
how the spike count variance decreases in comparison to
an equivalent neuron without adaptation, in an asymp-
totic limit. This reduction in the spike count variance
underlies the improvement in the decoding performance,
and we show how the dependence on the correlations and
on the intrinsic variability reduction due to the inhomo-
geneous driving shape the spike count variance reduction
for different spiking frequencies.
II. THEORETICAL FRAMEWORK
A. Basic model
We consider an integrate-and-fire (IF) neuron, where
subthreshold dynamics of the membrane potential V is
governed by
Cm
dV
dt
= f (V ) + Iadapt + Iext.
(1)
External, Iext, as well as internal currents, f (V ) and
Iadapt, drive the membrane potential whenever V < Vthr.
The internal current f (V ) takes into account different
interspike phenomena, such as leakage or spike initiation
onset. The simplest models are the perfect [f (V ) = 0]
and the leaky [f (V ) = −gLV ] IF neurons, where only
leakage is considered (the perfect IF model corresponds
to no leakage). The subthreshold dynamics is supple-
mented by a threshold condition, which simplifies the
highly nonlinear process of a spike excursion: whenever
the potential reaches Vthr a spike is defined and immedi-
ately, the membrane potential is set to a reset condition
Vr.
Several types of adaptation currents have been char-
acterized by experiments [8 -- 10]. According to previous
theoretical studies [4, 13, 15], we model the adaptation
current as a process x(t) that decays during spikes and is
incremented when a spike event occurs. In particular, we
consider the adaptation current as Iadapt(t) = −ga x(t),
where the interspike dynamics for the adaptation process
is given by
dx
dt
= − x
τa
,
(2)
and a fixed increase α > 0 is evoked at all spike times
tsp: x(tsp) → x(tsp) + α. This model could be considered
as an idealization of the Ca2+-dependent K+ current,
2
which is widely expressed in neurons exhibiting SFA [x(t)
would represent the calcium concentration in a current-
based scheme]. Without mathematical loss, we can set
ga = Cm/τa and use α to control the strength of the
adaptation current.
Since Iadapt(t) = −(Cm/τa) x(t), from Eq. (2) it follows
that, between the (n)th and the (n + 1)th spikes, the
evolution of the adaptation current is given by
Iadapt(t)
Cm
= − εn
τa
exp[−(t − tn)/τa],
(3)
where εn represents the state of the adaptation process
x immediately after the spike time tn.
Without a random component, the deterministic dy-
namics of the membrane potential in the IF model with
an adaptation current, Eqs. (1) and (3), evolves along
a prescribed trajectory and no correlations emerge since
each ISI is a replica of itself (however, even in the deter-
ministic regime, perturbations propagate in the sequence
of ISIs, which induces correlations [37]). We introduce
a stochastic component in the system through external
noise. In particular, we assume that the external current
is given by a constant deterministic component and an
additive Gaussian white noise representing fast external
fluctuations
Iext
Cm
= µ + ξ(t),
(4)
where (cid:104)ξ(t)(cid:105) = 0 and (cid:104)ξ(t)ξ(t(cid:48))(cid:105) = 2Dδ(t(cid:48)−t). Therefore,
the subthreshold dynamics between the (n)th and the
(n + 1)th spikes is determined by
dV
dt
= g(V ) + µ − εn
τa
exp[−(t − tn)/τa] + ξ(t).
(5)
Different (IF) models will have different g(V ) func-
tions. In particular, g(V ) = 0 for the perfect IF model
and g(V ) = −V /τm, with τm = Cm/gL, for the leaky
IF model. The threshold condition (at spike time tn+1)
resets the membrane potential to Vr and updates the
adaptation state to εn+1 = εn exp(−τn/τa) + α, where
τn = tn+1 − tn is the (n)th ISI.
In Fig. 1(a) we show a typical realization of the mem-
brane potential, V (t), and the adaptation process, x(t),
for the system described above. As shown in Fig. 1(b),
this model exhibits SFA, where a step input (bottom)
induces an initial firing rate f0 which decays to a lower
steady-state firing rate fss (top), with some typical time
scale. Spike times tn defining the spike train, Tsp(t) =
quence of FPT processes, {. . . , τn-1, τn, τn+1, . . .} ≡ {τn}
[Fig. 1(c)]. The ongoing (initial) state of the adaptation
process, εn, is a history-dependent random variable turn-
ing the sequence {τn} into non-Markovian.
(cid:80) δ(t − tn) [see top of Fig. 1(a)], also establish the se-
However, the bidimensional state s = (ε, τ ), defined
at spike times, supports a Markovian process since the
3
FIG. 1: (Color online) Neuron model with adaptation current. (a) A system realization, V (t) and x(t), and the resulting spike
train Tsp(t). (b) The system develops spike-frequency adaptation (top), usually characterized by the temporal evolution of the
firing frequency, f (t), stimulated with step currents (bottom). (c) A sample trace of the adaptation process x(t) and the basic
elements used to describe the stochastic nature of the system, ε and τ . (d) The hidden Markov model, composed by ε and the
observable τ , used for the analysis.
state at (spike) time tn is completely characterized from
the knowledge of the state at (spike) time tn-1: given εn-1
and τn-1, εn is defined by the deterministic relationship
εn = εn-1 exp(−τn-1/τa) + α, and τn is given by the FPT
problem of a certain stochastic process with an exponen-
tial time-dependent drift (−εn/τa) exp[−(t − tn)/τa]. In
particular, for the perfect (leaky) IF model, the underly-
ing stochastic process is a Wiener (Ornstein-Uhlenbeck)
diffusion process. In mathematical terms, the transition
probability density is given by
f (snsn-1, sn-2, . . . )
= f (εn, τnεn-1, τn-1)
= δ{εn − [εn-1 exp(−τn-1/τa) + α]} φ(τnεn),
B. Statistics of the (initial) adaptation strength
For one-dimensional systems as the IF models, the so-
lution to the FPT problem is given as an expansion in
terms of the strength of the exponential drift [43, 44]
φ(τε) =
εi φi(τ ).
(7)
i=0
Given the transition density, Eq. (6), the equilibrium
probability density for ε, ρeq(ε), should satisfy [42]
∞(cid:88)
(6)
ρeq(εn) =
fε(εnεn-1) ρeq(εn-1) dεn-1,
(8)
where φ(τnεn) is the FPT density function associated to
the (n)th ISI, which depends exclusively on εn and not
on previous outcomes (tn in the above notation for the
drift just set the initial time). This Markovian process is
represented in Fig. 1(d) and constitutes a hidden Markov
model. Based on Eq. (6), the key elements necessary to
analyze this stochastic system are the statistics of ε and
the time-inhomogeneous FPT density function.
where the transition density between substates is
fε(εnεn−1) =
f (εn, τnεn-1, τn-1)φ(τn-1εn-1)dτndτn-1.
(9)
Replacing Eq. (6) and the ε-expansion for φ(τn-1εn-1),
Eq. (7), in Eq. (9) we obtain a self-consistent integral
equation, hard to tackle analytically. However, from this
4
Laplace transform of the jth term in the expansion for
the FPT solution. The previous relationship relates
unconditional moments (subindexes are irrelevant), and
therefore it represents a set of infinite algebraic equations
for the (infinite) moments (cid:104)εm(cid:105).
In the slight adaptation regime, α (cid:54)= 0 but small, it is
easy to see that the moments (cid:104)εm(cid:105) ∼ O(αm). In partic-
ular, the first two moments read
(cid:104)ε(cid:105) =
(cid:104)ε2(cid:105) =
,
α
[1 − φL
0 (1/τa)]
α2 [1 + φL
0 (1/τa)]
0 (1/τa)] [1 − φL
[1 − φL
0 (2/τa)]
(11)
.
(12)
The α-normalized properties, (1/α)(cid:104)ε(cid:105) and (1/α2)(cid:104)(ε−
(cid:104)ε(cid:105))2(cid:105), are shown in Fig. 2(a) in black and gray arrows,
respectively (right margin), for the case of the perfect IF
neuron model. As expected, they agree with the results
obtained from simulations in the asymptotic limit (sym-
bols plus error bars: mean plus standard deviation of εn;
colored histogram: logarithm of the density).
For small α values, there is an additional interesting
viewpoint to derive (cid:104)ε(cid:105), also valid for any neuron model.
Multitrial experiments usually start from rest, where we
assume that there is no adaptation, ε0 = 0. Conditional
to this fact, the (n)th event satisfies
εn = α [1 +
exp(−τj/τa)].
(13)
The first moment is given by
n−1(cid:88)
n−1(cid:89)
i=1
j=i
n−1(cid:88)
(cid:104)n−1(cid:89)
i=1
j=i
(cid:104)εn(cid:105) = α [1 +
exp(−τj/τa)(cid:105)].
(14)
For a weak adaptation process we consider that (cid:104)n(cid:105) ∼
O(α), and then the conditional probabilities required in
the above expression are well approximated by their zero-
order expansion, fτ (τn/τn-1, τn-2, . . . ) = φ0(τn) for all n.
In this limit it is easy to see that
(cid:104)εn(cid:105) = α [1 +
(cid:104)exp(−τj/τa)(cid:105)φ0 ]
n−1(cid:88)
n−1(cid:89)
i=1
j=i
n−1(cid:88)
= α
[φL
0 (1/τa)]i,
(15)
i=0
where (cid:104)·(cid:105)φi indicates the average with respect to the func-
tion φi(τ ). This geometric series is readily obtained:
(cid:104)εn(cid:105) = α
1 − [φL
[1 − φL
0 (1/τa)]n
0 (1/τa)]
,
(16)
FIG. 2: (Color online) Statistics of ε for the perfect IF model.
(a) Density of normalized ε as a function of the spike number
(color bar in logarithmic scale). Symbols and error bars indi-
cate mean and standard deviation, respectively. The analyti-
cal expression for the mean as a function of the spike number,
Eq. (16), is represented by the dotted (blue) line. The analyt-
ical expressions for the stationary properties (mean/standard
deviation), Eqs. (11) and (12), are indicated by arrows at
the right (black/gray). Numerical results were obtained for
µ = 0.10 ms−1, D = 0.05 ms−1, Vthr−Vr = 1.0, τa = 100.0 ms,
and α = 0.01 (number of trials, Ntrials = 105). (b) The nor-
malized stationary mean as a function of α (semilogarithmic
plot). The analytical result is represented by a continuous
(red) line. Nonlinear effects for large α are fitted by a (dashed-
dotted) line, indicating that linear regime remains valid up to
α ≈ 0.1. Inset: plot of the (not normalized) stationary mean
as a function of α. Same parameters as in (a), except that
Ntrials = 103 for each point.
integral equation it is relatively simple to find a relation-
ship between the moments, which reads
(cid:104)εm
n (cid:105) =
εm
n ρeq(εn) dεn
(cid:18)m
(cid:19)
m(cid:88)
∞(cid:88)
= αm +
where (cid:0)a
(cid:1) is the binomial coefficient and φL
j=0
i=1
i
b
αm−i
j (i/τa) (cid:104)εi+j
φL
n-1(cid:105),(10)
j (s) is the
and converges asymptotically to α/[1− φL
0 (1/τa)], when-
ever φL
0 (1/τa) < 1, in concordance with the previous
analysis. The exponential growth of the normalized
adaptation strength (1/α)(cid:104)εn(cid:105), predicted by Eq. (16), is
shown in Fig. 2(a) as a function of the spike number
n (blue dotted line), together with the results obtained
from numerical simulations (symbols). The agreement is
remarkable for this case (simulation results were obtained
with α = 0.01).
In order to determine the range of α where the approx-
imation remains valid, we run several simulations and
calculate the asymptotic (equilibrium) (cid:104)ε(cid:105) for different α
values. In Fig. 2(b) we show the normalized asymptotic
mean value as a function of α (inset: not normalized
mean value). The average exhibits the linear behavior
indicated by the preceding results up to α ≈ 0.1, which
represents an adaptation of about 10% [(f0 − fss)/f0, see
Fig. 1(c)].
III. RESULTS
A. Onset of correlations
Correlations in nonrenewal point processes are usually
characterized by the serial correlation coefficient (SCC),
which is defined by
(cid:113)(cid:104)∆τ 2
(cid:104)τnτn+k(cid:105) − (cid:104)τn(cid:105)(cid:104)τn+k(cid:105)
n(cid:105)(cid:104)∆τ 2
n+k(cid:105)
ρk =
,
(17)
where k indicates the lag between successive ISIs, brack-
i = (τi − (cid:104)τi(cid:105))2
ets indicate ensemble average, and ∆τ 2
is the variance. Once the system reaches the station-
ary conditions (the firing frequency is adapted), the SCC
simplifies to
5
where dsi represents dεidτi and Dsi its integration do-
main.
At lag k = 1, the transition probability density be-
tween states is given directly by Eq. (6). The onset of cor-
relations is characterized by the smallest order in α which
produce finite SCC values. This order coincides with
the small adaptation regime, and therefore, the linear
ε-expansion suffices for relevant expressions in Eq. (19),
(cid:104)τnτn+1(cid:105) ∼ O(α). Consequently, we obtain
(cid:104)τnτn+1(cid:105) = (cid:104)τ(cid:105)2
+ α (cid:104)τ(cid:105)φ1
(cid:20)(cid:104)ε(cid:105)
φ0
α
(cid:18)
(cid:104)τ(cid:105)φ0 − dφL
0 (s)
ds
(cid:99)1/τa
(cid:19)
(cid:21)
,(20)
+ (cid:104)τ(cid:105)φ0
where (cid:104)·(cid:105)φi indicates the average with respect to the func-
tion φi(τ ). In order to keep the linear order in the nor-
malization required for the SCC, we have (cid:104)τ(cid:105)2 ∼ O(α)
and (cid:104)∆τ 2(cid:105) ∼ O(0):
(cid:104)τ(cid:105)2 = (cid:104)τ(cid:105)2
(cid:104)∆τ 2(cid:105) = (cid:104)∆τ 2(cid:105)φ0.
φ0
+ 2(cid:104)ε(cid:105)(cid:104)τ(cid:105)φ0(cid:104)τ(cid:105)φ1 ,
(21)
(22)
In Eqs. (20) and (21), the evaluation of (cid:104)ε(cid:105) should be
performed in the linear regime. Taking into account the
results obtained in the previous section, (cid:104)ε(cid:105)/α = 1/[1 −
φL
0 (1/τa)], the first SCC reads
(cid:104)τnτn+k(cid:105) − (cid:104)τ(cid:105)2
(cid:104)∆τ 2(cid:105)
.
ρk =
ρ1 = −α
(18)
(cid:104)τ(cid:105)φ1
(cid:20)
[1 − φL
×
0 (1/τa)] (cid:104)∆τ 2(cid:105)φ0
0 (1/τa)(cid:104)τ(cid:105)φ0 +
φL
(cid:21)
.
(23)
dφL
0 (s)
ds
(cid:99)1/τa
For the hidden Markov model defined by Eq. (6),
(cid:104)τnτn+k(cid:105) reads
(cid:104)τnτn+k(cid:105) =
Dsn
Dsn+k
τnτn+k f (εn+k, τn+kεn, τn)
× φ(τnεn) ρeq(εn) dsn dsn+k, (19)
To evaluate the SCC at higher lags, k > 1, we need
to obtain the transition probability density between the
states n and n + k which, based on the hidden Markov
model, reads
f (εn+k, τn+kεn, τn) =
Ds(k−1)
ds(k−1) f (εn+k, τn+k, . . . , εn+1, τn+1εn, τn)
=
Ds(k−1)
ds(k−1)
f (εn+i, τn+iεn+i−1, τn+i−1),
(24)
k(cid:89)
i=1
6
FIG. 3:
(Color online) Normalized serial correlation coefficients (SCC), ρk = ρk/α, for the perfect IF model. (a) Theoretical
expression for the normalized first SCC as a function of driving parameters, µ and D (ms−1) [see Eq. (23)]. (b) Numerical results
(symbols) for the normalized first SCC as a function of µ (ms−1) and different D values (ms−1). Corresponding theoretical
expressions are indicated by continuous lines. (c) Factor φL
0 (1/τa), which defines the relationship between successive SCC
values, as a function of the driving parameters. (d) Theoretical prediction of normalized SCC at higher lags as a function of
driving parameters [see Eq. (26)]. Continuous, dashed, and dashed-dotted (colored) lines represent ρ1, ρ2, and ρ3, respectively,
at different noise intensities (black, red, blue, and orange: D = 100, 10−2, 10−4, and 10−6 ms−1, respectively). Lines are
embedded in their respective continuous (semitransparent) surfaces. (e) Numerical results for the normalized SCC at higher
lags. ρ1, ρ3, and ρ5 as a function of µ (ms−1, D = 0.05 ms−1). Continuous lines represent the corresponding theoretical
expressions. Dashed black line indicates φL
0 (1/τa) (right scale). (f) Numerical results for the normalized SCC as a function of
the lag for the four cases indicated with arrows in (e). Dashed lines indicate theoretical expressions. Ratios between successive
SCCs are given by the value of φL
0 (1/τa) at each point [dashed line in (e) at the different values of µ]. Numerical results were
obtained from NISI = 106 consecutive ISIs, error bars were estimated from Nrepet = 10 repetitions. Remaining parameters:
Vthr − Vr = 1.0, τa = 100.0 ms, and α = 0.1.
where we have simplified the notation to ds(k−1) =
dsn+1 . . . dsn+k−1 and Ds(k−1) to the corresponding in-
tegration domain. Each of the factors under the product
symbol has the structure given by Eq. (6). After some
calculus [it is convenient to leave unevaluated all inte-
grals in τi in the transition density from sn to sn+k, for
posterior evaluation in Eq. (19)], the average between
successive ISIs is given, up to first order in α (and/or
(cid:104)ε(cid:105)), by
7
(cid:104)τnτn+k(cid:105) = (cid:104)τ(cid:105)2
φ0
(cid:34)(cid:16)
×
1 +
+ α (cid:104)τ(cid:105)φ1
(cid:104)ε(cid:105)
α
(cid:17)(cid:104)τ(cid:105)φ0 + (cid:104)τ(cid:105)φ0
−(cid:104)ε(cid:105)
k−1(cid:88)
i=1
0 (1/τa)]k−i
(cid:35)
[φL
0 (1/τa)]k−1 dφL
0 (s)
[φL
ds
(cid:99)1/τa
α
.(25)
Replacing the stationary value for (cid:104)ε(cid:105), and after normal-
ization, it is relatively simple to prove that SCC values
at superior lags are given by
ρk = [φL
0 (1/τa)]k−1 ρ1 ,
k > 1.
(26)
For the perfect IF model, the lowest orders in the ex-
pansion corresponding to the solution of the FPT prob-
lem, Eq. (7), are given [43], and therefore, the expressions
for ρk can be explicitly evaluated.
In Fig. 3(a) we show the normalized first SCC, ρ1 =
ρ1/α, as a function of the parameters governing the dy-
namics of the perfect IF neuron model, µ and D.
In
Fig. 3(b) we show the theoretical expression for the nor-
malized SCC ρ1 as a function of the driving force µ, for
different noise intensities, and compare it with numeri-
cal results. The agreement between both curves is re-
markable, since the limit of small adaptation is satisfied
(numerical results were obtained with α = 0.1 and then
normalized).
As expressed by Eq. (26), the SCC at higher lags, ρk for
0 (1/τa) ρk−1.
k > 1, have a geometric structure: ρk = φL
Since 0 < φL
0 (1/τa) < 1 [see Fig. 3(c)], SCC at higher
lags are scaled versions of the first SCC. In Fig. 3(d) we
show the first three normalized SCC, ρk = ρk/α, as a
function of the parameters µ and D. As stated by the
geometric relationship, the exact scaling between them is
given according to the precise location in the parameters
space [equivalent point in Fig. 3(c)]. The scaling can be
better appreciated in Fig. 3(e), which simply represents
a section of Fig. 3(d) along a particular D.
In this
case, theoretical (solid lines) as well as numerical results
(symbols) for ρ1, ρ3, and ρ5 are represented as a function
of the driving force µ (ρ2 and ρ4 are omitted for the sake
of clarity). The dashed line shows the respective section
of Fig. 3(c) (scale in the right margin), which governs
the scaling between consecutive SCCs. Actually, the
geometric structure represents an exponential decay of
the SCC as a function of the lag. This can be observed
in Fig. 3(f) for different µ values selected in Fig. 3(e).
The exponential decay is given from the scaling factor
φL
0 (1/τa), obtained from the intersection of the dashed
line in Fig. 3(e) and the particular value of µ considered.
For example, µ2 and µ4 were selected so ρ1 were
approximately the same for both cases. However, the
FIG. 4:
(Color online) Normalized serial correlation coeffi-
cients (SCCs), ρk = ρk/α, for the leaky IF model [g(V ) =
−V /τm]. (a) and (b) analogous to Figs. 3(e) and 3(f). Pa-
rameters: τm = 10.0 ms, D = 0.01 ms−1, Vthr − Vr = 1.0, and
τa = 100.0 ms.
scaling factor φL
the decay is correspondingly slower.
0 (1/τa) is higher for µ4 than for µ2, so
The onset of correlations, as characterized by the
SCC, has a general structure, Eqs. (23) and (26), that
relies critically on two factors: the Laplace transform
of the unperturbed distribution, φL
0 (s), and the linear
correction to the mean introduced by the exponential
temporal drift, (cid:104)τ(cid:105)φ1 . As shown in [43],
for small
intensities the main effect of the exponential drift on the
FPT statistics of a perfect IF model is to change the
mean consistently to what we have found in this work.
For other IF models, the complete FPT problem with
an exponential temporal drift can be addressed with a
similar procedure [44]. However, even when appealing
to set out the formalism, the complete statistics is not
required for computing the SCC and so we can proceed
with simpler approaches. For example, to compute
the linear correction to the mean FPT due to the
exponential temporal drift in generic IF models, we can
use the results obtained by Lindner using a perturbation
scheme [45, 46]. To illustrate the generality of the
results we have found, in Fig. 4 we show the normalized
SCC, ρk, predicted by Eqs. (23) and (26), using the
0 (s) and (cid:104)τ(cid:105)φ1 , obtained by Lindner
analytical results, φL
8
in Fig. 5(a) as a function of µ or D], but only a rough
estimate of the correct values of the correlations (cyan
symbols and line in Fig. 5).
C. Spike-count variance reduction
The presence of negative correlations affects the spike-
count variance. To analyze this effect, it is convenient to
introduce the Fano factor, FFT , which relates the mean
and the variance of the spike counts observed in a tem-
T(cid:105), respectively,
poral window of length T , (cid:104)nT(cid:105) and (cid:104)∆n2
as the ratio
FFT =
(cid:104)∆n2
T(cid:105)
(cid:104)nT(cid:105) .
(27)
For point processes, the asymptotic behavior of the
Fano factor reads [20]
(cid:32)
FF∞ = limT→∞ FFT = CV2
1 + 2
(cid:33)
ρk
,
(28)
∞(cid:88)
k=1
where CV is the coefficient of variation, defined as the
ratio between the standard deviation and the mean of
the unconditional ISI statistics,
CV =
.
(29)
Combining the preceding equations and given that
in the
(cid:104)nT(cid:105) = T /(cid:104)τ(cid:105), the spike-count variance reads,
asymptotic limit,
(cid:112)(cid:104)∆τ 2(cid:105)
(cid:104)τ(cid:105)
(cid:104)∆n2
T(cid:105) =
(cid:104)∆τ 2(cid:105)
(cid:104)τ(cid:105)3
(cid:32)
(cid:33)
∞(cid:88)
k=1
1 + 2
ρk
T.
(30)
The linear growth in T of the spike-count variance is
a characteristic of a diffusive process (and the reason
for the usefulness of the Fano factor). Inasmuch as the
asymptotic limit is reached, it is useful to analyze the
preceding factor, which we denote (cid:104) ∆n2
T(cid:105),
∞(cid:88)
(cid:32)
(cid:33)
1 + 2
ρk
.
(31)
(cid:104) ∆n2
T(cid:105) =
(cid:104)∆τ 2(cid:105)
(cid:104)τ(cid:105)3
k=1
Equation (31) highlights two contributions to the
spike-count variance: a contribution from the FPT statis-
tics related to a single spiking process, (cid:104)∆τ 2(cid:105)/(cid:104)τ(cid:105)3, and
a contribution from the ISI correlations provided by the
entire spike train, (1 + 2(cid:80)∞
k=1 ρk).
The effect of the adaptation strength α on the spike-
count variance is twofold; it changes the ISI statistics as
well as the correlations between the ISIs. In the slight
FIG. 5:
(Color online) Loss of linearity. (a) First SCC, ρ1,
as a function of the driving parameter µ in a perfect IF neu-
ron model, for different adaptation strengths, α. In all cases,
D = 0.02 ms−1, Vthr − Vr = 1.0, and τa = 100.0 ms. (b) As
the value of α increases, the stationary firing rate decreases
[asymptotic value of f (t) at the right margin] and correspond-
ingly, the adaptation effect is more prominent. In all cases,
the driving drift is µ = 0.10 ms−1 (which corresponds to a
firing rate of 100 Hz in the absence of adaptation, α = 0),
applied as a step function at time t = 0 [see Fig. 1(b)]. Re-
maining parameters as in (a).
for the leaky IF neuron model [45], and compare them
to numerical simulations. As expected, theoretical and
numerical results agree.
B. Loss of linearity
In the previous section we have shown that the on-
set of correlations has a specific structure and scales lin-
early with the adaptation strength, Eqs. (23) and (26).
This scaling enables us to consider normalized SCC, ρk,
as shown in the preceding figures. However, for a large
enough value of α, higher-order effects become important
and these equations are no longer applicable. In partic-
ular, given that the expression for ρ1, Eq. (23), does not
saturate we expect that higher-order effects oppose the
linear growth.
In Fig. 5(a) we show the (not normal-
ized) first SCC, ρ1, for different adaptation strengths.
As expected, for small values of α the theoretical ex-
pression, Eq. (23), properly accounts for the linear scal-
ing. Specifically, correlations grow linearly up to approxi-
mately α ≈ 0.25 [red symbols and line in Fig. 5(a)], which
represents a frequency adaptation of about 20% [see red
line in Fig. 5(b)]. This limit is slightly better than that
expected from the upper limit of the linear scaling in the
stationary mean adaptation strength, (cid:104)ε(cid:105) [see Fig. 2(b)].
For larger values, higher effects are non-negligible and
affect the correlations in the predicted manner. For a
realistic SFA (adaptation of about 50%), the analyti-
cal expressions provide a qualitative agreement regarding
the dependence on the parameters [shape of the curve
9
φ0
(cid:104)∆τ 2(cid:105)φ0
(cid:104)τ(cid:105)3
(cid:104)∆τ 2(cid:105)
(cid:104)τ(cid:105)3 =
+(cid:104)ε(cid:105) (cid:104)τ 2(cid:105)φ1 − 2(cid:104)τ(cid:105)φ0(cid:104)τ(cid:105)φ1 − 3(cid:104)∆τ 2(cid:105)φ0(cid:104)τ(cid:105)φ1 /(cid:104)τ(cid:105)φ0
where (cid:104)ε(cid:105) is given by Eq. (11). Replacing the com-
plete expressions for both factors [Eqs. (32) and (33)] in
Eq. (31), and keeping the first order in α [(cid:104) ∆n2
T(cid:105) ≈ O(α)],
we obtain
(cid:104)τ(cid:105)3
,(33)
φ0
(cid:104) ∆n2
+
[1 − φL
(cid:104)τ 2(cid:105)φ0(cid:104)τ(cid:105)φ1
φ0
(cid:104)∆τ 2(cid:105)φ0
(cid:34)
T(cid:105) =
(cid:104)τ(cid:105)3
(cid:104)τ 2(cid:105)φ1 − 3
(cid:18)
(cid:104)τ(cid:105)φ0
(cid:104)τ(cid:105)φ0 − 3(cid:104)τ(cid:105)φ0φL
×
×
α
φ0
0 (1/τa)](cid:104)τ(cid:105)3
(cid:104)τ(cid:105)φ1
0 (1/τa)
1 − φL
+
0 (1/τa) − 2
dφL
0 (s)
ds
(cid:99)1/τa
(cid:19)(cid:35)
.(34)
It is clear that the independent term on the right-hand
side of Eq. (34) corresponds to the asymptotic spike-
count variance (normalized by T ) of the neuron without
adaptation (α = 0). Even when general in the regime
we consider, the behavior of this rather complicated ex-
pression for α (cid:54)= 0 is not obvious at all. The expression
simplifies enormously for the case of the perfect IF neu-
ron model. For this neuron, we have
(cid:104) ∆n2
T(cid:105) =
2D
(Vthr − Vr)2 −
4D
(Vthr − Vr)3 α.
(35)
Surprisingly, this expression for the spike-count vari-
ance reduction does not depend on the driving param-
eter µ, which sets the firing frequency.
In Fig. 6 we
show this behavior for different values of the adaptation
strength. The case α = 0 corresponds to a perfect IF neu-
ron model without adaptation (black line and squares).
As α increases, the spike-count variance decreases, and
this effect does not depend on the firing frequency set
by varying µ. For α = 0.05, theoretical and numerical
results agree (red thick lines and empty circles, respec-
tively); whereas for α = 0.10 the flat reduction is numer-
ically observed, but the appropriate variance decrease is
just approximated by the theoretical prediction (green
thick lines and empty triangles). The same holds true for
higher adaptation strengths (not shown). As expected,
since the expression for ρ1 [Eq. (23)] and the exponential
structure for higher lags [Eq. (26)] are approximate, the
infinite sum in Eq. (32) amplifies a tiny error. In conse-
quence, even when we have observed that ρ1 is properly
described by Eq. (23) for the case α = 0.10 [Fig. 5(a)],
the expression for the spike-count variance [Eq. (35)] is
qualitatively good but approximate [note, however, that
there is a significant reduction even for small values of α,
and Eq. (35) has a moderate relative error for α = 0.1].
FIG. 6:
(Color online) Spike-count variability in the perfect
IF neuron model, as a function of the firing frequency. The
neuron without adaptation (α = 0) shows a spike-count vari-
ance that does not depend on the firing frequency [theoretical
(T ) and numerical (N ) results are represented by a black thick
line and open squares, respectively]. For α (cid:54)= 0 (α = 0.05 and
α = 0.10 in red and green, respectively) there is a reduction
in the variance, which also is independent of the firing rate
[theoretical (T ) and numerical (N ) results are shown by thick
lines and empty symbols, respectively]. By shuffling the orig-
inal spike train, from which the spike counts are drawn, we
destroy the correlations. For the shuffled spike train, theoret-
ical (T sh.) and numerical (N sh.) results are represented by
hidden lines and semifilled symbols, respectively. Parameters:
Vthr − Vr = 1.0, τa = 100.0 ms, D = 0.02 ms−1. The driving
parameter µ was varied from 0.005 to 0.2 ms−1. Numerical
results are presented as a function of the firing frequency ob-
served from the mean spike counts. Theoretical results are
shown as a function of the firing frequency computed from
f−1 = (cid:104)τ(cid:105)φ0 + (cid:104)ε(cid:105)(cid:104)τ(cid:105)φ1 . Numerical procedure: to compute
the spike-count statistics, the length of the temporal window
T was varied for different values of µ, checking in all cases
that the asymptotic regime was reached. Numerical results
were obtained from a single spike train with a temporal length
Ttotal = T Nsample, with Nsample = 100. The value of Nsample
is low, due to the computational cost of the shuffling pro-
cedure. To improve the accuracy, results are presented as
averages over Nrepet = 500 repetitions.
adaptation regime we consider, the contribution to the
spike-count variance due to the correlations is a linear
term in α. Explicitly, it is easy to show that
∞(cid:88)
k=1
1 + 2
ρk = 1 + 2
ρ1(α)
1 − φL
0 (1/τa)
,
(32)
where ρ1(α) is given by Eq. (23). To be consistent with
this description, the term provided by the ISI statistics
should be linearized, and reads
The linear spike-count variance reduction expressed by
Eq. (35) is unbounded, which is obviously unreasonable
(it enables negative values for the spike-count variance).
Higher-order terms should oppose this linear reduction,
as can be observed in Fig. 6 for the case α = 0.10.
0 [1 + 2(cid:80)∞
Since the spike-count variance reduction for a given
firing frequency arises from the interplay between the
FPT statistics and the presence of correlations, it would
be interesting to disentangle to what extent each effect
contributes to the observed reduction. To analyze this
question we shuffled the spike train, which maintains the
first-order statistics (FPT statistics) destroying ISI corre-
lations. The spike-count variance reduction for the shuf-
fled spike train is shown in Fig. 6 as semifilled symbols.
The theoretical analog corresponds to (cid:104) ∆n2
T(cid:105) given ex-
clusively by Eq. (33), since the sum of the SCC values is
k=1 ρk = 1 instead of Eq. (32)], and it is shown
as dashed lines in Fig. 6. As expected, theoretical and
numerical results are in good accordance. The behav-
ior observed for each contribution is reasonable. At low
firing frequencies, within each ISI the exponential evo-
lution of the adaptation process has decayed, meaning
that each εn = α and correlations disappear. Corre-
spondingly, the spike-count variance reduction is given
exclusively by the FPT statistics (this case corresponds
to τd → 0 in [43, 45]), and we denote this regime as an
intrinsic reduction in Fig. 6. At large firing frequencies,
the adaptation process within a single ISI is essentially
constant (i.e., τa → ∞). In this case, the collection of
ISIs satisfies a quasistatic approximation [47], and the
spike-count variance of the shuffled spike train is indis-
tinguishable from the case of no adaptation. However, in
0 (1/τa) → 1;
this limit, correlations decay very slowly [φL
see Fig. 3(c)], and even when each SCC is small because
ρ1 is, the sum of correlations accumulates over many lags,
building up a finite nonvanishing value. In Fig. 6, we de-
note this range as a reduction due to correlations ("corrs"
in the figure).
Equation (35) shows that the spike-count variance re-
duction is independent of the firing frequency, for the
perfect IF neuron model. As shown in the previous para-
graph, in the limit of low as well as large firing frequen-
cies, the mechanisms that account for the reduction are
different. The fact that both effects influence the spike-
count variance to the same extent, and furthermore, that
these mechanisms exactly compensate each other at in-
termediate firing frequencies is intriguing. Obviously,
these results hold for this particular IF model. It would
be interesting to analyze the contributions in other IF
models; we expect analogous spike-count variance reduc-
tions and the same limit behaviors, but not an indepen-
dence on firing frequencies (in general, in other models
the spike-count variance for α = 0 depends on the firing
frequency).
10
IV. DISCUSSION AND CONCLUDING
REMARKS
In this work we have analyzed the onset of correlations
for a general neuron model, where an external input as
well as an internal spike-based adaptation current drive
the membrane potential. The external current is com-
posed by a static input and fast fluctuations. For this
system, the dependence of the adaptation current on the
past history, through the initial states of the adaptation
process, facilitates the development of correlations be-
tween successive ISIs [6, 13 -- 15, 19].
In the regime of
slight adaptation, we have shown that correlations share
a general structure across different models. By means
of a hidden Markov model, we have explicitly derived
the dependence of the SCC on different properties of
the FPT statistics corresponding to the underlying time-
inhomogeneous stochastic process, Eqs. (23) and (26).
In this regime, for one-dimensional models such as IF
neuron models, the necessary properties are given [43 --
46]. For any other (high-dimensional) model, whenever
a slight exponential time-dependent current smoothly re-
shapes the FPT statistics in comparison to the unper-
turbed case (analogously to Fig. 1 in [43]), the expres-
sions derived here apply. The geometric structure and
exponential decay for the SCC is surprisingly simple and
general for the scenario considered here. This kind of
structure was first observed by Lindner and Schwalger
for successive escapes of an overdamped Brownian parti-
cle in a randomly modulated asymmetric double well,
which can be modeled as transitions between discrete
states [40, 41]. Posteriorly, these authors extended their
results to a situation with multiple internal states [39]
and, in particular, studied the case of negative corre-
lations in neurons with inhibitory feedback (resembling
current-based adaptation), with an appropriate scheme
of transitions. Even when general and very promising, a
quantitative evaluation of correlations in adapting neu-
rons under this framework requires a procedure for the es-
timation of transition rates, in a discrete version of adap-
tation, obtained from dynamical models and/or experi-
mental data. The results obtained here are in qualitative
agreement to those obtained by Schwalger and Lindner
[39], which implies that this estimation procedure could
be an interesting topic to study.
The development of negative correlations in successive
ISIs in a spike train influences the encoding capabilities
of a neural system [19, 21 -- 24]. Here, we have analyzed
the decrease in the asymptotic spike-count variance due
to the intrinsic variability reduction of the FPT statis-
tics and the presence of correlations [Eqs. (31)-(34)] in
comparison to the case without adaptation. In particu-
lar, for the perfect IF neuron model the decrease is ex-
tremely simple and does not depend on the firing fre-
quency, Eq. (35). This analysis is theoretically impor-
tant, but it should be put in the proper context. The
spike-count variance reduction given by Eq. (31) is valid
for the asymptotic limit, which can be unfeasible in real
neurons, especially at low firing rates. For systems oper-
ating with finite temporal windows, the framework pre-
sented here should be extended by using the complete for-
malism derived by van Vreeswijk in [42], and finely used
by Farkhooi et al. to analyze a population scheme [24].
In this case, the spike-count variance will be a function
of the length of the temporal window used to compute
the statistics, and obviously, the results presented here
should agree in the limit of large windows. This behav-
ior was outlined by Chacron et al. in a different version
of adaptation (see results of the model without slow noise
in Fig. 4 of [33]). That work also highlighted a possible
explanation to the interesting finding made by Ratnam
and Nelson [32], where the spike-count variance exhibits a
minimum as a function of the length of the counting win-
dow (a possible behaviorally important phenomenon). In
their work, Chacron et al. demonstrated that a slow ex-
ternal noise leaves relatively intact the short-range cor-
relations, while destroying negative correlations at large
lags, giving rise to small and slowly decaying positive
correlations which dominate the asymptotic regime. On
11
the other hand, across different neural systems it has
been observed that the only significant SCC corresponds
to the lag 1 [19, 23], which reinforces the idea that a
slow external noise would be an important part of the
incoming signal, in addition to fast fluctuations. Theo-
retically, the analysis of the spike-count variance for IF
neuron models driven by slow fluctuations were success-
fully carried out via a quasistatic approximation [34, 36].
This suggests that the hidden Markov model used here
to model negative correlations could be extended with
a similar quasistatic approximation in order to include
slow fluctuations in the analysis.
V. ACKNOWLEDGMENTS
The author thanks In´es Samengo for a critical read-
ing of the manuscript. This work was supported by the
Consejo de Investigaciones Cient´ıficas y T´ecnicas de la
Rep´ublica Argentina.
[1] E. R. Kandel, J. H. Schwartz, and T. M. Jessell, Princi-
ples of Neural Science (McGraw-Hill, New York, 2000).
[2] B. Wark, B. N. Lundstrom, and A. Fairhall, Curr. Opin.
Neurobiol. 17, 423 (2007).
[3] B. Ermentrout, Neural Comput. 10, 1721 (1998).
[4] J. Benda and A. V. M. Herz, Neural Comput. 15, 2523
(2003).
[21] M. P. Nawrot, C. Boucsein, V. Rodriguez-Molina, A.
Aertsen, S. Grun, and S. Rotter, Neurocomputing 70,
1717 (2007).
[22] M. P. Nawrot, in Analysis of Parallel Spike Trains, edited
by S. Grun and S. Rotter (Springer Series in Computa-
tional Neuroscience, Springer-Verlag, Berlin, 2010).
[23] F. Farkhooi, M. F. Strube-Bloss, and M. P. Nawrot,
[5] M. H. Higgs, S. J. Slee, and W. J. Spain, J. Neurosci.
Phys. Rev. E 79, 021905 (2009).
26(34), 8787 (2006).
[24] F. Farkhooi, E. Muller, and M. P. Nawrot, Phys. Rev. E
[6] S. A. Prescott and T. J. Sejnowski, J. Neurosci. 28(50),
83, 050905 (2011).
13649 (2008).
[25] M. J. Chacron, B. Lindner, and A. Longtin, Phys. Rev.
[7] S. Chung, X. Li, and S. B. Nelson, Neuron 34, 437 (2002).
[8] D. V. Madison and R. A. Nicoll, J. Physiol. 354, 319-331
Lett. 92(8), 080601 (2004).
[26] B. Lindner, M. J. Chacron, and A. Longtin, Phys. Rev.
(1984).
E 72, 021911 (2005).
[9] F. Helmchen, K. Imoto, and B. Sakmann, Biophys. J. 70,
[27] O. ´Avila Akerberg and M. J. Chacron, Phys. Rev. E 79,
1069-1081 (1996).
[10] P. Sah, Trends Neurosci. 19(4), 150-154 (1996).
[11] B. Hille, Ion Channels of Excitable Membranes, 2nd ed.
(Sinauer Associates, Sunderland, MA, 1992).
[12] J. Benda, A. Longtin, and L. Maler, J. Neurosci. 25(9),
2312 (2005).
011914 (2009).
[28] W. H. Nesse, L. Maler, and A. Longtin, Proc. Natl. Acad.
Sci. USA 107(51), 21973 (2010).
[29] S. B. Lowen and M. C. Teich, Fractal-Based Point Pro-
cesses (John Wiley & Sons, New Jersey, 2005).
[30] S. B. Lowen and M. C. Teich, J. Acoust. Soc. Am. 92(2),
[13] J. Benda, L. Maler, and A.Longtin, J. Neurophysiol. 104,
Pt. 1, 803 (1992).
2806 (2010).
[14] X. -J. Wang, J. Neurophysiol. 79, 1549 (1998).
[15] Y. -H. Liu and X. -J. Wang, J. Comput. Neurosci. 10, 25
[31] A. Longtin and D. M. Racicot, Biosystems 40, 111
(1997).
[32] R. Ratnam and M. E. Nelson, J. Neurosci. 20(17), 6672
(2001).
(2000).
[16] S. Peron and F. Gabbiani, Nature Neurosci. 12, 318
[33] M. J. Chacron, A. Longtin, and L. Maler, J. Neurosci.
(2009).
21(14), 5328 (2001).
[17] S. P. Peron and F. Gabbiani, Biol. Cybern. 100, 505
[34] J. W. Middleton, M. J. Chacron, B. Lindner, and A.
(2009).
[18] P. Dayan and L. F. Abbott, Theoretical Neuroscience:
Computational and Mathematical Modeling of Neural
Systems, (The MIT Press, Cambridge, MA, 2001).
[19] O. Avila-Akerberg and M. J. Chacron, Exp. Brain Res.
210, 353 (2011).
Longtin, Phys. Rev. E 68, 021920 (2003).
[35] B. Lindner, Phys. Rev. E 69, 022901 (2004).
[36] T. Schwalger and L. Schimansky-Geier, Phys. Rev. E 77,
031914 (2008).
[37] T. Schwalger, K. Fisch, J. Benda, and B. Lindner, PLoS
Comput. Biol. 6(12), e1001026 (2010).
[20] D. R. Cox and P. A. W. Lewis, The statistical analysis
[38] E. Muller, L. Buesing, J. Schemmel, and K. Meier, Neural
of series of events (Methuen, London, 1966).
12
Comput. 19(11), 2958 (2007).
[39] T. Schwalger and B. Lindner, Eur. Phys. J. Spec. Topics
187, 211 (2010).
[40] B. Lindner and T. Schwalger, Phys. Rev. Lett. 98, 210603
(2007).
[41] T. Schwalger and B. Lindner, Phys. Rev. E 78, 021121
(2008).
[42] C. van Vreeswijk, in Analysis of Parallel Spike Trains,
edited by S. Grun and S. Rotter (Springer Series in Com-
putational Neuroscience, Springer-Verlag, Berlin, 2010).
[43] E. Urdapilleta, Phys. Rev. E 83, 021102 (2011).
[44] In [43] we have found the FPT density function for the
perfect IF model (Wiener process) with the exponential
time-inhomogeneous drift supporting the subthreshold
dynamics of the adaptation current. As pointed out in
the conclusions of [43], the system of recurrence equa-
tions can be extended to other diffusion processes [here
characterized by g(V ) (cid:54)= 0]. In this case, the homoge-
neous part of the equations to solve, Eqs. (8) and (9) in
[43], changes respectively. Given that the homogeneous
part is the same for all order terms, the formal solution
of each term is obtained from the Green's function asso-
ciated to the unperturbed system and the corresponding
source.
[45] B. Lindner, J. Stat. Phys. 117(3/4), 703 (2004).
[46] B. Lindner and A. Longtin, J. Theor. Biol. 232, 505
(2005).
[47] E. Urdapilleta and I. Samengo, Phys. Rev. E 80, 011915
(2009).
|
1109.2239 | 1 | 1109 | 2011-09-10T17:36:38 | A sparse coding model with synaptically local plasticity and spiking neurons can account for the diverse shapes of V1 simple cell receptive fields | [
"q-bio.NC",
"cond-mat.dis-nn"
] | Sparse coding algorithms trained on natural images can accurately predict the features that excite visual cortical neurons, but it is not known whether such codes can be learned using biologically realistic plasticity rules. We have developed a biophysically motivated spiking network, relying solely on synaptically local information, that can predict the full diversity of V1 simple cell receptive field shapes when trained on natural images. This represents the first demonstration that sparse coding principles, operating within the constraints imposed by cortical architecture, can successfully reproduce these receptive fields. We further prove, mathematically, that sparseness and decorrelation are the key ingredients that allow for synaptically local plasticity rules to optimize a cooperative, linear generative image model formed by the neural representation. Finally, we discuss several interesting emergent properties of our network, with the intent of bridging the gap between theoretical and experimental studies of visual cortex. | q-bio.NC | q-bio | A sparse coding model with synaptically local plasticity and spiking neurons can
account for the diverse shapes of V1 simple cell receptive fields
Joel Zylberberg,1, 2, a) Jason Timothy Murphy,3 and Michael Robert DeWeese1, 2, 3
1)Department of Physics, University of California, Berkeley CA 94720
2)Redwood Center for Theoretical Neuroscience, University of California,
Berkeley CA 94720
3)Helen Wills Neuroscience Institute, University of California,
Berkeley CA 94720
(Dated: 2 September 2018)
Sparse coding algorithms trained on natural images can accurately predict the fea-
tures that excite visual cortical neurons, but it is not known whether such codes can
be learned using biologically realistic plasticity rules. We have developed a biophysi-
cally motivated spiking network, relying solely on synaptically local information, that
can predict the full diversity of V1 simple cell receptive field shapes when trained on
natural images. This represents the first demonstration that sparse coding principles,
operating within the constraints imposed by cortical architecture, can successfully re-
produce these receptive fields. We further prove, mathematically, that sparseness and
decorrelation are the key ingredients that allow for synaptically local plasticity rules
to optimize a cooperative, linear generative image model formed by the neural repre-
sentation. Finally, we discuss several interesting emergent properties of our network,
with the intent of bridging the gap between theoretical and experimental studies of
visual cortex.
1
1
0
2
p
e
S
0
1
]
.
C
N
o
i
b
-
q
[
1
v
9
3
2
2
.
9
0
1
1
:
v
i
X
r
a
a)[email protected]
1
I.
INTRODUCTION
A central goal in systems neuroscience is to determine what underlying principles might
shape sensory processing in the nervous system. Several coding optimization principles have
been proposed, including redundancy reduction1 -- 3, and information maximization4 -- 8,10,11,
which have both enjoyed some successes in predicting the behavior of real neurons3,12 -- 14.
Closely related to these notions of coding efficiency is the principle of sparseness15 -- 17, which
posits that few neurons are active at any given time (population sparseness), or that indi-
vidual neurons are responsive to few specific stimuli (lifetime sparseness).
Sparseness is an appealing concept, in part because it provides a simple code for later
stages of processing and it is in principle more quickly and easily modifiable by simple
learning rules compared with more distributed codes involving many simultaneously active
units17,18. There is some experimental evidence for sparse coding in the cortex18 -- 23, but there
are also reports of dense neural activity24 and mixtures of both25 as well. Compounding this,
it is not obvious what absolute standard should be used to assess the degree of sparseness in
cortex, but it is notable that the relative level of sparseness of cortical responses to natural
images increases when a larger fraction of the visual field is covered by the stimulus20 -- 22,
as a result of inhibitory interneuronal connections22. Interestingly, the correlations between
the neuronal activities also decreases when a larger area is stimulated, as a result of these
inhibitory connections22.
In a landmark paper, Olshausen and Field15 reproduced several qualitative features of
the receptive fields (RFs) of neurons in primary visual cortex (VI) without imposing any
biological constraints other than their hypothesis that cortical representations simultane-
ously minimize the average activity of the neural population while maximizing fidelity when
representing natural images. However, agreement with measured V1 simple cell receptive
fields was not perfect26. Recently, a more sophisticated version of Olshausen and Field's
algorithm27 has been developed that is capable of minimizing the number of active neurons
rather than minimizing the average activity level across the neural population. This algo-
rithm, called the sparse-set coding (SSC) network27, learns the full set of physiologically
observed RF shapes of simple cells in V1, which include small unoriented features, localized
oriented features resembling Gabor wavelets, and elongated edge-detectors. We note that,
under certain conditions28 not necessarily satisfied by the natural image coding problem,
2
minimizing the average activity across the neural population (L1-norm minimization), as is
done by Olshausen and Field's original Sparsenet algorithm, can be equivalent to minimizing
the number of active units (L0-norm minimization), as is achieved by Rehn and Sommer's
SSC algorithm.
The SSC model is the only sparse coding algorithm that has been shown to learn, from
the statistics of natural scenes alone, RFs that are in quantitative agreement with those
observed in V1.
It has also been found that sufficiently overcomplete representations (4
times more model neurons than image pixels) that minimize the L1 norm can display the
same qualitative variety of RF shapes, but these have not been quantitatively compared
with physiologically measured RFs29.
Unfortunately, the lack of work on biophysically realistic sparse coding models has left in
doubt whether V1 could actually employ a sparse code for natural scenes. Indeed, it is not
clear how Olshausen's original algorithm15, the highly overcomplete L1-norm minimization
algorithm29, or that of Rehn and Sommer27, could be implemented in the cortex. Rather than
employing local network modification rules such as the synaptic plasticity that is thought
to underly learning in cortex30, all three of these networks rely on learning rules requiring
that each synapse has access to information about the receptive fields of many other, often
distant, neurons in the network.
Furthermore, both the SSC sparse coding model that has successfully reproduced the
full diversity of V1 simple cell RFs27 as well as the L1-norm minimization algorithm that
achieved qualitatively similar RFs29 involve non-spiking computational units: continuous-
valued information is shared between units while inference is being performed. In cortex,
however, information is transferred in discrete, stereotyped pulses of electrical activity called
action potentials or spikes. Particularly for a sparse coding model with few or no spikes
elicited per stimulus presentation, approximating spike trains with a graded function may
not be justified. Spiking image processing networks have been studied31 -- 36, but none of them
have been shown to learn the full diversity of V1 RF shapes using local plasticity rules.
It remains to be demonstrated that sparse coding can be achieved within the limitations
imposed by biological architecture, and thus that it could potentially be an underlying
principle of neural comptutation.
Here we present a biologically-inspired variation on a network originally due to Foldi`ak17,37
that performs sparse coding with spiking neurons. Our model performs learning using only
3
synaptically local rules. Using the fact that constraints imposed by such mechanisms as
homeostasis and lateral inhibition cause the units in the network to remain sparse and in-
dependent throughout training, we prove mathematically that it learns to approximate the
optimal linear generative model of the input, subject to constraints on the average lifetime
firing rates of the units and the temporal correlations between the units' firing rates. This
is the first demonstration that synaptically local plasticity rules are sufficient to account for
the observed diversity of V1 simple cell RF shapes, and the first rigorous derivation of a
relationship between synaptically local network modification rules and the twin properties
of sparseness and decorrelation.
Finally, we describe several emergent properties of our image coding network, in order to
elucidate some experimentally testable hallmarks of our model. Interestingly, we observe a
lognormal distribution of inhibitory connection strengths between the units in our model,
when it is trained on natural images; such a distribution has previously been observed in the
excitatory connections between neurons in rat V138, but the inhibitory connection strength
distribution remains unknown.
II. RESULTS
Our Sparse And Independent Local network (SAILnet) learns receptive fields
that closely resemble those of V1 simple cells
Our primary goal is to develop a biophysically inspired network of spiking neurons that
learns to sparsely encode natural images, while employing only synaptically local learning
rules. Towards this end, we implement a network of spiking, leaky integrate-and-fire units30
as model neurons. As in many previous models17,31,37,39,40, each unit has a time dependent
internal variable ui(t) and an output yi(t) associated with it. The simulation of our network
operates in discrete time. The neuronal output at time t, yi(t), is binary-valued: it is either
1 (spike) or 0 (no spike), whereas the internal variable ui(t) is a continuous-valued function
of time that is analogous to the membrane potential of a neuron. When this internal variable
exceeds a threshold θi, the unit fires a punctate spike of output activity that lasts for one
time step. This thresholding feature plays the role of neuronal voltage-gated ion channels
(represented, as in Hopfield's40 circuit model, by a diode) whose opening allows cortical
4
neurons to fire. Other units in the network, and the inputs Xk, which are pixel intensities
in an image, modify the internal variable ui(t) by injecting current into the model neuron.
The structure of our network, and circuit diagram of our neuron model, are illustrated in
Fig. 1. The dynamics of SAILnet neurons are discussed in detail in the Methods section.
We assess the computational output of each neuron in response to a stimulus image X
by counting the number of spikes emitted by that neuron, ni =(cid:80)
t yi(t), following stimulus
onset for a brief period of time lasting five times the time constant τRC of the RC circuit.
Our simulation updates the membrane potential every 0.1 τRC , thus there are 50 steps
in the numerical integration following each stimulus presentation. Consequently, at least
in principle, 50 is the maximum number of spikes we could observe from one neuron in
response to any image. We note that one could instead use first-spike latencies to measure
the computational output32,35; these two measures are highly correlated in our network, with
shorter latencies corresponding to greater spike counts (data not shown). The network learns
via rules similar to those of Foldi´ak17,37. These rules drive each unit to be active for only a
small but non-zero fraction of the time (lifetime sparseness) and to maintain uncorrelated
activity with respect to all other units in the network:
∆Wim = α(ninm − p2)
∆Qik = βni(Xk − niQik)
∆θi = γ(ni − p),
(1)
where p is the target average value for the number of spikes per image, which defines each
neuron's lifetime sparseness, and α, β, and γ are learning rates -- small positive constants
that determine how quickly the network modifies itself. Updating the feed-forward weights
Qik in our model is achieved with Oja's implementation41 of Hebb's rule; this rule is what
drives the network to represent the input. Note that because the firing rates are low (p = 0.05
spikes per image, for the results shown in this paper), and spikes can only be emitted in
integer units, our model implicitly allows only small numbers of neurons to be active at any
given time (so called "hard" sparseness, or L0 sparseness), similar to what is achieved by
other means in some recent non-spiking sparse coding models27,39.
These learning rules can be viewed as an approximate stochastic gradient descent ap-
proach to the constrained optimization problem in which the network seeks to minimize the
5
error between the input pixel values {Xk}, and a linear generative model formed by all of
the neurons Xk = (cid:80)
i niQik, while maintaining fixed average firing rates and no firing rate
correlations. This constrained optimization interpretation of our learning rules, and the
approximations involved, are discussed in the Methods section.
In Fig. 2, we demonstrate that the activity of the SAILnet units can be linearly decoded
to recover (approximately) the input stimulus. The success of linear decoding in a model
that encodes stimuli in a non-linear fashion is a product of our learning rules, and it has been
observed in multiple sensory systems42 and spiking neuron models optimized to maximize
information transmission4,14.
Our learning rules encourage all neurons to have the same average firing rate of p spikes
per image, which may at first appear to be at odds with the observation19 that cortical
neurons display a broad distribution of activities -- firing rates vary from neuron to neuron.
However, when trained on natural images, neurons in SAILnet can actually exhibit a
fairly broad range of firing rates. Moreover, the mean firing rate distribution ranges from
approximately lognormal to exponential in response to natural image stimuli, depending on
the mean contrast of the stimulus ensemble with which they are probed. We discuss this
further in the Firing Rates section below.
We emphasize here that each of our learning rules is "synaptically" local: the information
required to determine the change in the connection strength at any synaptic junction between
two units is merely the activity of the pre- and post-synaptic units. The inhibitory lateral
connection strengths, for example, are modified according to how many spikes arrived at
the synapse, and how many times the post-synaptic unit spiked. The information required
for the unit to modify its firing threshold is the unit's own firing rate. Finally, the rule
for modifying the feed-forward connections requires only the pre-synaptic activity Xk, the
post-synaptic activity ni, and the present strength of that connection Qik. This locality is a
desirable model feature because learning in cortex is thought30 to occur by the modification
of synaptic strengths and thus by necessity should depend only upon information available
locally at the synapse.
By contrast, much previous work15,27,29,33 has used a different learning rule for the feed
forward weights: ∆Qik ∝ ni(Xk −(cid:80)
j njQjk). This rule is non-local because the update for
the connection strength between input pixel k and unit i requires information about the
activities and feed-forward weights of every other unit in the network (indexed by j). It is
6
unlikely that such information is available to individual synapses in cortex. Interestingly,
in the limit of highly sparse and uncorrelated neuronal activities, our local learning rule
approximates the non-local rule used by previous workers15,27,33, when averaged over several
input images; we provide a mathematical derivation of this result in the Methods section.
This suggests an additional reason why sparseness is beneficial for cortical networks, in which
plasticity is local, but cooperative representations may be desired.
We trained a 1536-unit SAILnet with sparseness p = 0.05 on 16× 16 pixel image patches
drawn randomly from whitened natural images from the image set of Olshausen and Field15.
The network is six-times overcomplete with respect to the number of input pixels. This
mimics the anatomical fact that V1 contains many more neurons than does LGN, from
which it receives its inputs. Owing to the computational complexity of the problem -- there
are O(N 2) parameters to be learned in a SAILnet model containing N neurons -- we found
it prohibitive to consider networks that are much more than 6× overcomplete.
Our six-times overcompleteness is in a sense analogous to the three-times overcomplete-
ness of the SSC network described by Rehn and Sommer27, since the outputs of their com-
putational units could be either positive or negative, while our model neurons can output
only one type of spike. Thus, each of their units can be thought of as representing a pair of
our neurons, with opposite-signed receptive fields.
The RFs of 196 randomly selected units from our SAILnet are shown in Fig. 3, as mea-
sured by their spike-triggered average activity in response to whitened natural images. These
are virtually identical to the feed-forward weights of the units; in the Methods section, we
discuss why this must be the case.
To facilitate a comparison between the SAILnet RFs, and those measured in macaque V1
(courtesy of D. Ringach), we fit both the SAILnet, and the macaque RFs to Gabor functions.
As in the SSC study of Rehn and Sommer27, only those RFs that could be sensibly described
by a Gabor function were included in Fig. 3; for example, we excluded RFs with substantial
support along the square boundary, suggesting that the RF is only partly visible. In the
Methods section, we discuss the Gabor fitting routine and the quality control measures we
used to define and identify meaningful fits.
Our SAILnet model RFs show the same diversity of shapes observed in macaque V1,
and in the non-local SSC model27. They consist of three qualitatively distinct classes of
neuronal RFs: small unoriented features, localized and oriented Gabor-like filters, and elon-
7
gated edge-detectors. Our SAILnet learning rules approximately minimize the same cost
function as the SSC model27, albeit with constraints as opposed to unconstrained optimiza-
tion, which explains how it is possible for SAILnet to learn similar RFs using only local
rules. Furthermore, in our model, the number of co-active units is small, owing to the low
average lifetime neuronal firing rates, and the fact that spikes can only be emitted in integer
numbers. This feature is similar to the L0-norm minimization used in the SSC model of
Rehn and Sommer27 and the LCA model of Rozell and colleagues39.
This is the first demonstration that a network of spiking neurons using only synaptically
local plasticity rules applied to natural images can account for the observed diversity of V1
simple cell RF shapes.
SAILnet units exhibit a broad distribution of mean firing rates in response to
natural images
Our learning rules (Eq. 1) encourage every unit to have the same target value, p, for
its average firing rate, which might appear to be inconsistent with observations19,44,45 that
cortical neurons exhibit a broad distribution of mean firing rates. However, we find that
SAILnet, too, can display a wide range of mean rates, as we now describe.
To determine the distribution of mean firing rates across the population of model neurons
in our network, we first trained a 1536-unit SAILnet on 16 × 16 pixel patches drawn from
whitened natural images, and then presented the network with 50, 000 patches taken from
the training ensemble. Our measurement was performed with all learning rates set to zero,
so that we were probing the properties of the network at one fixed set of learned parameter
values, rather than observing changes in network properties over time.
We then counted the number of spikes per image from each unit to estimate each neuron's
average firing rate, as it might be measured in a physiology experiment. The distribution
of these mean firing rates is fairly broad and well-described by a lognormal distribution
(Fig. 4a). This distribution is strongly non-monotonic, clearly indicating that it is poorly
fit by an exponential function.
Subsequently, we probed the same network (still with the learning turned off, so that the
network parameters were identical in both cases) with 50, 000 low-contrast images consisting
of patches from our training ensemble with all pixel values multiplied by 1/3. We found that
8
the firing rate distribution was markedly different than what we found when the network
was probed with higher-contrast stimuli. In particular, it became a monotonic decreasing
function that was similarly well-described by either a lognormal or an exponential function
(Fig. 4b).
From the dynamics of our leaky integrate-and-fire units, it is clear that the low con-
trast stimuli with reduced pixel values will cause the units to charge up more slowly and
subsequently to spike less in the allotted time the network is given to view each image.
Consequently, the firing rate distribution gets shifted towards lower firing rates. However,
negative firing rates are impossible, so in addition to being shifted, the low-firing-rate tail
of the distribution is effectively truncated. Note that truncating the lognormal distribution
anywhere to the right of the peak results in a distribution that looks qualitatively similar to
an exponential.
Mean firing rates in primary auditory cortex (A1) have been reported by one group19 to
obey a lognormal distribution, whether spontaneous or stimulus-evoked in both awake and
anesthetized animals. However, exponentially distributed spontaneous mean firing rates
have also been reported in awake rat A143. Although several groups have measured the
distribution of firing rates over time for individual neurons44,45, we are unaware of a published
claim regarding the distribution of mean firing rates in visual cortex.
Recall that our learning rules encourage the neurons to all have the same average firing
rate. This fact may be puzzling at first given the spread in mean firing rates apparent in the
distributions shown in Fig. 4. There are two main effects to consider when making sense of
this: finite measurement time, and non-zero step-sizes for plasticity.
The first effect relates to the fact that there is intrinsic randomness in the measurement
process -- which randomly selected image patches happen to fall in the ensemble of probe
stimuli -- so that the measured distribution tends to be broader than the "true" underlying
distribution of the system. To check that this effect is not responsible for the broad distribu-
tion in firing rates, we computed the variance in the measured firing rate distribution after
different numbers of images were presented to the network. The variance decreased until it
reached an asymptotic value after approximately 25, 000 -- 30, 000 image presentations (data
not shown). Thus, the 50, 000 image sample size in our experiment is large enough to see the
true distribution; finite sample-size effects do not affect the distributions that we observed.
The other, more interesting, effect that gives rise to a broad distribution of firing rates is
9
related to learning. While the network is being trained, the feed-forward weights, inhibitory
lateral connections, and firing thresholds get modified in discrete jumps, after every image
presentation (or every batch of images, see the Methods section for details). Since those
jumps are of a non-zero size -- as determined by the learning rates α, β, and γ -- there will
be times when the firing threshold gets pushed below the specific value that would lead to
the unit having exactly the target firing rate, and the unit will thus spike more than the
target rate. Similarly, some jumps will push the threshold above that specific value, and
the unit will fire less than the target amount. Even after learning has converged, and the
parameters are no longer changing on average in response to additional image presentations,
the network parameters are still bouncing around their average (optimal) values; any image
presentation that makes a neuron spike more than the target amount results in an increased
firing threshold, while any image that makes the neuron fire less than the target amount
leads to a decreased firing threshold. Recent results46 suggest that the sizes of these updates
(jumps) are quite large for real neurons. Interestingly, this indicates that the observed broad
distributions in firing rate19 do not rule out the possibility that homeostatic mechanisms are
driving each neuron to have the same average firing rate.
Reducing the SAILnet learning rates α, β and γ does reduce the variance of the firing rate
distributions, but our qualitative conclusions -- non-monotonic, approximately lognormal
firing rate distribution in response to images from the training set, and monotonic, expo-
nential/lognormal distribution in response to low contrast images -- are unchanged when
we use different learning rates for the network (data not shown).
Pairs of SAILnet units have small firing rate correlations.
Recent experimental work48,49 has shown that neurons in visual cortex tend to have
small correlations between their firing rates. In order to facilitate a comparison between
our model, and the physiological observations, we have measured the (Pearson's) linear
correlation coefficients between spike counts of SAILnet units, in response to an ensemble
30,000 natural images. These correlations (Fig. 5) tend to be near zero, as is observed
experimentally48, while the experimental data show a larger variance in the distribution of
correlation coefficients than we observe with SAILnet. We note that, like the firing rate
distribution (discussed above), the distribution of correlation coefficients is affected by the
10
update sizes (learning rates) in the simulation, with larger update sizes leading to a larger
variance of the measured distribution.
In Fig. 5, the distribution appears truncated on the left. This effect arises because there
is a lower bound on the correlation between the neuronal firing rates that arises when the
two neurons are never co-active. The low mean firing rate of p = 0.05 used in our simulation
means that this bound is not too far below zero.
Connectivity learned by SAILnet allows for further experimental tests of the
model
Several previous studies of sparse coding models15,17,27,29,31,37 have focused on the receptive
fields learned by adaptation to naturalistic inputs, but we are aware of only one published
study50 that investigated the connectivity in sparse coding models, albeit with a model
that lacked biological realism. One previous study51 investigated synaptic mechanisms that
could give rise to the measured distribution of connection strengths, but this work was
not performed in the context of a sensory coding model. No prior work has studied the
connectivity learned in a biophysically well-motivated sensory coding network, which would
provide additional testable predictions for physiology experiments.
Fig. 6 shows the distribution of non-zero connection strengths (non-zero elements of the
matrix Wim) learned by a 1536-unit SAILnet with p = 0.05 trained on 16× 16 pixel patches
drawn from whitened natural images (the same network whose receptive fields are shown
in Fig. 3). When trained on natural images, SAILnet learns an approximately lognormal
distribution of inhibitory connection strengths; a Gaussian best fit to the histogram of the
logarithms of the connection strengths accounts for 98% of the variance in the data.
Despite this close agreement, SAILnet shows some systematic deviations from the log-
normal fit, especially on the low-connection-strength tail of the distribution. Interestingly,
the experimental data38 show an approximately lognormal distribution of excitatory con-
nection strengths, with similar systematic deviations (Fig. 5b of Song et al.38). By contrast,
prior theoretical work38,51 has employed learning rules tailored to create exactly lognormal
connection strength distributions, and thus show no such deviations. Note also that neither
of these previous studies addressed the issue of how neurons might represent sensory inputs,
nor how they might learn those representations.
11
Whereas the experimental data of Song et al.38 show a roughly lognormal distribution in
the strengths of excitatory connections between V1 neurons, our model makes predictions
about the strengths of inhibitory connections in V1. The 1 ms time window for measuring
post-synaptic potentials in the experiment of Song et al.38 ensured that they measured only
direct synaptic connections. However, suppressive interactions between excitatory neurons
in cortex are mediated by inhibitory interneurons. Consequently, the inhibitory interactions
between pairs of excitatory neurons in V1 must involve two or more synaptic connections
between the cells. Thus, our model predicts that the inhibitory functional connections
between excitatory simple cells in V1, like the excitatory connections measured by Song
et al.38, should follow an approximately lognormal distribution (Fig. 6), but it does not
specify the extent to which this is achieved through variations in strength among dendritic
or axonal synaptic connections of V1 inhibitory interneurons. One recent theoretical study46
has uncovered some interesting relationships between coding schemes and connectivity in
cortex, but it did not make any statements about the anticipated distribution of inhibitory
connections.
Interestingly, there is a clear correlation between the strengths of the inhibitory connec-
tion between pairs of SAILnet neurons, and the overlap (measured by vector dot product)
between their receptive fields: neurons with significantly overlapping receptive fields tend
to have strong inhibitory connections between them (Fig. 6). This correlation is expected
because cells with similar RF's receive much common feed-forward input. Thus, in order to
keep their activities uncorrelated, significant mutual inhibition is required. This same fea-
ture was assumed by the LCA algorithm of Rozell39 and colleagues, but is naturally learned
by SAILnet, in response to natural stimuli.
Our connectivity predictions are amenable to direct experimental testing, although that
testing may be challenging, owing to the difficulty of measuring functional connectivity
mediated by two or more synaptic connections between pairs of V1 excitatory simple cells.
III. DISCUSSION
The present work represents the first demonstration that synaptically local plasticity
rules can be used to learn a sparse code for natural images that accounts for the diverse
shapes of V1 simple cell receptive fields. Our model uses purely synaptically local learning
12
rules -- connection strengths are updated based only on the number of spikes arriving at the
synapse and the number of spikes generated by the post-synaptic cell. By contrast, the local
competition algorithm (LCA) of Rozell and colleagues39 assumes that Wim =(cid:80)
k QikQmk, so
that the strength of the inhibitory connection between two neurons is equal to the overlap
(i.e., vector dot product) between their receptive fields. This non-local rule requires that
individual inhibitory synapses must somehow keep track of the changes in the receptive fields
of many neurons throughout the network in order to update their strengths. Moreover, the
LCA network does not contain spiking units, even though cortical neurons are known to
communicate via discrete, indistinguishable, spikes of activity30.
Similarly, the units in the networks of Falconbridge et al.37 and Foldi´ak17 communicate
via continuous-valued functions of time. Although these two models17,37 do use synaptically
local plasticity rules, neither of these groups demonstrated that such local plasticity rules
are sufficient to explain the diversity of simple cell RF shapes observed in V1.
We note that, independent of the present work, Rozell and Shapero have recently imple-
mented a spiking version of LCA39 that uses leaky integrate-and-fire units (S. Shapero, D.
Bruderle, P. Hasler, and C. Rozell, CoSyne 2011 abstract). However, that work does not
address the issue of how to train such a network using synaptically local plasticity rules.
Some groups have used spiking units to perform image coding31 -- 35, but those studies did
not address the question of whether synaptically local plasticity rules can account for the ob-
served diversity of V1 RF shapes. Interestingly, it has been demonstrated32 that orientation
selectivity can arise from spike timing dependent plasticity rules applied to natural scenes.
Previous work33 has also explored the addition of homeostatic mechanisms to sparse coding
algorithms and found it to improve the rate at which learning converges and to qualitatively
affect the shapes of the learned RFs; homeostasis is enforced in our model via modifiable
firing thresholds.
Finally, we note that one previous group36 has demonstrated that independent component
analysis (ICA) can be implemented with spiking neurons and local plasticity rules. That
work did not, however, account for the diverse shapes of V1 receptive fields, although they
did also demonstrate that homeostasis (a mean firing rate constraint) was critical to the
learning process.
Our model attempts to be biophysically realistic, but it is not a perfect model of vi-
sual cortex in all of its details.
In particular, like many previous models15,27,29,31,33, our
13
network alternates between brief periods of inference (the representation of the input by a
specific population activity pattern in the network) and learning (the modification of synap-
tic strengths), which may not be realistic. Indeed, it is unclear how cortical neurons would
"know" when the inference period is over and when the learning period should begin, though
it is interesting to note that these iterations could be tied to the onset of saccades, given
the 5τRC ≈ 100 ms inference period between "learning" stages in our model.
As in previous models, the inputs to our network Xj are continuous-valued, whereas the
actual inputs from the lateral geniculate nucleus to primary visual cortex (V1) are spiking.
As mentioned above, suppressive interactions between pairs of units in our model are me-
diated by direct, one-way, inhibitory synaptic connections between units, rather than being
mediated by a distinct population of inhibitory interneurons. We do not include the effects
of spike-timing dependent plasticity47, although this has been shown to have interesting the-
oretical implications for cortex46 in general and for image coding in particular32,34. We are
currently developing models that incorporate spike timing dependent learning rules, applied
to time-varying image stimuli such as natural movies.
Finally, the neurons in our model have no intrinsic noise in their activities, although that
noise may, in practice, be small52.
Interestingly, since our model neurons require a finite amont of time to update their
internal variables ui(t), there is a hysteresis effect if one presents the network with time-
varying image stimuli -- the content of previous frames affects how the network processes
and represents the current frame. Even if the features in a movie change slowly, the optimal
representation of one frame can be very different from the optimal representation of the
next frame in many coding models, so this hysteresis effect can provide stability to the
image representation compared to other models such as ICA8,9 or Olshausen and Field's
sparsenet15. This effect has previously been studied by Rozell and colleagues39, encouraging
our efforts to apply SAILnet to dynamic stimuli.
Though it is highly simplified, our model does captures many qualitative features of
V1, such as inhibitory lateral connections22, largely uncorrelated neuronal activities48,49,
sparse neuronal activity20 -- 22, a greater number of cortical neurons than input neurons (over-
complete representation), synaptically local learning rules, and spiking neurons.
Impor-
tantly, this model allows us to make several falsifiable experimental predictions about in-
terneuronal connectivity and population activity in cortex. We hope that these predictions
14
will help uncover the coding principles at work in the visual cortex.
IV. METHODS
SAILnet dynamics
Each of the neurons in our SAILnet follows leaky integrate-and-fire dynamics30. The
neurons, indexed by subscript i, each have a time-dependent, continuous-valued internal
variable ui(t), analogous to a neuronal membrane potential. We explicitly model each neuron
as an RC circuit (Fig. 1), where the internal variable ui(t) corresponds to the voltage across
the capacitor. Whenever this internal variable exceeds a threshold value θi specific to that
neuron, the neuron emits a punctate spike of activity. The unit's external variable yi(t),
which represents the spiking output that is communicated to other neurons throughout the
network, is 1 for a brief moment. At all other times, the unit's external variable is 0.
Since the thresholds θi are adapted slowly compared to the time scale of inference, they
are approximately constant during inference. The same is true for the feed-forward weights
Qik and the lateral connection strengths Wim, discussed below.
We model the effects of the input image {Xk} and the activities of other neurons in the
network ym(t) on the internal variable as a current, Iinput(t) =(cid:80)
m(cid:54)=i Wimym(t),
k QikXk −(cid:80)
that is impinging on the RC circuit; here the feed-forward weights Qik and lateral connection
strengths Wim describe how much a given input (either an image pixel value, or a spike from
another neuron in the network) should modify the neuron's internal variable. The internal
variable evolves in time via the differential equation for voltage across our capacitor, in
response to the input current dui(t)/dt + ui(t) = Iinput(t).
We simulate these dynamics in discrete time, performing numerical integration of the
differential equation dui(t)/dt+ui(t) = Iinput(t). Whenever the internal variable ui(t) exceeds
the threshold (at time t∗; ui(t∗) > θi), the output spike occurs at the next time step:
yi(t∗ + 1) = 1. In the subsequent time step, the external variable yi(t∗ + 2) returns to zero,
unless the internal variable ui(t) has again crossed the threshold.
After the unit spikes, the internal variable returns to its resting value of 0, from whence
the unit can again be charged up.
For simplicity, our differential equation assumes that the RC time constant of the model
15
neuron is one "unit" of time. Our simulated dynamics are allowed to run for five such units
of time (with the time step of numerical integration being 0.1 units in duration), in response
to each input image. At the start of these dynamics, the internal variables of all neurons
are set to their resting values: ui(t = 0) = 0 ∀ i.
SAILnet learning rules can be viewed as a gradient descent approach to a
constrained optimization problem
Unlike previous work15,27, which performed unconstrained optimization on a cost function
penalizing both reconstruction error and network activity, our learning rules can be viewed
as a gradient descent approach to a constrained optimization problem.
Given the neuronal activities ni in response to an image, and their feed-forward weights
Qik, one can form a linear generative model X of the input stimulus Xk = (cid:80)
k (Xk −(cid:80)
mean squared error between that model X and the true input X is E =(cid:80)
i niQik. The
i niQik)2,
and the creation of a high fidelity representation suggests that this error function E, or one
like it, be minimized by the learning process.
Let us suppose that the neuronal network is not free to choose any solution to this
problem; instead it must satisfy constraints that require the neurons to have a fixed average
firing rate of p and minimal correlation between neurons.
Indeed, neurons tend to have
low mean firing rates when averaged across many different images, and those firing rates
span a finite range of values19,44,45, motivating our first constraint. The second constraint is
justified by observations that neural systems tend to exhibit little or no correlation between
pairs of units48,49, and that the correlation between the activity of V1 neurons decreases
significantly as one increases the fraction of the visual field that is stimulated21.
We use the method of Lagrange multipliers to solve this problem, allowing our learning
rules to adapt the network so as to minimize reconstruction error while approximately sat-
isfying these constraints. To do this, we perform gradient descent on a Lagrange function L
that contains both the error function and the constraints:
(cid:88)
(cid:88)
k
i
L =
+
(cid:32)
Xk −(cid:88)
i
λi(ni − p) +
16
niQik
(cid:88)
i(cid:54)=k
(cid:33)2
τim(ninm − p2),
(2)
where the sets of values {λi} and {τim} are our (unknown) Lagrange multipliers. To
perform constrained optimization, gradient descent is performed with respect to all of the
free parameters in L: namely, the set of feed-forward weights {Qik}, and the Lagrange
multipliers {λi} and {τim}:
∂L
∂λi
∂L
∂τim
∂L
∂Qik
= ni − p
= ninm − p2
= −2ni(Xk −(cid:88)
r
nrQrk).
(3)
The first two equations lead to our learning rules for inhibitory connections and firing
thresholds, once we identify λi ∝ −θi and τim ∝ −Wim; these network parameters correspond
to the Lagrange multipliers of the constrained optimization problem. This reflects the fact
that the role of the variable thresholds and inhibitory connections is to enforce the sparseness
and non-correlation constraints in the network, which is the same as the role of the Lagrange
multipliers in the Lagrange function.
We emphasize that the terms of our objective function that effectively enforce these
constraints are critical for our algorithm's success. By contrast, consider the situation in
which the model units had no other possibility but to maintain their fixed firing rate and lack
of correlation, due to some clever parameterization of the model's state space. In that case,
one could simply minimize the reconstruction error, via gradient descent, and the existence
of these extra terms, or even of the analogous Lagrange multipliers, would be redundant.
However, in our model, each change of the feed-forward weights (Qik) could change the
neuron's firing rate, and the correlation between its activity and those of other neurons,
unless something forces the network back towards the constraint surface. The variable firing
thresholds and inhibitory inter-neuronal connection strengths in our model perform this
function.
The last equation from our gradient descent calculation gives the update rule for the
feed-forward weights ∆Qik ∝ ni(Xk −(cid:80)
r nrQrk). This rule, as written, is unacceptable
for our SAILnet because we wish to interpret the strengths of connections in that network
as the strengths of synaptic connections in cortex.
In that case, learning at any given
synapse should be accomplished using only information available locally, at that synapse.
17
For updating connection strength Qik, this could include the pre-synaptic activity Xk, the
post-synaptic activity ni, and the current value of the connection strength Qik, but should
not require information about the receptive fields of other neurons in the network, nor their
activities, because it is not clear that that information is available at each synapse. Hence,
r nrQrk term that arises from gradient descent on our objective function is a problem
for the biological interpretation of these learning rules. We will now show that, in the limit
the (cid:80)
that the neuronal activity is sparse and uncorrelated, when averaged over several input
images, the non-local gradient descent rule ∆Qik ∝ ni(Xk −(cid:80)
Consider the non-local update rule ∆Qik ∝ ni(Xk −(cid:80)
equivalent to a simpler rule, originally due to Oja41, that is synaptically local.
r nrQrk). Expanding the polyno-
r nrQrk) is approximately
mial, and averaging over image presentations, we find
(cid:104)∆Qik(cid:105) ∝ (cid:104)niXk(cid:105) −(cid:68)
n2
i Qik
(cid:69) −(cid:88)
r(cid:54)=i
(cid:104)ninrQrk(cid:105) .
(4)
If the learning rate β is small, such that the feed-forward weights change only slowly over
time, then we can approximate that they are constant over some (small) number of image
presentations, and take them outside of the averaging brackets;
(cid:104)∆Qik(cid:105) ∼ (cid:104)niXk(cid:105) −(cid:68)
n2
i
(cid:69)
Qik −(cid:88)
r(cid:54)=i
(cid:104)ninr(cid:105) Qrk.
(5)
Now, so long as the neuronal activities are uncorrelated, and all units have the same av-
erage firing rate (recall these constraints are enforced by our Lagrange multipliers), (cid:104)ninr(cid:105) =
(cid:104)ni(cid:105)(cid:104)nr(cid:105) = p2 ∀i, r, and thus the learning rule is
(cid:104)∆Qik(cid:105) ∼ (cid:104)niXk(cid:105) −(cid:68)
n2
i
(cid:69)
Qik − p2(cid:88)
r(cid:54)=i
Qrk.
(6)
This last term is small compared to the first two for a few reasons. First, the neurons
in the network have sparse activity, meaning they are selective to particular image features,
and thus (cid:104)n2
i(cid:105) (cid:29) (cid:104)ni(cid:105)2 = p2. This can be easily seen by that fact that we use small values
for p, meaning that the neurons fire, on average, much less than one spike per image. The
spikes, however, can only be emitted in integer numbers, so the neurons are silent in response
to most image presentations, and are thus highly selective.
18
Furthermore, the last term, p2(cid:80)
will be negative. These random signs mean that the sum (cid:80)
neurons in the network. Some of the RFs will be positive for a given pixel, whereas others
r(cid:54)=i Qrk, involves a sum over the receptive fields of many
r(cid:54)=i Qrk tends towards zero.
Thus, in the limit of sparse and uncorrelated neuronal activity (the limit in which our
network operates), gradient descent on the error function E yields approximately
(cid:104)∆Qik(cid:105) ∼ (cid:104)niXk(cid:105) −(cid:68)
n2
i
(cid:69)
Qik,
(7)
which is equivalent to the average update from Oja's implementation of Hebbian learning41,
which we use for learning in SAILnet. Thus, SAILnet learns to approximately solve the
same error minimization problem as did previous, non-local sparse coding algorithms15,27.
Interestingly, our result suggests that, despite the highly non-linear way in which our
model neurons' outputs (spikes ni) are generated from the input, a linear decoding of the
network activity should provide a good match to the input: Xk ≈ (cid:80)
i niQik . This linear
decodability has previously been observed in physiology experiments42, as well as models
designed to maximize the information rate about input stimulus conveyed by individual
spiking neurons4,14, and it is indeed a property of SAILnet.
We summarize the learning rules for SAILnet here.
∆θi ∝ ni − p
∆Wim ∝ ninm − p2
∆Qik ∝ ni(Xk − niQik)
(8)
The first two rules enforce the sparseness and correlation constraints, and arise from the
Lagrange multipliers in our Lagrange function. The final rule drives the SAILnet represen-
tation to form a better match to the input stimulus, as it adapts to the ensemble of training
images.
Receptive fields measured by spike-triggered average are proportional to the
feed-forward weights of the neurons when the probe stimulus statistics match
those of the training stimuli.
Consider the Oja-Hebb41 learning rule for the feed-forward weights in our model,
19
∆Qik ∝ ni (Xk − niQik) .
(9)
Once the learning has converged over some set of training stimuli, the feed-forward weights
are, on average, no longer changing in response to repeated presentations of examples from
the training set. Thus,
(cid:104)∆Qik(cid:105) ∝ (cid:104)ni (Xk − niQik)(cid:105) = 0.
Expanding the middle term in this expression, we find that
(cid:68)
n2
i Qik
(cid:69)
(cid:69)
(cid:68)
n2
i
=
Qik,
(cid:104)niXk(cid:105) =
(10)
(11)
where the second equality occurs because the learning has converged, and thus the feed-
forward weights are constant over repeated image presentations. Thus, we find that
i(cid:105) = Qik; the spike-triggered average (STA) stimulus is equivalent to the set
(cid:104)niXk(cid:105) /(cid:104)n2
of feed-forward weights, up to a multiplicative scaling factor that can be calculated from
the spike train.
Training SAILnet
We start out each simulation with all inhibitory connection strengths Wim set to zero, all
firing thresholds θi set to 5, and the feed-forward weights Qik initialized with Gaussian white
noise. To train the network, batches37 of 100 images with zero mean, and unit standard
deviation pixel values, are presented, and the number of spikes from each neuron are counted
separately for each image. After each batch, the average update for the network properties is
computed (following our learning rules) over the 100-image batch. This batch-wise training
lets us use matrix operations for computing the updates, which dramatically speeds up
the training process. After each update, all negative values for inhibitory connections Wim
(which would correspond to excitatory connections) are set to zero, as in the previous work
by Foldi´ak17. Relaxing this constraint, and allowing the recurrent weights to change sign
does not affect our qualitative conclusions. In that case, some of the recurrent connections
20
become excitatory, while the majority remain inhibitory, the RF's are qualitatively the
same as those shown in Fig. 3, and the distributions of inhibitory and excitatory connection
strengths are both approximately lognormal (data not shown).
The relative values of α, β and γ were chosen based on Foldi´ak's17 observation that β
must be much less than α or γ so that the neurons' activities remain sparse and uncorrelated,
even in the face of changing feed-forward weights.
We study the network after the properties stop changing macroscopically over time.
However, as noted in the firing rates section of this paper, the network parameters continue
to bounce around the final "target" state, with the size of the bounces determined by the
learning rates in the network. Empirically, we find that it takes on the order of 107 image
presentations (105 steps of 100 image presentations per step) for this dynamic equilibrium to
be established. For the results presented in this paper, we let the network train for roughly
2 × 108 image presentations.
To speed up the simulation, we start the training with large values for the learning
rates, and these are eventually reduced. For the last 104 batches of training (106 image
presentations), the learning rates were (α, β, γ) = (0.1, 0.001, 0.01).
All of the computer codes used to generate the results presented in this paper are available
upon request.
Fitting SAILnet RFs to Gabor functions
A Gabor function (G(x, y)) is a common model for visual cortical receptive fields26,27,
which consists of a two-dimensional Gaussian multiplied by a sinusoid:
−
(cid:32) xp√
G = A cos(2πf xp + ψ) exp
2σx
xp = (x − x0) cos(θ) + (y − y0) sin(θ),
yp = −(x − x0) sin(θ) + (y − y0) cos(θ).
(cid:33)2 −
(cid:32) yp√
2σy
(cid:33)2 ,
(12)
The center of the shape is defined by the coordinates x0 and y0, while the amplitude and
orientation of the pattern are defined by the parameters A and θ, respectively. ψ defines
the phase of the sinusoid, relative to the center of the Gaussian envelope, which has spatial
21
extent σx and σy in the direction along, and perpendicular to, the direction in which the
sinusoid oscillates (with frequency f ), respectively.
Given a neuronal RF, our code performs unconstrained optimization to choose the Gabor
parameters such that the mean squared error G − RF2 is minimized. We then perform
several quality control measures to ensure that our analysis only contains sensible Gabor
parameters that accurately describe our RFs.
The first such measure is to exclude any RF for which the deviation between the RF and
the Gabor fit is large; cells with G− RF2/RF2 > 0.5 were excluded. This is equivalent
to placing a (fairly mild) restriction on the minimum allowable signal-to-noise ratio.
The second quality control measure is to exclude those RFs for which the center of the
pattern (x0, y0) lies either outside the 16 × 16 pixel patch, or within one standard deviation
(of the Gaussian envelope) of the patch edge. As described by other workers27, when the
center of the pattern is outside of the visible 16×16 pixel patch, it is not clear that the shape
of the RF itself is well-described by the Gabor parameters, or even well-constrained, for that
matter. Our (more stringent) restriction also avoids the problem of biased shape estimates,
when fitting Gabors to RFs that are truncated by the edge of the patch; the model RFs
essentially tile the available space, so some of them will, by necessity, have centers that lie
right along or outside of the edges of the patch. Indeed, in a 16× 16 pixel space, many pixels
are near the edge, thus this cut excludes many RFs.
After making all of these cuts, we were left with 299 RFs, on which to perform subsequent
shape analysis.
We performed the same fitting and quality-control analysis on both the SAILnet and
the macaque physiology RFs, although we used a gentler goodness-of-fit restriction on the
macaque data, since the macaque RFs, as measured, have fairly large regions of zero support,
in which any measurement noise reduces the apparent goodness-of-fit. For the macaque data,
we excluded those RFs with G− RF2/RF2 > 0.8, leaving 116 of the 250 macaque RFs
for subsequent analysis.
One reason for the relatively low yield of well-fit RFs is that not all RFs are actually
well-described by Gabor functions. For example, there is no choice of Gabor parameters that
will accurately describe a center-surround receptive field; that RF is much better described
by a difference of Gaussians function, for example. We leave for future work the issue of
determining the best family of functions with which to describe visual cortical receptive
22
fields.
REFERENCES
1Attneave, F. (1954). Some informational aspects of visual perception. Psychol. Rev. 61,
18393.
2Barlow, H.B. (1961). Possible principles underlying the transformations of sensory mes-
sages. In Sensory Communication, W.A. Rosenblith, ed. (Cambridge, MA: MIT Press),
pp. 21734.
3Atick, J.J. and Redlich, A.N. (1992). What does the retina know about natural scenes?
Neural Comput. 4, 196-210.
4Bialek, W., DeWeese, M., Rieke, F. and Warland, D. (1993). Bits and brains: Information
fow in the nervous system. Physica A 200, 581-593.
5Bialek, W., Ruderman, D.L. and Zee, A. (1991). Optimal sampling of natural images: A
design principle for the visual system. In Advances in Neural Information Processing 3,
Lippman, R.P., Moody, J.E. and Touretzky, D.S. eds. (San Mateo, CA: Morgan Kauf-
mann), pp. 363-369.
6Linsker, R. (1989). An application of the principle of maximum information preservation
to linear systems. In Advances in Neural Information Processing 1., Touretzky, D. ed. (San
Mateo, CA: Morgan Kaufmann), pp.186-194.
7Atick, J.J. (1992). Could information theory provide an ecological theory of sensory pro-
cessing? In Princeton Lectures on Biophysics, Bialek, W. ed. (Singapore: World Scientific).
8Bell, A.J. and Sejnowski, T.J. (1997) The "Independent Components" of Natural Scenes
are Edge Filters. Vis. Res. 37, 3327-3338.
9Hyvarinen, A. and Hoyer, P.O. (2001). A two-layer sparse coding model learns simple and
complex cell receptive fields and topography from natural images. Vis. Res. 41, 2413-2423.
10Tkacik, G., Prentice, J.S., Balasubramanian, V. and Schneidman, E. (2010). Optimal
population coding by noisy spiking neurons. Pro. Natl. Acad. Sci. USA 107, 14419-14424.
11Laughlin, S.B. (1981). A simple coding procedure enhances a neuron's information capac-
ity. Z. Naturforsch. 36c, 910-912
12Rieke, F., Warland D., de Ruyter van Steveninck, R., and Bialek, W. (1999). Spikes:
Exploring the Neural Code (Cambridge, MA: MIT Press).
23
13Dan, Y., Atick, J.J. and Reid, C.R. (1996). Efficient coding of natural scenes in the lateral
geniculate nucleus: experimental test of a computational theory. J. Neurosci. 16, 3351-
3362.
14Deweese, M. (1996). Optimization principles for the neural code. Network 7, 325-31.
15Olshausen, B.A. and Field, D.J. (1996). Emergence of simple-cell receptive field properties
by learning a sparse code for natural images. Nature 381, 607-609.
16Olshausen, B.A. and Field, D.J. (1997). Sparse coding with an overcomplete basis set: A
strategy employed by V1? Vis. Res. 37, 3311-3325.
17Foldi´ak, P. (1990). Forming sparse representations by local anti-Hebbian learning. Biol.
Cybern. 64, 165-170.
18Graham, D.J. and Field, D.J. (2007). Sparse Coding in the Neocortex. In Evolution of
Nervous Systems, Vol. III, J. H. Kass et al., eds. (Oxford: Academic Press), pp. 181-187.
19Hrom´adka, T., Deweese, M.R. and Zador, A.M. (2008). Sparse representation of sounds
in the unanesthetized auditory cortex. PLoS Biol. 6, 124-137.
20Vinje, W.E. and Gallant, J.L. (2002). Natural stimulation of the nonclassical receptive
field increases information transmission efficiency in V1. J. Neurosci. 22, 2904-2915.
21Vinje, W.E. and Gallant, J.L. (2000). Sparse coding and decorrelation in primary visual
cortex during natural vision. Science 287, 1273-1276.
22Haider, B.A., Krause, M.R., Duque, A., Yu, Y., Touryan, J., Mazer, J.A. and McCormick,
D.A. (2010). Synaptic and network mechanisms of sparse and reliable visual cortical ac-
tivity during nonclassical receptive field stimulation. Neuron 65, 107-121.
23Lennie, P. (2003). The cost of cortical computation. Curr. Biol. 13, 493-497.
24Tolhurst, D.J., Smyth, D. and Thompson, I.D. (2009). The sparseness of neuronal re-
sponses in ferret primary visual cortex. J. Neurosci. 29, 2355-2370.
25Sakata, S. and Harris, K.D. (2009). Laminar structure of spontaneous and sensory-evoked
population activity in auditory cortex. Neuron 64, 404-18.
26Ringach, D.L. (2002). Spatial structure and symmetry of simple-cell receptive fields in
macaque primary visual cortex. J. Neurophysiol. 88, 455-63.
27Rehn, M. and Sommer, F.T. (2007). A network that uses few active neurones to code
visual input predicts the diverse shapes of cortical receptive fields. J. Comput. Neurosci.
22, 135-146.
28Donoho, D.L. (2004). Compressed Sensing. IEEE Tran. Inform. Theory. 52, 1289-1396.
24
29Olshausen, B.A., Cadieu, C.F., and Warland, D.K. (2009). Learning real and complex over-
complete representations from the statistics of natural images. Proc. SPIE 7446, 74460S-1
- 74460S-11.
30Dayan, P. and Abbott, L. (2001). Theoretical Neuroscience: Computational and Mathe-
matical Modelling of Neural Systems (Cambridge, MA: MIT Press).
31Perrinet, L., Sameulides, M. and Thorpe, S. (2004). Sparse spike coding in an asynchronous
feed-forward multi-layer neural network using matching pursuit. Neurocomputing 57, 125-
134.
32Delorme, A., Perrinet, L. and Thorpe, S.J. (2000). Networks of integrate-and-fire neurons
using rank order coding B: spike timing dependent plasticity and emergence of orientation
selectivity. Neurocomputing 38, 539-545.
33Perrinet, L. (2010). Role of homeostasis in learning sparse representations. Neural Comput.
22, 1812-1836.
34Masquelier, T. and Thorpe, S.J. (2007). Unsupervised learning of visual features through
spike timing dependent plasticity. PLoS Comput. Biol. 3, e31.
35VanRullen, R. and Thorpe, S.J. (2002). Surfing a spike wave down the ventral stream. Vis.
Res. 42, 2593-2615.
36Savin, S., Joshi, P. and Triesch, J. (2010). Independent Component Analysis in spiking
neurons. PLoS Comp. Biol. 6, e1000757.
37Falconbridge, M.S., Stamps, R.L. and Badcock, D.R. (2006). A Simple Hebbian/Anti-
Hebbian Network Learns the Sparse, Independent Components of Natural Images. Neural
Comput. 18, 415-429.
38Song, S., Sjostrom, P.J., Reigl, M., Nelson, S. and Chklovskii, D.B. (2005). Highly nonran-
dom features of synaptic connectivity in local cortical circuits. PLoS Biol. 3, 0507-0519.
39Rozell, C.J., Johnson, D.H., Baraniuk, R.G. and Olshausen, B.A. (2008). Sparse coding
via thresholding and local competition in neural circuits. Neural Comput. 20, 2526-2563.
40Hopfield, J. (1984). Neurons with graded response have collective computational properties
like those of two-state neurons. Proc. Natl. Acad. Sci. USA 81, 3088-3092.
41Oja, E. (1982). A simplified neuron model as a principal component analyzer. J. Math.
Biol. 15, 267-273.
42Bialek, W., Rieke, F., De Ruyter Van Steveninck, R.R. and Warland, D. (1991). Reading
a neural code. Science 252, 1854-1857.
25
43Gaese, B.H. and Ostwald, J. (2003). Complexity and temporal dynamics of frequency
coding in the awake rat auditory cortex. Eur. J. Neurosci. 18, 26382652.
44Baddeley, R., et al. (1997). Responses of neurons in primary and inferior temporal visual
cortices. Proc. R. Soc. Lon. B. 264, 1775-1783.
45Abeles, M., Vaadia, E. and Berman, H. (1990). Firing patterns of single units in the
prefrontal cortex and neural network models. Network 1, 13-25.
46Clopath, C., Busing, L., Vasilaki, E. and Gerstner, W. (2010). Connectivity reflects coding:
a model of voltage-based STDP with homeostasis. Nat. Neurosci. 13, 344-352.
47Dan, Y. and Poo, M.M. (2006). Spike timing-dependent plasticity: from synapse to per-
ception. Physiol. Rev. 86, 1033-1048.
48Ecker, A.S. et al. (2010). Decorrelated neuronal firing in cortical microcircuits. Science
327, 584-587.
49Renart, A. et al. (2010). The asynchronous state in cortical circuits. Science 327, 587-590.
50Garrigues, P.J. and Olshausen, B.A. (2008). Learning horizontal connections in a sparse
coding model of natural images. In Advances in Neural Information Processing Systems 20,
J.C. Platt, D. Koller, Y. Singer and S. Roweis eds. (San Mateo, CA: Morgan Kaufmann),
pp. 505-512.
51Koulakov, A.A., Hrom´adka, T. and Zador, A.M. (2009). Correlated connectivity and the
distribution of firing rates in the neocortex. J. Neurosci. 29, 3685-3694.
52Mainen, Z.F. and Sejnowski, T.J. (1995). Reliability of spike timing in neocortical neurons.
Science 286, 1503-1506.
ACKNOWLEDGMENTS
The authors are grateful to Bruno Olshausen for his role in inspiring this work and
to Bruno Olshausen, Jascha Sohl-Dickstein, Fritz Sommer, and the other members of the
Redwood Center for many helpful discussions. We are also very grateful to Dario Ringach for
providing the macaque physiology data, and to Chris Rozell and Jascha Sohl-Dickstein for
constructive comments on the manuscript. The work of JZ is supported by funds from the
University of California and a Fulbright international science and technology PhD fellowship.
MRD gratefully acknowledges support from the National Science Foundation, the McKnight
Foundation, the McDonnell Foundation, and the Hellman Family Faculty Fund.
26
AUTHOR CONTRIBUTIONS
JZ initiated the project, designed and implemented the simulations, and performed the
analytic calculations. JTM implemented the Gabor fitting routine. JZ performed the re-
mainder of the data analysis. JZ and MRD wrote the manuscript.
ADDITIONAL INFORMATION
Competing financial interests
The authors declare no competing financial interests.
27
FIG. 1. SAILnet network architecture and neuron model (A) Our network architecture
is based on those of Rozell et al.39 and Foldi´ak17,37, and inspired by recent physiology experi-
ments20,22,48. Inputs Xk to the network (from image pixels) contact the neuron at connections
(synapses) with strengths Qik, whereas inhibitory recurrent connections between neurons22 in the
network have strengths Wim. The outputs of the neurons are given by yi(t); these spiking out-
puts are communicated through the recurrent connections, and also on to subsequent stages of
sensory processing, such as cortical area V2, which we do not include in our model. (B) Circuit
diagram of our simplified leaky integrate-and-fire30 neuron model. The inputs from the stimulus
with pixel values Xk, and the other neurons in the network, combine to form the input current
Iinput(t) =(cid:80)
k QikXk −(cid:80)
m(cid:54)=i Wimym(t) to the cell. This current charges up the capacitor, while
some current can leak to ground through a resistor in parallel with the capacitor. The resistors
are shown as cylinders to highlight the fact that they model the collective action of ion channels in
the cell membrane. The internal variable evolves in time via the differential equation for voltage
across our capacitor, in response to input current Iinput: dui(t)/dt + ui(t) = Iinput(t), which we
simulate in discrete time. Once that voltage exceeds threshold θi, the diode, which models neuronal
voltage-gated ion channels, opens, causing the cell to fire a punctate action potential, or spike, of
activity. For sake of a complete circuit diagram, the output is denoted as the voltage, Vout, across
some (small: Rout (cid:28) R) resistance. After spiking, the unit's internal variable returns to the resting
value of 0, from whence it can again be charged up.
28
FIG. 2. SAILnet activity can be linearly decoded to approximately recover the input
stimulus (A) An example of an image that was whitened using the filter of Olshausen and Field15,
which is the same filter used to process the images in the training set. The image in panel (A)
was not included in the training set. (B) A reconstruction of the whitened image in (A), by linear
decoding of the firing rates of SAILnet neurons, which were trained on a different set of natural
images. The input image was divided into non-overlapping 16 × 16 pixel patches, each of which
was preprocessed so as to have zero-mean and unit variance of the pixel values (like the training
set). Each patch was presented to SAILnet, and the number of spikes were recorded from each unit
in response to each patch. A linear decoding of SAILnet activity for each patch Xk = (cid:80)
i niQik
was formed by multiplying each unit's activity by that unit's RF and summing over all neurons.
The preprocessing was then inverted, and the patches were tiled together to form the image in
panel (B). The decoded image resembles the original, but is not identical, owing to the severe
compression ratio; on average, each 16× 16 input patch, which is defined by 256 continuous-valued
parameters, is represented by only 75 binary spikes of activity, emitted by a small subset of the
neural population. Linear decodability is a product of our learning rules, and it is an observed
feature of multiple sensory systems42 and spiking neuron models optimized to maximize information
transmission4,14.
29
FIG. 3. SAILnet learns receptive fields (RFs) with the same diversity of shapes as
those of simple cells in macaque primary visual cortex (V1) (A) 98 randomly selected
receptive fields recorded from simple cells in macaque monkey V1 (courtesy of D. Ringach). Each
square in the grid represents one neuronal RF. The sizes of these RFs, and their positions within
the windows, have no meaning in comparison to the SAILnet data. The data to the right of the
break line have an angular scale (degrees of visual angle spanned horizontally by the displayed RF
window) of 0.94o, whereas those to the left of it span 1.88o. (B) RFs of 196 randomly selected model
neurons from a 1536-unit SAILnet trained on patches drawn from whitened natural images. The
gray value in all squares represents zero, whereas the lighter pixels correspond to positive values,
and the darker pixels correspond to negative values. All RFs are sorted by a size parameter,
determined by a Gabor function best fit to the RF. The SAILnet model RFs show the same
diversity of shapes as do the RFs of simple cells in macaque monkey V1 (A); both the model units
and the population of recorded V1 neurons consist of small unoriented features, oriented Gabor-
like wavelets containing multiple subfields, and elongated edge-detectors. (C) We fit the SAILnet
and macaque RFs to Gabor functions (see Methods section), in order to quantify their shapes.
Shown are the dimensionless width and length parameters (σx × f and σy × f , respectively) of
the 299 SAILnet RFs and 116 (out of 250 RFs in the dataset) macaque RFs for which the Gabor
fitting routine converged. These parameters represent the size of the Gaussian envelope in either
direction, in terms of the number of cycles of the sinusoid. The SAILnet data (open blue circles)
span the space of the macaque data (solid red squares) from our Gabor fitting analysis; SAILnet
is accounting for all of the observed RF shapes. We highlight four SAILnet RFs with distinct
shapes, which are identified by the large triangular symbols that are also displayed next to the
corresponding RFs in panel (B).
30
FIG. 4. Units in SAILnet exhibit a broad range of mean firing rates, which can be
lognormally or exponentially distributed depending on the choice of probe stimuli. (A)
Frequency histogram of firing rates averaged over 50, 000 image patches drawn from the training
ensemble for each of the 1536 units of a SAILnet trained on whitened natural images. All learning
rates were set to zero during probe stimulus presentation. A wide range of mean rates was observed,
but as expected, the distribution is peaked near p = 0.05 spikes per image, the target mean firing
rate of the neurons. The paucity of units with near-zero firing rates suggests that this distribution
is closer to lognormal than exponential. Accordingly, the lognormal least-squares (solid red curve)
fit accounts for R2 = 96 % of the variance in the data, whereas the exponential fit (black dashed
curve) accounts for only 2 %. (B) In response to low contrast stimuli, the firing rate distribution
across the units (every unit fired at least once) in the same network as in panel (A) was similarly
well fit by either an exponential (dashed black curve; accounting for R2 = 88 % of the variance
in the data) or a lognormal function (solid red curve; accounting for 90 % of the variance). The
low-contrast stimulus ensemble used to probe the network consisted of images drawn from the
training set, with all pixel values reduced by a factor of three.
31
FIG. 5. Pairs of SAILnet units have small firing rate correlations.
The probability
distribution function (PDF) of the Pearson's linear correlation coefficients between the spike-counts
of pairs of SAILnet neurons responding to an ensemble of 30,000 natural images is sharply peaked
near zero.
32
FIG. 6. Connectivity learned by SAILnet allows for further experimental tests of the
model. (A) Probability Density Function (PDF) of the logarithms of the inhibitory connection
strengths (non-zero elements of the matrix Wim) learned by a 1536 unit SAILnet trained on 16×16
pixel patches drawn from whitened natural images. The measured values (blue points) are well-
described by a Gaussian distribution (solid line), which accounts for R2 = 98 % of the variance
in the dataset. This indicates that the data are approximately lognormally distributed. Note
that there are some systematic deviations between the Gaussian best fit and the true distribution,
particularly on the low-connection strength tail, similar to what has been observed for excitatory
connections within V138. This plot was created using the binning procedure of Hrom´adka and
colleagues19. The histogram was normalized to have unit area under the curve. (B) The strengths of
the inhibitory connections between pairs of cells are correlated with the overlap between those cells'
receptive fields: cells with significantly overlapping RFs tend to have strong mutual inhibition. Data
shown in panel (B) are for 5,000 randomly selected pairs of cells. Pairs of cells with significantly
negatively overlapping RFs tend not to have inhibitory connections between them, hence the
apparent asymmetry in the RF overlap distribution obtained by marginalizing over connection
strengths in panel (B).
33
|
1607.00952 | 1 | 1607 | 2016-07-04T16:38:16 | Graph analysis of spontaneous brain network using EEG source connectivity | [
"q-bio.NC"
] | Exploring the human brain networks during rest is a topic of great interest. Several structural and functional studies have previously been conducted to study the intrinsic brain networks. In this paper, we focus on investigating the human brain network topology using dense Electroencephalography (EEG) source connectivity approach. We applied graph theoretical methods on functional networks reconstructed from resting state data acquired using EEG in 14 healthy subjects. Our findings confirmed the existence of sets of brain regions considered as functional hubs. In particular, the isthmus cingulate and the orbitofrontal regions reveal high levels of integration. Results also emphasize on the critical role of the default mode network (DMN) in enabling an efficient communication between brain regions. | q-bio.NC | q-bio | Graph analysis of spontaneous brain network using EEG source
Aya Kabbara a,b,c,d*
Wassim El Falou a,b
Mohamad Khalil a,b
Fabrice Wendling c,d
Mahmoud Hassan c,d
connectivity
a Azm research center in biotechnology, EDST, Lebanese University, Lebanon
b CRSI research center, Faculty of engineering, Lebanese University, Lebanon
c Université de Rennes 1, LTSI, F-35000, France
d INSERM, U1099, Rennes, F-35000, France
[email protected]
brain
network
topology
Abstract- Exploring the human brain networks during rest is
a topic of great interest. Several structural and functional
studies have previously been conducted to study the intrinsic
brain networks. In this paper, we focus on investigating the
human
dense
Electroencephalography (EEG) source connectivity approach.
We applied graph theoretical methods on functional networks
reconstructed from resting state data acquired using EEG in 14
healthy subjects. Our findings confirmed the existence of sets
of brain regions considered as ‗functional hubs'. In particular,
the isthmus cingulate and the orbitofrontal regions reveal high
levels of integration. Results also emphasize on the critical
role of the default mode network (DMN) in enabling an
efficient communication between brain regions.
using
Keywords- Brain networks; EEG source connectivity;
resting state; default mode network; hubs
I.
INTRODUCTION
Our brain is a complex network. It consists of distinct
regions which are anatomically and/or functionally connected
[1][2][3]. Over the past decade, several studies were interested
in exploring the human brain functional organization during
rest. In this context, a number of networks termed as ―Resting
State Networks (RSNs)‖ have been revealed and found to be
modalities
consistent
[1][4][5][6][7][3][8][9][10][11]. The networks
that are
frequently reported are the default mode network (DMN), the
dorsal attention network (DAN), the ventral attention network
(VAN), the salience network (SAN), the motor network, the
visual network and the auditory network.
subjects
over
and
as
as well
functional
the medial
region and
The high level of RSNs connectivity suggests the existence
of a set of crucial regions (hubs) particularly important in
providing an efficient communication between brain regions.
idea has been supported by numerous structural
This
[12][13][14][15][16]
studies
[17][18][19][20]. The most common identified regions include
the cingulate
region.
Furthermore, [20] has demonstrated that these ―hubs‖ are
highly interconnected with each other forming a ―rich-club‖
organization of the human brain network.
The RSNs have been explored using different neuroimaging
techniques such as functional magnetic resonance imaging
(fMRI), magnetoencephalography (MEG), positron emission
tomography (PET) [21][22][23][20]. However, exploring the
human brain architecture using electroencephalography (EEG)
recordings has not been well established yet, which is the main
purpose of the presented paper. For this end, we collected
frontal
dense-EEG data from 14 subjects at rest with eyes closed. We
then reconstructed the functional networks using the EEG
source connectivity approach [24]. This step has been
followed by a graph quantification of
the constructed
networks. Our results demonstrate the existence of crucial
nodes located in the cingulate region and in the medial frontal
region confirming the previous results obtained by fMRI and
MEG analyses [15][16][17][18][19][20]. Moreover, the results
insist on the importance of DMN in establishing efficient brain
connectivity, since the majority of the identified central nodes
belong particularly to the DMN.
II. MATERIALS AND METHODS
A. Data acquisition
Fourteen healthy subjects were asked to relax with their
eyes closed without falling asleep, while 10 minutes of EEG
were recorded. For each subject, the individual structural MRI
was acquired in addition to dense EEG (256-channels, EGI,
Electrical Geodesic Inc.). EEGs were sampled at 1000 Hz,
band-pass filtered within 3-45 Hz. The acquisition was
performed following the procedure approved by the National
Ethics Committee for the Protection of Persons (CPP)
(BrainGraph study, agreement number 2014-A01461- 46,
promoter: Rennes University Hospital). All subjects gave an
informed consent prior to their participation.
B. Task, procedure and design
We segmented the EEGs into non-overlapping 40s epochs.
A segment with amplitude ±80μV was considered as
artifactual and rejected after visual inspection.
C. Cortical network reconstruction
The reconstruction of functional networks from scalp
EEGs included three main steps:
C.1) Solving the EEG inverse problem: The EEG signals S(t)
are expressed as linear combinations of time-varying current
dipole sources D(t): S = G. D + B where G and B(t) are
respectively the matrix containing the lead fields of the dipolar
sources and the additive noise. The inverse method aims at
estimating the parameters of the dipolar sources D (t) (notably
the position, orientation and magnitude). Among the methods
available, we chose to use the weighted minimum norm
estimate method (wMNE) implemented in Brainstorm [25].
This method was chosen in the presented work based on a
comparative study reported in [26], where authors have
demonstrated the robustness of wMNE over other methods.
Figure 1. Distribution of the four graph measures on the 68 ROIs .The dotted line illustrated in each histogram presents
the sum of the average value and twice the standard deviation of nodes distribution. A bar is colored if its value is above
the dotted line.
C.2) Identifying the regions of interest (ROIs): We used the
Desikan-Killiany atlas to anatomically segment the brain into
68 cortical regions [27]. Dynamics of sources according to the
same ROI were averaged over time resulting in 68 regional
time series.
C.3) Measuring the functional connectivity: We quantified the
synchronization between the 68 regional time series using the
phase locking value (PLV). PLV has been also chosen based
on the comparative study performed by [26], who concluded
that the wMNE/PLV combination provides the best results
among many possible inverse/connectivity combinations. This
combination was also recently used to track dynamics of
functional brain networks during cognitive task [24]. The final
networks were obtained by applying a proportional threshold
(10%) to remove weak connections from the PLV matrices.
to strength metric, we observe that the ―right isthmus
cingulate‖, ―left/right medial orbitofrontal‖, the ―left lateral
orbitofrontal‖ are the most important nodes. While the right
―frontal pole‖, the ―left/right inferior temporal‖, the ―left
paracentral‖, the ―right parsorbitalis‖ and the ―right superior
parietal‖ were significant with regard to the clustering
coefficient.
The spatial distribution of node centrality, vulnerability,
strength and clustering coefficient on the cortical surface
across all participants is shown in figure 2. The figure shows
that most of the central nodes belong particularly to the DMN.
The same for the vulnerability and the strength figures where
the prefrontal cortex and the cingulate region are also
involved. However, one can notice that the regions that have
the highest clustering coefficient do not belong to the DMN.
D. Graph metrics extraction
We quantified the importance of each node in terms of four
network metrics:
1- Betweenness Centrality:
𝐵𝐶𝑖 =
𝑖,𝑗
𝜎(𝑖, 𝑢, 𝑗)
𝜎(𝑖, 𝑗)
where 𝜎(𝑖, 𝑢, 𝑗) is the number of shortest paths between
nodes 𝑖 and 𝑗 that pass through node 𝑢, 𝜎(𝑖, 𝑗) is the total
number of shortest paths between 𝑖 and 𝑗, and the sum is
over all pairs 𝑖, 𝑗 of distinct nodes.
𝑉𝑖 =
𝐸
2- Vulnerability: The vulnerability of a node can be defined as
the reduction in the efficiency of the network when the
node and all its edges are removed:
𝐸 − 𝐸𝑖
Where E is the global efficiency of the network before any
attach, and 𝐸𝑖 is the global efficiency of the network after
attacking the node 𝑖 [28].
3- Strength:
𝑆𝑖 = 𝑤𝑖𝑗
𝑗
Where 𝑤𝑖𝑗 is the weight of the edge linking the node 𝑖 to
the node 𝑗.
4- Clustering coefficient: The clustering coefficient of a node
in a graph quantifies how close its neighbors are to being a
clique.
III. RESULTS
For each segment of each subject, we extracted nodes with
highest centrality, strength, vulnerability, and clustering
coefficient values. Since the analysis was done for N=14
subjects, a node can be designated as the most critical node (in
terms of centrality, vulnerability, strength and clustering
coefficient) for a number of times varying from 0 to 14. The
four histograms shown in Figure 1 depict the number of times
each of the 68 nodes was considered as the most important
node in terms of centrality, vulnerability, strength and
clustering coefficient. A significant bar is colored if its value
exceeds the sum of the average value and twice the standard
deviation of nodes distribution. As illustrated, the most central
nodes were the ―left/right isthmus cingulate‖ and the ―right
posterior cingulate‖. The most vulnerable nodes were the
―left/right isthmus cingulate‖, the ―left posterior cingulate‖,
the ―right parahippocampal‖, the right ―entorhinal‖. According
Figure 2. Graph measures distribution on the cortical
surface for the 14 subjects.
IV. DISCUSSION
A node has been previously defined as hub if it has an
unusually high strength or centrality, and a low clustering
coefficient [29]. Based on this definition of ―hubness‖, we can
say that the ―isthmus cingulate‖ region plays the role of a hub.
Similar findings on the important role of cingulate gyrus
region have been reported by many structural and functional
analysis based MRI and MEG
[16][17][18][19][20].
Examining our obtained results, we also recognize the
importance of the orbito-frontal region, already considered as
critical [15] [17][18][19]. Moreover, [20] have insisted on the
importance of the posterior cingulate also depicted here as
central and vulnerable node with low clustering coefficient. In
addition, results suggest that the DMN plays a major role in
maintaining an efficient brain communication during rest,
since the majority of the relevant nodes are included in the
DMN. This confirms the fact that the DMN is highly activated
during rest [30][8]. Furthermore, EEG source connectivity
approach could offer a unique insight into the way the brain
network can be dynamically reconfigured and reorganized,
thanks to the excellent time resolution offered by EEG.
Further work will be
the dynamic
characteristics of RSNs and the analysis of how the dynamic
interactions across RSNs are spatially and
temporally
modified.
tracking of
the
V. CONCLUSION
In this paper, we used the dense EEG source connectivity
method to explore the human brain network architecture
during rest. The networks were characterized in terms of
node's centrality, vulnerability, strength and clustering. Our
results confirmed the existence of regions playing the role of
―functional hubs‖, consistent with the state-of-the art findings.
Moreover, we reported
the critical role of nodes that
correspond to the default mode network (DMN). Our findings
highlighted also the capacity of EEG source connectivity
method to reveal the brain network topology during rest.
ACKNOWLEDGMENT
This work was supported by the Rennes University Hospital
(COREC Project named BrainGraph, 2015-17). The work has
also received a French government support granted to the
CominLabs excellence
the
National Research Agency in the ―Investing for the Future‖
program under reference ANR-10-LABX-07-01. It was also
financed by Azm center for research in biotechnology and its
applications. Authors would like to thank Dufor O. for his help
in the data acquisition.
laboratory and managed by
REFERENCES
[1] B. Biswal, F. Z. Yetkin, V. M. Haughton, and J. S. Hyde, ―Functional
connectivity in the motor cortex of resting human brain using echo-
planar MRI,‖ Magn. Reson. Med., vol. 34, pp. 537–541, 1995.
[2] B. B. Biswal, J. Van Kylen, and J. S. Hyde, ―Simultaneous assessment of
flow and BOLD signals in resting-state functional connectivity maps.,‖
NMR Biomed., vol. 10, pp. 165–170, 1997.
[3] M. D. Fox and M. E. Raichle, ―Spontaneous fluctuations in brain activity
observed with functional magnetic resonance imaging,‖ Nat Rev
Neurosci, vol. 8, pp. 700–711, 2007.
[13] L. Harriger, M. P. van den Heuvel, and O. Sporns, ―Rich Club
Organization of Macaque Cerebral Cortex and Its Role in Network
Communication,‖ PLoS One, vol. 7, 2012.
[14] J. D. van Horn, A. Irimia, C. M. Torgerson, M. C. Chambers, R. Kikinis,
and A. W. Toga, ―Mapping connectivity damage in the case of phineas
gage,‖ PLoS One, vol. 7, 2012.
[15] P. Hagmann, L. Cammoun, X. Gigandet, R. Meuli, C. J. Honey, J. Van
Wedeen, and O. Sporns, ―Mapping the structural core of human cerebral
cortex,‖ PLoS Biol., vol. 6, pp. 1479–1493, 2008.
[16] Y. Iturria-Medina, R. C. Sotero, E. J. Canales-Rodr??guez, Y. Alem??n-
G??mez, and L. Melie-Garc??a, ―Studying the human brain anatomical
network via diffusion-weighted MRI and Graph Theory,‖ Neuroimage,
vol. 40, pp. 1064–1076, 2008.
[17] D. Tomasi and N. D. Volkow, ―Association between Functional
Connectivity Hubs and Brain Networks,‖ Cereb. Cortex, vol. 21, pp.
2003–2013, 2011.
[18] D. Tomasi and N. D. Volkow, ―Functional connectivity density
mapping.,‖ Proc. Natl. Acad. Sci. U. S. A., vol. 107, pp. 9885–9890,
2010.
[19] X. N. Zuo, R. Ehmke, M. Mennes, D. Imperati, F. X. Castellanos, O.
Sporns, and M. P. Milham, ―Network centrality in the human functional
connectome,‖ Cereb. Cortex, vol. 22, pp. 1862–1875, 2012.
[20] F. de Pasquale, S. Della Penna, A. Z. Snyder, L. Marzetti, V. Pizzella, G.
L. Romani, and M. Corbetta, ―A Cortical Core for Dynamic Integration
of Functional Networks in the Resting Human Brain,‖ Neuron, vol. 74,
pp. 753–764, 2012.
[21] D. Mantini, M. G. Perrucci, C. Del Gratta, G. L. Romani, and M.
Corbetta, ―Electrophysiological signatures of resting state networks in
the human brain.,‖ Proc. Natl. Acad. Sci. U. S. A., vol. 104, pp. 13170–5,
2007.
[22] M. J. Brookes, J. R. Hale, J. M. Zumer, C. M. Stevenson, S. T. Francis,
G. R. Barnes, J. P. Owen, P. G. Morris, and S. S. Nagarajan, ―Measuring
functional connectivity using MEG: Methodology and comparison with
fcMRI,‖ Neuroimage, vol. 56, pp. 1082–1104, 2011.
[4] D. Cordes, V. M. Haughton, K. Arfanakis, G. J. Wendt, P. A. Turski, C.
H. Moritz, M. A. Quigley, and M. E. Meyerand, ―Mapping functionally
related regions of brain with functional connectivity MR imaging,‖ Am.
J. Neuroradiol., vol. 21, pp. 1636–1644, 2000.
[23] Z. Liu, M. Fukunaga, J. A. de Zwart, and J. H. Duyn, ―Large-scale
spontaneous fluctuations and correlations in brain electrical activity
observed with magnetoencephalography,‖ Neuroimage, vol. 51, pp. 102–
111, 2010.
[5] D. Cordes, V. M. Haughton, K. Arfanakis, J. D. Carew, P. A. Turski, C.
H. Moritz, M. A. Quigley, and M. E. Meyerand, ―Frequencies
contributing to functional connectivity in the cerebral cortex in ‗resting-
state' data,‖ Am. J. Neuroradiol., vol. 22, pp. 1326–1333, 2001.
[6]
J. S. Damoiseaux, S. A. R. B. Rombouts, F. Barkhof, P. Scheltens, C. J.
Stam, S. M. Smith, and C. F. Beckmann, ―Consistent resting-state
networks across healthy subjects.,‖ Proc. Natl. Acad. Sci. U. S. A., vol.
103, pp. 13848–53, 2006.
[7] M. De Luca, S. Smith, N. De Stefano, A. Federico, and P. M. Matthews,
―Blood oxygenation level dependent contrast resting state networks are
relevant to functional activity in the neocortical sensorimotor system,‖
Exp. Brain Res., vol. 167, pp. 587–594, 2005.
[8] M. D. Greicius, B. Krasnow, A. L. Reiss, and V. Menon, ―Functional
connectivity in the resting brain: a network analysis of the default mode
hypothesis.,‖ Proc. Natl. Acad. Sci. U. S. A., vol. 100, pp. 253–8, 2003.
[9] M. J. Lowe, B. J. Mock, and J. A. Sorenson, ―Functional Connectivity in
Imaging Using Resting-State
Single and Multislice Echoplanar
Fluctuations,‖ Neuroimage, vol. 7, pp. 119–132, 1998.
[24] M. Hassan, P. Benquet, A. Biraben, C. Berrou, O. Dufor, and F.
Wendling, ―Dynamic reorganization of functional brain networks during
picture naming,‖ Cortex, vol. 73, pp. 276–288, 2015.
[25] F. Tadel, S. Baillet, J. C. Mosher, D. Pantazis, and R. M. Leahy,
―Brainstorm: A user-friendly application for MEG/EEG analysis,‖
Comput. Intell. Neurosci., vol. 2011, 2011.
[26] M. Hassan, O. Dufor, I. Merlet, C. Berrou, and F. Wendling, ―EEG
source connectivity analysis: From dense array recordings to brain
networks,‖ PLoS One, vol. 9, 2014.
[27] R. S. Desikan, F. S??gonne, B. Fischl, B. T. Quinn, B. C. Dickerson, D.
Blacker, R. L. Buckner, A. M. Dale, R. P. Maguire, B. T. Hyman, M. S.
Albert, and R. J. Killiany, ―An automated labeling system for
subdividing the human cerebral cortex on MRI scans into gyral based
regions of interest,‖ Neuroimage, vol. 31, pp. 968–980, 2006.
[28] V. Gol'dshtein, G. A. Koganov, and G. I. Surdutovich, ―Vulnerability
and Hierarchy of Complex Networks,‖ Physics (College. Park. Md).,
vol. 16, p. 4, 2004.
[29] O. Sporns, C. J. Honey, and R. K??tter, ―Identification and classification
[10] M. van den Heuvel, R. Mandl, and H. H. Pol, ―Normalized cut group
of hubs in brain networks,‖ PLoS One, vol. 2, 2007.
clustering of resting-state fMRI data,‖ PLoS One, vol. 3, 2008.
[11] J. Xiong, L. M. Parsons, J. H. Gao, and P. T. Fox, ―Interregional
connectivity to primary motor cortex revealed using MRI resting state
images.,‖ Hum. Brain Mapp., vol. 8, pp. 151–156, 1999.
[12] M. P. van den Heuvel and O. Sporns, ―Rich-Club Organization of the
Human Connectome,‖ J. Neurosci., vol. 31, pp. 15775–15786, 2011.
[30] M. E. Raichle, A. M. MacLeod, A. Z. Snyder, W. J. Powers, D. A.
Gusnard, and G. L. Shulman, ―A default mode of brain function.,‖
Proceedings of the National Academy of Sciences of the United States of
America, vol. 98, pp. 676–82, 2001.
|
1602.06827 | 1 | 1602 | 2016-02-22T15:44:06 | Lagged and instantaneous dynamical influences related to brain structural connectivity | [
"q-bio.NC"
] | Contemporary neuroimaging methods can shed light on the basis of human neural and cognitive specializations, with important implications for neuroscience and medicine. Different MRI acquisitions provide different brain networks at the macroscale; whilst diffusion-weighted MRI (dMRI) provides a structural connectivity (SC) coincident with the bundles of parallel fibers between brain areas, functional MRI (fMRI) accounts for the variations in the blood-oxygenation-level-dependent T2* signal, providing functional connectivity (FC).Understanding the precise relation between FC and SC, that is, between brain dynamics and structure, is still a challenge for neuroscience. To investigate this problem, we acquired data at rest and built the corresponding SC (with matrix elements corresponding to the fiber number between brain areas) to be compared with FC connectivity matrices obtained by 3 different methods: directed dependencies by an exploratory version of structural equation modeling (eSEM), linear correlations (C) and partial correlations (PC). We also considered the possibility of using lagged correlations in time series; so, we compared a lagged version of eSEM and Granger causality (GC). Our results were two-fold: firstly, eSEM performance in correlating with SC was comparable to those obtained from C and PC, but eSEM (not C nor PC) provides information about directionality of the functional interactions. Second, interactions on a time scale much smaller than the sampling time, captured by instantaneous connectivity methods, are much more related to SC than slow directed influences captured by the lagged analysis. Indeed the performance in correlating with SC was much worse for GC and for the lagged version of eSEM. We expect these results to supply further insights to the interplay between SC and functional patterns, an important issue in the study of brain physiology and function. | q-bio.NC | q-bio | Frontiers in Psychology
Research Article
1 October 2018
Lagged and instantaneous dynamical
influences related to brain structural
connectivity
Carmen Alonso-Montes 1,∗, Ibai Diez 2, Lakhdar Remaki 1, I naki Escudero 2,3,
Beatriz Mateos 2,3, Yves Rosseel 4, Daniele Marinazzo 4, Sebastiano
Stramaglia 1,5, and Jesus M. Cortes,2,6,7
1BCAM - Basque Center for Applied Mathematics, Bilbao, Spain
2 Biocruces Health Research Institute, Cruces University Hospital. Barakaldo,
Spain
3Radiology Service, Cruces University Hospital. Barakaldo, Spain
4Department of Data Analysis, Faculty of Psychological and Pedagogical
Sciences, Ghent University, Belgium
5Dipartimento di Fisica, Universit´a degli Studi di Bari and INFN. Bari, Italy
6Ikerbasque, The Basque Foundation for Science. Bilbao, Spain.
7Department of Cell Biology and Histology. University of the Basque Country.
Leioa, Spain
Correspondence*:
Carmen Alonso-Montes
BCAM - Basque Center for Applied Mathematics, Alameda Mazarredo 14,
E-48009, Bilbao, Spain, [email protected]
6
1
0
2
b
e
F
2
2
]
.
C
N
o
i
b
-
q
[
1
v
7
2
8
6
0
.
2
0
6
1
:
v
i
X
r
a
1
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
ABSTRACT
Contemporary neuroimaging methods can shed light on the basis of human neural and
cognitive specializations, with important implications for neuroscience and medicine. Indeed,
different MRI acquisitions provide different brain networks at the macroscale; whilst diffusion-
weighted MRI (dMRI) provides a structural connectivity (SC) coincident with the bundles
of parallel fibers between brain areas,
functional MRI (fMRI) accounts for the variations
in the blood-oxygenation-level-dependent T2* signal, providing functional connectivity (FC).
Understanding the precise relation between FC and SC, that is, between brain dynamics and
structure, is still a challenge for neuroscience. To investigate this problem, we acquired data at
rest and built the corresponding SC (with matrix elements corresponding to the fiber number
between brain areas) to be compared with FC connectivity matrices obtained by three different
methods: directed dependencies by an exploratory version of structural equation modeling
(eSEM), linear correlations (C) and partial correlations (PC). We also considered the possibility
of using lagged correlations in time series; in particular, we compared a lagged version of eSEM
and Granger causality (GC). Our results were two-fold: firstly, eSEM performance in correlating
with SC was comparable to those obtained from C and PC, but eSEM (not C, nor PC) provides
information about directionality of the functional interactions. Second, interactions on a time
scale much smaller than the sampling time, captured by instantaneous connectivity methods,
are much more related to SC than slow directed influences captured by the lagged analysis.
Indeed the performance in correlating with SC was much worse for GC and for the lagged
version of eSEM. We expect these results to supply further insights to the interplay between SC
and functional patterns, an important issue in the study of brain physiology and function.
* Keywords: structural equation modeling;
functional connectivity; structural connectivity; resting state;
functional magnetic
resonance imaging; tensor diffusion imaging
This is a provisional file, not the final typeset article
2
C. Alonso-Montes et al.
LIST OF ACRONYMS
SC: Structural Connectivity
FC: Functional Connectivity
EC: Effective Connectivity
eSEM vs SC in the resting brain
eSEM: Exploratory Structural Equation Modeling (a non-lagged model)
eSEM1: eSEM with lag=1
eSEM2: eSEM with lag=2
eSEM3: eSEM with lag=3
GC: Granger Causality
GC1: GC with lag=1
GC2: GC with lag=2
GC3: GC with lag=3
C: Correlation
PC: Partial Correlation
RSN: Resting State Network
DMN: Default Mode Network
SM: Sensory Motor
ExC: Executive Control
ROI: Region of Interest
CP: (Structurally) Connected Pairs
NCP: (Structurally) Non-Connected Pairs
Frontiers in Psychology
3
C. Alonso-Montes et al.
1 INTRODUCTION
eSEM vs SC in the resting brain
Three different main classes of brain networks are currently investigated (Sporns et al., 2005;
Bonifazi et al., 2009; Sporns et al., 2004; Friston, 1994, 2011): networks defined by their structural
connectivity (SC) refer to anatomical connections between brain regions; networks defined by their
functional connectivity (FC) account for statistical similarities in the dynamics between distinct neuronal
populations; and effective connectivity (EC) networks identify interactions or information flow between
regions.
Current magnetic resonance imaging (MRI) techniques have allowed SC, FC and EC brain networks
to be measured at the macroscale. Thus, SC networks have been obtained from diffusion tensor images
(DTI) and high-resolution tractography (Craddock et al., 2013) while FC networks have been obtained
from correlations between blood oxygen-level dependent (BOLD) time-series (Biswal et al., 1995).
Different methods can assess EC. One possibility is the dynamic causal modeling, addressing how
the activity in one brain area is affected by the activity in another area using explicit models of neural
populations (Friston et al., 2003; Penny et al., 2004). Other possibilities are data-driven approaches with
no further assumptions about the hemodynamic response, nor about the biophysics of the BOLD signal
from individual neuron to population level. Two popular existing data-driven methods to calculate EC are
Granger causality (GC) (Granger, 1969) and transfer entropy (Schreiber, 2000).
Another well-known method to calculate EC is the structural equation modeling (SEM). Although SEM
assumes an implicit model (ie. an influence matrix) (Bollen, 1989), in the present study we focus on
an exploratory version of SEM (labelled eSEM) where all variables might (a priori) interact with all the
others. Notice that, eSEM is by construction exploratory whilst SEM is largely confirmatory. Both SEM
and eSEM are methods to calculate EC, since both methods provide directed connectivity matrices.
In this paper, we aim to bring some light in a long lasting question: How brain structure is shaped by
its function, and viceversa? Or alternatively, using the language of networks: How are the three classes of
networks SC, FC and EC related to each other? It is important to emphasize that, this challenging problem
has not yet a clear answer for any general brain condition and data set. Here, to address this question, we
will focus here in the resting brain, i.e. when the brain is not performing any goal-oriented task.
Notice that despite the simplicity of the context where these patterns of brain activity are generated,
the resting brain dynamics is complex, encompassing a superposition of multiple resting state networks
(Raichle et al., 2001; Fox et al., 2005; Raichle and Mintun, 2006; Raichle and Snyder,
(RSNs)
This is a provisional file, not the final typeset article
4
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
2007; Raichle, 2009); each RSN underlying a different cognitive function e.g., there are visual networks,
sensory-motor networks, auditory networks, default mode networks, executive control networks, and
some others (for further details see for instance (Beckmann et al., 2005) and references therein).
Pioneering work showed that SC and FC are correlated to some extent (Hagmann et al., 2008; Honey
et al., 2009). After these fundamental papers, some other studies made use of the combined data sets to
address different aspects of brain dynamics (Fraiman et al., 2009; Cabral et al., 2011; Deco et al., 2011;
Haimovici et al., 2013; Marinazzo et al., 2014; Mess´e et al., 2014; Goni et al., 2013; Kolchinsky et al.,
2014). In this paper, both structural and functional data have been used to demonstrate to which extent the
EC obtained by eSEM and the FC obtained by C and PC are similar to SC.
Previous approaches analyzed fMRI data based on SEM (Bullmore et al., 2000; Schlosser et al., 2006;
Kim et al., 2007; Gates et al., 2010, 2011), dealing with subsets of candidate regions selected on the basis
of prior knowledge. However, the performance of these approaches depends strongly on the correctness
and completeness of the hypothesized model of connections. In the present work, eSEM is applied in
an exploratory fashion to a multivariate dataset corresponding to a specific brain system consisting in
15 different regions of interest (ROIs), fully covering (with no further assumptions about the underlying
connectivity) three of the well-known RSNs (Beckmann et al., 2005): The sensory-motor network (SM),
the executive-control network (ExC) and the default mode network (DMN). The application of eSEM
returns an influence matrix which is not symmetric (ie., a region A can influence B differently than how
B influences A) and describes fully connected directed dependencies between ROIs.
The performance achieved by eSEM in correlating with SC (thus, measusing the similarity between
eSEM and SC) is also compared with FC, obtained by two other methods: the linear correlation (C) and
partial correlation (PC). Unlike C, PC is commonly used to analyze direct relationships among fMRI time
series with good performance (Marrelec et al., 2006; Marrelec and Benali, 2009; Maki-Marttunen
et al., 2013), since network influences beyond the specific pair are removed.
Furthermore, eSEM is also applied to lagged time series to estimate a saturated, fully connected, but
recursive model. Notice that bi-directional influences here are detected as cross-lagged effects. The results
from this lagged version of eSEM are compared with those from GC. As a result, we will show that lagged
methods are less related to SC measures, which implies that the dependencies found in the data on slower
time scales (in comparison to instantaneous interactions) are less related to SC.
Frontiers in Psychology
5
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
2 MATERIAL & METHODS
2.1 SAME-SUBJECT STRUCTURE-FUNCTION ACQUISITIONS
This work was approved by the Ethics Committee at the Cruces University Hospital; all the methods were
carried out in accordance to approved guidelines. A population of n=12 (6 males) healthy subjects, aged
between 24 and 46 (33.5 ± 8.7), provided information consents before the imaging session. For all the
participants, we acquired same-subject structure-function data with a Philips Achieva 1.5T Nova scanner.
The total scan time for each session was less than 30 minutes and high-resolution anatomical MRI was
acquired using a T1-weighted 3D sequence with the following parameters: TR = 7.482 ms, TE = 3.425 ms;
parallel imaging (SENSE) acceleration factor=1.5; acquisition matrix size=256x256; FOV=26 cm; slice
thickness=1.1 mm; 170 contiguous sections. Diffusion weighted images (DWIs) were acquired using
pulsed gradient-spin-echo echo-planar-imaging (PGSE-EPI) under the following parameters: 32 gradient
directions, TR = 11070.28 ms, TE = 107.04 ms, 60 slices with thickness of 2 mm, no gap between slices,
128x128 matrix with an FOV of 23x23 cm. Changes in blood-oxygenation-level-dependent (BOLD) T2*
signals were measured using an interleaved gradient-echo EPI sequence. The subjects lay quietly for 7.28
minutes, during which 200 whole brain volumes were obtained under the following parameters: TR =
2200 ms, TE = 35 ms; Flip Angle 90, 24 cm field of view, 128x128 pixel matrix, and 3.12 x 3.19 x 4.00
mm voxel dimensions.
We have shown in (Diez et al., 2015a) that the relationship between SC and FC found with the data used
in this study is confirmed by the MGH-USC Human Connectome Project, of much higher quality. The
results we show here open the possibility to a generalization to many other data sets.
2.2 DATA PREPROCESSING
2.2.1
Structural data: To analyze the diffusion images (dMRI), the eddy current correction was
applied to overcome artifacts produced by changes in the gradient field directions of the MR scanner
and subject head movement. In particular, the eddy-correct tool from FSL was used to correct both eddy
current distortions, and simple head motion, using affine registration to a reference volume. After this,
DTIFIT was used to perform the fitting of the diffusion tensor for each voxel, using as an input the
eddy-correct output. No extra de-noising was applied in the data and our results were not wrapped to
any template. Two computations were performed to transform the atlas to each individual space: (1) the
transformation between the MNI template to the subject structural image (T1), and (2) the transformation
This is a provisional file, not the final typeset article
6
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
between the T1 to the diffusion image space. Combining both transformations, each atlas region is
transformed to the diffusion space, allowing to count the number of fibers connecting all ROIs pairs.
Using the corrected data, a local fitting of the diffusion tensor was applied to compute the diffusion tensor
model at each voxel. Then, a deterministic tractography algorithm (FACT) (Mori et al., 1999) was applied
using TrackVis (Wang et al., 2007), an interactive software for fiber tracking.
2.2.2 Functional data: The functional MRI (fMRI) data was preprocessed with FSL (FMRIB Software
Library v5.0). The first 10 volumes were discarded for correction of the magnetic saturation effect and
for the remaining volumes, first the movement is corrected and then, the slice-time is also corrected for
temporal alignment. All voxels were spatially smoothed with a 6mm FWHM isotropic Gaussian kernel
and after intensity normalization, a band pass filter was applied between 0.01 and 0.08 Hz (Cordes et al.,
2001). Finally, linear and quadratic trends were removed. We next regressed out the motion time courses,
the average CSF signal, the average white matter signal and the average global signal. Finally, fMRI data
was transformed to the MNI152 template, such that a given voxel had a volume of 3mm*3mm*3mm.
It is important to emphasize that to remove or not the average global signal in FC studies is currently
a controversial issue (Saad et al., 2012); see also http://rfmri.org/GSRDiscussion. Here,
following most of the studies addressing brain FC, we have applied the global signal removal; but the
situation of not applying the global signal removal has been also explored (figure S2).
2.2.3 HRF blind deconvolution:
In order to eliminate the confounding effect of HRF on temporal
precedence, we individuated point processes corresponding to signal fluctuations with a given signature
and extracted a voxel-specific HRF to be used for deconvolution, after following an alignment procedure.
The parameters for blind deconvolution were chosen with a physiological meaning according to (Wu
et al., 2013): for a TR equal to 2.2s, the threshold was fixed to 1 SD (standard deviation) and the maximum
time lag was fixed to 5 TR (for further details on the complete HRF blind deconvolution method and
the different parameters to be used, see (Wu et al., 2013)). The resulting time-series, after HRF blind
deconvolution, are the ones used for the calculation of EC and FC.
2.3 ROIS EXTRACTION
Regions of interest (ROIs) were defined by using the masks of the resting state networks (RSNs) reported
in (Beckmann et al., 2005), which can be downloaded from http://www.fmrib.ox.ac.uk/
Frontiers in Psychology
7
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
analysis/royalsoc8/. Note that, we are not dealing with the independent components per se,
but with the voxels time-series localized within the masks. Similar approaches based on the RSNs
masks to define ROIs have been widely used before (Tagliazucchi et al., 2012; Haimovici et al., 2013;
Carhart-Harris et al., 2014; Tagliazucchi et al., 2014; Diez et al., 2015b).
Specifically, the following three RSNs were selected: the default mode network (DMN), the executive
control (ExC) network and the sensory motor (SM) network. Next, these three networks were manually
subdivided in distinct spatially contiguous regions (see figure 1). For each region, a region growing
segmentation method was applied by manually selecting a seed region, thus obtaining a total of 15
different ROIs: 1 SM region, 6 DMNs and 8 ExCs regions. In particular, the "island effect" method
incorporated in 3D Slicer (http://www.slicer.org) was applied, which selects all the voxels of
the contiguous region given an initial seed. Visual representations of all ROIs are given in figure 1 and
their sizes in table 1.
2.4 CALCULATION OF STRUCTURAL, FUNCTIONAL AND EFFECTIVE CONNECTIVITY
MATRICES
Structural connectivity (SC): Matrices were obtained per each subject by counting the number
2.4.1
of fibers connecting two ROIs (that is, starting in one ROI and finalizing in another) for each individual
pair; thus, for a number of 15 ROIs, it gave 105 different values.
2.4.2
Functional connectivity (FC): Matrices were calculated by applying to the rs-fMRI time series
two methods: the linear correlation coefficient (C) and the partial correlation analysis (PC). Here, C was
calculated by using the corr function from Matlab (MathWorks Inc., Natick, MA). Assuming C to be a
non-singular matrix, the elements of the PC matrix satisfy that PCij ∝ (C−1)ij, so they are proportional
to the elements of the so-called precision matrix (Maki-Marttunen et al., 2013). Here, PC was computed
using the partialcorr function from Matlab (MathWorks Inc., Natick, MA). Thus, PC is an extension of
C to calculate direct interactions between pairs, as it achieves to remove for a given pair the correlation
contribution from other pairs.
2.4.3 Effective connectivity (EC) by the exploratory structural equation modeling (eSEM). This refers
to a statistical technique aiming to estimate Granger-causal relationships based on quantitative and
qualitative causal information, by means of linear regression-based models. Unlike regression, SEM is
This is a provisional file, not the final typeset article
8
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
formulated as a confirmatory model rather than a predictive model. Being interested in the description
of the directed dependencies between the 15 ROIs, avoiding any prior hypothesis on the connectivity
pattern, we here applied multiple regressions among all the variables. Therefore, our analysis by SEM
has neither structural model nor a measurement model, and provides a fully connected estimate of the
directed dependencies among all the pairs of ROIs. This exploratory analysis is referred as eSEM. This
model does not use temporal correlations in the data and it is applied to non-lagged time series.
To estimate the model parameters of eSEM, a standard maximum likelihood estimation was used using
the lavaan package in R (R Core Team, 2014; Rosseel, 2012). Notice that, this is justified since for
saturated linear models, the maximum likelihood estimates are identical to least squares estimates.
In a second part of this study, eSEM was also applied to lagged time series to estimate a saturated, fully-
connected, but recursive model. Notice that lagged eSEM is recursive but the non-lagged eSEM is not. The
observed variables are the time series for the 15 ROIs augmented with lagged versions of the same time
series. For eSEM1, only the time series accounting for lag=1 were added, resulting in 30 variables in total;
for eSEM2, both lag =1 and lag =2 time series were added, resulting in 45 observed variables in total;
finally, for eSEM3, lag=1, lag=2 and lag=3 time series were added, resulting in 60 observed variables
in total. Three types of parameters were included in the model: (1) all autoregressive regressions within
each ROI to take into account the time-dependencies; (2) all possible cross-lagged regressions between
the ROIs; (3) (residual) covariances for all other pairwise relations that were not included in the set of
regressions (for example, all contemporaneous connections). Importantly, contemporaneous regressions
between ROIs at the same time point were not included. Moreover, to estimate the model parameters of
eSEM1, eSEM2 and eSEM3, standard maximum likelihood estimation was used using the lavaan package
(Rosseel, 2012).
After estimation of all model parameters, an influence matrix was computed as follows: For each pair,
the evidence for this particular (directed) connection was collected. For eSEM1, this was simply the
regression coefficient corresponding to the cross-lagged effect of one ROI on another (controlling for
both auto-regressive effects and cross-lagged effects of other ROIs). That is, the effect of a ROI on the
previous time point on a target ROI at the current time point. For eSEM2, this was a function (here, the
product) of two regression coefficients: one for the effect of a ROI on the previous time point on the target
ROI at the current time point (just like eSEM1), and one for the effect of a ROI measured two time points
towards the target ROI at the current time point. This was done for all possible pairs, averaging all ROIs
of the influence matrix except for the diagonal, which was kept at zero.
Frontiers in Psychology
9
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Notice that the cross-lagged evidence is only used to determine the directed influence of one ROI on
another, while controlling for both auto-regressive effects and the cross-lagged effects of other ROIs. In
fact, the regression coefficients computed by eSEM1, eSEM2 and eSEM3 are identical to those that would
be computed when Granger causality (GC1, GC2, GC3, of order 1, 2 and 3 respectively) is employed
(Granger, 1969); see also suppl. material for further details. But instead of computing an F -statistics for
each pairwise connection as GC does, here, we use the product regression coefficient(s) to average the
influence matrix.
2.5 STATISTICAL ANALYSIS
The values of the average matrices across subjects eSEM, C and PC were compared into two groups:
values associated to structurally connected pairs (CP), meaning that two ROIs are connected with a non-
zero fiber number, and those ones associated to non-connected pairs (NCP), i.e., zero fibers existed
between the two ROIs. A one-way ANOVA test was performed using the MATLAB function anova1
(MathWorks Inc., Natick, MA) between CP and NCP (statistical significance is considered to have a p-
value < 0.01). Thus, small p-values show that the connectivity matrices calculated on the two groups CP
and NCP have a different mean, ie., they are different from each another which indicates that a given
method can separate connected pairs from non-connected ones. The same analysis was also applied to
eSEM1, eSEM2, eSEM3, GC1, GC2, GC3.
3 RESULTS
Firstly, the three different resting networks SM, DMN and ExC were selected. Next, the three networks
were divided in a total number of 15 different ROIs (see methods and figure 1 for details).
Next, the average across subjects SC matrix (figure 2A1) was computed by averaging the fiber number
between pairs of ROIs. Notice that SC is a matrix with many near-zero values. So, it is represented in
logarithmic scale just to improve visualization, but all the analyses were performed using the original SC
matrix.
Next, three connectivity matrices were calculated for each subject from the rs-fMRI time series: eSEM,
C and PC (details in Methods). Next, an average matrix across subjects was calculated for all matrices.
The values of eSEM, C and PC after normalization in the range [0,1] are represented in figures 2A2-A4.
Notice that, unlike C and PC, eSEM provides a non-symmetrical connectivity matrix.
This is a provisional file, not the final typeset article
10
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
To address the similarity between these matrices, and following previous work (Hagmann et al., 2008;
Honey et al., 2009), the Pearson's correlation between the SC entries (vector-wise using all matrix
elements) and the corresponding ones for eSEM, C and PC was computed. The three connectivity matrices
increased their similarity (based on correlation) with SC on connected pairs, pairs connected with non-zero
fibers between ROIs, compared to the situation when all pairs were used for the correlation calculation,
ie., values in figure 2B2 are bigger than in figure 2B1. The same results also hold when Spearman's
correlations were calculated (figure S1). It is important to emphasize that these results did not depend on
the effect of removing or not the global signal to the time series data. Indeed, similar results than in figures
2B1 and 2B2 were obtained without global signal removal (figure S2). Thus, after this simple analysis, we
show that the three measures (eSEM, C and PC) are dependent on SC.
We next investigated whether average values of eSEM, C and PC had significant differences between
CP and NCP (non-connected pairs). The three connectivity matrices showed bigger (significant) values
on CP compared to NCP (figure 2C), thus indicating that the three methods eSEM, C and PC separated
the groups of structurally connected links from those which were not connected. Moreover, PC performed
better than eSEM, whilst eSEM and C performed approximately equal.
Next, we addressed the effect that lagged interactions had on eSEM. Thus, when calculating eSEM
on lagged-time series, eSEM could not distinguish (ie., the p-value between the two groups was high)
between CP and NCP (see figure 5). And this occurred independently on using eSEM or a different
model accounting for lagged interactions, here, the method of multivariate GC was used (figure 5). These
results indicated that instantaneous measures of interactions (ie., approaches dealing only with equal-time
correlations) are better shaped by SC in comparison to algorithms using temporal information (and this
was observed both using eSEM and GC).
For a further analysis we looked at the values of SC, eSEM, C and PC on three specific links: the ones
with a highest value in each SC, FC and EC:
• The structural link, the pair of ROIs sharing the highest value of SC, which was ExC1-DMN2 (x-label
colored in magenta in figure 3).
• The functional link, the pair of ROIs with highest value of C, which was coincident with the pair with
maximum PC, that was ExC2-DMN5 (x-label colored in green).
• The effective link, the pair of ROIs with highest value of SEM: ExC6-DMN6 (x-label colored in
black).
Frontiers in Psychology
11
C. Alonso-Montes et al.
Table 1. ROI size (mm3).
eSEM vs SC in the resting brain
Network
ROIs
size (mm3)
Sensory Motor (SM) Network
SM
194.960
Default Mode Network (DMN) DMN1
DMN2
DMN3
DMN4
DMN5
DMN6
Executive Control (ExC)
ExC1
ExC2
ExC3
ExC4
ExC5
ExC6
ExC7
ExC8
97.091
44.115
28.374
22.330
8.343
10.826
79.956
43.313
52.225
48.483
15.769
15.745
11.723
32.752
From the structural link, and although the average value of eSEM performed similarly to C (figure 2B),
eSEM gave a significantly smaller value than C and PC, reflecting high relation between ExC1 and
DMN2 due to SC. By looking at the functional link, eSEM also provided a high value, indicating that
the two areas with neuronal activity most statistical similar each other, ExC2 and DMN5, also had a high
directed influence between them. Finally, results on the effective link showed that the link with the highest
dynamical influence, from DMN6 to ExC6, also had a high value of C and PC.
Beyond results at the level of individual links, scatter plots between the different connectivity matrices
(SC, eSEM, C and PC) for all the pairs are shown in figure 4. The matrices resulting from eSEM, C
and PC were significantly correlated with the structural one, SC (rounded green rectangles in figure 4).
Correlation coefficients were 0.44, 0.43 and 0.50 for respectively eSEM, C and PC.
We also found that eSEM matrix was highly correlated with C and PC matrices for both CP and NCP
(rounded red circles); indeed, for CP the correlation was equal to 0.86 (for C) and 0.93 (for PC). Thus,
on CP pairs, PC and eSEM were approximately equivalent to each other. When looking to NCP, this
correlation between eSEM and PC went down to 0.88, still a very high value.
Finally, correlation between C and PC matrices were high for both CP (corr=0.86) and NCP (corr=0.63).
This is represented by the rounded blue rectangles in figure 4.
This is a provisional file, not the final typeset article
12
C. Alonso-Montes et al.
4 DISCUSSION
eSEM vs SC in the resting brain
Multiple evidence have shown brain topology (ie., structure) supporting dynamics (ie., function) and
brain dynamics reinforcing structure via synaptic plasticity (or punishing it via synaptic prunning), but
the precise relationship between the two (structure and function) is still challenging (Damoiseaux and
Greicius, 2015; Park and Friston, 2013).
A powerful method to approach this problem at the large-scale brain organization is to calculate
structural and functional networks and address their mutual relationships (Park and Friston, 2013).
Following this strategy, here, we calculated SC, FC and EC for a very specific brain parcellation,
with ROIs covering the entirety of three well-know resting networks, the executive control, the default
mode and the sensory motor network. After this brain division, we obtained 15 different ROIs and by
performing to the same subject two classes of MRI acquisitions (one structural, one functional) we made
a careful comparison between SC (ie. fiber number connectivity between ROIs), FC (pairwise C and PC
connectivities) and EC (by generalizing SEM to its exploratory version eSEM).
We have made use of eSEM for the inference of functional integration; eSEM, although rooted in the
SEM framework, is exploratory and can assess influences between brain regions without assuming any
implicit model. We have studied how much similar eSEM was to SC, an compared these results with
equal-time correlational analysis by calculating both C and PC, which are the leading methods to estimate
FC.
In the first part of this study, our results showed that eSEM, in addition to C and PC, were able to
significantly separate the set of non-connected pairs in the structural network from the set of connected
pairs. Although the PC analysis is slightly the best one in correlating with the strength of structural links,
interestingly, for the specific situation of restricting to connected pairs, the eSEM estimation was almost
identical to PC (correlation value of 0.93). The fact that eSEM provided a similar correlation with SC to
the one achieved by C and PC makes the use of eSEM equally valid as C and PC for FC brain studies.
On the other hand it must be stressed that eSEM also provided information about the case of fiber
pairs where information preferably flowed in one direction. These results showed the usefulness of fully
connected eSEM inference of directed dependencies between structurally connected ROIs in the human
brain.
It is important to emphasize that there are other studies also relating SEM with C and/or PC. Thus, it
was shown that PC performed better than SEM in identifying local patterns of interaction detected by SC
Frontiers in Psychology
13
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
(Marrelec and Benali, 2009) and that C and PC were suitable candidates to simultaneously analyse SC
and FC in the entire brain (Horn et al., 2013); furthermore, this evidence was even stronger when focused
on the Default Mode Network, an important RSN with important implications in memory performance.
In another study, when SEM was used in combination with DTI data (Voineskos et al., 2012), the authors
approached aging and cognitive performance using SC to analyse tract degeneration and SEM to address
white matter tract integrity.
In the second part of our study, we have applied eSEM and multivariate Granger Causality to show that,
when lagged time series are considered to estimate EC, the results are much less correlated with SC (figure
5). This suggests that fast interactions (captured by instantaneous measures of connectivity) are shaped by
the structural strength, whilst slower directed functional interactions (those captured by methods relying
on temporal correlations) are less shaped by the structural strength. In other words, at slow time scales,
the statistical dependencies among ROIs appear to be less related to the details of the underlying structural
connectivity.
The fact that the lagged methods found influences between brain regions acting at a time scale equal to
the sampling time suggests that the lagged algorithms may be seen as complementary to the standard
correlational analysis. The eSEM method, here described,
is suitable tool to detect those directed
functional interactions which cannot be described merely to the presence of a strong structural connection
between brain areas.
To summarize, based on the evidence that RSNs are functionally integrated by structural connections
(van den Heuvel and Sporns, 2013) here, by building a very simple large-scale brain system consisting
of three of those RSNs, and without assuming any implicit connectivity between them, we have shown that
eSEM can perform equally well than C and PC in correlating with SC, thus encouraging the use of eSEM
for FC studies at rest. Whether this statement still holds during task paradigms needs to be investigated.
DISCLOSURE/CONFLICT-OF-INTEREST STATEMENT
The authors declare that the research was conducted in the absence of any commercial or financial
relationships that could be construed as a potential conflict of interest.
This is a provisional file, not the final typeset article
14
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
AUTHOR CONTRIBUTIONS
IE and BM performed MRI acquisitions; IE preprocessed and postprocessed the MRI data; CAM
performed all simulations; CAM, ID, LR, YR, DM, SS and JMC developed the methods; CAM, ID,
LR, YR, DM, SS and JMC wrote the paper; DM, SS and JMC designed the research; all the authors
reviewed the manuscript.
ACKNOWLEDGEMENTS
This work has been conducted within the Computational Neuroimaging Lab in the Biocruces Institute,
and using the computing resources provided by BCAM.
Funding: Financial support from the Basque Government (BERC 2014-2017) and the Spanish Ministry
of Economy and Competitiveness (MINECO) : BCAM Severo Ochoa accreditation (SEV-2013-0323) to
CAM and LR; from Ikerbasque: The Basque Foundation for Science and Euskampus at UPV/EHU to
JMC; from Bizkaia Talent (AYD-000-285) to SS.
DATA AVAILABILITY
The SC matrices for each subject and the time series rs-fMRI data are available upon the reader's request.
SUPPLEMENTARY MATERIAL
EFFECTIVE CONNECTIVITY (EC) BY MULTIVARIATE GRANGER CAUSALITY (GC).
Let us first describe bivariate Granger causality (Granger, 1969). Suppose we model the temporal
dynamics of a stationary time series {ξn}n=1,.,N +m by an autoregressive model of order m:
m(cid:88)
j=1
ξn =
Aj ξn−j + En,
and by a bivariate autoregressive model which takes into account also a simultaneously recorded time
series {ηn}n=1,.,N +m:
m(cid:88)
ξn =
m(cid:88)
A(cid:48)
j ξn−j +
Bj ηn−j + E(cid:48)
n.
Frontiers in Psychology
15
j=1
j=1
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
The coefficients of the models are calculated by standard least squares regression; m is usually selected
according to the Akaike criterion applied to the VAR modeling of the multivariate time series, providing
an optimal order of the model (Akaike, 1974).
It can be said that η Granger-causes ξ if the variance of residuals E(cid:48) is significantly smaller than the
variance of residuals E, as it happens when coefficients Bj are jointly significantly different from zero.
This can be tested by performing an either F-test or Levene's test for equality of variances (Geweke,
1982). An index measuring the strength of the causal interaction is δ = 1 − (cid:104)E(cid:48)2(cid:105)
(cid:104)E2(cid:105) , where (cid:104)·(cid:105) means
averaging over n (note that (cid:104)E(cid:105) = (cid:104)E(cid:48)(cid:105) = 0). Exchanging the roles of the two time series, one may
equally test causality in the opposite direction, i.e. to check whether ξ Granger-causes η.
n}n=1,.,N +m, a = 1, . . . , M, be M other simultaneously recorded time
In the conditioned case, let {ψa
series. When several variables are present in the system, the Granger influence η → ξ must take into
account their possible conditioning effect. In this case, it is recommended to treat the data-set as a whole,
including the ψ times series in both the autoregressive models for ξ described above. To assess causality
in GC, another VAR is learned from data excluding one variable (the candidate driver) from the input set
of variables. Then, an F-test is applied to assess significance of the variance reduction due to the inclusion
of the candidate driver variable. The conditioned Granger causality η → ξ measures the reduction in
the variance of residuals going from one to other of the following two conditions: (i) all variables ψ
are included in the model and (ii) all variables ψ and the variable η are included. Conditioning on the
remaining variables allows to discard indirect interactions that would be recognized as direct by the
pairwise approach. We refer the reader to (Stramaglia et al., 2014) for a discussion on advantages and
pitfalls of pairwise and conditioned Granger causality. In this paper we will refer to GC1, GC2, GC3 when
discussing the application of GC with m=1,2,3 respectively.
REFERENCES
Akaike, H. (1974), A new look at the statistical model identification, IEEE Trans Autom Control, 19,
716 -- 723
Beckmann, C. F., DeLuca, M., Devlin, J. T., and Smith, S. M. (2005), Investigations into resting-state
connectivity using independent component analysis., Philos. Trans. R. Soc. Lond. B. Biol. Sci., 360,
1457, 1001 -- 13
This is a provisional file, not the final typeset article
16
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Biswal, B., Zerrin Yetkin, F., Haughton, V. M., and Hyde, J. S. (1995), Functional connectivity in the
motor cortex of resting human brain using echo-planar mri, Magn. Reson. Med., 34, 4, 537 -- 541
Bollen, K. (1989), Structural Equations with Latent Variables (Hoboken: John Wiley & Sons)
Bonifazi, P., Goldin, M., Picardo, M. A., Jorquera, I., Cattani, A., Bianconi, G., et al. (2009), GABAergic
hub neurons orchestrate synchrony in developing hippocampal networks., Science, 326, 5958, 1419 -- 24
Bullmore, E., Horwitz, B., Honey, G., Brammer, M., Williams, S., and T., S. (2000), How good is good
enough in path analysis of fMRI data?., Neuroimage, , 11, 289 -- 301
Cabral, J., Hugues, E., Sporns, O., and Deco, G. (2011), Role of local network oscillations in resting-state
functional connectivity., Neuroimage, 57, 1, 130 -- 9
Carhart-Harris, R., Leech, R., Hellyer, P., Shanahan, M., Feilding, A., Tagliazucchi, E., et al. (2014),
The entropic brain: A theory of conscious states informed by neuroimaging research with psychedelic
drugs, Front. Hum. Neurosci., 8, 20
Cordes, D., Haughton, V. M., Arfanakis, K., Carew, J. D., Turski, P. A., Moritz, C. H., et al. (2001),
Frequencies contributing to functional connectivity in the cerebral cortex in resting-state data, American
Journal of Neuroradiology, 22, 7, 1326 -- 1333
Craddock, R. C., Jbabdi, S., Yan, C.-G., Vogelstein, J. T., Castellanos, F. X., Di Martino, A., et al. (2013),
Imaging human connectomes at the macroscale., Nat. Methods, 10, 6, 524 -- 39
Damoiseaux, J. and Greicius, M. (2015), Greater than the sum of its parts: a review of studies combining
structural connectivity and resting-state functional connectivity, Brain Struct. Funct., 213, 525 -- 533
Deco, G., Jirsa, V. K., and McIntosh, A. R. (2011), Emerging concepts for the dynamical organization of
resting-state activity in the brain., Nat. Rev. Neurosci., 12, 1, 43 -- 56
Diez, I., Bonifazi, P., Escudero, I., Mateos, B., Munoz, M., Stramaglia, S., et al. (2015a), A novel brain
partition highlights the modular skeleton shared by structure and function, Sci. Rep., arXiv:1410.7959
Diez, I., Erramuzpe, A., Escudero, I., Mateos, B., Marinazzo, D., Sanz-Arigita, E., et al. (2015b),
Information flow between resting state networks, Brain Connect., arXiv:1505.03560
Fox, M. D., Snyder, A. Z., Vincent, J. L., Corbetta, M., Van Essen, D. C., and Raichle, M. E. (2005), The
human brain is intrinsically organized into dynamic, anticorrelated functional networks., Proc. Natl.
Acad. Sci. U. S. A., 102, 27, 9673 -- 8
Fraiman, D., Balenzuela, P., Foss, J., and Chialvo, D. R. (2009), Ising-like dynamics in large-scale
functional brain networks., Phys. Rev. E. Stat. Nonlin. Soft Matter Phys., 79, 6 Pt 1, 061922
Friston, K. J. (1994), Functional and effective connectivity in neuroimaging: A synthesis, Hum. Brain
Mapp., 2, 1-2, 56 -- 78
Frontiers in Psychology
17
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Friston, K. J. (2011), Functional and effective connectivity: a review., Brain Connect., 1, 1, 13 -- 36
Friston, K. J., Harrison, L., and Penny, W. (2003), Dynamic causal modelling., Neuroimage, 19, 4,
1273 -- 302
Gates, K. M., Molenaar, P. C. M., Hillary, F. G., Ram, N., and Rovine, M. J. (2010), Automatic search
for fMRI connectivity mapping: an alternative to Granger causality testing using formal equivalences
among SEM path modeling, VAR, and unified SEM., Neuroimage, 50, 3, 1118 -- 25
Gates, K. M., Molenaar, P. C. M., Hillary, F. G., and Slobounov, S. (2011), Extended unified SEM
approach for modeling event-related fMRI data., Neuroimage, 54, 2, 1151 -- 8
Geweke, J. (1982), Measurement of linear dependence and feedback between multiple time series, J Am
Stat Assoc, 77, 304 -- 313
Goni, J., van den Heuvel, M. P., Avena-Koenigsberger, A., Velez de Mendizabal, N., Betzel, R. F., Griffa,
A., et al. (2013), Resting-brain functional connectivity predicted by analytic measures of network
communication, Proc. Natl. Acad. Sci., 111, 2, 833 -- 838
Granger, C. W. (1969), Investigating causal relations by econometric models and cross-spectral methods,
Econometrica: Journal of the Econometric Society, 424 -- 438
Hagmann, P., Cammoun, L., Gigandet, X., Meuli, R., Hone, C., Wedeen, V. J., et al. (2008), Mapping the
structural core of human cerebral cortex, PLoS Biol, 6, e159
Haimovici, A., Tagliazucchi, E., Balenzuela, P., and Chialvo, D. (2013), Brain Organization into Resting
State Networks Emerges at Criticality on a Model of the Human Connectome, Phys. Rev. Lett., 110,
178101
Honey, C., Sporns, O., Cammoun, L., Gigandet, X., Thiran, J., Meuli, R., et al. (2009), Predicting human
resting-state functional connectivity from structural connectivityppro, Proc Natl Acad Sci USA, 106,
2035 -- 2040
Horn, A., Ostwald, D., Reisert, M., and Blankenburg, F. (2013), The structural-functional connectome
and the default mode network of the human brain., Neuroimage, 102, 142 -- 151
Kim, J., Zhu, W., Chang, J., Bentler, P., and Ernst, T. (2007), Unified structural equation modeling
approach for the analysis of multisubject, multivariate functional MRI data., Human Brain Mapping, ,
28, 85 -- 93
Kolchinsky, A., van den Heuvel, M. P., Griffa, A., Hagmann, P., Rocha, L. M., Sporns, O., et al. (2014),
Multi-scale integration and predictability in resting state brain activity, Front. Neuroinform., 8, 66
This is a provisional file, not the final typeset article
18
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Maki-Marttunen, V., Diez, I., Cortes, J. M., Chialvo, D. R., and Villarreal, M. (2013), Disruption
of transfer entropy and inter-hemispheric brain functional connectivity in patients with disorder of
consciousness, Frontiers in Neuroinformatics, 7, 24
Marinazzo, D., Pellicoro, M., Wu, G., Angelini, L., Cort´es, J. M., and Stramaglia, S. (2014), Information
transfer and criticality in the Ising model on the human connectome., PLoS One, 9, 4, e93616
Marrelec, G. and Benali, H. (2009), A theoretical investigation of the relationship between structural
equation modeling and partial correlation in functional MRI effective connectivity., Comput. Intell.
Neurosci., 369341
Marrelec, G., Krainik, A., Duffau, H., P´el´egrini-Issac, M., Leh´ericy, S., Doyon, J., et al. (2006), Partial
correlation for functional brain interactivity investigation in functional MRI., Neuroimage, 32, 1, 228 --
37
Mess´e, A., Rudrauf, D., Benali, H., and Marrelec, G. (2014), Relating structure and function in the human
brain: relative contributions of anatomy, stationary dynamics, and non-stationarities., PLoS Comput.
Biol., 10, 3, e1003530
Mori, S., Crain, B. J., Chacko, V. P., and van Zijl, P. C. (1999), Three-dimensional tracking of axonal
projections in the brain by magnetic resonance imaging., Ann. Neurol., 45, 2, 265 -- 9
Park, H. and Friston, K. (2013), Structural and functional brain networks: from connections to cognition,
Science, 342, 1238411
Penny, W. D., Stephan, K. E., Mechelli, A., and Friston, K. J. (2004), Modelling functional integration: a
comparison of structural equation and dynamic causal models., Neuroimage, 23 Suppl 1, S264 -- 74
R Core Team (2014), R: A Language and Environment for Statistical Computing, R Foundation for
Statistical Computing, Vienna, Austria
Raichle, M. E. (2009), A paradigm shift in functional brain imaging., J. Neurosci., 29, 41, 12729 -- 34
Raichle, M. E., MacLeod, A. M., Snyder, A. Z., Powers, W. J., Gusnard, D. A., and Shulman, G. L.
(2001), A default mode of brain function., Proc. Natl. Acad. Sci. U. S. A., 98, 2, 676 -- 82
Raichle, M. E. and Mintun, M. A. (2006), Brain work and brain imaging., Annu. Rev. Neurosci., 29,
449 -- 76
Raichle, M. E. and Snyder, A. Z. (2007), A default mode of brain function: a brief history of an evolving
idea., Neuroimage, 37, 4, 1083 -- 90; discussion 1097 -- 9
Rosseel, Y. (2012), lavaan: An R Package for Structural Equation Modeling, J. Stat. Softw., 48, 2, 1 -- 36
Saad, Z., Gotts, S., Murphy, K., Chen, G., Jo, H., Martin, A., et al. (2012), Trouble at rest: how correlation
patterns and group differences become distorted after global signal regression, Brain Connect., 2, 25 -- 32
Frontiers in Psychology
19
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Schlosser, R. G. M., Wagner, G., and Sauer, H. (2006), Assessing the working memory network: studies
with functional magnetic resonance imaging and structural equation modeling., Neuroscience, 139, 1,
91 -- 103
Schreiber, T. (2000), Measuring Information Transfer, Phys. Rev. Lett., 85, 2, 461 -- 464
Sporns, O., Chialvo, D., Kaiser, M., and Hilgetag, C. (2004), Organization, development and function of
complex brain networks, Trends in Cognitive Sciences, 8, 9, 418 -- 425
Sporns, O., Tononi, G., and Kotter, R. (2005), The human connectome: A structural description of the
human brain., PLoS Comput. Biol., 1, 4, e42
Stramaglia, S., Cortes, J. M., and Marinazzo, D. (2014), Synergy and redundancy in the Granger causal
analysis of dynamical networks, N J Phys, 16, 105003
Tagliazucchi, E., Balenzuela, P., Fraiman, D., and Chialvo, D. (2012), Criticality in large-scale brain fmri
dynamics unveiled by a novel point process analysis, Front. Physiol., 3, 15
Tagliazucchi, E., Leech, R. C.-H. R., Nutt, D., and Chialvo, D. (2014), Enhanced repertoire of brain
dynamical states during the psychedelic experience, Hum. Brain Mapp., 35, 5442 -- 5456
van den Heuvel, M. and Sporns, O. (2013), An anatomical substrate for integration among functional
networks in human cortex, J Neurosci, 33, 14489 -- 14500
Voineskos, A. N., Rajji, T. K., Lobaugh, N. J., Miranda, D., Shenton, M. E., Kennedy, J. L., et al. (2012),
Age-related decline in white matter tract integrity and cognitive performance: a DTI tractography and
structural equation modeling study., Neurobiol. Aging, 33, 1, 21 -- 34
Wang, R., Benner, T., Sorensen, A., and Wedeen, V. (2007), Diffusion toolkit: a software package for
diffusion imaging data processing and tractography, in Proc Intl Soc Mag Reson Med, volume 15,
volume 15, 3720
Wu, G.-R., Liao, W., Stramaglia, S., Ding, J.-R., Chen, H., and Marinazzo, D. (2013), A blind
deconvolution approach to recover effective connectivity brain networks from resting state fMRI data.,
Med. Image Anal., 17, 3, 365 -- 74
This is a provisional file, not the final typeset article
20
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
FIGURES
Figure 1. Sketch for regions of interest (ROIs). 15 different ROIs were extracted from three different resting state networks: 1 ROI in the sensory motor
(SM), 6 ROIs in the default mode network (DMN) and 8 ROIs in the executive control (ExC).
Frontiers in Psychology
21
Sensory Motor (SM)DMN1DMN2DMN3DMN4DMN5DMN6Default Mode Network (DMN)ExC1ExC2ExC3ExC4ExC5ExC6ExC7ExC8Executive Control (ExC) C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Figure 2. Structural, effective and functional connectivity matrices (SC,EC and FC, respectively): A1: SC matrix calculated by the fiber number.
Because many of the values in this matrix are very small, we plotted it in logarithmic scale only to enhance visibility. A2-A4: EC (eSEM) and FC matrices
(C and PC), all of them normalized in the [0, 1] range for comparison purposes. B: Correlation-based similarity between SC and eSEM, C and PC, calculated
either over all pairs or only on connected pairs. C: Mean values of connectivity matrices separated in two groups: pairs such that they have non-zero fibers
between them (structurally connected pairs, CP) and non-connected pairs (NCP). * p − value < 0.01, otherwise means no statistical significance.
This is a provisional file, not the final typeset article
22
C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Figure 3. Connectivity values on specific links. All matrices eSEM, C, PC and SC were normalized in the range [0, 1] for visualization purposes. The
maximum values used for normalization in each case are shown, as well as the mean (µ) and the standard deviation (σ) values for all matrices.
Frontiers in Psychology
23
Max0.0573030.178081.4401e+03SC0.142660.161750.3888eSEM Normalized0.15840.180350.7084PC0.211420.191620.7636C00.20.40.60.81Normalized link strengthExC1-DMN2ExC2-DMN5ExC6-DMN6 C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Figure 4. Scatter plots between different connectivity matrices and separating in two groups: structurally connected pairs (CP) and non-connected pairs
(NCP). Different panels are showing scatter plots of A: (green rectangles) SC with eSEM, C and PC, B: (red rectangles) eSEM with C and PC, C: (blue
rectangles) C with PC.
This is a provisional file, not the final typeset article
24
eSEM matrix (NCP)CCorrcoef=0.80621P-value=9.8105e−310.80.60.40.20110.80.60.40.2eSEMCorrcoef=0.88943P-value=5.215e−45eSEMPC0.80.60.40.201eSEM matrix (CP)BCCorrcoef=0.86859P-value=2.0436e−3010.80.60.40.20.80.60.40.201PCCorrcoef=0.93842P-value=3.6004e−450.80.60.40.201eSEMeSEMC matrix (CP)CCorrcoef=0.86135P-value=0.21387e-29PC0.80.60.40.20110.80.60.40.2CPCC matrix (NCP)Corrcoef=0.63877P-value=3.8028e-16PC10.80.60.40.20.80.60.40.201CCorrcoef=0.4483P-value=4.6485e−06SC matrix (CP)AeSEMSC10.80.60.40.20.80.60.40.201Corrcoef=0.4336P-value=1.0189e−05CSC0.80.60.40.201Corrcoef=0.5073P-value=1.3308e−07PCSC0.80.60.40.201 C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Figure 5. Mean values of structurally connected pairs (CP) and not connected pairs (NCP) across several lags in A: Granger Causality and B: eSEM.
eSEM1, eSEM2 and eSEM3 (the same as GC1, GC2 and GC3) refers to lag={1,2,3} for both eSEM and GC. Notice that, in all the cases, the differences
found between the two groups were not significant according to the p-value. So, neither eSEM nor GC distinguished between CP and NCP.
Frontiers in Psychology
25
NCPGC1GC2GC300.20.40.60.81NCPNCPA. Granger Causality p-value=0.3586p-value=0.9029p-value=0.1948NCP00.20.40.60.81NCPNCPeSEM1eSEM2eSEM3B. Exploratory Structural Equation Modeling p-value=0.1377p-value=0.6097p-value=0.2849CPCPCPCPCPCP C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Figure S1. Pearson's vs Spearman's correlations. The same figures 2B1 and 2B2 are now plotted together with the results from Spearman's correlations
for comparison purposes.
This is a provisional file, not the final typeset article
26
Pearson's Correlation−0.100.10.20.30.40.5SC-eSEMSC-CSC-PC0.38400.36590.4239ALL PAIRSAB−0.100.10.20.30.40.5SC-eSEMSC-CSC-PC0.44830.43360.5073ONLY CONNECTED PAIRS−0.100.10.20.30.40.5SC-eSEMSC-CSC-PCSpearman's Correlation0.20190.28380.1758C−0.100.10.20.30.40.5SC-eSEMSC-CSC-PC0.39940.46050.4303D C. Alonso-Montes et al.
eSEM vs SC in the resting brain
Figure S2. The effect of applying global signal regression vs not appliying it to the time-series rs-fMRI data. Similar to figures 2B1 and 2B2, here we
compared the results of applying global signal removal to the time-series data (panels A and B, and all other figures in this manuscript), to the results without
global signal removal (panels C and D).
Frontiers in Psychology
27
Global Signal RemovalSC-eSEMSC-CSC-PC0.38400.36590.4239ABSC-eSEMSC-CSC-PC0.44830.43360.5073Without Global Signal RemovalALL PAIRSONLY CONNECTED PAIRSSC-eSEMSC-CSC-PCDSC-eSEMSC-CSC-PCC0.41930.41940.50220.49940.54100.596700.10.20.30.40.50.600.10.20.30.40.50.600.10.20.30.40.50.600.10.20.30.40.50.6 |
1106.3616 | 5 | 1106 | 2015-07-31T16:34:45 | When can dictionary learning uniquely recover sparse data from subsamples? | [
"q-bio.NC",
"math.CO"
] | Sparse coding or sparse dictionary learning has been widely used to recover underlying structure in many kinds of natural data. Here, we provide conditions guaranteeing when this recovery is universal; that is, when sparse codes and dictionaries are unique (up to natural symmetries). Our main tool is a useful lemma in combinatorial matrix theory that allows us to derive bounds on the sample sizes guaranteeing such uniqueness under various assumptions for how training data are generated. Whenever the conditions to one of our theorems are met, any sparsity-constrained learning algorithm that succeeds in reconstructing the data recovers the original sparse codes and dictionary. We also discuss potential applications to neuroscience and data analysis. | q-bio.NC | q-bio | When can dictionary learning uniquely
recover sparse data from subsamples?
Christopher J. Hillar, Friedrich T. Sommer
1
5
1
0
2
l
u
J
1
3
]
.
C
N
o
i
b
-
q
[
5
v
6
1
6
3
.
6
0
1
1
:
v
i
X
r
a
is a useful
Abstract -- Sparse coding or sparse dictionary learning has
been widely used to recover underlying structure in many
kinds of natural data. Here, we provide conditions guaranteeing
when this recovery is universal; that is, when sparse codes and
dictionaries are unique (up to natural symmetries). Our main
tool
lemma in combinatorial matrix theory that
allows us to derive bounds on the sample sizes guaranteeing such
uniqueness under various assumptions for how training data are
generated. Whenever the conditions to one of our theorems are
met, any sparsity-constrained learning algorithm that succeeds
in reconstructing the data recovers the original sparse codes and
dictionary. We also discuss potential applications to neuroscience
and data analysis.
Index Terms -- Dictionary learning, sparse coding, sparse ma-
trix factorization, uniqueness, compressed sensing, combinatorial
matrix theory
I. INTRODUCTION
I NDEPENDENT component analysis [1], [2] and dictionary
learning with a sparse coding scheme [3] have become
important tools for revealing underlying structure in many
different types of data. The common goal of these algorithms
is to produce a latent representation of data that exposes
underlying structure. Such a map from data to representations
is often called coding or inference and the reverse map from
representations to estimated data is called reconstruction. In
the coding step of [3], for instance, given a data point y in a
dataset Y , one computes a sparse code vector b = f (y) with
a small number of nonzero coordinates. This code vector b is
used to linearly reconstruct y as
y = Bb,
(1)
using a reconstruction matrix B (often called a basis or
dictionary for Y ).
The code map f (y) and the matrix B are fit to training
data using unsupervised learning. If reconstruction succeeds
and y = y for all data in Y , then representations b and the
dictionary B capture essential structure in the data. Otherwise,
(y − y) provides an error signal to improve the two steps of
coding and reconstruction. This control loop for optimizing
a data representation is often referred to as "self-supervised
learning" or "auto-encoding" (e.g., [4]). In the context of
theoretical neuroscience, the sparse vector b is also sometimes
called an "efficient representation" of y because it exposes
C. J. Hillar conducted the research while at the Mathematical Sciences
Research Institute (MSRI), Berkeley, CA, USA and the Redwood Center for
Theoretical Neuroscience, UC Berkeley, CA, e-mail: [email protected]. F.
T. Sommer is with the Redwood Center for Theoretical Neuroscience at UC
Berkeley, CA, e-mail: [email protected].
the coefficients of the "independent components",
minimizing redundancy of description [5], [6].
thereby
In the literature, self-supervised learning with a sparseness
constraint is called sparse coding or sparse dictionary learning
(SDL). It has been found empirically that SDL succeeds at
revealing (unique) structure for a wide range of sensory data,
most notably natural images [3], [7], [8], [9], [10], natural
sounds [11], [12], [13], and even artistic output [14]. Im-
portantly for applications, SDL can also reveal overcomplete
representations; that is, the dimension of b can exceed that of
the data. For example, SDL can reveal structure in data that
has been sparsely composed from multiple dictionaries [15].
Assume that observed data Y were generated as
y = Aa,
(2)
using a fixed n × m generation matrix A and sparse m-
dimensional
latent representations a. Here we investigate
conditions under which the generation matrix and the sparse
representations can be recovered from samples Y . Specifically,
we ask: How much data is required for self-supervised learning
to uniquely discover A and the corresponding latent represen-
tations a? How overcomplete can the latent representation be
and still be uniquely identified?
This study extends earlier pioneering work on uniqueness
of sparse matrix factorizations in dictionary learning [16]
although our initial motivation was to explain findings in
theoretical [17], [18] and computational neuroscience [19].
Specifically,
it was proposed in [17] and [18] that SDL
could be used as a model for how neurons in sensory areas
communicate high-dimensional sparse representations through
wiring bottlenecks. Assume a rather small number of pro-
jecting neurons subsample the sensory representations in one
brain region and send the signals downstream to another brain
region. In this case, uniqueness of SDL guarantees that self-
supervised learning in the downstream region automatically
recovers the full original sensory representations.
Before addressing our main questions, it is informative to
first explain their relationship to compressed sensing (CS), a
recent advance in signal processing [20], [21]. The theory of
CS provides techniques to recover data vectors x with sparse
structure after they have been subsampled as y = Φx by a
known compression matrix Φ. The sparsity usually enforced
is that x can be expressed as x = Ψa using a known dictionary
matrix Ψ and an m-dimensional vector a with at most k (cid:28) m
nonzero entries. Such vectors a are called k-sparse.
Under mild CS conditions on the generation matrix:
A = ΦΨ,
(3)
2
[23]. More precisely, if P is a permutation matrix2 and D is
an invertible diagonal matrix, then
Aa = (AD−1P (cid:62))(P Da)
for each sample y = Aa. Thus, without access to A, one could
not discriminate which of a or P Da (resp. A or AD−1P (cid:62))
was the original sparse vector (resp. generation matrix). This
discussion motivates the following definition and problem.
Definition 1: When a dataset Y = {Aa1, . . . , AaN} with
k-sparse ai has the property that any other k-sparse coding
of it is the same up to a uniform permutation and invertible
scaling, we say that Y has a unique sparse coding.
Problem 1: Let Y = {y1, . . . , yN} be generated by N
subsamplings yi = Aai as in (2), where A ∈ Rn×m satisfies
the CS spark condition (5) and the ai are k-sparse. When does
Y have a unique sparse coding?
if
In equations, Problem 1 asks when we can guarantee that
Aai = yi = yi = Bbi,
i = 1, . . . , N,
(6)
is another coding of Y using some other matrix B (not
necessarily satisfying the spark condition) with k-sparse codes
bi, then there is a permutation matrix P and an invertible
diagonal matrix D with
A = BP D and bi = P Dai,
i = 1, . . . , N.
(7)
Theorem 1: There exist N = k(cid:0)m
Here, we determine bounds on the number of samples N
guaranteeing that datasets Y have unique sparse codings under
various assumptions for how sparse ai are produced; Table I
gives a summary. Our main result is the following.
k-sparse vectors
a1, . . . , aN ∈ Rm such that any matrix A ∈ Rn×m sat-
isfying spark condition (5) gives rise to a dataset Y =
{Aa1, . . . , AaN} having a unique sparse coding.
(cid:1)2
k
In fact,
there are many such sets of (deterministically
produced) ai; we give a parameterized family in Section II.
Our next result exploits randomness to give a different
construction requiring fewer samples. Fix a matrix A sat-
isfying CS condition (5) and a small β > 0. There is a
simple random drawing procedure (Definition 2 below) for
generating N = k+1
that Y = {Aa1, . . . , AaN} has a unique sparse coding
β
(with probability at least 1 − β). In particular, sparse matrix
factorization of sparsely coded data is typically unique for a
large enough number of random samples.
(cid:1) k-sparse vectors a1, . . . , aN such
(cid:0)m
k
A subtle difference between these two results is that the
a1, . . . , aN in Theorem 1 are independent of the generation
matrix A. The following statement is the best we are able to
do in this direction using random methods. With probability
one, a random draw of (k + 1) k-sparse ai from each of the
set of matrices Z ⊂ Rn×m with Lebesgue measure zero such
that if A /∈ Z, then Y = {Aa1, . . . , AaN} has a unique sparse
(cid:1) support sets satisfies the following property. There is a
(cid:0)m
k
2A permutation matrix P has binary entries and precisely one 1 in each
row and column (thus, P v for a column vector v permutes its entries). Note
that P P (cid:62) = P (cid:62)P = I, where I denotes the m × m identity matrix, and
M(cid:62) for a matrix M is its transpose. In particular, we have P −1 = P (cid:62).
a)
b)
Sparse coding in a compressed space reveals original sparse
Fig. 1.
representations but not original dictionaries [17], [18]. a) Dictionary columns
trained in 2× compressed space y = Φx starting from 16 × 16 natural
image patches x appear random (compression matrix Φ was formed from
i.i.d. standard normals). b) Cross-correlating 16× 16 natural image patches x
with the coordinates of inferred sparse vectors b = f (y) after compression
recovers original sparse coding "Gabor" dictionary Ψ for natural patches [3].
which in this case involves both the compression matrix and
the sparse dictionary for the data, the theory gives accurate
recovery of a in equation (2) (and thus x) from the compressed
vector y as long as the dimension n of y satisfies:
n ≥ Ck log(m/k).
(4)
independent of m, n, and k.1 In
Here, C is a constant
other words, one can easily recover sparse high-dimensional
representations a from "projections" (2) into spaces with a
dimension (almost) as small as the count of a's active entries.
A typical CS assumption on the linear transformation A ∈
Rn×m is the spark condition:
Aa1 = Aa2 for k-sparse a1, a2 ∈ Rm =⇒ a1 = a2. (5)
Note that a generic square matrix A is invertible; thus, (5)
trivially holds for almost all matrices A whenever n = m. In
the interesting regimes of compressed sensing, however, the
sample dimension n is significantly smaller than the original
data dimension m. Thus, condition (5) supplants invertibility
of the matrix A with an "incoherence" among every 2k of
its columns. Rather remarkably, even in the critical regime
close to equality of (4), condition (5) holds with very high
probability for many ensembles of randomly generated A. In
particular, it is easy to produce matrices satisfying (5).
Conditions (4) and (5) are cornerstones of CS theory,
describing when it is possible to recover a code vector a from
a subsampled measurement y generated by (2) using a known
generation matrix A. The dictionary learning problem is even
more difficult: to recover code vectors and the generation
matrix from measurements. As we shall see, however, the
same spark assumption for A in CS theory also guarantees
universality of SDL, given enough training samples.
Taken literally, of course, uniqueness in SDL is ill-posed
1For a more detailed discussion of these facts (including proofs) and their
relationship to approximation theory, we refer to [22] and its references.
k
k(cid:0)m
given the matrix A satisfying (5), draw k + 1 randomly from each support set
given the matrix A satisfying (5), draw randomly uniformly over supports
Samples Y = {Aa1, . . . , AaN} with k-sparse ai ∈ Rm chosen as follows:
(cid:1) in general position from each support set, independent of A satisfying (5)
k(cid:0)m
(k + 1)(cid:0)m
(cid:1) (with probability one)
(cid:0)m
(cid:1) (with probability 1 − β)
(k + 1)(cid:0)m
(cid:1) (with probability one)
(cid:0)m
(cid:1) (with probability 1 − β)
randomly draw k + 1 from each support set, independent of A /∈ Z
randomly draw uniformly over supports, independent of A /∈ Z
(cid:1) different support sets specifying which k entries of a ∈ Rm are nonzero. Given a support set, a "randomly drawn" sparse vector a has k
i.i.d. uniform [0, 1] random variables placed as the nonzero components determined by the support set. The set Z ⊂ Rn×m has Lebesgue measure zero.
There are(cid:0)m
Sufficient samples N for unique recovery:
(cid:1)2
TABLE I
k+1
k+1
β
β
k
k
k
k
k
k
3
coding. In other words, almost surely such a set of ai will have
almost every A give rise to Y with unique sparse codings.
The most general previous uniqueness result on this problem
is given in [16, Theorem 3] with an argument using the
singular value decomposition for matrices. Assuming addi-
tionally that the matrix B also obeys CS condition (5) and
that b1, . . . , bN satisfy certain assumptions3, the authors of
(cid:1) many equations (6) imply (7).
[16] show that (k + 1)(cid:0)m
With Theorem 1, we close a gap in the literature by finding
the weakest possible assumptions on recovery codes and
dictionaries that guarantee uniqueness (although there is room
for improvement in the size N of our set of samples).
k
image patches x. It
As a motivating example of Theorem 1, consider applying
SDL on grayscale natural
is known
empirically that after SDL converges, any image patch can
be represented as x = Ψa where a is a sparse vector and
the matrix Ψ contains two-dimensional Gabor functions [3]
(as displayed in Fig. 1b). Consider now SDL on subsampled
natural image patches y = Φx = ΦΨa, with Φ a compression
matrix. If A = ΦΨ satisfies the spark condition, then any
successful sparse coding y = Bb of enough such y should
produce codes b = f (y) that are equal (up to symmetry) to
the sparse vectors a that resulted from learning on the full
image patches. Indeed, this is the empirical finding [18], even
when we choose x uniformly from a database of natural image
patches [9] rather than generating them from special a. Note
that although the dictionary trained on compressed images
appears uninformative (Fig. 1a), the cross-correlation of the
inferred sparse vectors b with the full image patches contains
two-dimensional Gabor functions (Fig. 1b). In fact, the cross-
correlation matrix is equal to the dictionary Ψ for constructing
the original patches (again, up to natural symmetries).
The organization of this paper is as follows. In Section
II, we state our main tool from combinatorial matrix theory
(Lemma 1) and then derive from it Theorem 1. We next
provide precise statements in Section III of our randomized
theorems and then prove them in Section IV. The final section
gives a short discussion of how our results fit into the general
sparse dictionary learning literature and how they might apply
to neuroscience. An appendix contains the proof of Lemma 1.
3For completeness, we list them here: (i) the supports (the indices of
nonzero components) of each ai and bi consist of exactly k elements, (ii)
for each possible k-element support, there are at least k + 1 vectors ai (resp.
bi) having this support, (iii) any k + 1 vectors ai (resp. bi) having the same
support span a k-dimensional space, and (iv) any k + 1 vectors ai (resp. bi)
having different supports span a (k + 1)-dimensional space.
(cid:41)
(cid:40) (cid:96)(cid:88)
II. DETERMINISTIC UNIQUENESS THEOREM
{1, . . . , m}, and(cid:0)[m]
(cid:1) for the set of k-element subsets of [m].
In what follows, we will use the notation [m] for the set
Also, recall that Span{v1, . . . , v(cid:96)} for real vectors v1, . . . , v(cid:96)
is the vector space consisting of their R-linear span:
k
Span{v1, . . . , v(cid:96)} =
tivi : t1, . . . , t(cid:96) ∈ R
.
Finally, for a subset S ⊆ [m] and a matrix A with columns
{A1, . . . , Am}, we define
i=1
Span{AS} = Span{As : s ∈ S}.
Before proving Theorem 1 in full generality, it is illustrative
to consider when k = 1. In this simple case, we only need
N = m samples for uniqueness. Set ai = ei (i = 1, . . . , m)
to be the standard basis column vectors in Rm. Assuming that
(6) holds for some matrix B and 1-sparse bi, it follows that
(8)
for some map π : {1, . . . , m} → {1, . . . , m} and ci ∈ R. Note
that if ci = 0 for some i, then Aei = 0, contradicting (5).
Aei = B cieπ(i),
i = 1, . . . , m,
Next, we show that π is necessarily injective (and thus is a
permutation). Suppose that π(i) = π(j); then,
Aei = ciBeπ(i) = ciBeπ(j) =
ci
cj
Bcjeπ(j) =
ci
cj
Aej.
Again by (5) this is only possible if i = j. Thus, π is a
permutation.
Let P and D denote the following permutation and diagonal
matrices, respectively:
P =(cid:0)eπ(1) ··· eπ(m)
.
(cid:1) , D =
c1
(cid:0)c1Beπ(1) ··· cmBeπ(m)
(cid:1) = BP D.
···
...
···
...
0
0
...
cm
The matrix formed by stacking left-to-right the column vectors
on the right-hand side of (8) is easily seen to satisfy:
(9)
On the other hand, the columns Aei form the matrix A. Taken
together, therefore, equations (8) imply that A = BP D.
Note that the identity A = BP D already implies in general
the recovery result bi = P Dai for all i. This follows since
Aai = Bbi = A(D−1P (cid:62)bi)
k
k
k
gives bi = P Dai from (5) because both ai and D−1P (cid:62)bi
are k-sparse vectors.
map such that
for all S ∈
(cid:1) is a
Span{AS} = Span{Bπ(S)},
Unfortunately, the proof for larger k is more challenging.
The difficulty is that in general it is nontrivial to produce P
and D as in (9) using only assumptions (5) and (6). For this,
our main results depend on combinatorial linear algebra.
Lemma 1 (Main Lemma): Fix positive integers n and k <
m. Let A ∈ Rn×m and B ∈ Rn×m have columns
{A1, . . . , Am} and {B1, . . . , Bm}, respectively. Suppose that
A satisfies spark condition (5) and that π :(cid:0)[m]
(cid:1) →(cid:0)[m]
(cid:19)
(cid:18)[m]
(10)
Then there exists a permutation matrix P ∈ Rm×m and an
invertible diagonal matrix D ∈ Rm×m such that A = BP D.
In particular, the matrix B also satisfies the spark condition.
We defer the proof of this lemma for k > 1 to the Appendix.
Proof of Theorem 1: First, we produce a set of N =
k(cid:0)m
(cid:1)2 vectors si ∈ Rk in general linear position (i.e., any
of the k × N matrix V = (cid:0)αi
where each possible support set is represented k(cid:0)m
subset of k of them are linearly independent). Specifically,
let α1, . . . , αN be any distinct numbers. Then the columns
are in general linear
position (since the αj are distinct, any k × k "Vandermonde"
sub-determinant is nonzero). Next, form the k-sparse vectors
a1, . . . , aN ∈ Rm by taking si for the support values of ai
(cid:1) times.
(cid:1)k,N
i,j=1
We claim that these ai always produce unique sparse cod-
ings. To prove this, suppose that A satisfies the spark condition
and set yi = Aai. By general linear position and the spark
condition, for every subset of k vectors yi1, . . . , yik generated
using the same support S, we have Span{yi1, . . . , yik} =
Span{AS} is k-dimensional. Now suppose an alternate fac-
torization yi = Bbi, with bi k-sparse. Since there are
(cid:1) vectors y for each support S, the "pigeon-hole prin-
k(cid:0)m
k
k
k
.
j
there are at
ciple" implies that
least k vectors b in the
new factorization that have the same support S(cid:48). Hence,
Span{AS} ⊆ Span{BS(cid:48)}; and since dim(Span{BS(cid:48)}) ≤ k,
we have Span{AS} = Span{BS(cid:48)}. Applying Lemma 1, we
obtain that B differs from A by a permutation and invertible
scaling.
to solve a problem in recreational mathematics [24].
We close this section by using a special case of Theorem 1
Corollary 1: Fix positive integers k < n. Those A ∈ Rn×n
that satisfy (5) and have the property that Aa is k-sparse for
all k-sparse a are the matrices P D, where P and D run over
permutation and invertible diagonal matrices, respectively.
Proof: Let a1, . . . , aN be the k-sparse vectors from
Theorem 1. Let A be any matrix with the stated property
and define k-sparse vectors bi = Aai. Then, with B as
the identity matrix, the set {b1, . . . , bN} is another k-sparse
coding of Aai. Theorem 1 now implies that A = P D for
some permutation matrix P and invertible diagonal matrix D.
4
Definition 2 (Random drawing of sparse vectors): Given
the support set for its k nonzero entries, a random draw of a
is the k-sparse vector with support entries chosen uniformly
from the interval [0, 1] ⊂ R, independently. When a support
set is not specified, a random draw is a choice of one support
set uniformly from all T =(cid:0)m
(cid:1) of them and then a random
draw.
Theorem 2: Fix n, k < m, and a generation matrix
A ∈ Rn×m satisfying spark condition (5). If (k + 1) k-
sparse ai are randomly drawn from each support set, then
Y = {Aa1, . . . , AaN} has a unique sparse coding with
probability one.
k
The following is a direct application of this result, partially
answering one of our questions from the introduction.
Corollary 2: Suppose m, n, and k satisfy inequality (4).
With probability one, a random4 n × m generation matrix A
satisfies (5). Fixing such an A, with probability one a dataset
Y = {Aa1, . . . , AaN} generated from N = (k + 1)(cid:0)m
(cid:1) k-
sparse samples ai uniquely prescribes A and these sparse
vectors ai up to a fixed permutation and scaling.
k
We now state our third theorem. Note that an algebraic set
is a solution to a finite set of polynomial equations.
Theorem 3: Fix positive integers n and k < m. If (k + 1)
k-sparse ai are randomly drawn from each support set, then
with probability one the following holds. There is an algebraic
set Z ⊂ Rn×m of Lebesgue measure zero with the following
property: if A /∈ Z, then Y = {Aa1, . . . , AaN} has a unique
sparse coding.5
With probability one, a random draw of N = (k + 1)(cid:0)m
(cid:1)
Corollary 3: Suppose m, n, and k obey inequality (4).
k
k-sparse samples ai satisfies: almost every matrix A gives
Y = {Aa1, . . . , AaN} a unique sparse coding.
It is somewhat surprising that Theorems 2 and 3 do not
automatically prove Theorem 1. We illustrate some of the
subtlety between these three results as follows. Set
g(A, a) = A − a, A ∈ [0, 1], a ∈ [0, 1],
and consider the following three statements. (i) There is an
a satisfying: for every A, we have g(A, a) (cid:54)= 0. (ii) Fix A.
With probability one, a random a will satisfy g(A, a) (cid:54)= 0.
(iii) With probability one, a random a will satisfy: for almost
every A (i.e., outside of a set of Lebesgue measure zero), we
have g(A, a) (cid:54)= 0. While (ii) and (iii) are easily seen to be true,
statement (i) is false (for this particular g). Given this possible
technicality, it is interesting that a deterministic statement of
the form (i) can be made for the general uniqueness problem
we study in this paper.
Experimental verification that SDL robustly satisfies both
implications (7) of the above theorems appears in [16], [18].
In particular, even if samplings (2) are not exact but contain
measurement inaccuracy, SDL is still observed to succeed.
These findings suggest that noisy versions of the above results
and [16, Theorem 3] hold, a focus of future work.
III. STATEMENTS OF PROBABILISTIC THEOREMS
We next give precise statements of our probabilistic versions
of Theorem 1, all of which rely on the following construction.
4Many natural ensembles of random matrices work, e.g., [22, Section 4].
5The spark condition itself defines an algebraic set of matrices. Thus, the
measure zero set Z here can be defined without mention of (5).
IV. PROOFS OF PROBABILISTIC THEOREMS
In this section, we prove Theorems 2 and 3. Our general
perspective is algebraic. We consider the n × m generation
matrix A as a matrix of nm indeterminates Aij (i = 1, . . . , n;
j = 1, . . . , m). Such matrices become actual real-valued
matrices when real numbers are substituted for all the inde-
with S1 < ··· < Sk, we also consider the following k-sparse
vectors of indeterminates:
terminates. For each support set S = {S1, . . . , Sk} ∈ (cid:0)[m]
(cid:1)
k
S,(cid:96)eS1 + ··· + tk
aS,(cid:96) = t1
(cid:96) = 1, . . . , k + 1.
S,(cid:96)eSk ,
In a moment, each indeterminate tj
draw from the uniform distribution on the interval [0, 1].
(k + 1)(cid:0)m
(cid:1) subsampled vectors of indeterminates:
(cid:27)
(cid:26)
Our main object of interest is the following dataset of N =
S,(cid:96) will represent an i.i.d.
(cid:19)
(cid:18)[m]
(11)
k
Y =
AaS,(cid:96) : S ∈
, (cid:96) = 1, . . . , k + 1
.
(12)
k
We would like to show that under appropriate substitutions for
the indeterminates, the dataset Y has a unique sparse coding.
This construction follows the intuition in [16, Theorem 3],
which is to cover every subspace spanned by k columns of A
with a sufficient number of generic vectors.
Consider any k samples {AaS1,(cid:96)1 , , . . . , AaSk,(cid:96)k}, and let
M be the n× k matrix formed by stacking them horizontally:
M = (AaS1,(cid:96)1 ··· AaSk,(cid:96)k ).
S1,(cid:96)1
Sk,(cid:96)k
, . . . , tj
Let g(A, t) be the polynomial defined as the sum of the
squares of all k × k sub-determinants of M. This polynomial
involves the indeterminates tj
(j = 1, . . . , k),
as well as indeterminates contained in columns AS1, . . . , ASk.
By construction, the polynomial g evaluates to a nonzero num-
ber for a substitution of real numbers for the indeterminates if
and only if the k columns of M span a k-dimensional space.6
Fix a real-valued matrix A satisfying spark condition (5).
We first claim that for almost every substitution for the inde-
terminates tj
S,(cid:96) with real numbers, the polynomial g evaluates
to a nonzero real number. Note this would imply that with
probability one all subsets of Y with k elements are linearly
independent (as a finite union of sets with measure zero also
has measure zero). Somewhat surprisingly, to verify this claim
about g, it is enough to show that some substitution of real
numbers for the indeterminates gives a nonzero real value for
g. Perhaps because we feel this fact should be more well-
known, we state it here. It can be proved by induction on the
number of indeterminates and Fubini's theorem in real analysis
(the base case being that a nonzero univariate polynomial has
finitely many roots). We note that the same statement also
holds for real analytic g.
its zeroes are a set with Lebesgue measure zero.
Proposition 1: If a polynomial g is not the zero polynomial,
Lemma 2: Fix A ∈ Rn×m satisfying (5). With probability
one, no subset of k vectors from Y is linearly dependent.
Proof: By Proposition 1, we need only show that each of
the above g can be nonzero. However, this is clear since we
6If column rank of M is k, then (since row rank equals column rank) there
are k linearly independent rows; the determinant of this submatrix is nonzero.
5
can choose tj
S,(cid:96) so that M consists of k different columns of
A, which are linearly independent by assumption.
Lemma 3: Fix A ∈ Rn×m satisfying (5). With probability
one, if k + 1 vectors yj ∈ Y are linearly dependent, then all
of the k + 1 vectors aj have the same supports.
Proof: Consider the polynomial h which is the sum of
the squares of all (k + 1) × (k + 1) minors of the n × (k + 1)
matrix of columns y1, . . . , yk+1. Suppose that the supports
S1, . . . , Sk+1 of the k + 1 vectors in Y have more than k
different indices. Then since any k + 1 columns of A are
linearly independent, there is a setting for the tj
S,(cid:96) such that
h is a nonzero real number. As before, it follows that for
almost every substitution of the tj
S,(cid:96) with real numbers, the
polynomial h evaluates to a nonzero number. In particular,
with probability one, if the chosen y are linearly dependent,
then S1 = ··· = Sk+1.
Given a support set S ∈(cid:0)[m]
(cid:1), let J(S) = {j : support(bj) ⊆
Proof of Theorem 2: Fix a real matrix A satisfying (5)
and consider any alternate factorization yi = Bbi for Y .
S}, and note that those y indexed by J(S) span at most a k-
dimensional space. By Lemma 3, with probability one, either
J(S) ≤ k or all aj with j ∈ J(S) have the same support
S(cid:48). Since the number of aj with each support pattern S is
k + 1, we have J(S) ≤ k + 1, and since there are a total
(cid:1) samples, it follows that J(S) = k + 1
(cid:1) that takes each subset S to the unique S(cid:48)
for each S. Thus, by Lemma 3 again, we can define a map
which is the support of aj for each j ∈ J(S). One easily
checks that φ is injective and thus bijective, and also that
(10) holds where π = φ−1. Thus, by Lemma 1 there is a
permutation P and invertible diagonal D such that A = BP D.
of N = (k + 1)(cid:0)n
φ :(cid:0)[m]
(cid:1) →(cid:0)[m]
k
k
k
k
In the proof above, we first picked a matrix A and then
verified that a random setting of the tj
S,(cid:96) determine a dataset
Y with a unique sparse coding. Ideally, our sparse vectors
that sample to Y should be independent of the CS matrix A,
as in Theorem 1. The closest we come to this using random
methods is Theorem 3 from Section III, which we now prove.
Proof of Theorem 3: Consider a polynomial g before
the proof of Lemma 2 or a polynomial h as in the proof of
Lemma 3. By Proposition 1, we have for almost every setting
of the tj
S,(cid:96) that g (resp. h) is not the zero polynomial in the
indeterminates Aij. In particular, for almost every A, it is not
the real number zero. The proof remainder is as above.
(cid:1) sparse supports. How many samples N are
(cid:1) is sufficient.
needed if samples are not systematically generated to meet this
requirement but simply drawn randomly from a large set data?
If we require the failure probability to be bounded above by
some small β, then as we explain N = k+1
β
Our probabilistic theorems require k + 1 points for each
of the T = (cid:0)m
Consider first the number of ways for each of the T support
sets to contain at least k + 1 points of the N samples: it is the
number of nonnegative integer solutions di to the equation
(cid:0)m
k
k
(d1 + k + 1) + ··· + (dT + k + 1) = N.
d1 + ··· + dT = M is well-known to number (cid:0)M +T−1
(cid:1). It
More generally, the number of solutions to such an equation
follows that N samples (with supports chosen uniformly) fail
T−1
(cid:18)
T−1(cid:89)
i=1
1
T−1(cid:88)
i=1
to have k+1 points in each of T support sets with a probability
(cid:0)N +T−1
T−1
(cid:1) −(cid:0)N−T k−1
(cid:1)
(cid:0)N +T−1
(cid:1)
T−1
T−1
q =
= 1 − 1/r.
Our goal therefore reduces to finding a sufficient quantity
of samples N = αT guaranteeing a failure probability at most
β ≥ q. To achieve this, we shall show that for large α, the
quantity r is asymptotic to 1 + (k + 1)/α; in particular, q ≈
(k + 1)/α and to obtain q ≤ β, we need N ≥ k+1
β T . First,
we approximate r as follows:
r =
(N + T − i)
(N − T k − i)
=
1 +
k + 1
(N/T − k − i/T )
(cid:19)
T−1(cid:89)
i=1
= 1 + (k + 1)
α − k − i/T
+ O(α−2).
And then we estimate the sum on the right using calculus:
α − k − 1/T
α − k − 1
α − k − i/T
= ln
≈
dx
1
T−1(cid:88)
i=1
(cid:90) 1
1/T
α − k − x
α − k
α − k − 1
≈ 1
α
.
≈ ln
We close this section by sketching proofs for the corollaries
stated in Section III.
Proofs of Corollaries 2, 3: From CS theory (e.g., [22]),
there exist matrices with the restricted isometry property (RIP)
for any m, n, and k satisfying (4). Thus, using the ideas
presented here, one can construct a nonzero polynomial in
the entries of a general matrix A that vanishes if and only if
A fails to obey the spark condition. In particular, almost every
matrix A satisfies (5). The result now follows directly from
Theorem 2; Corollary 3 is proved similarly.
V. DISCUSSION
In this article, we have proved theorems specifying bounds
on the number of samples N sufficient for guaranteeing
uniqueness in sparse dictionary learning. Such bounds depend
critically on the precise problem statement. In particular, a
guarantee with certainty that any generation matrix A is re-
coverable (Theorem 1) requires more samples than a guarantee
with probability one for a particular A (Theorem 2) or for all
but a Lebesgue measure zero set (Theorem 3). We note that the
lower bound for achieving the deterministic guarantee given
here is a drastic tightening of an earlier bound [25], which
used Ramsey theory.
We emphasize that our theorems rely on systematic con-
struction of training samples, which for Theorems 2 and 3
include random drawing. It is an open problem how much
our bounds can be strengthened. In regimes of moderate
compression, computer experiments (e.g., [16], [18]) suggest
that the amount of data needed for successful SDL is smaller
than our estimates.
Another important item left unaddressed here is to find con-
ditions under which sparse dictionary learning is guaranteed to
converge. Although widely used in practice, sparse dictionary
learning algorithms are typically non-convex, and finding a
6
proof of convergence for an SDL scheme can be challenging
[26], [27], [28], [29], [30], [31], [32], [33].
Comments about the mathematics: A mathematical tech-
nique initiated by Szele and Erdos [34], [35] called "the
probabilistic method" [36] produces combinatorial structures
using randomness that are difficult to find deterministically,
much as we have done here. We note that the general problem
of finding deterministic constructions of objects easily formed
using randomness can be very difficult. For instance, how to
deterministically construct optimal compressed sensing RIP
matrices is still open, with the best results so far being [37],
[38] (see also [39] for a recent, large experimental study).
Consequences for applications: A practical consequence of
our work is the description of a feasible regime of over-
completeness for universality in sparse dictionary learning
(Corollaries 2 and 3). Another interesting implication is that
obvious structure in a learned dictionary should not be the only
criterion of success for SDL. For instance, if compression by
a matrix Φ is involved, then columns of a learned dictionary
B = ΦΨD−1P (cid:62) might not reveal visually salient structure
even though learning has converged and the resulting sparse
codes accurately represent the underlying sparse structure of
original data (see Fig. 1 for an example of this phenomenon).
Even though the dictionary might be difficult to interpret in the
case of compression, extraction of sparse components might
still be useful for subsequent steps of analysis. For example,
recent work has demonstrated that sparse coding of sensory
signals can significantly improve performance in classification
tasks [40], [41], [42], [43]. It is possible that compressing data
with a random matrix in SDL can reduce the number of model
parameters, thereby speeding training.
Implications for neuroscience: Our work was originally
motivated by the problem of communication between brain re-
gions in theoretical neuroscience [18]. Neural representations
of sensory input can employ large numbers of local neurons
in a brain area. However, the local neurons with axonal fibers
projecting into a second brain region usually comprise a small
fraction of all local neurons. Thus, there is potentially an
information bottleneck in the communication between brain
areas. Assume that the firing patterns in the sender region
are large sparse vectors a and that the transmission to the
receiver brain area can be described by compression with a
linear transformation Φ. Decoding of the compressed signals y
in the receiving brain area can now be performed by overcom-
plete SDL, implemented by synaptic learning in local neural
networks. As shown in this paper, a decoding network that
successfully reconstructs enough compressed data necessarily
recovers original sparse representations. Such a theory of brain
communication is compatible with early theories of efficient
coding in the brain [5], [6], [3] with one important difference.
The learning objective of traditional efficient coding is the
faithful reconstruction of a sensory signal. In the adjusted
theory [18],
the objective is reconstruction of the locally
available compressed signals. Our results here describe con-
ditions under which learning driven by this weaker objective
can still uniquely recover the sparse components of the full
sensory signal. Of course, without compression, the learning
objective coincides with that of traditional efficient coding. For
some other examples of compressed sensing being applied to
neuroscience, see the survey [44].
Our final application of uniqueness in SDL is to the
analysis of neuroscience data. With measurement techniques
improving, the analysis of signals from dense electrode arrays
placed in the brain has become a pressing issue. In particular,
if stimulus features are not topographically organized in a
neuronal circuit, each electrode might sense a mixture of
signals from many neurons tuned to different features. Such
a situation occurs in multi-electrode recordings from the hip-
pocampal brain region in navigating rats. Individual neurons
in this region encode different locations of the animal in its
environment. Because neighboring cells do not encode similar
locations, the local field potentials have no pronounced place
selectivity and have been regarded as largely uninformative
about the animal's location. Recently, however, performing
SDL on these vastly subsampled data has led to the discov-
ery of sparse place-selective field components. The extracted
components tile the entire environment of the animal and
encode its precise instantaneous localization [19]. Uniqueness
guarantees of the type presented here are important to ensure
that
the extracted behaviorally relevant components reflect
structure of the data and are independent of the details of
the particular SDL algorithm used for data analysis.
APPENDIX
I. COMBINATORIAL MATRIX THEORY
In this section, we prove Lemma 1, which was a main ingre-
dient in proofs of our main theorems. We suspect that there is
an appropriate generalization to matroids. First, however, we
state the following easily deduced facts.
Lemma 4: Let M ∈ Rn×m. If every set of (cid:96) + 1 columns
of M are linearly independent, then for S, S(cid:48) ∈(cid:0)[m]
(cid:1),
(cid:1), then
If M satisfies condition (5) and S1, S2 ∈(cid:0)[m]
Span{MS} = Span{MS(cid:48)} =⇒ S = S(cid:48).
(13)
(cid:96)
Span{MS1 ∩ S2} = Span{MS1} ∩ Span{MS2}.
Proof of Lemma 1: We shall induct on k; the base case
k = 1 is worked out at the beginning of Section II. We first
prove that π is injective (and thus bijective). Suppose that
(cid:1) have π(S1) = π(S2); then by (10),
S1, S2 ∈(cid:0)[m]
(14)
Span{AS1} = Span{Bπ(S1)} = Span{Bπ(S2)} = Span{AS2}.
In particular, using (13) from Lemma 4 above with (cid:96) = k
and M = A, it follows that S1 = S2 and thus π is bijective.
Moreover, from this bijectivity of π and the fact that every k
columns of A are linearly independent, it follows that every
k columns of B are also linearly independent.
k
k
We complete the proof, inductively, by producing a map:
(cid:18) [m]
(cid:19)
(cid:18) [m]
(cid:19)
→
τ :
k − 1
which satisfies Span{AS} = Span{Bτ (S)} for S ∈(cid:0) [m]
(cid:1). Let
(cid:1), and set S1 = S∪{r} and S2 = S∪{s} for some r, s /∈
(cid:0) [m]
α = π−1 denote the inverse of π. Fix S = {i1, . . . , ik−1} ∈
k−1
k − 1
(15)
k−1
7
S with r (cid:54)= s (so that α(S1) (cid:54)= α(S2) by injectivity of α).7
Intersecting equations (10) with S = α(S1) and S = α(S2)
and then applying identity (14) with M = A, it follows that
Span{BS, Br} ∩ Span{BS, Bs} = Span{Aα(S1) ∩ α(S2)}.
(16)
Since the left-hand side of (16) is at least k − 1 dimensional,
the number of elements in the set α(S1)∩α(S2) is either k−1
or k. But α(S1) (cid:54)= α(S2) so that α(S1) ∩ α(S2) consists of
k− 1 elements. Moreover, Span{BS} ⊆ Span{Aα(S1) ∩ α(S2)}
implies that Span{BS} = Span{Aα(S1) ∩ α(S2)}.
The association S (cid:55)→ α(S1) ∩ α(S2) above defines a
(cid:1) with Span{BS} = Span{Aγ(S)}.
function γ :(cid:0) [m]
(cid:1) →(cid:0) [m]
Finally, we show that γ is injective, which implies that
τ = γ−1 is the map desired in (15) for the induction. If
γ(S) = γ(S(cid:48)), then Span{BS} = Span{BS(cid:48)}. By (13) in
Lemma 4 with (cid:96) = k− 1 and M = B, we have S = S(cid:48). Thus,
γ is injective.
the case n = m = 3, k = 2. Suppose that π :(cid:0)[3]
Example 1: We show how the proof of Lemma 1 works in
(cid:1) →(cid:0)[3]
(cid:1) is
k−1
k−1
2
2
π({1, 2}) = {2, 3}, π({1, 3}) = {1, 2}, π({2, 3}) = {1, 3}.
Following the proof of Lemma 1, one can check that
γ({1}) = {3}, γ({2}) = {1}, γ({3}) = {2},
and thus we obtain the map τ = γ−1 as desired in (15). The
resulting permutation P is the cycle 1 (cid:55)→ 2 (cid:55)→ 3 (cid:55)→ 1.
ACKNOWLEDGMENT
We thank the following people for helpful discussions:
Charles Cadieu, Melody Chan, Will Coulter, Jack Culpepper,
Mike DeWeese, Rina Foygel, Guy Isely, Amir Khosrowshahi,
Matthias Mnich, and Chris Rozell. We also thank the anony-
mous referees for comments that considerably improved this
work, including supplying ideas for proofs of several results.
This work was supported by grant IIS-1219212 from the
National Science Foundation. CJH was also partially supported
by an NSF All-Institutes Postdoctoral Fellowship administered
by the Mathematical Sciences Research Institute through its
core grant DMS-0441170. FTS was also supported by the
Applied Mathematics Program within the Office of Science
Advanced Scientific Computing Research of the U.S. Depart-
ment of Energy under contract No. DE-AC02-05CH11231.
REFERENCES
[1] P. Comon, "Independent component analysis, a new concept?" Signal
processing, vol. 36, no. 3, pp. 287 -- 314, 1994.
[2] A. Bell and T. Sejnowski, "An information-maximization approach to
blind separation and blind deconvolution," Neural computation, vol. 7,
no. 6, pp. 1129 -- 1159, 1995.
[3] B. Olshausen and D. Field, "Emergence of simple-cell receptive field
properties by learning a sparse code for natural images," Nature, vol.
381, no. 6583, pp. 607 -- 609, 1996.
[4] G. Hinton, "Connectionist learning procedures," Artificial intelligence,
vol. 40, no. 1-3, pp. 185 -- 234, 1989.
[5] F. Attneave, "Informational aspects of visual perception," Psychol. Rev.,
vol. 61, pp. 183 -- 93, 1954.
7Here we use the assumption that k < m so that such a pair r (cid:54)= s exists.
[6] H. B. Barlow, "Sensory mechanisms,
the reduction of redundancy,
and intelligence," in Proc. of the Symposium on the Mechanization of
Thought Processes, D. Blake and A. Utlley, Eds., vol. 2, 1959, pp. 537 --
574.
[7] J. Hurri, A. Hyvarinen, J. Karhunen, and E. Oja, "Image feature
extraction using independent component analysis," in Proc. NORSIG,
1996, pp. 475 -- 478.
[8] A. Bell and T. Sejnowski, "The independent components of natural
scenes are edge filters," Vision research, vol. 37, no. 23, pp. 3327 -- 3338,
1997.
[9] J. van Hateren and A. van der Schaaf, "Independent component filters
of natural images compared with simple cells in primary visual cortex,"
Proceedings of
the Royal Society of London. Series B: Biological
Sciences, vol. 265, no. 1394, pp. 359 -- 366, 1998.
[10] M. Rehn and F. Sommer, "A network that uses few active neurones to
code visual input predicts the diverse shapes of cortical receptive fields,"
Journal of Computational Neuroscience, vol. 22, no. 2, pp. 135 -- 146,
2007.
[11] A. Bell and T. Sejnowski, "Learning the higher-order structure of a
natural sound," Network: Computation in Neural Systems, vol. 7, no. 2,
pp. 261 -- 266, 1996.
[12] E. Smith and M. Lewicki, "Efficient auditory coding," Nature, vol. 439,
no. 7079, pp. 978 -- 982, 2006.
[13] N. L. Carlson, V. L. Ming, and M. R. DeWeese, "Sparse codes for speech
predict spectrotemporal receptive fields in the inferior colliculus," PLoS
Comput Biol, vol. 8, no. 7, p. e1002594, 2012.
[14] J. M. Hughes, D. J. Graham, and D. N. Rockmore, "Quantification of
artistic style through sparse coding analysis in the drawings of Pieter
Bruegel the Elder," Proc. of the National Academy of Sciences, vol. 107,
no. 4, pp. 1279 -- 1283, 2010.
[15] S. Chen, D. Donoho, and M. Saunders, "Atomic decomposition by basis
pursuit," SIAM review, vol. 43, p. 129, 2001.
[16] M. Aharon, M. Elad, and A. Bruckstein, "On the uniqueness of overcom-
plete dictionaries, and a practical way to retrieve them," Linear Algebra
and its Applications, vol. 416, no. 1, pp. 48 -- 67, 2006.
[17] W. Coulter, C. Hillar, G. Isley, and F. Sommer, "Adaptive compressed
sensing: A new class of self-organizing coding models for neuroscience,"
in Acoustics Speech and Signal Processing (ICASSP), 2010 IEEE
International Conference on.
IEEE, 2010, pp. 5494 -- 5497.
[18] G. Isely, C. Hillar, and F. T. Sommer, "Deciphering subsampled data:
adaptive compressive sampling as a principle of brain communication,"
in Advances in Neural Information Processing Systems 23, J. Lafferty,
C. K. I. Williams, J. Shawe-Taylor, R. Zemel, and A. Culotta, Eds.,
2010, pp. 910 -- 918.
[19] G. Agarwal, I. H. Stevenson, A. Ber´enyi, K. Mizuseki, G. Buzs´aki, and
F. T. Sommer, "Spatially distributed local fields in the hippocampus
encode rat position," Science, vol. 344, no. 6184, pp. 626 -- 630, 2014.
[20] D. Donoho, "Compressed sensing," Information Theory, IEEE Transac-
tions on, vol. 52, no. 4, pp. 1289 -- 1306, 2006.
[21] E. Candes and T. Tao, "Decoding by linear programming," Information
Theory, IEEE Transactions on, vol. 51, no. 12, pp. 4203 -- 4215, 2005.
[22] R. Baraniuk, M. Davenport, R. DeVore, and M. Wakin, "A simple proof
of the restricted isometry property for random matrices," Constructive
Approximation, vol. 28, no. 3, pp. 253 -- 263, 2008.
[23] S. Gleichman and Y. Eldar, "Blind compressed sensing," Information
Theory, IEEE Transactions on, vol. 57, no. 10, pp. 6958 -- 6975, 2011.
[24] C. Hillar, "Problem 11534," American Mathematical Monthly, vol. 117,
no. 9, p. 835, 2010.
[25] C. Hillar and F. T. Sommer, "Ramsey theory reveals the conditions when
sparse coding on subsampled data is unique," Tech. Rep., 2011.
[26] J. Mairal, F. Bach, J. Ponce, and G. Sapiro, "Online learning for matrix
factorization and sparse coding," The Journal of Machine Learning
Research, vol. 11, pp. 19 -- 60, 2010.
[27] R. Gribonval and K. Schnass, "Dictionary identification -- sparse matrix-
factorization via (cid:96)1-minimization," Information Theory, IEEE Transac-
tions on, vol. 56, no. 7, pp. 3523 -- 3539, 2010.
[28] A. Balavoine, J. Romberg, and C. Rozell, "Convergence and rate analysis
of neural networks for sparse approximation," Neural Networks and
Learning Systems, IEEE Transactions on, vol. 23, no. 9, pp. 1377 -- 1389,
2012.
[29] D. A. Spielman, H. Wang, and J. Wright, "Exact recovery of sparsely-
used dictionaries," in Proceedings of the Twenty-Third international joint
conference on Artificial Intelligence. AAAI Press, 2013, pp. 3087 -- 3090.
[30] S. Arora, R. Ge, and A. Moitra, "New algorithms for learning incoherent
and overcomplete dictionaries," arXiv preprint arXiv:1308.6273, 2013.
8
[31] A. Agarwal, A. Anandkumar, and P. Netrapalli, "Exact
recov-
sparsely used overcomplete dictionaries," arXiv preprint
ery of
arXiv:1309.1952, 2013.
[32] A. Agarwal, A. Anandkumar, P. Jain, P. Netrapalli, and R. Tandon,
"Learning sparsely used overcomplete dictionaries via alternating min-
imization," arXiv preprint arXiv:1310.7991, 2013.
[33] S. Arora, A. Bhaskara, R. Ge, and T. Ma, "More algorithms for provable
dictionary learning," arXiv preprint arXiv:1401.0579, 2014.
[34] T. Szele, "Kombinatorikai vizsgalatok az iranyitott teljes graffal kapc-
solatban," Mat. Fiz. Lapok, vol. 50, pp. 223 -- 256, 1943.
[35] P. Erdos, "Some remarks on the theory of graphs," Bull. Amer. Math.
Soc, vol. 53, no. 2, pp. 292 -- 294, 1947.
[36] N. Alon and J. Spencer, The probabilistic method. Wiley-Interscience,
2011, vol. 73.
[37] J. Bourgain, S. Dilworth, K. Ford, S. Konyagin, and D. Kutzarova,
"Explicit constructions of RIP matrices and related problems," Duke
Mathematical Journal, vol. 159, no. 1, pp. 145 -- 185, 2011.
[38] S. Li, F. Gao, G. Ge, and S. Zhang, "Deterministic construction of
compressed sensing matrices via algebraic curves," Information Theory,
IEEE Transactions on, vol. 58, no. 8, pp. 5035 -- 5041, 2012.
[39] H. Monajemi, S. Jafarpour, M. Gavish, Stat 330 / CME 362 Col-
laboration, and D. L. Donoho, "Deterministic matrices matching the
compressed sensing phase transitions of gaussian random matrices,"
Proc. of the National Academy of Sciences, vol. 110, no. 4, pp. 1181 --
1186, 2013.
[40] J. Yang, K. Yu, Y. Gong, and T. Huang, "Linear spatial pyramid
matching using sparse coding for image classification," in Computer
Vision and Pattern Recognition (CVPR), 2009 IEEE Conference on.
IEEE, 2009, pp. 1794 -- 1801.
[41] Y.-L. Boureau, F. Bach, Y. LeCun, and J. Ponce, "Learning mid-level
features for recognition," in Computer Vision and Pattern Recognition
(CVPR), 2010 IEEE Conference on.
IEEE, 2010, pp. 2559 -- 2566.
[42] A. Coates and A. Ng, "The importance of encoding versus training with
sparse coding and vector quantization," in Proc. of the 28th International
Conference on Machine Learning (ICML), L. Getoor and T. Scheffer,
Eds., New York, NY, 2011, pp. 921 -- 928.
[43] R. Rigamonti, M. Brown, and V. Lepetit, "Are sparse representations
really relevant for image classification?" in Computer Vision and Pattern
Recognition (CVPR), 2011 IEEE Conference on. IEEE, 2011, pp. 1545 --
1552.
[44] S. Ganguli and H. Sompolinsky, "Compressed sensing, sparsity, and
dimensionality in neuronal information processing and data analysis,"
Annual Review of Neuroscience, vol. 35, pp. 485 -- 508, 2012.
Christopher J. Hillar completed a B.S. in Mathematics and a B.S. in Com-
puter Science at Yale University. Supported by an NSF Graduate Research
Fellowship, he received his Ph.D. in Mathematics from the University of
California, Berkeley in 2005. From 2005 - 2008, he was a Visiting Assistant
Professor and NSF Postdoctoral Fellow at Texas A & M University. From
2008 - 2010, he was an NSF Mathematical Sciences Research Institutes
Postdoctoral Fellow at MSRI in Berkeley, CA. In 2010, he joined the Redwood
Center for Theoretical Neuroscience at UC Berkeley.
Friedrich T. Sommer received the diploma in physics from the University of
Tuebingen, in 1987, the Ph.D. in physics from the University of Duesseldorf,
in 1993, and his habilitation in computer science from University of Ulm,
in 2002. He is an Adjunct Professor at the Redwood Center for Theoretical
Neuroscience and at the HelenWills Neuroscience Institute, University of Cal-
ifornia, Berkeley, since 2005. In 2003 - 2005, he was a Principal Investigator
at the Redwood Neuroscience Institute, Menlo Park, CA. In 1998 - 2002, he
was an Assistant Professor at the Department of Computer Science, University
of Ulm, after completing Postdocs at Massachusetts Institute of Technology,
Cambridge, MA, and the University of Tuebingen.
|
1511.00083 | 2 | 1511 | 2015-12-01T21:20:26 | Why Neurons Have Thousands of Synapses, A Theory of Sequence Memory in Neocortex | [
"q-bio.NC",
"cs.AI"
] | Neocortical neurons have thousands of excitatory synapses. It is a mystery how neurons integrate the input from so many synapses and what kind of large-scale network behavior this enables. It has been previously proposed that non-linear properties of dendrites enable neurons to recognize multiple patterns. In this paper we extend this idea by showing that a neuron with several thousand synapses arranged along active dendrites can learn to accurately and robustly recognize hundreds of unique patterns of cellular activity, even in the presence of large amounts of noise and pattern variation. We then propose a neuron model where some of the patterns recognized by a neuron lead to action potentials and define the classic receptive field of the neuron, whereas the majority of the patterns recognized by a neuron act as predictions by slightly depolarizing the neuron without immediately generating an action potential. We then present a network model based on neurons with these properties and show that the network learns a robust model of time-based sequences. Given the similarity of excitatory neurons throughout the neocortex and the importance of sequence memory in inference and behavior, we propose that this form of sequence memory is a universal property of neocortical tissue. We further propose that cellular layers in the neocortex implement variations of the same sequence memory algorithm to achieve different aspects of inference and behavior. The neuron and network models we introduce are robust over a wide range of parameters as long as the network uses a sparse distributed code of cellular activations. The sequence capacity of the network scales linearly with the number of synapses on each neuron. Thus neurons need thousands of synapses to learn the many temporal patterns in sensory stimuli and motor sequences. | q-bio.NC | q-bio | Why Neurons Have Thousands of Synapses, A Theory of
Sequence Memory in Neocortex
Jeff Hawkins*, Subutai Ahmad
Numenta, Inc, Redwood City, California, United States of America
*Corresponding author
Emails: [email protected], [email protected]
Keywords: neocortex, prediction, neocortical theory, active dendrites, sequence memory.
A version of this manuscript has been submitted for publication as of October 30, 2015. Note that figures and tables are at the
end of this PDF.
Please contact the authors for updated citation information.
1
Abstract
Neocortical neurons have thousands of excitatory synapses.
It is a mystery how neurons integrate the input from so many
synapses and what kind of large-scale network behavior this
enables. It has been previously proposed that non-linear
properties of dendrites enable neurons to recognize multiple
patterns. In this paper we extend this idea by showing that a
neuron with several thousand synapses arranged along active
dendrites can learn to accurately and robustly recognize
hundreds of unique patterns of cellular activity, even in the
presence of large amounts of noise and pattern variation. We
then propose a neuron model where some of the patterns
recognized by a neuron lead to action potentials and define
the classic receptive field of the neuron, whereas the majority
of the patterns recognized by a neuron act as predictions by
slightly depolarizing
immediately
generating an action potential. We then present a network
model based on neurons with these properties and show that
the network learns a robust model of time-based sequences.
Given the similarity of excitatory neurons throughout the
neocortex and the importance of sequence memory in
inference and behavior, we propose that this form of
sequence memory is a universal property of neocortical
tissue. We further propose that cellular layers in the
neocortex implement variations of the same sequence
memory algorithm to achieve different aspects of inference
and behavior. The neuron and network models we introduce
are robust over a wide range of parameters as long as the
network uses a sparse distributed code of cellular activations.
The sequence capacity of the network scales linearly with the
number of synapses on each neuron. Thus neurons need
thousands of synapses to learn the many temporal patterns in
sensory stimuli and motor sequences.
the neuron without
1. Introduction
Excitatory neurons in the neocortex have thousands of
excitatory synapses. The proximal synapses, those closest to
the cell body, have a relatively large effect on the likelihood
of a cell generating an action potential. However, a majority
of the synapses are distal, or far from the cell body. The
activation of a single distal synapse has little effect at the
soma, and for many years it was hard to imagine how the
thousands of distal synapses could play an important role in
determining a cell’s responses (Major et al., 2013). We now
know that dendrite branches are active processing elements.
The activation of several distal synapses within close spatial
and temporal proximity can lead to a local dendritic NMDA
spike and consequently a significant and sustained
depolarization of the soma (Antic et al., 2010; Major et al.,
2013). This has led some researchers to suggest that dendritic
branches act as independent pattern recognizers (Poirazi et
al., 2003; Polsky et al., 2004). Yet, despite the many
advances in understanding the active properties of dendrites,
it remains a mystery why neurons have so many synapses
and what their precise role is in memory and cortical
processing.
Lacking a theory of why neurons have active dendrites,
almost all artificial neural networks, such as those used in
deep learning (LeCun et al., 2015) and spiking neural
networks (Maass, 1997), use artificial neurons without active
dendrites and with unrealistically few synapses, strongly
suggesting they are missing key functional aspects of real
neural tissue. If we want to understand how the neocortex
works and build systems that work on the same principles as
the neocortex, we need an understanding of how biological
neurons use their thousands of synapses and active dendrites.
Of course, neurons cannot be understood in isolation. We
also need a complementary theory of how networks of
neurons, each with thousands of synapses, work together
towards a common purpose.
In this paper we introduce such a theory. First, we show how
a typical pyramidal neuron with active dendrites and
thousands of synapses can recognize hundreds of unique
patterns of cellular activity. We show that a neuron can
recognize hundreds of patterns even in the presence of large
amounts of noise and variability as long as overall neural
activity is sparse. Next we introduce a neuron model where
the inputs to different parts of the dendritic tree serve
different purposes. In this model the patterns recognized by a
neuron’s distal synapses are used for prediction. Each neuron
learns to recognize hundreds of patterns that often precede
the cell becoming active. The recognition of any one of these
learned patterns acts as a prediction by depolarizing the cell
without directly causing an action potential. Finally, we show
how a network of neurons with this property will learn and
recall sequences of patterns. The network model relies on
depolarized neurons firing quickly and inhibiting other
nearby neurons,
the network’s activation
towards its predictions. Through simulation we illustrate that
the sequence memory network exhibits numerous desirable
properties such as on-line learning, multiple simultaneous
predictions, and robustness.
Given the similarity of neurons throughout the neocortex and
the importance of sequence memory for inference and
behavior, we propose that sequence memory is a property of
neural tissue throughout the neocortex and thus represents a
new and important unifying principle for understanding how
the neocortex works.
2. Results
thus biasing
2.1. Neurons Recognize Multiple Patterns
It is common to think of a neuron as recognizing a single
pattern of activity on its synapses. This notion, sometimes
called a “point neuron”, forms the basis of almost all
artificial neural networks (Fig. 1A).
[Figure 1 about here – see end of manuscript]
Active dendrites suggest a different view of the neuron,
where neurons recognize many unique patterns (Larkum and
Nevian, 2008; Poirazi et al., 2003; Polsky et al., 2004).
Experimental results show that the coincident activation of
eight to twenty synapses in close spatial proximity on a
dendrite will combine in a non-linear fashion and cause an
NMDA dendritic spike (Larkum et al., 1999; Major et al.,
2013; Schiller and Schiller, 2001; Schiller et al., 2000). Thus,
a small set of neighboring synapses acts as a pattern detector.
It follows that the thousands of synapses on a cell’s dendrites
2
act as a set of independent pattern detectors. The detection of
any of these patterns causes an NMDA spike and subsequent
depolarization at the soma.
It might seem that eight to twenty synapses could not reliably
recognize a pattern of activity in a large population of cells.
However, robust recognition is possible if the patterns to be
recognized are sparse; i.e. few neurons are active relative to
the population (Olshausen and Field, 2004). For example,
consider a population of 200K cells where 1% (2,000) of the
cells are active at any point in time. We want a neuron to
detect when a particular pattern occurs in the 200K cells. If a
section of the neuron’s dendrite forms new synapses to just
10 of the 2,000 active cells, and the threshold for generating
an NMDA spike is 10, then the dendrite will detect the target
pattern when all 10 synapses receive activation at the same
time. Note that the dendrite could falsely detect many other
patterns that share the same 10 active cells. However, if the
patterns are sparse, the chance that the 10 synapses would
become active for a different random pattern is small. In this
example it is only 9.8 x 10-21.
The probability of a false match can be calculated precisely
as follows. Let 𝑛 represent the size of the cell population and
𝑎 the number of active cells in that population at a given
point in time, for sparse patterns 𝑎≪𝑛. Let 𝑠 be the number
of synapses on a dendritic segment and 𝜃 be the NMDA
at least 𝜃 synapses become active, i.e. at least 𝜃 of the 𝑠
spike threshold. We say the segment recognizes a pattern if
synapses match the currently active cells.
Assuming a random distribution of patterns, the exact
probability of a false match is given by:
!!!!
𝑠𝑏 × 𝑛−𝑠
𝑎−𝑏
𝑛𝑎
(1)
The denominator is simply the total number of possible
patterns containing 𝑎 active cells in a population of 𝑛 total
would connect to 𝜃 or more of the 𝑠 synapses on one
cells. The numerator counts the number of patterns that
dendritic segment. A more detailed description of this
equation can be found in (Ahmad and Hawkins, 2015).
The equation shows that a non-linear dendritic segment can
robustly classify a pattern by sub-sampling (forming
synapses to only a small number of the cells in the pattern to
be classified). Table A in S1 Text lists representative error
probabilities calculated from Eq. (1).
By forming more synapses than necessary to generate an
NMDA spike, recognition becomes robust to noise and
variation. For example, if a dendrite has an NMDA spike
threshold of 10, but forms 20 synapses to the pattern it wants
to recognize, twice as many as needed, it allows the dendrite
to recognize the target pattern even if 50% of the cells are
changed or inactive. The extra synapses also increase the
likelihood of a false positive error. Although the chance of
error has increased, Eq. (1) shows that it is still tiny when the
patterns are sparse. In the above example, doubling the
number of synapses and hence introducing a 50% noise
tolerance, increases the chance of error to only 1.6 x 10-18.
Table 1B in S1 Text lists representative error rates when the
number of synapses exceeds the threshold.
The synapses recognizing a given pattern have to be co-
located on a dendritic segment. If they lie within 40µm of
each other then as few as eight synapses are sufficient to
create an NMDA spike (Major et al., 2008). If the synapses
are spread out along the dendritic segment, then up to twenty
synapses are needed (Major et al., 2013). A dendritic
segment can contain several hundred synapses; therefore
each segment can detect multiple patterns. If synapses that
recognize different patterns are mixed together on the
dendritic segment, it introduces an additional possibility of
error by co-activating synapses from different patterns. The
probability of this type of error depends on how many sets of
synapses share the dendritic segment and the sparsity of the
patterns to be recognized. For a wide range of values the
chance for this type of error is still low (Table C in S1 Text).
Thus the placement of synapses to recognize a particular
pattern is somewhat precise (they must be on the same
dendritic segment and ideally within 40µm of each other),
but also somewhat imprecise (mixing with other synapses is
unlikely to cause errors).
If we assume an average of 20 synapses are allocated to
recognize each pattern, and that a neuron has 6,000 synapses,
then a cell would have the ability to recognize approximately
300 different patterns. This is a rough approximation, but
makes evident that a neuron with active dendrites can learn
to reliably recognize hundreds of patterns within a large
population of cells. The recognition of any one of these
patterns will depolarize the cell. Since all excitatory neurons
in the neocortex have thousands of synapses, and, as far as
we know, they all have active dendrites, then each and every
excitatory neocortical neuron recognizes hundreds of patterns
of neural activity.
In the next section we propose that most of the patterns
recognized by a neuron do not directly lead to an action
potential, but instead play a role in how networks of neurons
make predictions and learn sequences.
2.1.1. Three Sources of Synaptic Input to Cortical
Neurons
Neurons receive excitatory input from different sources that
are segregated on different parts of the dendritic tree. Fig. 1B
shows a typical pyramidal cell, the most common excitatory
neuron in the neocortex. We show the input to the cell
divided into three zones. The proximal zone receives
feedforward input. The basal zone receives contextual input,
mostly from nearby cells in the same cortical region
(Petreanu et al., 2009; Rah et al., 2013; Yoshimura et al.,
2000). The apical zone receives feedback input (Spruston,
2008). (The second most common excitatory neuron in the
neocortex is the spiny stellate cell; we suggest they be
considered similar to pyramidal cells minus the apical
dendrites.) We propose
three zones of synaptic
integration on a neuron (proximal, basal, and apical) serve
the following purposes.
Proximal Synapses Define the Classic Receptive Field of a
Cell
The synapses on the proximal dendrites (typically several
hundred) have a relatively large effect at the soma and
the
3
dendrite/coincidence detectors with up to 40 synapses per
dendrite. For clarity, Fig. 1C shows only a few dendrites and
synapses.
2.2. Networks of Neurons Learn Sequences
inference, prediction,
Because all tissue in the neocortex consists of neurons with
active dendrites and thousands of synapses, it suggests there
are common network principles underlying everything the
neocortex does. This leads to the question, what network
property is so fundamental that it is a necessary component
of sensory
language, and motor
planning?
We propose that the most fundamental operation of all
neocortical tissue is learning and recalling sequences of
patterns (Hawkins and Blakeslee, 2004), what Karl Lashley
famously called “the most important and also the most
neglected problem of cerebral physiology” (Lashley, 1951).
More specifically, we propose that each cellular layer in the
neocortex implements a variation of a common sequence
memory algorithm. We propose cellular layers use sequence
memory for different purposes, which is why cellular layers
vary in details such as size and connectivity. In this paper we
illustrate what we believe is the basic sequence memory
algorithm without elaborating on its variations.
We started our exploration of sequence memory by listing
several properties required of our network in order to model
the neocortex.
1) On-line learning
Learning must be continuous. If the statistics of the world
change, the network should gradually and continually adapt
with each new input.
2) High-order predictions
Making correct predictions with complex sequences requires
the ability to incorporate contextual information from the
past. The network needs to dynamically determine how much
temporal context is needed to make the best predictions. The
term “high-order” refers to “high-order Markov chains”
which have this property.
3) Multiple simultaneous predictions
Natural data streams often have overlapping and branching
sequences. The sequence memory therefore needs to make
multiple predictions at the same time.
4) Local learning rules
The sequence memory must only use learning rules that are
local to each neuron. The rules must be local in both space
and time, without the need for a global objective function.
5) Robustness
The memory should exhibit robustness to high levels of
noise, loss of neurons, and natural variation in the input.
Degradation in performance under these conditions should be
gradual.
All these properties must occur simultaneously in the context
of continuously streaming data.
inputs. We will explain
therefore are best situated to define the basic receptive field
response of the neuron (Spruston, 2008). If the coincident
activation of a subset of the proximal synapses is sufficient to
generate a somatic action potential and If the inputs to the
proximal synapses are sparsely active, then the proximal
synapses will recognize multiple unique feedforward patterns
in the same manner as discussed earlier. Therefore, the
feedforward receptive field of a cell can be thought of as a
union of feedforward patterns.
Basal Synapses Learn Transitions in Sequences
We propose that basal dendrites of a neuron recognize
patterns of cell activity that precede the neuron firing, in this
way the basal dendrites learn and store transitions between
activity patterns. When a pattern is recognized on a basal
dendrite it generates an NMDA spike. The depolarization
due to an NMDA spike attenuates in amplitude by the time it
reaches the soma, therefore when a basal dendrite recognizes
a pattern it will depolarize the soma but not enough to
generate a somatic action potential (Antic et al., 2010; Major
et al., 2013). We propose this sub-threshold depolarization is
an important state of the cell. It represents a prediction that
the cell will become active shortly and plays an important
role in network behavior. A slightly depolarized cell fires
earlier than it would otherwise if it subsequently receives
sufficient feedforward input. By firing earlier it inhibits
neighboring cells, creating highly sparse patterns of activity
for correctly predicted
this
mechanism more fully in a later section.
Apical Synapses Invoke a Top-down Expectation
The apical dendrites of a neuron also generate NMDA spikes
when they recognize a pattern (Cichon and Gan, 2015). An
apical NMDA spike does not directly affect the soma.
Instead it can lead to a Ca2+ spike in the apical dendrite
(Golding et al., 1999; Larkum et al., 2009). A single apical
Ca2+ spike will depolarize the soma, but typically not
enough to generate a somatic action potential (Antic et al.,
2010). The interaction between apical Ca2+ spikes, basal
NMDA spikes, and somatic action potentials is an area of
ongoing research (Larkum, 2013), but we can say that under
many conditions a recognized pattern on an apical dendrite
will depolarize the cell and therefore have a similar effect as
a recognized pattern on a basal dendrite. We propose that the
depolarization caused by the apical dendrites is used to
establish a top-down expectation, which can be thought of as
another form of prediction.
2.1.2. The HTM Model Neuron
Fig. 1C shows an abstract model of a pyramidal neuron we
use in our software simulations. We model a cell’s dendrites
as a set of threshold coincidence detectors; each with its own
synapses.
the number of active synapses on a
dendrite/coincidence detector exceeds a threshold the cell
detects a pattern. The coincidence detectors are in three
groups corresponding to the proximal, basal, and apical
dendrites of a pyramidal cell. We refer to this model neuron
as an “HTM neuron” to distinguish it from biological
neurons and point neurons. HTM is an acronym for
Hierarchical Temporal Memory, a term used to describe our
models of neocortex (Hawkins et al., 2011). HTM neurons
used
this paper have 128
in
If
the simulations
for
4
2.2.1. Mini-columns and Neurons: Two
Representations
High-order sequence memory requires two simultaneous
representations. One represents the feedforward input to the
network and the other represents the feedforward input in a
particular temporal context. To illustrate this requirement,
consider two abstract sequences “ABCD” and “XBCY”,
where each letter represents a sparse pattern of activation in a
population of neurons. Once these sequences are learned the
network should predict “D” when presented with sequence
“ABC” and it should predict “Y” when presented with
sequence “XBC”. Therefore, the internal representation
during the subsequence “BC” must be different in the two
cases; otherwise the correct prediction can’t be made after
“C” is presented.
Fig. 2 illustrates how we propose these two representations
are manifest in a cellular layer of cortical neurons. The
panels in Fig. 2 represent a slice through a single cellular
layer in the neocortex (Fig. 2A). The panels are greatly
simplified for clarity. Fig. 2B shows how the network
represents two input sequences before the sequences are
learned. Fig. 2C shows how the network represents the same
input after the sequences are learned. Each feedforward input
to the network is converted into a sparse set of active mini-
columns. (Mini-columns in the neocortex span multiple
cellular layers. Here we are only referring to the cells in a
mini-column in one cellular layer.) All the neurons in a mini-
column share the same feedforward receptive fields. If an
unanticipated input arrives, then all the cells in the selected
mini-columns will recognize the input pattern and become
active. However, in the context of a previously learned
sequence, one or more of the cells in the mini-columns will
be depolarized. The depolarized cells will be the first to
generate an action potential, inhibiting the other cells nearby.
Thus a predicted input will lead to a very sparse pattern of
cell activation that is unique to a particular element, at a
particular location, in a particular sequence.
[Figure 2 about here – see end of manuscript]
2.2.2. Basal Synapses Are the Basis of Sequence
Memory
In this theory, cells use their basal synapses to learn the
transitions between
input patterns. With each new
feedforward input some cells become active via their
proximal synapses. Other cells, using their basal synapses,
learn to recognize this active pattern and upon seeing the
pattern again, become depolarized, thereby predicting their
own feedforward activation in the next input. Feedforward
input activates cells, while basal input generates predictions.
As long as the next input matches the current prediction, the
sequence continues, Fig. 3. Fig. 3A shows both active cells
and predicted cells while the network follows a previously
learned sequence.
[Figure 3 about here – see end of manuscript]
Often
the network will make multiple simultaneous
predictions. For example, suppose that after learning the
sequences “ABCD” and “XBCY” we expose the system to
just the ambiguous sub-sequence “BC”. In this case we want
the system to simultaneously predict both “D” and “Y”. Fig.
3B illustrates how the network makes multiple predictions
5
when the input is ambiguous. The number of simultaneous
predictions that can be made with low chance of error can
again be calculated via Eq. (1). Because the predictions tend
to be highly sparse, it is possible for a network to predict
dozens of patterns simultaneously without confusion. If an
input matches any of the predictions it will result in the
correct highly-sparse representation. If an input does not
match any of the predictions all the cells in a column will
become active, indicating an unanticipated input.
Although every cell in a mini-column shares the same
feedforward
recognize
different patterns. Therefore cells within a mini-column will
respond uniquely in different learned temporal contexts, and
overall levels of activity will be sparser when inputs are
anticipated. Both of these attributes have been observed
(Martin and Schröder, 2013; Vinje and Gallant, 2002; Yen et
al., 2007).
For one of the cells in the last panel of Fig. 3A, we show
three connections the cell used to make a prediction. In real
neurons, and in our simulations, a cell would form 15 to 40
connections to a subset of a larger population of active cells.
2.2.3. Apical Synapses Create a Top-Down
their basal synapses
response,
Expectation
that
Feedback axons between neocortical regions often form
synapses (in layer 1) with apical dendrites of pyramidal
neurons whose cell bodies are in layers 2, 3, and 5. It has
long been speculated
these feedback connections
implement some form of expectation or bias (Lamme et al.,
1998). Our neuron model suggests a mechanism for top-
down expectation in the neocortex. Fig. 4 shows how a stable
feedback pattern to apical dendrites can predict multiple
elements in a sequence all at the same time. When a new
feedforward input arrives it will be interpreted as part of the
predicted sequence. The feedback biases the input towards a
particular interpretation. Again, because the patterns are
sparse, many patterns can be simultaneously predicted.
[Figure 4 about here – see end of manuscript]
elements
sequence
simultaneously.
Thus there are two types of prediction occurring at the same
time. Lateral connections to basal dendrites predict the next
input, and top-down connections to apical dendrites predict
multiple
The
physiological interaction between apical and basal dendrites
is an area of active research (Larkum, 2013) and will likely
lead to a more nuanced interpretation of their roles in
inference and prediction. However, we propose that the
mechanisms shown in Figs. 2, 3 and 4 are likely to continue
to play a role in that final interpretation.
2.2.4. Synaptic Learning Rule
Our neuron model requires two changes to the learning rules
by which most neural models learn. First, learning occurs by
growing and removing synapses from a pool of “potential”
synapses (Chklovskii et al., 2004). Second, Hebbian learning
and synaptic change occur at the level of the dendritic
segment, not the entire neuron (Stuart and Häusser, 2001).
Potential Synapses
For a neuron to recognize a pattern of activity it requires a set
of co-located synapses (typically fifteen to twenty) that
connect to a subset of the cells that are active in the pattern to
be recognized. Learning to recognize a new pattern is
accomplished by the formation of a set of new synapses
collocated on a dendritic segment.
Figure 5 shows how we model the formation of new
synapses in a simulated HTM neuron. For each dendritic
segment we maintain a set of “potential” synapses between
the dendritic segment and other cells in the network that
could potentially form a synapse with
the segment
(Chklovskii et al., 2004). The number of potential synapses is
larger than the number of actual synapses. We assign each
potential synapse a scalar value called “permanence” which
represents stages of growth of the synapse. A permanence
value close to zero represents an axon and dendrite with the
potential to form a synapse but that have not commenced
growing one. A 1.0 permanence value represents an axon and
dendrite with a large fully formed synapse.
[Figure 5 about here – see end of manuscript]
The permanence value is incremented and decremented using
a Hebbian-like rule. If the permanence value exceeds a
threshold, such as 0.3, then the weight of the synapse is 1, if
the permanence value is at or below the threshold then the
weight of the synapse is 0. The threshold represents the
establishment of a synapse, albeit one that could easily
disappear. A synapse with a permanence value of 1.0 has the
same effect as a synapse with a permanence value at
threshold but is not as easily forgotten. Using a scalar
permanence value enables on-line learning in the presence of
noise. A previously unseen input pattern could be noise or it
could be the start of a new trend that will repeat in the future.
By growing new synapses, the network can start to learn a
new pattern when it is first encountered, but only act
differently after several presentations of the new pattern.
Increasing permanence beyond the threshold means that
patterns experienced more than others will take longer to
forget.
HTM neurons and HTM networks rely on distributed
patterns of cell activity, thus the activation strength of any
one neuron or synapse is not very important. Therefore, in
HTM simulations we model neuron activations and synapse
weights with binary states. Additionally, it is well known that
biological synapses are stochastic (Faisal et al., 2008), so a
neocortical theory cannot require precision of synaptic
efficacy. Although scalar states and weights might improve
performance, they are not required from a theoretical point of
view and all of our simulations have performed well without
them. The formal learning rules used in our HTM network
simulations are presented in the Materials and Methods
section.
3. Simulation Results
Fig. 6 illustrates the performance of a network of HTM
neurons implementing a high-order sequence memory. The
network used in Fig. 6 consists of 2048 mini-columns with
32 neurons per mini-column. Each neuron has 128 basal
dendritic segments, and each dendritic segment has up to 40
actual synapses. Because this simulation is designed to only
illustrate properties of sequence memory it does not include
apical synapses. The network exhibits all five of the desired
properties for sequence memory listed earlier.
[Figure 6 about here – see end of manuscript]
illustrates on-line
Although we have applied HTM networks to many types of
real-world data, in Fig. 6 we use an artificial data set to more
clearly illustrate the network’s properties. The input is a
stream of elements, where every element is converted to a
2% sparse activation of mini-columns (40 active columns out
of 2048 total). The network learns a predictive model of the
data based on observed transitions in the input stream. In Fig.
6 the data stream fed to the network contains a mixture of
random elements and repeated sequences. The embedded
sequences are six elements long and require high-order
temporal context for full disambiguation and best prediction
accuracy, e.g. “XABCDE” and “YABCFG”. For
this
simulation we designed the input data stream such that the
maximum possible average prediction accuracy is 50% and
this is only achievable by using high-order representations.
Fig. 6A
learning and high-order
predictions. The prediction accuracy of the HTM network
over time is shown in red. The prediction accuracy starts at
zero and increases as the network discovers the repeated
temporal patterns mixed within the random transitions. For
comparison, the accuracy of a first-order network (created by
using only one cell per column) is shown in blue. After
sufficient learning, the high-order HTM network achieves the
maximum possible prediction accuracy of 50% whereas the
first-order network only achieves about 33% accuracy. After
the networks reached their maximum performance the
embedded sequences were modified. The accuracy drops at
that point, but since the network is continually learning it
recovers by learning the new high-order patterns.
Fig. 6B illustrates the robustness of the network. After the
network reached stable performance we inactivated a random
selection of neurons. At up to about 40% cell death there was
minimal impact on performance. This robustness is due to
the noise tolerance described earlier that occurs when a
dendritic segment forms more synapses than necessary to
generate an NMDA spike. At higher levels of cell death the
network performance initially declines but then recovers as
the network relearns the patterns using the remaining
neurons.
4. Discussion
We presented a model cortical neuron that is substantially
different than model neurons used in most artificial neural
networks. The key feature of the model neuron is its use of
active dendrites and thousands of synapses, allowing the
neuron to recognize hundreds of unique patterns in large
populations of cells. We showed that a neuron can reliably
recognize many patterns, even in the presence of large
amounts of noise and variation. In this model, proximal
synapses define the feedforward receptive field of a cell. The
basal and apical synapses depolarize the cell, representing
predictions.
We showed that a network of these neurons will learn a
predictive model of a stream of data. Basal synapses detect
contextual patterns that predict the next feedforward input.
Apical synapses detect feedback patterns that predict entire
sequences. The operation of the neuron and the network rely
on neural activity being sparse. The sequence memory model
learns continuously, uses variable amounts of context to
make predictions, makes multiple simultaneous predictions,
6
relies on local learning rules, and is robust to failure of
network elements, noise, and variation.
Although we refer to the network model as a “sequence
memory”, it is actually a memory of transitions. There is no
representation or concept of the length of sequences or of the
number of stored sequences. The network only learns
transitions between inputs. Therefore, the capacity of a
network is measured by how many transitions a given
network can store. This can be calculated as the product of
the expected duty cycle of an individual neuron (cells per
column/column sparsity) times the number of patterns each
neuron can recognize on its basal dendrites. For example, a
network where 2% of the columns are active, each column
has 32 cells, and each cell recognizes 200 patterns on its
basal dendrites, can store approximately 320,000 transitions
((32/0.02)*200). The capacity scales linearly with the
number of cells per column and the number of patterns
recognized by the basal synapses of each neuron.
Another important capacity metric is how many times a
particular input can appear in different temporal contexts
without confusion. This is analogous to how many times a
particular musical interval can appear in melodies without
confusion, or how many times a particular word can be
memorized in different sentences. If mini-columns have 32
cells it doesn’t mean a particular pattern can have only 32
different representations. For example, if we assume 40
active columns per input, 32 cells per column, and one active
cell per column, then there are 3240 possible representations
of each input pattern, a practically unlimited number.
Therefore, the practical limit is not representational but
memory-based. The capacity is determined by how many
transitions can be learned with a particular sparse set of
columns.
So far we have only discussed cellular layers where all cells
in the network can potentially connect to all other cells with
equal likelihood. This works well for small networks but not
for large networks. In the neocortex, it is well known that
most regions have a topological organization. For example
cells in region V1 receive feedforward input from only a
small part of the retina and receive lateral input only from a
local area of V1. HTM networks can be configured this way
by arranging the columns in a 2D array and selecting the
potential synapses for each dendrite using a 2D probability
distribution centered on the neuron. Topologically organized
networks can be arbitrarily large.
There are several testable predictions that follow from this
theory.
1) The theory provides an algorithmic explanation for the
experimentally observed phenomenon
that overall cell
activity becomes sparser during a continuous predictable
sensory stream (Martin and Schröder, 2013; Vinje and
Gallant, 2002; Yen et al., 2007). In addition, it predicts that
unanticipated inputs will result in higher cell activity, which
should be
correlated vertically within mini-columns.
Anticipated inputs on the other hand will result in activity
that is uncorrelated within mini-columns. It is worth noting
that mini-columns are not a strict requirement of this theory.
The model only requires the presence of small groups of cells
that share feedforward responses and that are mutually
inhibitory. We refer to these groups as mini-columns, but the
the existence of a mechanism
columnar aspect is not a requirement, and the groupings
could be independent of actual mini-columns.
2) A second core prediction of the theory is that the current
pattern of cell activity contains information about past
stimuli. Early experimental results supporting this prediction
have been reported in (Nikolić et al., 2009). Further studies
are required to validate the exact nature of dynamic cell
activity and the role of temporal context in high order
sequences.
3) Synaptic plasticity should be localized to dendritic
segments that have been depolarized via synaptic input
followed a short time later by a back action potential. This
effect has been reported (Losonczy et al., 2008), though the
phenomenon has yet to be widely established.
4) There should be few, ideally only one, excitatory synapses
formed between a given axon and a given dendritic segment.
If an excitatory axon made many synapses in close proximity
onto a single dendrite then the presynaptic cell would
dominate in causing an NMDA spike. Two, three, or even
four synapses from a single axon onto a single dendritic
segment could be tolerated, but if axons routinely made more
synapses to a single dendritic segment it would lead to errors.
Pure Hebbian learning would seem to encourage forming
multiple synapses. To prevent this from happening we
predict
that actively
discourages the formation of a multiple synapses after one
has been established. An axon can form synapses onto
different dendritic segments of the same neuron without
causing problems, therefore we predict this mechanism will
be spatially localized within dendritic segments or to a local
area of an axonal arbor.
5) When a cell depolarized by an NMDA spike subsequently
generates an action potential via proximal input, it needs to
inhibit all other nearby excitatory cells. This requires a fast,
probably single spike,
inhibition. Fast-spiking basket
inhibitory cells are the most likely source for this rapid
inhibition (Hu et al., 2014).
6) All cells in a mini-column need to learn common
feedforward responses. This requires a mechanism
to
encourage all the cells in a mini-column to become active
simultaneously while learning feedforward patterns. This
requirement for mutual excitation seems at odds with the
prior requirement for mutual inhibition when one or more
cells are slightly depolarized. We don’t have a specific
proposal for how these two requirements are met but we
predict a mechanism where sometimes cells in a column are
mutually excited and at other times they are mutually
inhibited.
Pyramidal neurons are common in the hippocampus. Hence,
parts of our neuron and network models might apply to the
hippocampus. However, the hippocampus is known for fast
learning, which is incompatible with growing new synapses,
as synapse formation can take hours in an adult (Holtmaat
and Svoboda, 2009; Knott et al., 2002; Niell et al., 2004;
Trachtenberg et al., 2002). Rapid learning could be achieved
in our model if instead of growing new synapses, a cell had a
multitude of inactive, or “silent” synapses (Kerchner and
Nicoll, 2008). Rapid learning would then occur by turning
silent synapses into active synapses. The downside of this
approach is a cell would need many more synapses, which is
7
metabolically expensive. Pyramidal cells in hippocampal
region CA2 have several times the number of synapses as
pyramidal cells in neocortex (Megías et al., 2001). If most of
these synapses were silent it would be evidence to suggest
that region CA2 is also implementing a variant of our
proposed sequence memory.
It is instructive to compare our proposed biological sequence
memory mechanism to other sequence memory techniques
used in the field of machine learning. The most common
technique is Hidden Markov Models (HMMs) (Rabiner and
Juang, 1986). HMMs are widely applied, particularly in
speech recognition. The basic HMM is a first-order model
and its accuracy would be similar to the first-order model
shown in Fig. 6A. Variations of HMMs can model restricted
high order sequences by encoding high-order states by hand.
More recently, recurrent neural networks, specifically long
short-term memory (LSTM) (Hochreiter and Schmidhuber,
1997), have become popular, often outperforming HMMs.
Unlike HTM networks, neither HMMs nor LSTMs attempt
to model biology in any detail; as such they provide no
insights into neuronal or neocortical functions. The primary
functional advantages of the HTM model over both these
techniques are its ability to learn continuously, its superior
robustness, and its ability to make multiple simultaneous
predictions. A more detailed comparison can be found in S1
Table.
A number of papers have studied spiking neuron models
(Ghosh-Dastidar and Adeli, 2009; Maass, 1997) in the
context of sequences. These models are more biophysically
detailed than the neuron models used in the machine learning
literature. They show how spike-timing-dependent plasticity
(STDP) can lead to a cell becoming responsive to a particular
sequence of presynaptic spikes and to a specific time delay
between each spike (Rao and Sejnowski, 2000; Ruf and
Schmitt, 1997). These models are at a lower level of detail
than the HTM model proposed in this paper. They explicitly
model integration times of postsynaptic potentials and the
corresponding time delays are typically sub-millisecond to a
few milliseconds. They also typically deal with a very small
subset of the synapses and do not explicitly model non-linear
active dendrites. The focus of our work has been at a higher
level. The work presented in this paper is a model of the full
set of synapses and active dendrites on a neuron, of a
networked layer of such neurons and the emergence of a
computationally
sequence memory. An
interesting direction for future research is to connect these
two levels of modeling, i.e. to create biophysically detailed
models that operate at the level of a complete layer of cells.
Some progress is reported in (Billaudelle and Ahmad, 2015),
but there remains much to do on this front.
A key consideration in learning algorithms is the issue of
generalization, or the ability to robustly deal with novel
patterns. The sequence memory mechanism we have outlined
learns by forming synapses to small samples of active
neurons in streams of sparse patterns. The properties of
sparse representations naturally allow such a system to
generalize. Two randomly selected sparse patterns will have
very little overlap. Even a small overlap (such as 20%) is
highly significant and implies that the representations share
significant semantic meaning. Dendritic thresholds are lower
than the actual number of synapses on each segment, thus
sophisticated
8
calculated as follows:
segments will recognize novel but semantically related
patterns as similar. The system will see similarity between
different sequences and make novel predictions based on
analogy.
Recently we showed that our sequence memory method can
learn a predictive model of sensory-motor sequences (Cui et
al., 2015). We also see it is likely that cortical motor
sequences are generated using a variation of the same
network model. Understanding how layers of cells can
perform these different functions and how they work together
is the focus of our current research.
5. Materials and Methods
Here we formally describe the activation and learning rules
for an HTM sequence memory network. There are three
basic aspects to the rules: initialization, computing cell states,
and updating synapses on dendritic segments. These steps are
described below,
some
implementation details.
Notation: Let N represent the number of mini-columns in the
layer, M the number of cells per column, and NM the total
along with notation
and
number of cells in the layer. Each cell can be in an active
state, in a predictive (depolarized) state, or in a non-active
state. Each cell maintains a set of segments each with a
number of synapses. (In this figure we use the term
“synapse” to refer to “potential synapses” as described in the
body of the paper. Thus at any point in time some of the
synapses will have a weight of 0 and some will have a weight
of 1.) At any time step t, the set of active cells is represented
by the M×N binary matrix 𝐀!, where a!"! is the activity of the
i’th cell in the j’th column. Similarly, the M×N binary matrix
𝚷! denotes cells in a predictive state at time t, where π!!! is
the predictive state of the i’th cell in the j’th column.
Each cell is associated with a set of distal segments, 𝐃!", such
that 𝐃!"! represents the d’th segment of the i’th cell in the j’th
NM−1 cells. Each synapse has an associated permanence
value (see Supplemental Fig. 2). Therefore, 𝐃!"! itself is also
an M×N sparse matrix. If there are s potential synapses
associated with the segment, the matrix contains s non-zero
connection threshold. We use 𝐃!"! to denote a binary matrix
elements representing permanence values. A synapse is
considered connected if its permanence value is above a
column. Each distal segment contains a number of synapses,
representing lateral connections from a subset of the other
containing only the connected synapses.
1) Initialization: the network is initialized such that each
segment contains a set of potential synapses (i.e. with non-
zero permanence value) to a randomly chosen subset of cells
in the layer. The permanence values of these potential
synapses are chosen randomly: initially some are connected
(above threshold) and some are unconnected.
2) Computing cell states: All the cells in a mini-column share
the same feed forward receptive fields. We assume that an
inhibitory process has already selected a set of k columns
that best match the current feed forward input pattern. We
denote this set as 𝐖!. The active state for each cell is
(6)
(5)
and
𝜋!"!!!=0
!
𝑫!"!∘𝑨!!! !=𝑚𝑎𝑥! (𝑫!"!∘𝑨!!! !)
∀!∈!!
permanence values by a small value p! and increase the
by a larger value p!:
Reinforcing the above segments is straightforward: we wish
to reward synapses with active presynaptic cells and punish
synapses with inactive cells. To do that we decrease all the
permanence values corresponding to active presynaptic cells
∆𝑫!"!=𝑝!𝑫!"!∘𝑨!!! −𝑝!𝑫!"!
The above rules deal with cells that are currently active. We
also apply a very small decay to active segments of cells that
did not become active. This can happen if segments were
mistakenly reinforced by chance:
∆𝑫!"!=𝑝—𝑫!"! where
The matrices ∆𝐃!"! are added to the current matrices of
𝑎!"! =0 and 𝑫!"!∘𝑨!!! !>𝜃
permanence values at every time step.
Implementation details: in our software implementation, we
make some simplifying assumptions that greatly speed up
simulation time for larger networks. Instead of explicitly
initializing a complete set of synapses across every segment
and every cell, we greedily create segments on a random cell
and initialize potential synapses on that segment by sampling
from currently active cells. This happens only when there is
no match to any existing segment. In our simulations
N=2048,M=32,k=40. We typically connect between
20 and 40 synapses on a segment, and θ is around 15.
threshold of 0.5. p!and p! are small values that are tuned
Permanence values vary from 0 to 1 with a connection
based on the individual dataset but typically less than 0.1.
The full source code for the implementation is available on
Github at https://github.com/numenta/nupic
(7)
𝑎!"! = 1 if 𝑗∈𝑾! and 𝜋!"!!!=1
1 if 𝑗∈𝑾! and
𝜋!"!!!
=0
0 otherwise
!
(2)
The first line will activate a cell in a winning column if it
was previously in a predictive state. If none of the cells in a
winning column were in a predictive state, the second line
will activate all cells in that column. The predictive state for
the current time step is then calculated as follows:
𝜋!"! = 1 if ∃! 𝑫!"!∘𝑨! !>𝜃
0 otherwise
Threshold θ represents the NMDA spiking threshold and ∘
time, if there are more than θ connected synapses with active
represents element-wise multiplication. At a given point in
(3)
presynaptic cells, then that segment will be active (generate
an NMDA spike). A cell will be depolarized if at least one
segment is active.
3) Updating segments and synapses: the HTM synaptic
plasticity rule is a Hebbian-like rule. If a cell was correctly
predicted (i.e. it was previously in a depolarized state and
subsequently became active via feedforward input), we
reinforce the dendritic segment that was active and caused
the depolarization. Specifically, we choose those segments
𝐃!"! such that:
∀!∈!!𝜋!"!!!>0 and 𝑫!"!∘𝑨!!! !>𝜃
(4)
The first term selects winning columns that contained correct
predictions. The second
term selects
those segments
specifically responsible for the prediction.
If a winning column was unpredicted, we need to select one
cell that will represent the context in the future if the current
sequence transition repeats. To do this we select the cell with
the segment that was closest to being active, i.e. the segment
that had the most input even though it was below threshold.
Let D!"! denote a binary matrix containing only the positive
entries in D!"!. We reinforce a segment where the following is
true:
9
6. REFERENCES
Ahmad, S., and Hawkins, J. (2015). Properties of Sparse Distributed Representations and their Application to Hierarchical
Temporal Memory. arXiv:1503.07469 [q-bio.NC]
Antic, S. D., Zhou, W. L., Moore, A. R., Short, S. M., and Ikonomu, K. D. (2010). The decade of the dendritic NMDA spike. J.
Neurosci. Res. 88, 2991–3001. doi:10.1002/jnr.22444.
Billaudelle, S., and Ahmad, S. (2015). Porting HTM Models to the Heidelberg Neuromorphic Computing Platform.
arXiv:1505.02142 [q-bio.NC]
Chklovskii, D. B., Mel, B. W., and Svoboda, K. (2004). Cortical rewiring and information storage. Nature 431, 782–8.
doi:10.1038/nature03012.
Cichon, J., and Gan, W.-B. (2015). Branch-specific dendritic Ca2+ spikes cause persistent synaptic plasticity. Nature 520, 180–5.
Cui, Y., Ahmad, S., Surpur, C., and Hawkins, J. (2015). Maintaining stable perception during active exploration. in Cosyne
Abstracts (Salt Lake City).
Faisal, A. A., Selen, L. P. J., and Wolpert, D. M. (2008). Noise in the nervous system. Nat. Rev. Neurosci. 9, 292–303.
Ghosh-Dastidar, S., and Adeli, H. (2009). Spiking neural networks. Int. J. Neural Syst. 19, 295–308.
Golding, N. L., Jung, H. Y., Mickus, T., and Spruston, N. (1999). Dendritic calcium spike initiation and repolarization are
controlled by distinct potassium channel subtypes in CA1 pyramidal neurons. J. Neurosci. 19, 8789–98.
Hawkins, J., Ahmad, S., and Dubinsky, D. (2011). Cortical learning algorithm and hierarchical temporal memory.
http://numenta.org/resources/HTM_CorticalLearningAlgorithms.pdf
Hawkins, J., and Blakeslee, S. (2004). On Intelligence. New York: Henry Holt and Company.
Hochreiter, S., and Schmidhuber, J. (1997). Long Short-Term Memory. Neural Comput. 9, 1735–1780.
Holtmaat, A., and Svoboda, K. (2009). Experience-dependent structural synaptic plasticity in the mammalian brain. Nat. Rev.
Neurosci. 10, 647–58.
Hu, H., Gan, J., and Jonas, P. (2014). Fast-spiking, parvalbumin+ GABAergic interneurons: From cellular design to microcircuit
function. Science (80-. ). 345, 1255263–1255263.
Kerchner, G. A., and Nicoll, R. A. (2008). Silent synapses and the emergence of a postsynaptic mechanism for LTP. Nat. Rev.
Neurosci. 9, 813–25.
Knott, G. W., Quairiaux, C., Genoud, C., and Welker, E. (2002). Formation of dendritic spines with GABAergic synapses
induced by whisker stimulation in adult mice. Neuron 34, 265–73.
Lamme, V. A., Supèr, H., and Spekreijse, H. (1998). Feedforward, horizontal, and feedback processing in the visual cortex. Curr.
Opin. Neurobiol. 8, 529–35.
Larkum, M. (2013). A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends
Neurosci. 36, 141–51. doi:10.1016/j.tins.2012.11.006.
Larkum, M. E., and Nevian, T. (2008). Synaptic clustering by dendritic signalling mechanisms. Curr. Opin. Neurobiol. 18, 321–
31. doi:10.1016/j.conb.2008.08.013.
Larkum, M. E., Nevian, T., Sandler, M., Polsky, A., and Schiller, J. (2009). Synaptic integration in tuft dendrites of layer 5
pyramidal neurons: a new unifying principle. Science 325, 756–760. doi:10.1126/science.1171958.
Larkum, M. E., Zhu, J. J., and Sakmann, B. (1999). A new cellular mechanism for coupling inputs arriving at different cortical
layers. Nature 398, 338–41.
10
Lashley, K. (1951). “The problem of serial order in behavior,” in Cerebral mechanisms in behavior, ed. L. Jeffress (New York:
Wiley), 112–131.
LeCun, Y., Bengio, Y., and Hinton, G. (2015). Deep learning. Nature 521, 436–444.
Losonczy, A., Makara, J. K., and Magee, J. C. (2008). Compartmentalized dendritic plasticity and input feature storage in
neurons. Nature 452, 436–41. doi:10.1038/nature06725.
Maass, W. (1997). Networks of spiking neurons: the third generation of neural network models. Neural networks 10, 1659–1671.
doi:10.1016/S0893-6080(97)00011-7.
Major, G., Larkum, M. E., and Schiller, J. (2013). Active properties of neocortical pyramidal neuron dendrites. Annu. Rev.
Neurosci. 36, 1–24. doi:10.1146/annurev-neuro-062111-150343.
Major, G., Polsky, A., Denk, W., Schiller, J., and Tank, D. W. (2008). Spatiotemporally graded NMDA spike/plateau potentials
in basal dendrites of neocortical pyramidal neurons. J. Neurophysiol. 99, 2584–601.
Martin, K. A. C., and Schröder, S. (2013). Functional heterogeneity in neighboring neurons of cat primary visual cortex in
response to both artificial and natural stimuli. J. Neurosci. 33, 7325–44.
Megías, M., Emri, Z., Freund, T. F., and Gulyás, A. I. (2001). Total number and distribution of inhibitory and excitatory synapses
on hippocampal CA1 pyramidal cells. Neuroscience 102, 527–40.
Niell, C. M., Meyer, M. P., and Smith, S. J. (2004). In vivo imaging of synapse formation on a growing dendritic arbor. Nat.
Neurosci. 7, 254–60.
Nikolić, D., Häusler, S., Singer, W., and Maass, W. (2009). Distributed fading memory for stimulus properties in the primary
visual cortex. PLoS Biol. 7, e1000260.
Olshausen, B. A., and Field, D. J. (2004). Sparse coding of sensory inputs. Curr. Opin. Neurobiol. 14, 481–487.
doi:10.1016/j.conb.2004.07.007.
Petreanu, L., Mao, T., Sternson, S. M., and Svoboda, K. (2009). The subcellular organization of neocortical excitatory
connections. Nature 457, 1142–5.
Poirazi, P., Brannon, T., and Mel, B. W. (2003). Pyramidal neuron as two-layer neural network. Neuron 37, 989–99.
Polsky, A., Mel, B. W., and Schiller, J. (2004). Computational subunits in thin dendrites of pyramidal cells. Nat. Neurosci. 7,
621–7. doi:10.1038/nn1253.
Rabiner, L., and Juang, B. (1986). An introduction to hidden Markov models. IEEE ASSP Mag. 3, 4–16.
Rah, J.-C., Bas, E., Colonell, J., Mishchenko, Y., Karsh, B., Fetter, R. D., et al. (2013). Thalamocortical input onto layer 5
pyramidal neurons measured using quantitative large-scale array tomography. Front. Neural Circuits 7, 177.
Rao, R. P. N., and Sejnowski, T. J. (2000). Predictive Sequence Learning in Recurrent Neocortical Circuits. in in Advances in
Neural Information Processing Systems 12, 164–170. doi:10.1.1.45.9739.
Ruf, B., and Schmitt, M. (1997). Learning temporally encoded patterns in networks of spiking neurons. Neural Process. Lett., 9–
18. doi:10.1023/A:1009697008681.
Schiller, J., Major, G., Koester, H. J., and Schiller, Y. (2000). NMDA spikes in basal dendrites of cortical pyramidal neurons.
Nature 404, 285–9.
Schiller, J., and Schiller, Y. (2001). NMDA receptor-mediated dendritic spikes and coincident signal amplification. Curr. Opin.
Neurobiol. 11, 343–8.
Spruston, N. (2008). Pyramidal neurons: dendritic structure and synaptic integration. Nat. Rev. Neurosci. 9, 206–21.
Stuart, G. J., and Häusser, M. (2001). Dendritic coincidence detection of EPSPs and action potentials. Nat. Neurosci. 4, 63–71.
11
Trachtenberg, J. T., Chen, B. E., Knott, G. W., Feng, G., Sanes, J. R., Welker, E., et al. (2002). Long-term in vivo imaging of
experience-dependent synaptic plasticity in adult cortex. Nature 420, 788–94.
Vinje, W. E., and Gallant, J. L. (2002). Natural Stimulation of the Nonclassical Receptive Field Increases Information
Transmission Efficiency in V1. J. Neurosci. 22, 2904–2915.
Yen, S.-C., Baker, J., and Gray, C. M. (2007). Heterogeneity in the responses of adjacent neurons to natural stimuli in cat striate
cortex. J. Neurophysiol. 97, 1326–1341. doi:10.1167/7.9.326.
Yoshimura, Y., Sato, H., Imamura, K., and Watanabe, Y. (2000). Properties of horizontal and vertical inputs to pyramidal cells in
the superficial layers of the cat visual cortex. J. Neurosci. 20, 1931–40.
12
Figure 1: Comparison of neuron models. A) The neuron model used in most artificial neural networks has few synapses and no
dendrites. B) A neocortical pyramidal neuron has thousands of excitatory synapses located on dendrites (inset). The co-activation
of a set of synapses on a dendritic segment will cause an NMDA spike and depolarization at the soma. There are three sources of
input to the cell. The feedforward inputs (shown in green) which form synapses proximal to the soma, directly lead to action
potentials. NMDA spikes generated in the more distal basal and apical dendrites depolarize the soma but typically not sufficiently
to generate a somatic action potential. C) An HTM model neuron models dendrites and NMDA spikes with an array of
coincident detectors each with a set of synapses (only a few of each are shown).
13
Figure 2: Representing sequences in cortical cellular layers. A) The neocortex is divided into cellular layers. The panels in this
figure show part of one generic cellular layer. For clarity, the panels only show 21 mini-columns with 6 cells per column. B)
Input sequences ABCD and XBCY are not yet learned. Each sequence element invokes a sparse set of mini-columns, only three
in this illustration. All the cells in a mini-column become active if the input is unexpected, which is the case prior to learning the
sequences. C) After learning the two sequences, the inputs invoke the same mini-columns but only one cell is active in each
column, labeled B’, B’’, C’, C’’, D’ and Y’’. Because C’ and C’’ are unique, they can invoke the correct high-order prediction of
either Y or D.
14
Figure 3: Basal connections to nearby neurons predict the next input. A) Using one of the sequences from Fig. 2, both active
cells (black) and depolarized/predicted cells (red) are shown. The first panel shows the unexpected input A, which leads to a
prediction of the next input B’ (second panel). If the subsequent input matches the prediction then only the depolarized cells will
become active (third panel), which leads to a new prediction (fourth panel). The lateral synaptic connections used by one of the
predicted cells are shown in the rightmost panel. In a realistic network every predicted cell would have 15 or more connections to
a subset of a large population of active cells. B) Ambiguous sub-sequence “BC” (which is part of both ABCD and XBCY) is
presented to the network. The first panel shows the unexpected input B, which leads to a prediction of both C’ and C’’. The third
panel shows the system after input C. Both sets of predicted cells become active, which leads to predicting both D and Y (fourth
panel). In complex data streams there are typically many simultaneous predictions.
15
Figure 4: Feedback to apical dendrites predicts entire sequences.
This figure uses the same network and representations as Fig. 2. Area labeled “apical dendrites” is equivalent to layer 1 in
neocortex; the apical dendrites (not shown) from all the cells terminate here. In the figure, the following assumptions have been
made. The network has previously learned the sequence ABCD as was illustrated in Fig. 2. A constant feedback pattern was
presented to the apical dendrites during the learned sequence, and the cells that participate in the sequence B’C’D’ have formed
synapses on their apical dendrites to recognize the constant feedback pattern.
After the feedback connections have been learned, presentation of the feedback pattern to the apical dendrites is simultaneously
recognized by all the cells that would be active sequentially in the sequence. These cells, shown in red, become depolarized (left
pane). When a new feedforward input arrives it will lead to the sparse representation relevant to the predicted sequence (middle
panel). If a feedforward pattern cannot be interpreted as part of the expected sequence (right panel) then all cells in the selected
columns become active indicative of an anomaly. In this manner apical feedback biases the network to interpret any input as part
of an expected sequence and detects if an input does not match any one of the elements in the expected sequence.
16
Figure 5: Learning by growing new synapses. Learning in an HTM neuron is modeled by the growth of new synapses from a set
of potential synapses. A “permanence” value is assigned to each potential synapse and represents the growth of the synapse.
Learning occurs by incrementing or decrementing permanence values. The synapse weight is a binary value set to 1 if the
permanence is above a threshold.
17
Figure 6: Simulation results of the sequence memory network. The input stream used for this figure contained high-order
sequences mixed with random elements. The maximum possible average prediction accuracy of this data stream is 50%. A)
High-order on-line learning. The red line shows the network learning and achieving maximum possible performance after about
2500 sequence elements. At element 3000 the sequences in the data stream were changed. Prediction accuracy drops and then
recovers as the model learns the new temporal structure. For comparison, the lower performance of a first-order network is
shown in blue. B) Robustness of the network to damage. After the network reached stable performance we inactivated a random
selection of neurons. At up to 40% cell death there is almost no impact on performance. At greater than 40% cell death the
performance of the network declines but then recovers as the network relearns using remaining neurons.
18
∑
!!!!
This table demonstrates the effect of sub-sampling on the probability of a false match using the above equation. The chance of an
error drops rapidly as the sampling size increases. A small number of synapses is sufficient for reliable matching.
𝑛= cell population size
𝑎= number of active cells
𝑠= number of synapses on segment
𝜃= NMDA spike threshold
S1 Text. Chance of Error When Recognizing Large Patterns with a Few Synapses
Formula for calculating chance of error
A non-linear dendritic segment can robustly classify a pattern by sub-sampling (forming synapses to) a small number of cells
from a large population. Assuming a random distribution of patterns, the exact probability of a false match, following is given by
the following equation:
!𝑠𝑏!×!𝑛−𝑠
𝑎−𝑏!
!𝑛𝑎!
Table A: Chance of error due to sub-sampling
𝒔
Table B: Chance of error with addition of 50% noise immunity
to large amounts of noise and pattern variation and still have low probability of a false match. For example, with 𝑠=2𝜃 the system
will be immune to 50% noise. The chance of an error drops rapidly as 𝜃 increases; even with noise a small number of synapses is
𝜽
This table demonstrates robustness to noise. By forming more synapses than required for an NMDA spike, a neuron can be robust
𝑛=200,000
𝑎=2,000
𝜃=𝑠
9.9 × 10!!"
9.8 × 10!!"
9.8 × 10!!"
sufficient for reliable matching.
Probability of false match
6
8
10
𝒔
12
16
20
24
6
8
10
12
Probability of false match
8.7 × 10!!"
1.2 × 10!!"
1.6 × 10!!"
2.3 × 10!!"
𝑛=200,000
𝑎=2,000
Table C: Chance of error with addition of mixing synapses on a dendritic segment
This table demonstrates that mixing synapses for 𝑚 different patterns on a single dendritic segment will still not cause
unacceptable errors. By setting 𝑠=2𝑚𝜃 we can see how a segment can recognize 𝑚 independent patterns and still be robust to
50% noise. It is possible to get very high accuracy with larger 𝑚 by using a slightly higher threshold.
𝜽
𝑛=200,000
𝑎=2,000
Probability of false match
𝒎
𝒔
10
10
10
15
2
4
6
6
40
80
120
120
6.3 × 10!!"
8.5 × 10!!
4.2 × 10!!
1.7 × 10!!"
19
High order sequences
Discovers high order sequence structure
Local learning rules
Continuous learning
Multiple simultaneous predictions
Unsupervised learning
Robustness and fault tolerance
Detailed mapping to neuroscience
Probabilistic model
HTM
Yes
Yes
Yes
Yes
Yes
Yes
Very high
Yes
No
HMMs
Limited
No
No
No
No
Yes
No
No
Yes
LSTM
Yes
Yes
No*
No
No
No
Yes
No
No
S1 Table, Comparison of Common Sequence Memory Algorithms
Table comparing two common sequence memory algorithms (HMM and LSTM) to proposed model (HTM).
* Although weight updated rules are local, LSTMs require computing a global error signal that is then back propagated.
20
|
1903.09671 | 1 | 1903 | 2019-03-22T18:38:12 | Axonal Conduction Velocity Impacts Neuronal Network Oscillations | [
"q-bio.NC",
"cs.NE"
] | Increasing experimental evidence suggests that axonal action potential conduction velocity is a highly adaptive parameter in the adult central nervous system. Yet, the effects of this newfound plasticity on global brain dynamics is poorly understood. In this work, we analyzed oscillations in biologically plausible neuronal networks with different conduction velocity distributions. Changes of 1-2 (ms) in network mean signal transmission time resulted in substantial network oscillation frequency changes ranging in 0-120 (Hz). Our results suggest that changes in axonal conduction velocity may significantly affect both the frequency and synchrony of brain rhythms, which have well established connections to learning, memory, and other cognitive processes. | q-bio.NC | q-bio | Axonal Conduction Velocity Impacts
Neuronal Network Oscillations
Vladimir A. Ivanov
The Computational Brain Lab
Ioannis E. Polykretis
The Computational Brain Lab
Konstantinos P. Michmizos
The Computational Brain Lab
Rutgers, The State University of New Jersey
Rutgers, The State University of New Jersey
Rutgers, The State University of New Jersey
New Brunswick, NJ 08854 USA
[email protected]
New Brunswick, NJ 08854 USA
New Brunswick, NJ 08854 USA
[email protected]
[email protected]
Abstract -- Increasing experimental evidence suggests that
axonal action potential conduction velocity is a highly adaptive
parameter in the adult central nervous system. Yet, the effects of
this newfound plasticity on global brain dynamics is poorly
understood. In this work, we analyzed oscillations in biologically
plausible neuronal networks with different conduction velocity
distributions. Changes of 𝟏 − 𝟐 (ms) in network mean signal
transmission time resulted in substantial network oscillation
frequency changes ranging in 𝟎 − 𝟏𝟐𝟎 (Hz). Our results suggest
that changes in axonal conduction velocity may significantly affect
both the frequency and synchrony of brain rhythms, which have
well established connections to learning, memory, and other
cognitive processes.
Keywords -- Axonal Action Potential Conduction Velocity, Spike
Timing, Oscillations, Synchronization
I. INTRODUCTION
The timing of action potential (AP) arrival is of great
importance to information processing in brain circuits [1].
Experimental studies have revealed that a number of pathways,
such as thalamic pathways to the cortex or the auditory
brainstem, maintain fine-tuned spike time arrival, with sub-
millisecond precision [2,3]. Individual axons primarily control
spike time arrival through their AP conduction velocity, which
is modifiable through various mechanisms including ion
channel densities, axonal structure, and myelinating glia, which
wrap their processes around axons and form the myelin sheath
that changes AP propagation speeds [1]. The prevailing
hypothesis has always been that these mechanisms actively
shape neuronal circuits during development of the central
nervous system (CNS) [4]. Aside from slow homeostatic
adjustments, conduction velocities in the adult brain have been
considered static [4]. However,
the emergence of new
visualization tools has revealed that AP conduction velocities
are highly adaptive in adult neuronal circuits [4].
Axons exhibit conduction velocity plasticity in response to
neuronal activity through multiple mechanisms operating at
different time scales [1]. Unmyelinated axons can adjust their
conduction velocity through axon depolarization and diameter
adjustment, while myelinated axons tend to rely more on their
interaction with the myelin sheath [5,6]. Interestingly, some of
these mechanisms rapidly alter conduction velocity in a matter
of seconds or minutes. For example, depolarization of an
Ioannis E. Polykretis was partially funded by the Onassis Foundation
Scholarship
unmyelinated axonal cell membrane, by prior APs, results in
conduction velocity slowing within a matter of minutes [5].
Similarly, depolarization of oligodendrocytes, the primary
myelinating glial cell in the CNS, increases conduction velocity
in the axons they myelinate from 20 seconds to 3 hours [9].
Moreover, new myelin formation is driven by local neuronal
activity, thereby changing local axonal conduction velocities
over the course of several hours [10]. This last form of myelin
plasticity has been shown to play a critical role in motor
learning, suggesting that the effects of adaptive axonal AP
conduction velocities may translate to network level dynamics.
local
interneuron
In this work, we computationally investigated the effects of
axonal AP conduction velocity distributions on the oscillatory
behavior of neuronal networks. To do so we developed an
Izhikevich-type network that conformed to several generally
observed characteristics of cortical networks including the ratio
between excitatory and inhibitory neurons, sparseness of
connectivity,
lognormal
synaptic weight distributions [8,11,13]. Most essential to our
goal, is that such an Izhikevich network has been shown to
exhibit brain-like rhythms [12]. Our analysis revealed multiple,
nontrivial relationships between network conduction velocity
statistics and the corresponding network oscillation frequency,
and synchrony level. Our results suggest a fascinating possibility
that neuronal networks may change their oscillatory activity in
response to precisely tuned AP propagation delays.
inhibition, and
II. METHODS
We developed Izhikevich-type neuronal networks with all
parameters being static. Each network was initialized and
simulated with a different conduction velocity distribution that
was biologically constrained within the reported range [14]. The
networks were analyzed for the ability of all their neurons to get
entrained
into synchronized activity, namely oscillation
frequencies and levels(extent) of neuronal synchrony.
Our network architecture followed the one previously
reported in [12]. Briefly, our network consisted of 1000
Izhikevich neurons, a simple, semi-empirical model of a cortical
neuron [12]. The neuronal membrane fluctuation was described
by a set of two differential equations:
𝑑𝑣 = 0.04𝑣2 + 5𝑣 + 140 − 𝑢 + 𝐼
𝑑𝑢 = 𝑎(𝑏𝑣 − 𝑢)
(1)
distributions to normal distributions with 𝜇 ∈ [0,20] (ms) and
𝜎2 ∈ [0.5,5.0].
where 𝑣, 𝑢 ∈ ℛ are the fast activation and slow recovery
variables, and 𝑎, 𝑏, 𝑐, 𝑑 ∈ ℛ determined neuron type. Neurons
spiked if 𝑣 > 30 (mV), and variables 𝑣, 𝑢 were reset as,
𝑣 = 𝑐
𝑢 = 𝑢 + 𝑑
(2)
The neuronal population consisted of 80% excitatory, regular
spiking (RS), and 20% inhibitory, fast spiking (FS), neuron
types, per experimental data [11]. Neuron parameters were
obtained from [12]. Network connectivity depended on neuron
type, with each neuron uniformly connected to 10% of the
network [8]. Outgoing connections from excitatory neurons
were connected to both excitatory and inhibitory neurons, while
inhibitory neurons were connected only to excitatory neurons.
We initialized excitatory and inhibitory synaptic weights with
(μ, σ2)exc = (1.67,0.5) and (μ, σ2)inh = (1.49,0.5) lognormal
distributions, preserving the experimentally observed functional
form of the distribution [13]. Lastly, axonal AP conduction
velocities were modeled as discrete time delays, with a
resolution of 1 (ms); same as the network integration timestep.
All inhibitory connections were assigned the minimum delay of
1
[11].
Experimental data shows that the distributions of cortico-
cortical axonal AP conduction delays from various cortical areas
to the midline can be approximated using normal distributions
described by moments ranging in 𝜇 = [3,6] (ms) and σ2 =
[0.5,5.0] with all measured delays largely limited to 0 − 10
(ms) [14]. We doubled this range to 0 − 20 (ms) to account for
the fact that experimental data measured pathway delays only up
to the midline, or half of possible total length. Therefore, for
simulation purposes we restricted excitatory conduction delay
(ms), mimicking
interneuron
inhibition
local
The purpose of the simulation process was to relate
conduction delay distribution to global network dynamics.
Simulations consisted of four steps, repeated for each unique
delay distribution. First, a neuronal network was generated with
a unique conduction delay distribution using the previously
described rules and statistics. Second, the network was launched
and stimulated continuously for 5 (s). Stimulation consisted of
randomly selecting a neuron every millisecond and injecting 20
(pA) of current into it. During network run time, all network
parameters remained constant. Third, spike data was collected
for the last 4 (s) of simulation time. This simulation process was
repeated several times for each conduction delay distribution
simulated to ensure robustness and generalization of our results.
We found the results were repeatable with little to no variation.
Lastly, spectral and synchronization analysis were performed on
the recorded data.
Network oscillation frequency was obtained from recorded
network spike data through spectral analysis. First, each
simulated network's 2-dimensional, 800 × 4000 , spike data
was averaged over all neurons resulting in the 1-dimensional
network activity time series, 𝑆(t), signal. We then applied the
Fast Fourier Transform (FFT) algorithm to 1000 (ms) segments
of the 𝑆(t) signal, to extract its frequency components. Lastly,
the largest frequency component was taken to represent the
oscillation frequency of the corresponding network.
Network synchrony was measured on spike data using a
metric based on variance of time-averaged, neuronal spike
fluctuations [15]. This measure computes the variance of
network level fluctuations, normalized by the average variance
of fluctuations of individual neurons, as described by the
following equation:
)
z
H
(
y
c
n
e
u
q
e
r
F
n
o
i
t
a
l
l
i
c
s
O
k
r
o
w
t
e
N
)
z
H
(
y
c
n
e
u
q
e
r
F
n
o
i
t
a
l
l
i
c
s
O
k
r
o
w
t
e
N
Mean Network Delay (ms)
Mean Network Delay (ms)
Fig. 1. Network oscillation frequency as a function of connection delay
average, displayed for delay variances, 𝜎2 ≥ 2. The relationship is nonlinear
and step-like, with delay variance controlling its shift.
Fig. 2. Network oscillation frequency as a function of connection delay
average, displayed for delay variances, 𝜎2 < 2. The relationship is highly
nonlinear with a double peak form. Delay variance both shifted and noticeably
altered the response curve.
r
e
w
o
P
y
c
n
e
u
q
e
r
F
k
r
o
w
t
e
N
y
n
o
r
h
c
n
y
S
k
r
o
w
t
e
N
Network Delay Variance
Network Delay Variance
Fig. 3. Relative power of network oscillation frequency as a function of delay
variance. Lower variance values resulted in higher relative power levels and
vice versa.
Fig. 4. Extent of network synchrony as a function of delay variance. Lower
variance values resulted in higher increased synchrony and vice versa. (Inset)
Each point corresponds to the rapid network frequency change shown in Fig. 1.
χn =
2
σ𝑆
σ𝑆𝑖
2
=
1
𝑁
〈[𝑆(𝑡)]2〉𝑡 − [〈𝑆(𝑡)〉𝑡]2
∑〈[𝑆𝑖(𝑡)]2〉𝑡 − [〈𝑆𝑖(𝑡)〉𝑡]2
(3)
where 〈. . . 〉𝑡 represents an average over time, 𝑆𝑖(𝑡) is the spike
train over time 𝑡 for neuron 𝑖, and 𝑆(𝑡) is the network spike train
signal averaged over 𝑁 neurons. This measure positively
quantifies network synchrony on a scale 0 to 1.
III. RESULTS
We analyzed the effects of different axonal conduction delay
distributions on network oscillations and synchronization. Our
results revealed a nonlinear relationship between network
oscillation frequency and mean delay of network connections.
Additionally, network delay variance impacted both the form of
the oscillation frequency response and network synchronization.
Network oscillation frequency exhibited a highly sensitive
response to the mean delay of network connections. This
nonlinear relationship is depicted in Fig. 1 and 2, depicting
network oscillation as a function of average network delay.
Oscillations ranged in the biologically relevant 0 - 120 (Hz).
Interestingly, sub-millisecond changes in mean delay produced
frequency changes on the order of tens of Hertz, shown in Fig.
1 and 2. For example, more than 80 (Hz) of oscillation
frequency range is covered by a mere 1 − 2 (ms) change in
network mean conduction delay.
Variance of network delay distribution appeared to control
both the position and form of the network oscillation frequency
response curve. First, variance determined the shift of the
response curve along the mean delay axis. Fig. 1 demonstrates
this relationship most clearly, where higher variance values tend
to shift the response curve towards higher mean delay values and
lower frequency ranges. Conversely, Fig. 2 shows that for
variance values, 𝜎2 < 2 , the form of the response curve
completely changes resulting in multiple high frequency peaks.
Interestingly, the low variance regime resulted in highest
network frequency of approximately 125 (Hz). Additionally,
variance impacted network synchrony.
Our model exhibited an inverse relationship between the
variance of network delay distributions and the degree of
network synchronization. This phenomenon was initially
observed through network oscillation frequency power levels
and network raster plots. Fig. 3 shows that the relative power
.
p
m
A
n
o
r
u
e
N
.
p
m
A
n
o
r
u
e
N
Time (ms)
Fig. 5. Raster plots comparing network oscillations and amplitude for two
different AP conduction delay distributions. [Top] Delay distribution of
(𝜇, 𝜎2) = (8,1), resulted in network oscillation frequency ≈120 (Hz) with high
network activity. [Bottom] Delay distribution of (𝜇, 𝜎2) = (8,3), resulted in
network oscillation frequency ≈ 35 (Hz) with low network activity.
level increases with decreasing delay variance, suggesting that
network dynamics tended towards greater order with minor
frequency components becoming less pronounced. Analysis of
individual network raster plots confirmed this through large
amplitude oscillations for low delay variance values, and small
amplitudes for higher delay variance values, shown in Fig. 5. To
conclusively verify this relationship, we measured network
synchrony, shown in Fig. 4. Network synchrony appeared to be
inversely related to delay variance, while the mean delay
positively shifted the synchrony curve. Interestingly, for delay
distributions with variance, σ2 ≥ 2, the abrupt rise in network
frequency occurred precisely at the cusp of rapidly increasing
synchronization. This can be seen in the inset of Fig. 4, where
the point on each curve corresponds to the mean at which
network frequency rapidly rose as seen in Fig. 1.
IV. DISCUSSION
and
In this paper, we demonstrated the existence of a nonlinear
relationship between neuronal network conduction delay
distribution
and
synchronization. Both the distribution mean and variance
inflicted non-trivial effects on network behavior. Conforming to
experimental evidence our model proposed a new perspective
on the origins of brain rhythms.
frequency,
oscillation
network
Intriguingly,
Our modeling results suggested that the entire biologically
observed network oscillation frequency range of approximately
0 − 100 (Hz), could be partly driven by precise, sub-
millisecond changes in the neuronal network's average axonal
conduction delay.
evidence
corroborates
signal
transmission, where it has been shown that the AP transmission
speed between any two specific neurons is maintained at the
sub-millisecond time scale with high degree of reproducibility
[1]. This hints at the possibility that adaptive signal velocity
mechanisms play a significant role in observed network level
phenomena.
such precise
in neuronal
experimental
tuning
Our computational results further suggested a fascinating
possibility that AP propagation speeds impact global network
dynamics. Given that oscillations and synchronization are
fundamental components of information processing in the brain
[16], understanding the role that neuronal and non-neuronal cells
have in higher cognitive functions is crucial. This challenges the
long-held notion of glial passivity in information processing and
reveals potential roles for non-neuronal cells proposed by us and
others [18,19]. For instance, since oligodendrocytes are now
known to adaptively affect AP velocity, through actively
restructuring white matter [1,17], this study paves the way for
computationally studying the interaction of neuronal and non-
neuronal cells in brain health and disease.
This work supports our ongoing effort to investigate network
AP conduction velocity distributions in the context of other
network parameters such as connectivity, excitation vs.
inhibition ratio, and synaptic weight distributions all of which
are known to affect network level properties. This will enable us
to study the prevalence of our findings in more comprehensive
models of brain cells and networks.
REFERENCES
[1] Seidl, A. H., & Bloedel, V. M. (2014). Regulation of Conduction Time
along Axons. Neuroscience, 0, 126 -- 134.
[2] Salami M., Itami C., Tsumoto T., & Kimura F. (2003). Change of
conduction velocity by regional myelination yields constant latency
irrespective of distance between thalamus and cortex. Proceedings of the
National Academy of Sciences of the United States of America, 100,
6174 -- 6179.
[3] Seidl, A. H., Rubel, E. W., & Harris, D. M. (2010). Mechanisms for
adjusting interaural time differences to achieve binaural coincidence
detection. The Journal of Neuroscience : The Official Journal of the
Society for Neuroscience, 30(1), 70 -- 80.
[4] Debanne D., Campanac E., Bialowas A., Carlier E., & Alcaraz G .(2011)
Axon physiology. Physiol Rev 91(2):555 -- 602.
[5] De Col, R., Messlinger, K., & Carr, R. W. (2012). Repetitive activity
slows axonal conduction velocity and concomitantly
increases
mechanical activation threshold in single axons of the rat cranial dura. The
Journal of Physiology C, 590, 725 -- 735.
[6] Chéreau, R., Saraceno, G. E., Angibaud, J., Cattaert, D., & Nägerl, U. V.
(2017). Superresolution imaging reveals activity-dependent plasticity of
axon morphology linked to changes in action potential conduction
velocity. Proceedings of the National Academy of Sciences of the United
States of America, 114(6), 1401 -- 1406.
[7] Schnitzler, A., & Gross, J. (2005). Normal and pathological oscillatory
communication in the brain. Nature Reviews Neuroscience, 6(4), 285 --
296.
[8] Costa, L. da F., Kaiser, M., & Hilgetag, C. C. (2007). Predicting the
connectivity of primate cortical networks from topological and spatial
node properties. BMC Systems Biology, 1(1), 16.
[9] Yamazaki Y, Fujiwara H, & Kaneko K et al (2014) Short- and long-term
functional plasticity of white matter induced by oligodendrocyte
depolarization in the hippocampus. Glia 62:1299 -- 1312.
[10] Mckenzie, I. A., Ohayon, D., Li, H., Paes De Faria, J., Emery, B.,
Tohyama, K., & Richardson, W. D. (n.d.). Motor skill learning requires
active central myelination. Science 346(6207), 318-322.
[11] Isaacson, J. S., & Scanziani, M. (2011). How Inhibition Shapes Cortical
Activity. Neuron, 72(2), 231 -- 243.
[12] Izhikevich, E. M. (2003). Simple Model of Spiking Neurons. IEEE
Transactions on Neural Networks, 14(6).
[13] Song, S., Sjöström, P. J., Reigl, M., Nelson, S., & Chklovskii, D. B.
(2005). Highly Nonrandom Features of Synaptic Connectivity in Local
Cortical Circuits. PLoS Biology, 3(3), e68.
[14] Tomasi, S., Caminiti, R., & Innocenti, G. M. (2012). Areal Differences in
Diameter and Length of Corticofugal Projections. Cerebral Cortex, 22(6),
1463 -- 1472.
[15] Ginzburg I. & Sompolinsky H., (1994) Theory of correlations in
stochastic neural networks. Phys. Rev. E 50, 3171.
[16] Schnitzler, A., & Gross, J. (2005). Normal and pathological oscillatory
communication in the brain. Nature Reviews Neuroscience, 6(4), 285 --
296.
[17] Chorghay Z., Karadottir R., & Ruthazer E. (2018) White Matter Plasticity
Keeps the Brain in Tune: Axons Conduct While Glia Wrap. Frontiers in
Cellular Neuroscience 12
[18] Polykretis, I., Ivanov, V., and Michmizos, K. P. (2018). A Neural-
Astrocytic Network Architecture: Astrocytic calcium waves modulate
synchronous neuronal activity. Proceedings of
the International
Conference on Neuromorphic Systems, 6, pp. 1-8, Nashville, TN, USA.
[19] Polykretis, I., Ivanov, V., and Michmizos, K. P. (2018). The Astrocytic
Microdomain as a Generative Mechanism for Local Plasticity.
International Conference on Brain Informatics, pp. 1-10, Arlington, TX,
USA.
|
1710.11227 | 2 | 1710 | 2018-01-27T11:28:38 | High-dimensional brain. A tool for encoding and rapid learning of memories by single neurons | [
"q-bio.NC"
] | Codifying memories is one of the fundamental problems of modern Neuroscience. The functional mechanisms behind this phenomenon remain largely unknown. Experimental evidence suggests that some of the memory functions are performed by stratified brain structures such as, e.g., the hippocampus. In this particular case, single neurons in the CA1 region receive a highly multidimensional input from the CA3 area, which is a hub for information processing. We thus assess the implication of the abundance of neuronal signalling routes converging onto single cells on the information processing. We show that single neurons can selectively detect and learn arbitrary information items, given that they operate in high dimensions. The argument is based on Stochastic Separation Theorems and the concentration of measure phenomena. We demonstrate that a simple enough functional neuronal model is capable of explaining: i) the extreme selectivity of single neurons to the information content, ii) simultaneous separation of several uncorrelated stimuli or informational items from a large set, and iii) dynamic learning of new items by associating them with already "known" ones. These results constitute a basis for organization of complex memories in ensembles of single neurons. Moreover, they show that no a priori assumptions on the structural organization of neuronal ensembles are necessary for explaining basic concepts of static and dynamic memories. | q-bio.NC | q-bio | Bull Math Biol manuscript No.
(will be inserted by the editor)
High-dimensional brain
A tool for encoding and rapid learning of memories by single
neurons
Ivan Tyukin · Alexander N. Gorban ·
Carlos Calvo · Julia Makarova ·
Valeri A. Makarov
Received: date / Accepted: date
Abstract Codifying memories is one of the fundamental problems of modern
Neuroscience. The functional mechanisms behind this phenomenon remain
largely unknown. Experimental evidence suggests that some of the memory
functions are performed by stratified brain structures such as, e.g., the hip-
pocampus. In this particular case, single neurons in the CA1 region receive a
highly multidimensional input from the CA3 area, which is a hub for informa-
tion processing. We thus assess the implication of the abundance of neuronal
signalling routes converging onto single cells on the information processing.
We show that single neurons can selectively detect and learn arbitrary infor-
I. Tyukin
University of Leicester, Department of Mathematics, University Road, LE1 7RH, United
Kingdom
Tel.: +44-116-2525106
E-mail: [email protected]
Saint-Petersburg State Electrotechnical University, Saint-Petersburg, Prof. Popova str. 5,
Russia
Alexander N. Gorban
University of Leicester, Department of Mathematics, University Road, LE1 7RH, United
Kingdom
Carlos Calvo
Instituto de Matem´atica Interdisciplinar, Faculty of Mathematics, Universidad Complutense
de Madrid, Avda Complutense s/n, 28040 Madrid, Spain
Julia Makarova
Department of Translational Neuroscience, Cajal Institute CSIC, Madrid, Spain;
Lobachevsky State University of Nizhny Novgorod, Gagarin Ave. 23, 603950 Nizhny Nov-
gorod, Russia
Valeri A. Makarov
Instituto de Matem´atica Interdisciplinar, Faculty of Mathematics, Universidad Complutense
de Madrid, Avda Complutense s/n, 28040 Madrid, Spain;
Lobachevsky State University of Nizhny Novgorod, Gagarin Ave. 23, 603950 Nizhny Nov-
gorod, Russia
E-mail: [email protected]
8
1
0
2
n
a
J
7
2
]
.
C
N
o
i
b
-
q
[
2
v
7
2
2
1
1
.
0
1
7
1
:
v
i
X
r
a
2
Ivan Tyukin et al.
mation items, given that they operate in high dimensions. The argument is
based on Stochastic Separation Theorems and the concentration of measure
phenomena. We demonstrate that a simple enough functional neuronal model
is capable of explaining: i) the extreme selectivity of single neurons to the in-
formation content, ii) simultaneous separation of several uncorrelated stimuli
or informational items from a large set, and iii) dynamic learning of new items
by associating them with already "known" ones. These results constitute a
basis for organization of complex memories in ensembles of single neurons.
Moreover, they show that no a priori assumptions on the structural organiza-
tion of neuronal ensembles are necessary for explaining basic concepts of static
and dynamic memories.
Keywords Neural memories · Single-neuron learning · Perceptron ·
Stochastic Separation Theorems
1 Introduction
The human brain is arguably amongst the most sophisticated and enigmatic
nature creations. Over millions of years it has evolved to amass billions of
neurons, featuring on average 86 × 109 cells [22]. This remarkable figure is
several orders of magnitude higher than that of the most mammals and several
times larger than in primates [21]. Whilst measuring roughly 2% of the body
mass, the human brain consumes about 20% of the total energy [8].
The significant metabolic cost associated with a larger brain in humans,
as opposed to mere body size - a path that great apes might have evolved
[21], must be justified by evolutionary advantages. Some of the benefits may
be related to the development of a remarkably important social life in hu-
mans. This, in particular, requires extensive abilities in formation of complex
memories. Indirectly this hypothesis is supported by the significant difference
among species in the number of neurons in the cortex [20] and the hippocam-
pus [3]. For example, in the CA1 area of the hippocampus there are 0.39× 106
pyramidal neurons in rats, 1.3 × 106 in monkeys, and 14 × 106 in humans.
Evolutionary implications in relation to cognitive functions have been widely
discussed in the literature (see, e.g., [36,42, 43]). Recently, it has been shown
that in humans new memories can be learnt very rapidly by supposedly in-
dividual neurons from a limited number of experiences [25]. Moreover, some
neurons can exhibit remarkable selectivity to complex stimuli, the evidence
that has led to debates around the existence of the so-called "grand mother"
and "concept" cells [38,49,37], and their role as elements of a declarative mem-
ory. These findings suggest that not only the brain can learn rapidly but also
it can respond selectively to "rare" individual stimuli. Moreover, experimen-
tal evidence indicates that such a cognitive functionality can be delivered by
single neurons [25,38,49]. The fundamental questions, hence, are: How is this
possible? and What could be the underlying functional mechanisms?
Recent theoretical advances achieved within the Blue Brain Project show
that the brain can operate in many dimensions [39]. It is claimed that the
High-dimensional brain
3
brain has structures operating in up to eleven dimensions. Groups of neu-
rons can form the so called cliques, i.e., networks of specially interconnected
neurons that generate precise representations of geometric objects. Then the
dimension grows with the number of neurons in the clique. Multidimensional
representation of spatiotemporal information in the brain is also implied in the
concept of generalized cognitive maps (see, e.g., [47, 7, 46]). Within this theory,
spatiotemporal relations between objects in the environment are encoded as
static (cognitive) maps and represented as elements of an n-dimensional space
(n (cid:29) 1). The cognitive maps as information items can be learnt, classified, and
retrieved on demand [48]. However, the questions concerning how the brain
or individual neurons can distinguish among a huge number of different maps
and select an appropriate one remain unknown.
In this work we propose that brain areas with a predominant laminar topol-
ogy and abundant signalling routes simultaneously converging on individual
cells (e.g., the hippocampus) are propitious for a high-dimensional processing
and learning of complex information items. We show that a canonical neuronal
model, the perceptron [41], in combination with a Hebbian-type of learning
may provide answers to the above mentioned fundamental questions. In par-
ticular, starting from stochastic separation theorems [14, 15] we demonstrate
that individual neurons gathering multidimensional stimuli through a suffi-
ciently large number of synaptic inputs can exhibit extreme selectivity either
to individual information items or to groups of items. Moreover, neurons are
capable of associating and learning uncorrelated information items. Thus, a
large number of signalling routes simultaneously converging on a large number
of single cells, as it is widely observed in laminar brain structures, translates
into a natural environment for rapid formation and maintenance of extensive
memories. This is vital for social life, and hence may constitute a significant
evolutionary advantage, albeit, at the cost of high metabolic expenditure.
2 Fundamental problems of encoding memories
Different brain structures, such as, e.g., the hippocampus, have a pronounced
laminar organization. For example the CA1 region of the hippocampus is con-
stituted by a palisade of morphologically similar pyramidal cells oriented with
their main axis in parallel and forming a monolayer (Fig. 1A). The major ex-
citatory input to these neurons comes through Schaffer collaterals from the
CA3 region [1,24,50], which is a hub routing information among many brain
structures. Each CA3 pyramidal neuron sends an axon that bifurcates and
leaves multiple collaterals in the CA1 with dominant parallel orientation (Fig.
1B). This topology allows multiple parallel axons conveying multidimensional
"spatial" information from one area (CA3) simultaneously leave synaptic con-
tacts on multiple neurons in another area (CA1). Thus, we have simultaneous
convergence and divergence of the information content (Fig. 1B, right).
Experimental findings show that multiple CA1 pyramidal cells distributed
in the rostro-caudal direction are activated near-synchronously by assemblies
4
Ivan Tyukin et al.
Fig. 1 General principles of encoding memories by single neurons in laminar structures.
A) Laminar organization of the CA3 and CA1 areas in the hippocampus facilitates multiple
parallel synaptic contacts between neurons in these areas by means of Schaffer collaterals. B)
Axons from CA3 pyramidal neurons bifurcate and pass through the CA1 area in parallel (left
panel) giving rise to the convergence-divergence of the information content (right panel).
Multiple CA1 neurons receive multiple synaptic contacts from CA3 neurons. C) Schematic
representation of three memory encoding schemes. 1) Selectivity. A neuron (shown in yellow)
receives inputs from multiple presynaptic cells that code different information items. It
detects (responds to) only one stimulus (purple trace), whereas rejecting the others. 2)
Clustering. Similar to 1, but now a neuron (shown in pink) detects a group of stimuli
(purple and blue traces) and ignores the others. 3) Acquiring memories. A neuron (shown
in green) learns dynamically a new memory item (blue trace) by associating it with a know
one (purple trace).
of simultaneously firing CA3 pyramidal cells [24,30,5]. Thus, an ensemble of
single neurons in the CA1 can receive simultaneously the same synaptic input
(Fig. 1B, left). Since these neurons have different topology and functional
connectivity [12], their response to the same input can be different. Moreover,
experimental in-vivo results show that long term potentiation can significantly
increase the spike transfer rate in the CA3-CA1 pathway [11]. This suggests
that the efficiency of individual synaptic contacts can be increased selectively.
In this work we will follow conventional and rather general functional repre-
sentation of signalling in the neuronal pathways. We assume that upon receiv-
ing an input, a neuron can either generate a response or remain silent. Forms
CAB123stimulusresponseCA3CA1outputSch. inputCA3Schaffer collateralsone itemlearningretrieval High-dimensional brain
5
of the neuronal responses as well as the definitions of synaptic inputs vary from
one model to another. Therefore, here we adopt a rather general functional
approach. Under a stimulus we understand a number of excitations simulta-
neously (or within a short time window) arriving to a neuron through several
axones and thus transmitting some "spatially coded" information items [6]. If
a neuron responds to a stimulus (e.g., generates output spikes or increases its
firing rate), we then say that the neuron detects the informational content of
the given stimulus.
We follow the standard machine learning assumptions [45], [9]. The stimuli
are generated in accordance with some distribution or a set of distributions
("Outer World Models"). All stimuli that a neuron may receive are samples
from this distribution. The sampling itself may be a complicated process, and
for simplicity we assume that all samples are identically and independently
distributed (i.i.d.). Once a sample is generated, a stimuli sub-sample is inde-
pently selected for testing purposes. If more than one neuron is considered,
we will assume that a rule (or a set of rules) is in place that determines how
a neuron is selected from the set. The rules can be both deterministic and
randomized. In the latter case we will specify this process.
Let us now pose the following fundamental questions related to the in-
formation encoding and formation of memories by single neurons and their
ensembles in laminated brain structures:
1. Selectivity: Detection of one stimulus from a set (Fig. 1C.1). Pick an arbi-
trary stimulus from a reasonably large set such that a single neuron from a
neuronal ensemble detects this stimulus. Then what is the probability that
this neuron is stimulus-specific, i.e., it rejects all the other stimuli from the
set?
2. Clustering: Detection of a group of stimuli from a set (Fig. 1C.2). Within
a set of stimuli we select a smaller subset, i.e., a group of stimuli. Then
what is the probability that a neuron detecting all stimuli from this subset
stays silent for all remaining stimuli in the set?
3. Acquiring memories: Learning new stimulus by associating it with one al-
ready known (Fig. 1C.3). Let us consider two different stimuli s1 and s2
such that for t ≤ t0 they do not overlap in time and a neuron detects s1,
but not s2. In the next interval (t0, t1], t1 > t0 the stimuli start to overlap
in time (i.e., they stimulate the neuron together). For t > t1 the neuron re-
ceives only stimulus s2. Then what is the probability that for some t2 ≥ t1
the neuron detects s2?
These questions are in the core of a broad range of puzzling phenomena re-
ported in [25,38,49]. In what follows we will show that, remarkably, these three
non-trivial fundamental questions can be answered within a simple classical
modeling framework, whereby a neuron is represented by a mere perceptron
equipped with a Hebbian-type of learning.
6
Ivan Tyukin et al.
3 Formal statement of the problem
In this section we specify the information content to be processed by neurons
and define a mathematical model of a generic neuron equipped with synaptic
plasticity. Before going any further' let us first introduce notational agreements
used throughout the text. Given two vectors x, y ∈ Rn, their inner product
(cid:104)x, y(cid:105) is: (cid:104)x, y(cid:105) =(cid:80)n
i=1 xiyi. If x ∈ Rn then (cid:107)x(cid:107) stands for the usual Euclidean
norm of x: (cid:107)x(cid:107) = (cid:104)x, x(cid:105)1/2. By Bn(1) = {x ∈ Rn (cid:107)x(cid:107) ≤ 1} we denote a unit
n-ball centered at the origin; V(Ξ) is the Lebesgue volume of Ξ ⊂ Rn, and
M is the cardinality of a finite set M. Symbol C(D), D ⊆ Rm stands for the
space of continuous real-valued functions on D.
3.1 Information content and classes of stimuli
We assume that a neuron receives and processes a large but finite set of dif-
ferent stimuli codifying different information items:
S = {si}.
(1)
Fig. 2 Codification of high-dimensional information by a neuron. Each of M stimuli com-
prises of the "spatial" information, xi ∈ Rn, (e.g., M images) conducted through n axons
(in yellow) and the temporal part, c(t − τi,j ), reflecting the times of stimuli presentation. A
neuron (in blue) receives the stimuli and generates responses determined by some transfer
function f .
Figure 2 illustrates schematically the information flow. Each individual
stimulus i is modeled by a function s : R × Rn → Rn:
c(t − τi,j),
s(t, xi) = xi(cid:88)j
(2)
times1s2sMx1=x2=...xM=S(t)=MXi=1si(t,xi)... T⌧ij+ T⌧ijf High-dimensional brain
7
where xi ∈ Rn \ {0} is the stimulus content codifying the information to be
transmitted over n individual "axons". An example of an information item
could be an l × k image (see Fig. 2). In this case the dimension of each infor-
mation item is n = l × k.
In Eq. (2) the function c(·) defines the stimulus context, i.e., the time
window when the stimulus arrives to the neuron. For the sake of simplicity we
use a rectangular window:
c(t) =(cid:26) 1,
if t ∈ [0, ∆T ]
0, otherwise,
(3)
where ∆T > 0 is the window length. The time instants of the stimulus pre-
sentations, τi,j, are ordered and satisfy:
τi,j+1 > τi,j + ∆T, ∀j.
(4)
Different stimuli arriving to the neuron are added linearly on the neuronal
membrane. Thus, the overall neuronal input S can be written as:
S(t) =(cid:88)i,j
xic(t − τi,j).
(5)
We assume that the information content of stimuli (5) and (2), i.e., vectors
xi are drawn i.i.d. from some distribution. For convenience, we partition all
information items into two sets:
M = {x1, . . . , xM}, Y = {xM +1, . . . , xM +m},
(6)
where M is large but finite and m ≥ 1 is in general smaller than M . The
set M contains a background content for a given neuron, whereas the set Y
models the informational content relevant to the task at hand. In other words,
to accomplish a static memory task the neuron should be able to detect all
elements from Y and to reject all elements from M.
The sets M and Y give rise to the corresponding subsets of stimuli:
S(M) = {si ∈ S si(·) = s(·, xi), xi ∈ M},
S(Y) = {si ∈ S si(·) = s(·, xi), xi ∈ Y}.
(7)
3.2 Neuronal model
To stay within functional description of the information processing let us con-
sider the most basic class of model neurons, a perceptron [41]. A single neuron
receives a stimulus s(t, x) through n synaptic inputs (Fig. 2) and its membrane
potential, y ∈ R, is given by
y(s, w) = (cid:104)w, s(cid:105),
(8)
8
Ivan Tyukin et al.
where w ∈ Rn is a vector of the synaptic weights. The neuron generates a
response, v ∈ R, according to:
v(s, w, θ) = f (y(s, w) − θ),
(9)
where θ ∈ R is the "firing" threshold and f : R → R is the transfer function
(Fig. 2): f ∈ C(R), f is locally Lipschitz, f (u) = 0 for u ∈ (−∞, 0], and
f (u) > 0 for u ∈ (0,∞).
Model (8), (9) captures the summation of postsynaptic potentials and the
threshold nature of the neuronal activation but disregards the specific dynam-
ics accounted for in other more advanced models. Nevertheless, as we will
show in Sect. 4, this phenomenological model is already sufficient to explain
the fundamental properties of information processing discussed in Sect. 2.
3.3 Synaptic plasticity
In addition to the basic neuronal response mechanism (Sect. 3.2), we also
model the synaptic plasticity. The description adopted here relies on the neu-
ronal firing rate and Hebbian learning. Such a learning rule implies that the
dynamics of w should depend on the product of the input signal, s, and the
neuronal output, v. We thus arrive to a modified classical Oja rule [34]:
w = αv(s, w, θ)y(s, w) (s − wy(s, w)) ,
w(t0) = w0 ∈ Rn, w0 (cid:54)= 0,
(10)
where α > 0 defines the relaxation time. The multiplicative term v in (10)
ensures that plastic changes of w occur only when an input stimulus evokes a
non-zero neuronal response. The fact that w0 (cid:54)= 0 reflects the assumption that
synaptic connections have already been established, albeit their efficacy could
be subjected to plastic changes. In addition to capturing general principle
of the classical Hebbian rule, model (10) guarantees that synaptic weights
w are bounded in forward time (see Appendix A) and hence conforms with
physiological plausibility.
4 Formation of memories in high-dimensions
In Sect. 2 we formulated three fundamental problems of organization of mem-
ories in laminar brain structures. Let us now show how they can be treated
given that pyramidal neurons operate in high dimensions.
To formalize the analysis let U be a subset of the stimulus set S. A neuron
(8), (9) parameterized by (w, θ) partitions the set U into the following subsets:
Activated(U, (w, θ)) ={si ∈ U ∃ t≥t0 : v(si(t), w, θ) > 0},
Silent(U, (w, θ)) ={si ∈ U v(si(t), w, θ) = 0 ∀ t ≥ t0}.
(11)
The first set corresponds to the stimuli detected by the neuron, while the
second one collects background stimuli.
High-dimensional brain
9
4.1 Extreme selectivity of a single neuron to single stimuli
Consider the case when the set Y in (6) contains only one element, i.e. Y = 1,
Y = {xM +1}, whereas the set M is allowed to be sufficiently large (M =
M (cid:29) 1). Let us also assume that the stimuli with different information content,
s(·, xi), do not overlap in time, i.e., we present them to a neuron one by one.
For a given non-zero xM +1 ∈ Y and stimulus s(·, xM +1) such that it
is not identically zero for t ≥ t0 we can always construct a neuron which
would generate a non-zero response to the stimulus s(·, xM +1) at some t ≥ t0.
In other words, s(·, xM +1) ∈ Activated(S(Y), (w, θ)). Mathematically such a
neuron can be defined as follows. Let
w∗ =
(12)
xM +1
(cid:107)xM +1(cid:107)
.
Then the space from which the synaptic weights are chosen can be represented
as a direct sum of the one-dimensional linear subspace L(cid:107)(w∗) spanned by w∗
and an (n − 1)-dimensional subspace L⊥(w∗) of Rn that is orthogonal to w∗.
In this representation, if a neuron with the synaptic weight w generates a non-
zero response to s(·, xM +1), then the coupling weight w∗ = (cid:104)w, w∗(cid:105) should
satisfy the following condition (Fig. 3, green area):
w∗ >
Indeed, such a choice is equivalent to
θ
(cid:107)xM +1(cid:107)
.
v(xM +1, w, θ) = f (w∗(cid:107)xM +1(cid:107) − θ) > 0,
which in turn implies that v(s(t, xM +1), w∗, θ) > 0 at some t and vice-versa.
Fig. 3 Selection of neuronal parameters ¯θ = θ/(cid:107)xM +1(cid:107) and w∗, such that the neuron re-
sponds to the relevant information xM +1. Neurons corresponding to points within the green
area detect the stimulus xM +1. Brown areas show projections of hypercylinders defined in
Theorem 1 for D1 = 0.3, D2 = 0.1, D3 = 0.03 and (cid:107)xM +1(cid:107) = 0.6.
Once a neuron that detects relevant information item, i.e. xM +1, is speci-
fied we can proceed with assessing its selectivity properties.
⌦D2⌦D11/211⌦D3¯✓~xw⇤ 10
Ivan Tyukin et al.
Definition 1 (Neuronal Selectivity) We say that a neuron is selective to
the information content Y iff it detects the relevant stimuli from the set S(Y)
and ignores all the others from the set S(M).
The notion of selectivity, as stated in Definition 1, could be relaxed to ac-
count for partial detection and rejection of information content from Y and
M, respectively. This naturally gives rise to various levels of neuronal selec-
tivity determined, for instance, by the proportion of elements from M that
correspond to stimuli that have been rejected. As we will see below, different
admissible pairs (w, θ) (Fig. 3) produce different selectivity levels. The closer
to the bisector, the higher the selectivity. One can pick an arbitrary firing
threshold θ ≥ 0 and select the synaptic efficiency at t = t0 as:
> 0, w⊥ ∈ L⊥.
w∗ + w⊥,
w(t0) =
(13)
θ +
(cid:107)xM +1(cid:107)
It can be shown (see Appendix A) that if the stimulus s(·, xM +1) is persistent
over time and w(t0) satisfies (13) then synaptic efficiency w(t, w0) converges
asymptotically (as t → ∞) to:
w∞ =(cid:40) w∗,
θ
(cid:107)xM +1(cid:107)
∞ is an element of L⊥.
where w⊥
w∗ + w⊥
∞,
if θ < (cid:107)xM +1(cid:107)
if θ ≥ (cid:107)xM +1(cid:107),
(14)
Fig. 4 Example of selective neuronal responses to stimulation with different (30×38)-pixels
images (only first few stimulus are shown in the time line). Each neuron responds to its own
(relevant) stimulus only and rejects the other (background) stimuli.
Figure 4 shows typical responses of neurons parameterized by different pairs
(w, θ) and subjected to stimulation by different information items xi. Here xi
correspond to (30× 38)-pixels color images (i.e., xi ∈ R3420). Firing thresholds
θ have been chosen at random, and weights w have been set in accordance
ty3(t)y2(t)y1(t)✓1✓3✓2S(t)ttt High-dimensional brain
11
with (13) with the first three images serving as the relevant information items
for the three corresponding neurons. No plastic changes in w were allowed.
The neurons detect their own (relevant) stimuli, as expected. Moreover, they
do not respond to the stimulation by other background information items (4
out of 103 images are shown in Fig. 4). Thus, the neurons indeed exhibit high
stimulus selectivity.
The following theorem provides theoretical justification for these observa-
tions.
Theorem 1 Let elements of the sets M and Y be i.i.d. random vectors drawn
from the equidistribution in Bn(1). Consider the sets of stimuli S(M) and
S(Y) specified by (7). Let (w, θ) be the neuron parameters such that
sM +1 ∈ Activated(S(Y), (w, θ)) and 0 < θ < (cid:107)w(cid:107).
Then:
1. The probability that the neuron is silent for all background stimuli si ∈
S(M) is bounded from below by:
P (si ∈ Silent(S(M), (w, θ)) ∀si ∈ S(M)(cid:12)(cid:12) w, θ) ≥
1
2(cid:18)1 −
θ2
2(cid:35)M
(cid:107)w(cid:107)2(cid:19) n
.
≥(cid:34)1 −
(15)
2. There is a family of sets parametrized by D (0 < D < min{ 1
2 ,(cid:107)xM +1(cid:107)}):
where w∗ = xM +1/(cid:107)xM +1(cid:107), such that sM +1 ∈ Activated(S(Y), (w, θ)), for
(w, θ) ∈ ΩD and
ΩD =(cid:110)(w, θ)(cid:12)(cid:12) (cid:107)w − w∗(cid:107) < D, D ≤ (cid:107)xM +1(cid:107) − θ ≤ 2D(cid:111),
P(cid:0)si ∈ Silent(S(M), (w, θ)) ∀si ∈ S(M)(cid:12)(cid:12) ∀(w, θ) ∈ ΩD(cid:1) ≥
ρ(ε, D)
(16)
(17)
n
≥ max
ε∈(0,1−2D)
1
2
(1 − (1 − ε)n)(cid:20)1 −
ρ(ε, D) = 1 −(cid:18) 1 − ε − 2D
1 + D (cid:19)2
.
2(cid:21)M
where
The proof is provided in Appendix B.
Remark 1 For an admissible fixed D > 0, the volume V(ΩD) > 0. Therefore,
the estimate provided by Theorem 1 is robust to small perturbations of (w, θ),
and slight fluctuations of neuronal characteristics are not expected to affect
neuronal functionality.
12
Ivan Tyukin et al.
Remark 2 Theorem 1 (part 2) specifies a non-iterative procedure for construct-
ing sets of selective neurons. Such neurons detect given stimuli and reject the
others, with high probability. Figure 3 (in brown) shows examples of three pro-
jections of the hypercylinders (16) ensuring robust selective stimulus detection.
The smaller is the cylinder, the higher is the selectivity.
To illustrate Theorem 1 numerically we fixed the neuronal dimensionality
parameter n and generated two random sets of information items comprising of
103 elements each, i.e. {xi}103
i=1. One set was sampled from the equidistribution
in a unit ball Bn(1) centered at the origin (i.e. (cid:107)xi(cid:107)2 ≤ 1), and the other from
the equidistribution in the hypercube (cid:107)xi(cid:107)∞ ≤ 1 (a product distribution).
For each set of informational items, a neuronal ensemble of 103 single neurons
parameterized by (wi, θi) was created. Each neuron was assigned fixed firing
threshold θi = 0.5, i = 1, . . . , 103, whereas the synaptic efficiencies were set
as wi = (θi + )xi/(cid:107)xi(cid:107), = 0.05. For these neuronal ensembles and their
corresponding stimuli sets we evaluated output of each neuron and assessed
the neuronal selectivity (see Def. 1). The procedure was repeated 10 times.
This was followed by evaluation of the frequencies of selective neurons in the
pool for each n.
Fig. 5 Extreme selectivity to stimuli and memory capacity of single neurons. A) Stimulus
selectivity vs the neuron dimension. The selectivity index steeply increases for n ∈ [10, 20].
For n > 20 practically all neurons become selective to a set of 103 random stimuli. B)
Memory capacity with reliability 0.95 of a neuronal ensemble vs the neuron dimension.
For both types of stimuli the memory capacity grows exponentially (straight lines show
regressions).
Figure 5A shows frequencies of selective neurons in an ensemble, for 103
stimuli taken from: i) a unit ball (red), ii) a hypercube (blue), and iii) the es-
timate provided by Theorem 1 (dashed). For n small (n < 6) neurons exhibit
no selectivity, i.e., they confuse different stimuli and generate nonspecific re-
sponses. As expected, when neuronal dimensionality, n, increases, the neuronal
selectivity increases rapidly; and at around n = 20 it approaches 100%.
ABselectivity(%)neurondimension,nneurondimension,nmemorycapacity10110210310410548121620102030020406080100hypercubeunitball100.33n100.14nhypercubetheoreticalunitball High-dimensional brain
13
4.2 Extreme selectivity of a single neuron and ensemble memory capacity
The property of a neuron to respond selectively to a single element from a large
set of stimuli can be related to the notion of memory capacity of a neuronal
ensemble comprising of a set of selective neurons.
Recall that in the framework of associative memory [23], for each informa-
tional item (pattern) xi from the set M there is a vicinity Vi associated with
xi and corresponding to all admissible perturbations of xi. Suppose that for
each xi there is a neuron in the ensemble that is activated for all stimuli with
informational content x in Vi and is silent for all other stimuli, i.e. for stimuli
with x in ∪j(cid:54)=iVj. The maximal size of the set M for which this property holds
will be referred to as the (absolute) memory capacity of the ensemble (cf. [23],
[4], [29]).
This conventional mechanistic definition of memory capacity, however, is
too restrictive to account for variability and uncertainty that biological neu-
ronal ensembles and systems are to deal with. Indeed, informational items
themselves may bear a degree of uncertainty resulting in that Vi ∩ Vj (cid:54)= ∅
for some j, i, i (cid:54)= j. Furthermore, errors in memory retrievals are known to
occur in classical artificial associative memory models too (see e.g., [23], [2],
[29]). To be able to formally quantify such errors in relation to the number of
informational items an ensemble is to store, we extend the classical notion as
follows.
Suppose that for each xi there is a neuron in the ensemble that is activated
for all stimuli with informational content x ∈ Vi and, with probability φ, is
silent for all stimuli with x ∈ Vj, j (cid:54)= i. The maximal size of the set M
for which this property holds will be referred to as the memory capacity with
reliability φ of the ensemble.
Assuming that Vi are sufficiently small an estimate of the memory capacity
with reliability φ of a neuronal ensemble follows from Theorem 1.
Corollary 1 Let elements of the sets M and Y be i.i.d. random vectors drawn
from the equidistribution in Bn(1). Consider the set of stimuli S(M) as defined
in (7). Then for a given fixed φ ∈ (0, 1) the maximal size M of the stimuli set
S(M) for which the following holds
grows at least exponentially with the neuronal dimension n:
P (si ∈ Silent(S(M), (w, θ)) ∀si ∈ S(M)(cid:12)(cid:12) w, θ) ≥ φ
(cid:112)(cid:107)w(cid:107)2 − θ2(cid:35) > 0.
M > − ln (φ) (2eαn − 1) , where α = ln(cid:34)
(cid:107)w(cid:107)
(18)
The proof is given in Appendix C.
Figure 5B illustrates how the memory capacity with reliability φ grows
with neuronal dimension n. For each neuronal dimension n we generated i.i.d.
samples M with M = M from the equidistribution in Bn(1) and the n-
cube [−1, 1]n. For each sample, we defined neuronal ensembles comprising of
14
Ivan Tyukin et al.
M neurons with synaptic weights wi = xi/(cid:107)xi(cid:107) and thresholds θi = 0.5,
and calculated the proportion of neurons in the ensemble that are activated
by each stimulus. If the proportion was smaller than 0.05 of the total num-
ber of neurons, we incremented the value of M , generated a new sample M
with increased cardinality M , and repeated the experiment. The values of M
corresponding to samples at which the process stopped have been recorded
and retained. These constituted empirical estimates of the maximal number
of stimuli for which the proportion of neurons responding to a single stimulus
is at most 0.05 = 1− φ. Figure 5B shows empirical means of such numbers for
the unit ball and in the hypercube. As follows from these observations, mem-
ory capacity grows exponentially with the neuron dimension in both cases.
Such a fast growth can easily cover quite exigent memory necessities.
4.3 Selectivity of a single neuron to multiple stimuli
To organize memories, the ability to associate different information items is
essential (Fig. 1C2). To determine if such associations are feasible at the level
of single neurons we assess neuronal selectivity to multiple stimuli. In par-
ticular, we consider the set Y [Eq. (6)] containing m > 1 random vectors:
Y = {xM +1, . . . , xM +m}. As in Sect. 4.1, here we assume that all stimuli do
not overlap in time and arrive to the neuron separately. The question of interest
is: Can we find a neuron [i.e., parameters (w, θ)], such that it would generate
a non-zero response to all si ∈ S(Y) and, with high enough probability, would
be silent to all si ∈ S(M)?
Below we will show that this is indeed possible, provided that the neuronal
dimensionality, n, is large enough. Moreover, the separation can be achieved
by a neuron with the vector of synaptic weights, w = w∗, closely aligned with
the mean vector of the stimulus set Y:
xM +i, w∗ =
.
(19)
¯x =
1
m
m(cid:88)i=1
¯x
(cid:107)¯x(cid:107)
This vector points to the center of the group to be separated from the set M. In
low dimensions, e.g. when n = 2, such functionality appears to be extremely
unlikely. However, high dimensional neurons can accomplish this task with
probability close to one. Formal statement of this property is provided in
Theorem 2.
Theorem 2 Let elements of the sets M and Y be i.i.d. random vectors drawn
from the equidistribution in Bn(1). Consider the sets of stimuli S(M) and
S(Y) specified by (7) and let D, ε, δ ∈ (0, 1) be chosen such that
Let w∗ = ¯x/(cid:107)¯x(cid:107) and consider the set:
θ∗ =
(1 − ε)3 − δ(m − 1)
(cid:112)m(1 − ε)[1 − ε + δ(m − 1)] ∈ (D, 1).
ΩD =(cid:110)(w, θ)(cid:12)(cid:12) (cid:107)w − w∗(cid:107) < D, θ ∈ (0, θ∗ − D](cid:111).
(20)
High-dimensional brain
Then
P(cid:16)[si ∈ Activated(S(Y), w, θ) ∀ si ∈ S(Y)] &
[si ∈ Silent(S(M), w, θ) ∀ si ∈ S(M)](cid:12)(cid:12)(cid:12) (w, θ) ∈ ΩD(cid:17) ≥ p(ε, δ, D, m),
2(cid:21)M
m−1(cid:89)d=1(cid:16)1 − d(cid:0)1 − δ2(cid:1) n
p(ε, δ, D, m) =(1 − (1 − ε)n)m
θ2
2(cid:17)(cid:20)1 −
1
2
∆
n
where
15
(21)
,
∆ = 1 −
(1 + D)2 .
The proof is provided in Appendix D. The theorem admits the following corol-
lary.
Corollary 2 Suppose that the conditions of Theorem 2 hold. Let θ∗ > 2D and
consider the set:
Then
Ω∗D =(cid:110)(w, θ)(cid:12)(cid:12) (cid:107)w − w∗(cid:107) < D, θ ∈ [θ∗ − 2D, θ∗ − D](cid:111).
P(cid:16)[si ∈ Activated(S(Y), w, θ) ∀ si ∈ S(Y)] &
[si ∈ Silent(S(M), w, θ) ∀ si ∈ S(M)](cid:12)(cid:12)(cid:12)(w, θ) ∈ Ω∗D(cid:17) ≥
2(cid:17)(cid:20)1 −
2(cid:21)M
m−1(cid:89)d=1(cid:16)1 − d(cid:0)1 − δ2(cid:1) n
(1 − (1 − ε)n)m
∆ = 1 −(cid:18) θ∗ − 2D
1 + D (cid:19)2
1
2
∆
.
,
n
(22)
Remark 3 Estimates (21), (22) hold for all feasible values of ε and δ. Maxi-
mizing the r.h.s of (21), (22) over feasible domain of ε, δ provides lower-bound
"optimistic" estimates of the neuron performance.
Remark 4 The term θ∗ in Theorem 2 and Corollary 2 is an upper bound for
the firing threshold θ. The larger is the value of θ, the higher is the neuronal se-
lectivity to multiple stimuli. The value of θ∗, however, decays with the number
of stimuli m.
The extent to which the decay mentioned in Remark 4 affects neuronal se-
lectivity to a group of stimuli depends largely on the neuronal dimension, n.
Note also that the probability of neuronal selective response to multiple stim-
uli, as provided by Theorem 2, can be much larger if elements of the set Y are
spatially close to each other or positively correlated [44] (see also Lemma 4 in
Appendix F).
16
Ivan Tyukin et al.
Remark 5 Similarly to the case considered in Corollary 1, the maximal size
of the stimuli set S(M) for which selective response is ensured, with some
fixed probability, grows exponentially with dimension n. Indeed, denoting φ =
(1 − z)M , letting z = 1/2∆n/2 (with ∆ defined in Theorem 2) and invoking
(34), (35) from the proof of Corollary 1, we observe that
M > − ln(φ)(z−1 − 1) = − ln(φ)(2eβn − 1), β = ln
1 + D
Similar estimate can be provided for the case considered in Corollary 2.
Thus, for M = S(M) ≤ M , the r.h.s. of (21) is bounded from below by
.
(cid:112)(1 + D)2 − θ2
m−1(cid:89)d=1(cid:16)1 − d(cid:0)1 − δ2(cid:1) n
2(cid:17) φ.
(1 − (1 − ε)n)m
Fig. 6 Selectivity of a single neuron to multiple stimuli. Panel (A) corresponds to the case
when the informational content vectors, xi, are sampled from the equidistribution in the
unit ball Bn(1), and panel (B) corresponds to the equidistribution in the n-cube centered
in the origin. In both cases the neuronal selectivity approaches 100% when the dimension n
grows. In (A) dashed curves show the estimates provided by Theorem 2. Parameter values:
ε = 0.01, D = 0.001, δ = (1 − ε)/2(m − 1), θ = θ∗ − D.
To illustrate Theorem 2 we conducted several numerical experiments. For
each n we generated M = 103 of background information items xi (the set M)
and m = 2, 5, 8 relevant vectors (the sets Y). In the first group of experiments
all M +m i.i.d. random vectors were chosen from the equidistribution in Bn(1).
Neuronal parameters were set in accordance with Theorem 2 (i.e., Eqs. (19) –
(21)). Figure 6A illustrates the results.
Similarly to the case of neuronal selectivity to a single item (Fig. 5A), we
observe a steep growth of the selectivity index with the neuronal dimension.
The sharp increase occurs, however, at significantly higher dimensions. The
100300400neuron dimension,n020406080100100200300400200020406080100m=2m=5m=8m=2m=5m=8selectivity in ball, (%)selectivity in cube, (%)neuron dimension,nAB High-dimensional brain
17
number of random and uncorrelated stimuli, m, to which a neuron should be
able to respond selectively is fundamentally linked to the neuron dimension-
ality. For example, the probability that a neuron is selective to m = 5 random
stimuli becomes sufficiently high only at n > 400. This contrasts sharply with
n = 120 for m = 2.
Our numerical experiments also show that the firing threshold specified
in Theorem 2 for arbitrarily chosen fixed values of δ and ε is not optimal in
the sense of providing the best possible probability estimates. Playing with θ
one can observe that the values of n at which neuronal selectivity to multiple
stimuli starts to emerge are in fact significantly lower than those predicted by
Eq. (22). This is not surprising. First, since estimate (22) holds for all admis-
sible values of δ and ε, it should also hold for the maximizer of p(ε, δ, D, m).
Second, the estimate is conservative in the sense that it is based on conserva-
tive estimates of the volume of spherical cups Cn (see, e.g., proof of Theorem
1). Deriving more accurate numerical expressions for the latter is possible,
although at the expense of simplicity.
To demonstrate that dependence of the selectivity index on the firing
threshold is likely to hold qualitatively for broader classes of distributions
from which the sets M and Y are drawn, we repeated the simulation for the
equidistribution in an n-cube centered at the origin. In this case, Theorem 2
does not formally apply. Yet, an equivalent statement can still be produced
(cf. [14]). In these experiments synaptic weights were set to w = ¯x/(cid:107)¯x(cid:107) and
θ = 0.5(cid:107)¯x(cid:107). The results are shown in Fig. 6B. The neuron's performance in the
cube is markedly better than that of in Bn(1). Interestingly, this is somewhat
contrary to expectations that might have been induced by our earlier experi-
ments (shown in Fig. 5) in which neuronal selectivity to a single stimulus was
more pronounced for Bn(1).
Overall, these results suggest that single neurons can indeed separate ran-
dom uncorrelated information items from a large set of background items with
probability close to one. This gives rise to a possibility for a neuron to respond
selectively to various arbitrary uncorrelated information items simultaneously.
The latter property provides a natural mechanism for accurate and precise
grouping of stimuli in single neurons.
4.4 Dynamic memory: Learning new information items by association
In the previous sections we dealt with a static model of neuronal functions, i.e.
when the synaptic efficiency w either did not change at all or the changes were
negligibly small over large intervals of stimuli presentation. In the presence of
synaptic plasticity (10), the latter case corresponds to 0 ≤ α (cid:28) 1 in (10).
In this section we explicitly account for the time evolution of the synaptic
efficiency, w(t, w0) [Eq. (10)]. As we will see below, this may give rise to
dynamic memories in single neurons.
As before, we will deal with two sets of stimuli, the relevant one, S(Y),
and the background one, S(M). We will consider two time epochs: i) Learning
18
Ivan Tyukin et al.
phase and ii) Retrieval phase. Within the learning phase we assume that all
stimuli from the set S(Y) arrive to a neuron completely synchronized, i.e.:
τM +1,j = τM +2,j = ··· = τM +m,j, ∀ j.
(23)
Such a synchronization could be interpreted as a mechanism for associating or
grouping different uncorrelated information items for the purposes of memo-
rizing them at a later stage.
The dynamics of the synaptic weights for t ≥ t0 is given Eq. (10) with the
input signal s replaced with:
¯s(t) =
m(cid:88)i=1
sM +i(t).
Let w0 = w(t0) and θ satisfy the following condition:
∃ sk ∈ S(Y) such that sk ∈ Activated(S(Y), w0, θ)
si ∈ Silent(S, w0, θ) for all si ∈ S \ {sk}.
(24)
(25)
Thus, at t = t0 only one information item is "known" to the neuron. All other
relevant items from the set Y are "new" in the sense that the neuron rejects
them at t = t0. Theorem 1 specifies the sets of neuronal parameters w0, θ for
which condition (25) holds with probability close to one if n is large enough.
The question is: What is the probability that, during the learning phase the
synaptic weights w(t, w0) evolve in time so that the neuron becomes responsive
to all si ∈ S(Y) whilst remaining silent to all si ∈ S(M) (Fig. 1C.3)? In other
words, the neuron learns new items and recognizes them in the retrieval phase.
The following theorem provides an answer to this question.
Theorem 3 Let elements of the sets M and Y be i.i.d. random vectors drawn
from the equidistribution in Bn(1). Consider the sets of stimuli S(M) and
S(Y) specified by (7). Let (23) hold, the dynamics of neuronal synaptic weights
satisfy (10), (24), and (w0, θ) be chosen such that condition (25) is satisfied.
Pick ε, δ ∈ (0, 1) such that
(1 − ε)3 > δ(m − 1).
Moreover, suppose that
1. There exist L, κ > 0 such that
(cid:90) t+L
t
v(¯s(τ ), w(τ, w0), θ)(cid:104)¯s(τ ), w(τ, w0)(cid:105)2dτ > κ, ∀ t ≥ t0.
2. The firing threshold, θ, satisfies
0 < θ <
(1 − ε)3 − δ(m − 1)
(cid:112)m(1 − ε)[(1 − ε) + δ(m − 1)]
= θ∗.
High-dimensional brain
19
Then for, any 0 < D ≤ θ∗ − θ, there is t1(D) > t0 such that
P ([S(Y) ∈ Activated(S, w(t, w0), θ)] & [S(M) ∈ Silent(S, w(t, w0), θ)]) ≥
(1 − (1 − ε)n)m
2(cid:35)M
(1 + D)2(cid:19) n
m−1(cid:89)d=1(cid:16)1 − d(cid:0)1 − δ2(cid:1) n
θ2
2(cid:17)(cid:34)1 −
1
2(cid:18)1 −
for all t ≥ t1(D).
The proof is provided in Appendix E.
Fig. 7 Dynamic memory: Learning new information items by association. A) Example of
the dynamic association of a known stimulus (neuron's response to the known stimulus is
shown by green curve) and a new one (neuron's response shown by orange curve). Two
relevant stimuli out of 502 are learnt by the neuron. At t ≈ 2 (red circle) the orange curve
crosses the threshold (red dashed line) and stays above it for t > 2. Thus the neuron detects
the corresponding stimulus for t > 2. B) Same as in A but for m = 4 and m = 12. Parameter
values: ε = 0.01, D = 0.001, δ = (1 − ε)3/2(m − 1), α = 1, M = 500, θ = θ∗ − D, n = 400.
Figure 7 illustrates the theorem numerically. First we assumed that the
relevant set Y consists of m = 2 items. One of them is considered as "known"
to the neuron (Fig. 7A, green). Its informational content, xM +1, satisfies the
condition (cid:104)w0, xM +1(cid:105) > θ, i.e., this stimulus evokes membrane potential above
the threshold at t = t0. Consequently, the neuron detects this stimulus selec-
tively as described in Sect. 4.1. For the second relevant stimulus (Fig. 7A,
orange), however, we have (cid:104)w0, xM +2(cid:105) < θ. Therefore, the neuron cannot de-
tect such a stimulus alone. The background stimuli from the set S(M) are
also sub-threshold (Fig. 7A, back curves).
013420 0.20.20.40.6learningstimuli retrievalm=2013420 0.20.20.40.6m=12y(t)0 0.20.20.40.613420y(t)m=4time (a.u.)time (a.u.)membrane potential (a.u.)response thresholdrelevant stimuliAB 20
Ivan Tyukin et al.
During the learning phase, the neuron receives M = 500 background and
m = 2 relevant stimuli. The relevant stimuli from the set S(Y) appear simulta-
neously, i.e., they are temporarily associated. The synaptic efficiency changes
during the learning phase by action of the relevant stimuli. Therefore, the
membrane potential, y(t) = (cid:104)w(t, w0), ¯s(t)(cid:105), progressively increases when the
relevant stimuli arrive (Fig. 7A, green area). These neuronal adjustments give
rise to a new functionality.
At some time instant (marked by red circle in Fig. 7A) the neuron becomes
responsive to the new relevant stimulus (Fig. 7A, orange), which is synchro-
nized with the "known" one. Note that all other background stimuli, that
show no temporal associativity, remain below the threshold (Fig. 7A, black
traces). Thus, after a transient period, the neuron learns new stimulus. Once
the learning is over, the neuron detects selectively either of the two relevant
stimuli.
The procedure just described can be used to associate together more than
two relevant stimuli. Figure 7B shows examples for m = 4 and m = 12. In
both cases the neuron was able to learn all relevant stimuli, whilst rejecting
all background ones. We observed, however, that increasing the number of
uncorrelated information items to be learnt, i.e. the value of m, reduces the
gap between firing thresholds and the membrane potentials evoked by back-
ground stimuli. In other words, the neuron does detect the assigned group
of new stimuli, but with lower accuracy. This behavior is consistent with the
theoretical bound on θ prescribed in the statement of Theorem 3.
5 Discussion
Theorems 1–3 and our numerical simulations demonstrate that the extreme
neuronal selectivity to single and multiple stimuli, and the capability to learn
uncorrelated stimuli observed in a range of empirical studies [38], [49], [25] can
be explained by simple functional mechanisms implemented in single neurons.
The following basic phenomenological properties have been used to arrive to
this conclusion: i) the dimensionality n of the information content and neurons
is sufficiently large, ii) a perceptron neuronal model, Eq. (9), is an adequate
representation of the neuronal response to stimuli, and iii) plasticity of the
synaptic efficiency is governed by Hebbian rule (10). A crucial consequence of
our study is that no a priori assumptions on the structural organization of
neuronal ensembles are necessary for explaining basic concepts of static and
dynamic memories.
Our approach does not take into account more advanced neuronal behav-
iors reproduced by, e.g., models of spike-timing dependent plasticity [33] and
firing threshold adaptation [13]. Nevertheless, our model captures essential
properties of neuronal dynamics and as such is generic enough for the purpose
of functional description of memories.
Firing threshold adaption, as reported in [13], steers firing activity of a
stimulated neuron to a homeostatic state. In this state, the value of the
High-dimensional brain
21
threshold is just large/small enough to maintain reasonable firing rate without
over/under-excitation. In our model, such a mechanism could be achieved by
setting the value of θ sufficiently close to the highest feasible values specified
in Theorems 1 and 2.
In addition to rather general model of neuronal behavior, another major
theoretical assumption of our work was the presumption that stimuli infor-
mational content is drawn from an equidistribution in a unit ball Bn(1). This
assumption, however, can be relaxed, and results of Theorems 1–3 generalized
to product measures. Key ingredients of such generalizations are provided in
[14], and their practical feasibility is illustrated by numerical simulations with
information items randomly drawn from a hypercube (Figs. 5–7).
Our theoretical and numerical analysis revealed an interesting hierarchy
of cognitive functionality implementable at the level of single neurons. We
have shown that cognitive functionality develops with the dimensionality or
connectivity parameter n of single neurons. This reveals explicit relationships
between levels of the neural connectivity in living organisms and different
cognitive behaviors such organisms can exhibit (cf. [32]). As we can see from
Theorems 1, 2 and Figs. 5, 6, the ability to form static memories increases
monotonically with n. The increase of cognitive functionality, however, occurs
in steps.
For n small (n ∈ [1, 10]), neuronal selectivity to a single stimulus does
not form. It emerges rapidly when the dimension parameter n exceeds some
critical value, around n = 10 ÷ 20 (see Fig. 5A). This constitutes the first
critical transition. Single neurons become selective to single information items.
The second critical transition occurs at significantly larger dimensions, around
n = 100 − 400 (see Fig. 6). At this second stage the neuronal selectivity to
multiple uncorrelated stimuli develops. The ability to respond selectively to
a given set of multiple uncorrelated information items is apparently crucial
for rapid learning "by temporal association" in such neuronal systems. This
learning ability as well as formation of dynamic memories are justified by
Theorem 3 and illustrated in Fig. 7.
In the core of our mathematical arguments are the concentration of measure
phenomena exemplified in [17,15] and stochastic separation theorems [14,16].
Some of these results, which have been central in the proofs of Theorem 2 and
3, namely, the statements that random i.i.d. vectors from equidistributions
in Bn(1) and product measures are almost orthogonal with probability close
to one, are tightly related to the notion of effective dimensionality of spaces
based on -quasiorthogonality introduced in [19,27]. In these works the authors
demonstrated that in high dimensions there exist exponentially large sets of
quasiorthogonal vectors. In [17], however, as well as in our current work (see
Lemma 3) we demonstrated that not only such sets exist, but also that they
are typical.
Finally, we note that the number of multiple stimuli that can be selec-
tively detected by single neurons is not extraordinarily large. In fact, as we
have shown in Figs. 6 and 7, memorizing 8 information items at the level of sin-
gle neurons requires more than 400 connections. This suggests, that not only
22
Ivan Tyukin et al.
new memories are naturally packed in quanta, but also that there is a limit on
this number that is associated with the cost of implementation of such a func-
tionality. This cost is the number of individual functional synapses. Balancing
the costs in living beings is of course a subject of selection and evolution. Nev-
ertheless, as our study have shown, there is a clear functional gain that these
costs may be paid for.
6 Conclusion
In this work we analyzed the striking consequences of the abundance of sig-
nalling routes for functionality of neural systems. We demonstrated that com-
plex cognitive functionality derived from extreme selectivity to external stimuli
and rapid learning of new memories at the level of single neurons can be ex-
plained by the presence of multiple signalling routes and simple physiological
mechanisms. At the basic level, these mechanisms can be reduced to a mere
perceptron-like behavior of neurons in response to stimulation and a Hebbian-
type learning governing changes of the synaptic efficiency.
The observed phenomenon is robust. Remarkably, a simple generic model
offers a clearcut mathematical explanation of a wealth of empirical evidence
related to in-vivo recordings of "Grandmother" cells, "concept" cells, and rapid
learning at the level of individual neurons [38, 49,25]. The results can also shed
light on the question why Hebbian learning may give rise to neuronal selectivity
in prefrontal cortex [31] and explain why adding single neurons to deep layers
of artificial neural networks is an efficient way to acquire novel information
while preserving previously trained data representations [10].
Finding simple laws explaining complex behaviours has always been the
driver of progress in Mathematical Biology and Neuroscience. Numerous ex-
amples of such simple laws can be found in the literature (see e.g. [40,26,18,
35]). Our results not only provide a simple explanation of the reported em-
pirical evidence but also suggest that such a behavior might be inherent to
neuronal systems and hence organisms that operate with high-dimensional in-
formational content. In such systems, complex cognitive functionality at the
level of elementary units, i.e., single neurons, occurs naturally. The higher the
dimensionality, the stronger the effect. In particular, we have shown that the
memory capacity in ensembles of single neurons grows exponentially with the
neuronal dimension. Therefore, from the evolutionary point of view, accom-
modating large number of signalling routes converging onto single neurons is
advantageous despite the increased metabolic costs.
The considered class of neuronal models, being generic, is of course a sim-
plification. It does not capture spontaneous firing, signal propagation in den-
dritic trees, and many other physiologically relevant features of real neurons.
Moreover, in our theoretical assessments we assumed that the informational
content processed by neurons is sampled from an equidistribution in a unit
ball. The results, however, can already be generalized to product measure dis-
High-dimensional brain
23
tributions (see, e.g., [14]). Generalizing the findings to models offering better
physiological realism is the focus of our future works.
Acknowledgements This work has been supported by Innovate UK grants KTP009890
and KTP010522, by the Spanish Ministry of Economy and Competitiveness under grant
FIS2014-57090-P, the Russian Federation Ministry of Education state assignment (No.
8.2080.2017/4.6), "Initiative scientific project" of the main part of the state plan of the
Ministry of Education and Science of Russian Federation (task No. 2.6553.2017/BCH Basic
Part), and by the Russian Science Foundation project 15-12-10018 (numerical assessment
and results). Alexander N. Gorban was supported by the Ministry of Education and Science
of Russian Federation (Project No. 14.Y26.31.0022).
A Dynamics of coupling weights
The following results demonstrate that the neuronal model provided in Section 3 is well-
posed.
Lemma 1 Consider (9), (10) with the function s(·, x), x ∈ Rn defined as in (2). Then
1) solutions w(·, w0) of (10) are defined for all t ≥ t0, and are unique and bounded in
forward time.
If, in addition, θ ≥ 0 and there exist numbers L, δ > 0 such that:
v(s(τ, x), w(τ, w0), θ)(cid:104)s(τ, x), w(τ, w0)(cid:105)2 dτ > δ,
∀t ≥ t0,
(26)
t
then
2) x/(cid:107)x(cid:107) is an attractor, that is:
lim
t→∞ w(t, w0) =
x
(cid:107)x(cid:107) .
(27)
Proof of Lemma 1.
1. The right-hand side of (10) is continuous in w and piece-wise continuous in t with finite
number of discontinuities of the first kind in any finite interval containing t0, independently
on the values of w. Hence, in accordance with Peano Theorem, solutions of (10) are defined
on some non-empty interval containing t0. Let T be the maximal interval of this solution's
definition (to the right of t0). Since the right-hand side of (10) is locally Lipschitz in w the
solution w(·, w0) is uniquely defined on T .
To show that T = [t0, ∞) consider
V (w) = 1 − (cid:107)w(cid:107)2.
In the interval T we have:
V = −2αvy2V.
Given that vy2 ≥ 0, the above expression implies that
(cid:90) t+L
Consequently,
1 − (cid:107)w0(cid:107)2 ≥ 1 − (cid:107)w(t, w0)(cid:107)2 ≥ (cid:107)w(t, w0)(cid:107)2 − 1.
(cid:107)w(t, w0)(cid:107) ≤(cid:0)1 + 1 − (cid:107)w0(cid:107)2(cid:1) 1
(28)
for all t ≥ t0, t ∈ T . Let t1 be an arbitrary point in the interval T . Recall that the right-hand
side of (10) is continuous and locally Lipschitz with respect to w (uniformly in t). Thus (28)
implies existence of some ∆(w0, x) > 0, independent on t1, such that the solution w(·, w0)
2
24
Ivan Tyukin et al.
is defined on the interval [t0, t1 + ∆(w0, x)]. Given that t1 was chosen arbitrarily in T , we
can conclude that T = [t0, ∞) (cf. Theorem 3.3 [28]).
2. For the sake of convenience, we denote
p(t) = v(s(t, x), w(t, w0), θ)(cid:104)s(t, x), w(t, w0)(cid:105)2.
Condition (26) assures that both x (cid:54)= 0, w0 (cid:54)= 0. Moreover, since V (w(t, w0)) is defined for
all t ≥ t0, we can conclude that
(cid:12)(cid:12)(cid:12)(cid:12)V0e
−2α(cid:82) t
t0
V (t) =
(cid:12)(cid:12)(cid:12)(cid:12) ≤ V0e
p(τ )dτ
−2αδ
(cid:106) t−t0
(cid:107)
L
.
Hence
Consider:
−α(cid:82) t
(cid:104)(cid:82) t
t0
t0
w(t, w0) = e
α
t→∞ (cid:107)w(t, w0)(cid:107) = 1.
lim
p(τ )dτ
e−α(cid:82) t
τ p(s)dsv(s(τ, x), w(τ, w0), θ)(cid:104)s(τ, x), w(τ, w0)(cid:105)(cid:80)
w0+
(29)
(cid:105)
x.
j c(τ − τj ) dτ
Observe that the first term decays exponentially to 0, whereas the second term is propor-
tional to x. Moreover, since θ ≥ 0, the term v(s(τ, x), w(τ, w0), θ)(cid:104)s(τ, x), w(τ, w0)(cid:105) ≥ 0 for
all τ ≥ t0. Hence the coefficient in front of x is non-negative. This, combined with (29),
implies that (27) holds. (cid:3)
Note that Lemma 1 apply to stimuli classes that are broader than the one defined by (2),
(3). The results hold e.g. for the functions c(·) in (2) that are non-negative, piece-wise contin-
uous, and bounded. On the other hand, to determine convergence and asymptotic properties
of w(·, w0) for t ≥ t0 (part 2 of the lemma) one needs to check that condition (26) holds.
A drawback of this condition is that it requires availability of signals v(s(t, x), w(t, w0), θ),
(cid:104)s(t, x), w(t, w0)(cid:105) for all t ≥ t0.
For c(·) specified by (2) this latter condition can be drastically simplified. To see this,
let us get a somewhat deeper geometrical insight into the dynamics of w governed by (10).
In order to bring the discussion in line with the question of neuronal selectivity, consider
the stimuli sets (6), (7) with Y = {xM +1}, and suppose that stimuli s(·, xi), i = 1, . . . , M
do not evoke any neuronal responses, i.e., v(s(·, xi), w(·, w0), θ) = 0 for all i = 1, . . . , M .
Hence no changes in w occur if the stimulus s in (10) is any of s(·, xi), i = 1, . . . , M .
Consider system (10) with s(·, xM +1). The variable w may change only over those
intervals of t when s(·, xM +1) (cid:54)= 0. Between these intervals w(t, w0) is constant. Let the
stimulus be persistent in the sense that for any t(cid:48) ≥ t0 there is a t(cid:48)(cid:48) such that s(t(cid:48)(cid:48), xM +1) (cid:54)=
0. Thus, without loss of generality and for the purposes of assessing asymptotic behavior of
w(t, w0) at t → ∞ variable s(t, xM +1) in (8) – (10) may be replaced with xM +1.
Recall that w(t, w0) can be represented as a sum
w(t, w0) = w∗(t, w0)w∗ + w⊥(t, w0), w∗(t, w0) = (cid:104)w(t, w0), w∗(cid:105),
where w∗ is defined in (12) and w⊥ ∈ L⊥. In this representation,
(cid:105)
w = w∗w∗ + w⊥ = αf ((cid:104)xM +1, w∗w∗ + w⊥(cid:105) − θ)(cid:104)xM +1, w∗w∗ + w⊥(cid:105)(xM +1−
(cid:104)
(cid:104)xM +1, w∗w∗ + w⊥(cid:105)[w∗w∗ + w⊥]) =
αf (w∗(cid:107)xM +1(cid:107) − θ)(cid:107)xM +1(cid:107)2(1 − w∗2)
w⊥
αf (w∗(cid:107)xM +1(cid:107) − θ)(cid:107)xM +1(cid:107)2w∗2(cid:105)
(cid:104)
or, equivalently,
w⊥ = −(cid:104)
w∗ = α(cid:107)xM +1(cid:107)2f (w∗(cid:107)xM +1(cid:107) − θ)(1 − w∗2)w∗
w⊥.
αf (w∗(cid:107)xM +1(cid:107) − θ)(cid:107)xM +1(cid:107)2w∗2(cid:105)
w∗w∗−
(30)
(31)
High-dimensional brain
Obviously, L(cid:107), L⊥, and the set
25
W(xM +1, θ) = {(w∗, w⊥), w∗ ∈ R, w⊥ ∈ L⊥ w∗(cid:107)xM +1(cid:107) − θ ≤ 0}
are invariant with respect to (10). Let xM +1 (cid:54)= 0, θ ≥ 0, and w0 /∈ W(xM +1, θ). Then two
non-trivial alternatives (Fig. 8) are possible:
Fig. 8 Sketch of the dynamics of w∗. Thick black curve shows the r.h.s. of (30) as a function
of w∗ for two cases: θ < (cid:107)xM +1(cid:107) (left) and θ > (cid:107)xM +1(cid:107) (right). Blue (red) dots correspond
to stable (unstable) equilibria. Green arrows mark trajectories. In the first case (left) w∗
tends to 1, whereas in the second (right) it goes asymptotically to θ/(cid:107)xM +1(cid:107).
A: If θ < (cid:107)xM +1(cid:107) then w∗(t, w0) → 1 and, according to (31), w⊥(t, w0) → 0 as t → ∞.
Thus,
lim
t→∞ w(t) =
xM +1
(cid:107)xM +1(cid:107) = w∗.
B: If θ ≥ (cid:107)xM +1(cid:107) then w∗(t, w0) → θ/(cid:107)xM +1(cid:107) as t → ∞. There is no guarantee, however
that w⊥(t, w0) converges to the origin asymptotically. Thus, there is a w⊥∞ ∈ L⊥:
lim
t→∞ w(t) =
θ
(cid:107)xM +1(cid:107) w∗ + w⊥
∞.
The above result can now be formalized as
Lemma 2 Consider (9), (10) with the function s(·, x), x ∈ Rn defined as in (2). Let θ ≥ 0
and (cid:104)w0, x(cid:105) > θ. Furthermore, let the stimulus s(·, x) be persistent in the sense that for
any t(cid:48) ≥ t0 there is a t(cid:48)(cid:48) > t(cid:48) such that s(t(cid:48)(cid:48), x) (cid:54)= 0. Then the following alternatives hold:
1) If θ < (cid:107)x(cid:107) then limt→∞ w(t, w0) = x/(cid:107)x(cid:107).
2) If θ ≥ (cid:107)x(cid:107) then limt→∞(cid:104)w(t, w0), x/(cid:107)x(cid:107)(cid:105) = θ/(cid:107)x(cid:107).
Note that alternative 1) in Lemma 2 is equivalent to the second statement of Lemma 1.
Alternative 2) corresponds to the case when condition (26) of Lemma 1 is not satisfied.
B Proof of Theorem 1
1. Let us first assume that (cid:107)w(cid:107) = 1. Notice that the condition
(cid:104)w, xi(cid:105) ≤ θ ∀xi ∈ M,
(32)
assures that v = 0 and hence si ∈ Silent(S(M), (w, θ)) ∀si ∈ S(M).
by construction. Therefore, it is sufficient to estimate the probability that (32) holds.
In this case the neuron is silent for all stimuli except sM +1 that does evoke a response
Let Cn(w, θ) be the spherical cap:
Cn(w, θ) = {x ∈ Bn(1) (cid:104)w, x(cid:105) > θ}.
10✓kxM+1k10✓kxM+1kw⇤w⇤w⇤w⇤ 26
Ivan Tyukin et al.
Then the ratio of volumes V(Cn(w, θ))/V(Bn(1)) is the probability that a random vector
xi ∈ Cn(w, θ). Observe that
V(Cn(w, θ))
V(Bn(1))
(1 − θ2)
n
2 .
≤ 1
2
(cid:20)
(cid:20)
(cid:21)M ≥
(cid:21)M
Thus, the probability that all xi ∈ M are outside the cap Cn(w, θ) is bounded from below:
P =
1 − V(Cn(w, θ))
V(Bn(1))
1 − 1
2
(1 − θ2)
n
2
,
(33)
which is equivalent to (15), given that (cid:107)w(cid:107) = 1.
Let (cid:107)w(cid:107) (cid:54)= 1. Noticing that, for (cid:107)w(cid:107) > 0
(cid:104)w, xi(cid:105) ≤ θ ∀xi ∈ M ⇔ (cid:104)w/(cid:107)w(cid:107), xi(cid:105) ≤ θ/(cid:107)w(cid:107) ∀xi ∈ M,
and substituting θ/(cid:107)w(cid:107) in place of θ in (33) results in (15).
2. Let us show that for (w, θ) ∈ ΩD the neuron detects the relevant stimulus sM +1, i.e.,
v > 0. Using (16) we observe that
(cid:104)w, xM +1(cid:105) − θ = (cid:104)w − w∗, xM +1(cid:105) + (cid:107)xM +1(cid:107) − θ ≥ (cid:104)w − w∗, xM +1(cid:105) + D ≥
≥ −(cid:107)w − w∗(cid:107)(cid:107)xM +1(cid:107) + D > D(1 − (cid:107)xM +1(cid:107)) ≥ 0,
implying that sM +1 ∈ Activated(Y, (w, θ)).
(w, θ) ∈ ΩD. According to (16) the following holds:
Let us evaluate the probability that the neuron rejects all background stimuli for all
θ
(cid:107)w(cid:107) ≥ (cid:107)xM +1(cid:107) − 2D
1 + D
∀(w, θ) ∈ ΩD.
,
Moreover, (cid:107)xM +1(cid:107) ≥ 1 − ε with probability p = 1 − (1 − ε)n. Therefore, with probability
larger or equal to p, the ratio
θ(cid:107)w(cid:107) us bounded from below as:
θ
(cid:107)w(cid:107) ≥ 1 − ε − 2D
1 + D
.
Finally, since the value of ε can be chosen arbitrarily in the interval (0, 1 − 2D) and taking
into account that the right-hand side of (33) is a monotone and increasing function with
respect to θ in the interval [0, 1], estimate (17) immediately follows from (33) and (15). (cid:3)
C Proof of Corollary 1
Consider (15) and denote
(cid:20)
1 − θ2
(cid:107)w(cid:107)2
(cid:21) n
2
z =
1
2
, φ = (1 − z)M .
(34)
According to (34), (1 − z)M ≥ φ for all 0 < M ≤ M . Given that z ∈ (0, 1), from Eq. (34)
we get ln(φ) = M ln(1 − z). Recall that ln(1 − z) > −z/(1 − z), ∀z ∈ (0, 1). Thus, we can
conclude that
1 − z
z
M > − ln(φ)
= − ln(φ)(z−1 − 1) = − ln(φ) (2eαn − 1) ,
(35)
where α is given by (18). Thus, according to (35), for 0 < M ≤ − ln(φ) (2ean − 1) < M the
following holds
P (si ∈ Silent(S(M), (w, θ)) ∀si ∈ S(M)(cid:12)(cid:12) w, θ) ≥ φ.
(cid:3)
High-dimensional brain
27
D Proof of Theorem 2
The proof of the Theorem is essentially contained in Lemmas 3 and 4 (Sect. F). Consider
the set Y. With probability (1 − (1 − ε)n)m, all elements xi ∈ Y satisfy the condition
(cid:107)xi(cid:107) ≥ 1 − ε. Hence, using Lemma 3 we have that the following inequality
holds with probability
(cid:104)xi, xj(cid:105) ≤ δ
1 − ε
, ∀xi, xj ∈ Y, i (cid:54)= j
p0 ≥ (1 − (1 − ε)n)m
(cid:17)
.
1 − d(1 − δ2)
n
2
m−1(cid:89)
(cid:16)
d=1
This implies that, with probability p0, the following conditions are met
(cid:107)xi(cid:107) ≥ 1 − ε, − (m − 1)δ
1 − ε
≤
(cid:104)xi, xj(cid:105) ≤ (m − 1)δ
1 − ε
, ∀ xi ∈ Y.
m(cid:88)
j=1, j(cid:54)=i
Consider (cid:96)(x) = (cid:104)w∗, x(cid:105) − θ∗ + D. Invoking Lemma 4 and setting β1 = δ/(1 − ε), β2 =
−δ/(1 − ε), we can conclude that, with probability p0,
(cid:96)(x) ≥ D, ∀x ∈ Y.
In fact, we can conclude that with probability p0
(cid:96)0(x) = (cid:104)w, x(cid:105) − θ = (cid:96)(x) + (cid:104)w − w∗, x(cid:105) − θ + (θ∗ − D) > 0, ∀ (w, θ) ∈ ΩD, x ∈ Y.
Thus, the probability that (cid:96)0(x) > 0 for all x ∈ Y and that (cid:96)0(x) ≤ 0 for all x ∈ M is
bounded from below by
(1 − (1 − ε)n)m
1 − d(1 − δ2)
n
2
m−1(cid:89)
(cid:16)
d=1
(cid:17)(cid:34)
(cid:18)
1 − 1
2
1 − θ2
(cid:107)w(cid:107)2
(cid:35)M
(cid:19) n
2
.
Noticing that (cid:107)w(cid:107) ≤ 1 + D, we can conclude that (21) holds. (cid:3)
E Proof of Theorem 3
According to Lemma 1, solutions w(t, w0) are defined for all t ≥ t0. Moreover, condition 1
of the theorem and Lemma 1 imply that
(cid:80)m
(cid:107)(cid:80)m
i=1 xM +i(cid:107) = ¯x/(cid:107)¯x(cid:107) = w∗.
i=1 xM +i
lim
t→∞ w(t, w0) =
Let D > 0 be chosen so that
0 < θ + D ≤ θ∗.
Given that 0 < θ < θ∗, such Ds always exist. Equation (36) implies that there is a t1(D) > t0
such that
(cid:107)w(t, w0) − w∗(cid:107) < D, θ ∈ (0, θ∗ − D] ∀ t ≥ t1(D).
The theorem now follows immediately from Theorem 2. (cid:3)
(36)
(37)
28
Ivan Tyukin et al.
F Auxiliary results
Lemma 3 (cf. Gorban et. al. 2016, [17]) Let Y = {x1, x2, . . . , xk} be a set of k i.i.d.
random vectors from the equidistribution in the unit ball Bn(1). Let δ, r ∈ (0, 1), and suppose
that (cid:107)xi(cid:107) ≥ r, for all i ∈ {1, . . . , k}.
Then the probability that the elements of Y are pair-wise δ/r-orthogonal, that is
cos(∠(xi, xj )) ≤ δ
r
for all i (cid:54)= j i, j ∈ {1, . . . , k},
is bonded from below as
(cid:18)
P
cos(∠(xi, xj )) ≤ δ
r
∀ i, j ∈ {1, . . . , k}, i (cid:54)= j (cid:107)xi(cid:107) ≥ r, 1 ≤ i ≤ k
≥ k−1(cid:89)
(cid:16)
1 − d(cid:0)1 − δ2(cid:1) n
2
(cid:17)
.
(cid:19)
d=1
Proof of Lemma 3. Let xi, i = 1, . . . , k be random vectors satisfying conditions of the lemma.
Let Eδ(xi) be the delta-thickening of the largest equator of Bn(1) that is orthogonal to xi.
There is only one such equator, and it is uniquely determined by xi. Consider the following
probabilities:
P (x2 ∈ Eδ(x1))
P ([x3 ∈ Eδ(x2)]&[x3 ∈ Eδ(x1)])
P ([x4 ∈ Eδ(x3)]&[x4 ∈ Eδ(x2)]&[x4 ∈ Eδ(x1)])
· · ·
P ([xk ∈ Eδ(xk−1)]& · · · &[xk ∈ Eδ(x1)]).
Pick xi, xj ∈ Y, i (cid:54)= j. Recall that, for any random events A1, . . . , Ak, the probability
P (A1&A2& · · · &Ak) ≥ 1 − k(cid:88)
(1 − P (Ai)).
According to (38), the probability that xi ∈ Eδ(xj ) is bounded from below by 1−(cid:0)1 − δ2(cid:1) n
i=1
2
(cf. [17], Proposition 3; see also Fig. 1 in [17] for illustration). Then
P (x2 ∈ Eδ(x1)) ≥ 1 −(cid:0)1 − δ2(cid:1) n
P ([x3 ∈ Eδ(x2)]&[x3 ∈ Eδ(x1)]) ≥ 1 − 2(cid:0)1 − δ2(cid:1) n
P ([x4 ∈ Eδ(x3)]&[x4 ∈ Eδ(x2)]&[x4 ∈ Eδ(x1)]) ≥ 1 − 3(cid:0)1 − δ2(cid:1) n
P ([xk ∈ Eδ(xk−1)]& · · · &[xk ∈ Eδ(x1)]) ≥ 1 − (k − 1)(cid:0)1 − δ2(cid:1) n
· · ·
2 .
2
2
2
(38)
(39)
The fact that xi ∈ Eδ(xj ) combined with the condition that (cid:107)xi(cid:107) ≥ r, (cid:107)xj(cid:107) ≥ r imply:
cos(∠(xi, xj )) ≤ δ
r
.
Finally, given that x1, . . . , xk are drawn independently and that the distribution is rotation-
ally invariant, the probability that all vectors in Y are pair-wise orthogonal is the product
of all probabilities in the left-hand side of (39). Thus the statement follows. (cid:3)
High-dimensional brain
29
Lemma 4 Let Y = {x1, . . . , xm} be a finite set from Bn(1). Let (cid:107)xi(cid:107) ≥ 1 − ε, ε ∈ (0, 1)
for all xi ∈ Y, and β1, β2 ∈ R be such that the following condition holds:
(cid:104)xi, xj(cid:105) ≤ β1(m − 1) for all i = 1, . . . , m.
(40)
β2(m − 1) ≤ (cid:88)
(cid:29)
(cid:28) ¯y
(cid:96)(x) =
(cid:107)¯y(cid:107) , x
− 1√
m
j∈{1,...,m}, j(cid:54)=i
Consider
(cid:32)
(1 − ε)2 + β2(m − 1)
(cid:112)1 + (m − 1)β1
(cid:33)
m(cid:88)
i=1
xi,
, ¯y =
1
m
and suppose that parameters β1, β2 satisfy:
(1 − ε)2 + β2(m − 1) > 0, 1 + (m − 1)β1 > 0.
Then
(41)
Proof of Lemma 4. Consider the set Y. According to the lemma assumptions, (cid:107)xi(cid:107) ≥ 1 − ε
for some given ε ∈ (0, 1) and all i = 1, . . . , m. Consider now the mean vector ¯y
(cid:96)(xi) ≥ 0 for all xi ∈ Y.
¯y =
1
m
xi,
m(cid:88)
i=1
(cid:88)
and evaluate the following inner products
(cid:28) ¯y
(cid:29)
(cid:107)¯y(cid:107) , xi
=
1
m(cid:107)¯y(cid:107)
(cid:107)xi(cid:107)2 +
(cid:29)
≥ 1
(cid:28) ¯y
(cid:107)¯y(cid:107) , xi
m(cid:107)¯y(cid:107)
According to assumption (40), the following holds
j∈{1,...,m}, j(cid:54)=i
(cid:0)(1 − ε)2 + β2(m − 1)(cid:1) ,
, i = 1, . . . , m.
(cid:104)xi, xj(cid:105)
and, respectively,
1
m
(1 + (m − 1)β1) ≥ (cid:104)¯y, ¯y(cid:105) = (cid:107)¯y(cid:107)2 ≥ 1
m
(cid:0)(1 − ε)2 + β2(m − 1)(cid:1)
Let (1 − ε)2 + β2(m − 1) > 0 and 1 + β1(m − 1) > 0. It is clear that for (cid:96), as defined by
(41), the following holds for all i = 1, . . . , m: (cid:96)(xi) ≥ 0. (cid:3)
References
1. D.G. Amaral and M.P. Witter. The three-dimensional organization of the hippocampal
formation: a review of anatomical data. Neuroscience, 31:571–591, 1989.
2. Daniel J Amit, Hanoch Gutfreund, and Haim Sompolinsky. Storing infinite numbers of
patterns in a spin-glass model of neural networks. In Spin Glass Theory and Beyond:
An Introduction to the Replica Method and Its Applications, pages 428–431. World
Scientific, 1987.
3. P. Andersen, R. Morris, D. Amaral, T. Bliss, and J. O'Keefe, editors. The hippocampus
book. 2007.
4. Lisa Feldman Barrett, Michele M Tugade, and Randall W Engle.
Individual differ-
ences in working memory capacity and dual-process theories of the mind. Psychological
bulletin, 130(4):553, 2004.
30
Ivan Tyukin et al.
5. N. Benito, A. Fernandez-Ruiz, V.A. Makarov, J. Makarova, A. Korovaichuk, and O. Her-
reras. Spatial modules of coherent activity in pathway-specific lfps in the hippocampus
reflect topology and different modes of presynaptic synchronization. Cerebral Cortex,
11(7):1738–1752, 2014.
6. N. Benito, G. Martin-Vazquez, J. Makarova, V.A. Makarov, and O. Herreras. The right
hippocampus leads the bilateral integration of gamma-parsed lateralized information.
eLife, 5:e16658, doi: 10.7554/eLife.16658, 2016.
7. C. Calvo, J.A. Villacorta-Atienza, V.I. Mironov, V. Gallego, and V.A. Makarov. Waves
in isotropic totalistic cellular automata: Application to real-time robot navigation. Ad-
vances in Complex Systems, 19(4):1650012–18, 2016.
8. D.D. Clark and L. Sokoloff. Circulation and energy metabolism of the brain.
In
G.J. Siegel, B.W. Agranoff, R.W. Albers, S.K. Fisher, and M.D. Uhler, editors, Basic
Neurochemistry: Molecular, Cellular and Medical Aspects, pages 637–670. Lippincott,
Philadelphia, 1999.
9. F. Cucker and S. Smale. On the mathematical foundations of learning. Bulletin of the
American mathematical society, 39(1):1–49, 2002.
10. T. J. Draelos, N. E. Miner, C. C. Lamb, C. M. Vineyard, K. D. Carlson, C. D. James,
and J. B. Aimone. Neurogenesis deep learning. arXiv preprint arXiv:1612.03770, 2016.
11. A. Fernandez-Ruiz, V.A. Makarov, and O. Herreras. Sustained increase of spontaneous
input and spike transfer in the ca3-ca1 pathway following long term potentiation in
vivo. Frontiers in Neural Circuits, 6:71, 2012.
12. C.T. Finnerty and J.G.R. Jefferys. Functional connectivity from ca3 to the ipsilateral
and contralateral ca1 in the rat dorsal hippocampus. Neuroscience, 56(1), 1993.
13. B. Fontaine, J.L. Pena, and R. Brette. Spike-threshold adaptation predicted by mem-
brane potential dynamics in vivo. PLoS computational biology, 10(4):e1003560, 2014.
14. A.N. Gorban and I.Y. Tyukin. Stochastic separation theorems. Neural Networks,
94:255–259, 2017.
15. A.N. Gorban and I.Y. Tyukin. Blessing of dimensionality: mathematical foundations
of the statistical physics of data. Phiolosphical Transactions of the Royal Society A.
doi:10.1098/rsta.2017.0237, 2018.
16. A.N. Gorban, I.Y. Tyukin, and I. Romanenko. The blessing of dimensionality: Separa-
tion theorems in the thermodynamic limit. IFAC-PapersOnLine, 49(24):64–69, 2016.
2th IFAC Workshop on Thermodynamic Foundations for a Mathematical Systems The-
ory TFMST 2016.
17. A.N. Gorban, I.Yu. Tyukin, D.V. Prokhorov, and K.I. Sofeikov. Approximation with
random bases: Pro et contra. Information Sciences, 364–365:129–145, 2016.
18. A.N. Gorban, T.A. Tyukina, E.V. Smirnova, and L.I. Pokidysheva. Evolution of adapta-
tion mechanisms: Adaptation energy, stress, and oscillating death. Journal of theoretical
biology, 405:127–139, 2016.
19. R. Hecht-Nielsen. Context vectors: General-purpose approximate meaning representa-
tions self-organized from raw data. In J. Zurada, R. Marks, and C. Robinson, editors,
Computational Intelligence: Imitating Life, pages 43 – 56. IEEE Press, 1994.
20. S. Herculano-Houzel. The human brain in numbers: a linearly scaled-up primate brain.
Front Hum Neurosci, 2009.
21. S. Herculano-Houzel. Gorilla and orangutan brains conform to the primate cellular
scaling rules: Implications for human evolution. Brain, Behavior and Evolution, 77:33–
44, 2011.
22. S. Herculano-Houzel. The remarkable, yet not extraordinary, human brain as a scaled-
up primate brain and its associated cost. Proceedings of National Academy of Science,
109:10661–10668, 2012.
23. John J Hopfield. Neural networks and physical systems with emergent collective com-
putational abilities. In Spin Glass Theory and Beyond: An Introduction to the Replica
Method and Its Applications, pages 411–415. World Scientific, 1987.
24. N. Ishizuka, J. Weber, and D.G. Amaral. Organization of intrahippocampal projections
riginating from ca3 pyramidal cells in the rat. J Comp Neurol, 295(580-623), 1990.
25. M.J. Ison, R. Quian Quiroga, and I. Fried. Rapid encoding of new memories by indi-
vidual neurons in the human brain. Neuron, 87(1):220–230, 2015.
26. P. Jurica, S. Gepshtein, I. Tyukin, and C. van Leeuwen. Sensory optimization by stochas-
tic tuning. Psychological Review, 120(4):798–816, 2013.
High-dimensional brain
31
27. P.C. Kainen and V. Kurkova. Quasiorthogonal dimension of euclidian spaces. Appl.
Math. Lett., 6(3):7–10, 1993.
28. H. Khalil. Nonlinear Systems (3d edition). Prentice Hall, 2002.
29. Chi-Sing Leung, Lai-Wan Chan, and Edmund Lai. Stability, capacity, and statisti-
cal dynamics of second-order bidirectional associative memory. IEEE Transactions on
Systems, Man, and Cybernetics, 25(10):1414–1424, 1995.
30. X.G. Li, P. Somogyi, A. Ylinen, and G. Buzsaki. The hippocampal ca3 network: an in
vivo intracellular labeling study. J Comp Neurol, 339:181–208, 1994.
31. G. W. Lindsay, M. Rigotti, M. R. Warden, E. K. Miller, and S. Fusi. Hebbian learning
in a random network captures selectivity properties of prefrontal cortex. bioRxiv, page
133025, 2017.
32. S.A. Lobov, M.O. Zhuravlev, V.A. Makarov, and V.B. Kazantsev. Noise enhanced
signaling in stdp driven spiking-neuron network. Mathematical Modelling of Natural
Phenomena, 12(4):109–124, 2017.
33. H. Markram, J. Lubke, M. Frotscher, and B. Sakmann. Regulation of synaptic efficacy
by coincidence of postsynaptic aps and epsps. Science, 275(5297):213–215, 1997.
34. E. Oja. A simplified neuron model as a principal component analyzer. Journal of
Mathematical Biology, 15:267–273, 1982.
35. L. I. Perlovsky. Toward physics of the mind: Concepts, emotions, consciousness, and
symbols. Physics of Life Reviews, 3(1):23–55, 2006.
36. M. Platek, J.P. Keenan, and T. K. Shackelford. Evolutionary Cognitive Neuroscience.
MIT Press, 2007.
37. R. Quian Quiroga. Concept cells: the building blocks of declarative memory functions.
Nature Reviews Neuroscience, 13(8):587–597, 2012.
38. R. Quian Quiroga, L. Reddy, G. Kreiman, C. Koch, and I. Fried.
Invariant visual
representation by single neurons in the human brain. Nature, 435(7045):1102–1107,
2005.
39. M.W. Reimann, M. Nolte, M. Scolamiero, K. Turner, R. Perin, G. Chindemi, P. Dlotko,
R. Levi, K. Hess, and H. Markram. Cliques of neurons bound into cavities provide a
missing link between structure and function. Front. Comput. Neurosci., 11:48, 2017.
40. A. Roberts, D. Conte, M. Hull, R. Merrison-Hort, A. K. al Azad, E. Buhl, R. Borisyuk,
and S.R. Soffe. Can simple rules control development of a pioneer vertebrate neuronal
network generating behavior? Journal of Neuroscience, 34(2):608–621, 2014.
41. F. Rosenblatt. Principles of Neurodynamics: Perceptrons and the Theory of Brain
Mechanisms. Spartan Books, 1962.
42. C.C. Sherwood, A.L. Bauernfeind, S. Bianchi, M.A. Raghanti, and P.R. Hof. Human
brain evolution writ large and small. Prog Brain Res, 195:237–254, 2012.
43. A.M. Sousa, K.A. Meyer, G. Santpere, F.O. Gulden, and N. Sestan. Evolution of the
human nervous system function, structure, and development. Cell, 170(2):226–247,
2017.
44. I.Y. Tyukin, A. N. Gorban, K. Sofeikov, and I. Romanenko. Knowledge transfer between
artificial intelligence systems. arXiv preprint arXiv:1709.01547, 2017.
45. V. Vapnik and O. Chapelle. Bounds on error expectation for support vector machines.
Neural Computation, 12(9):2013–2036, 2000.
46. J.A. Villacorta-Atienza, C. Calvo, S. Lobov, and V.A. Makarov. Limb movement in
dynamic situations based on generalized cognitive maps. Mathematical Modelling of
Natural Phenomena, 12(4):15–29, 2017.
47. J.A. Villacorta-Atienza, C. Calvo, and V.A. Makarov. Prediction-for-compaction: Nav-
igation in social environments using generalized cognitive maps. Biological Cybernetics,
109(3):307–320, 2015.
48. J.A. Villacorta-Atienza and V.A. Makarov. Neural network architecture for cognitive
IEEE Transactions on Neural Networks and
navigation in dynamic environments.
Learning Systems, 24(12):2075–2087, 2013.
49. I.V. Viskontas, R. Quian Quiroga, and I. Fried. Human medial temporal lobe neurons re-
spond preferentially to personally relevant images. Proc. Nat. Acad. Sci., 106(50):21329–
21334, 2009.
50. L. Wittner, D.A. Henze, L. Zaborszky, and G. Buzsaki. Three-dimensional reconstruc-
tion of the axon arbor of a ca3 pyramidal cell recorded and filled in vivo. Brain Struct
Funct, 212(1):75–83, 2007.
|
1512.01197 | 2 | 1512 | 2016-05-13T12:42:43 | The topology of large Open Connectome networks for the human brain | [
"q-bio.NC",
"cond-mat.dis-nn",
"physics.bio-ph"
] | The structural human connectome (i.e.\ the network of fiber connections in the brain) can be analyzed at ever finer spatial resolution thanks to advances in neuroimaging. Here we analyze several large data sets for the human brain network made available by the Open Connectome Project. We apply statistical model selection to characterize the degree distributions of graphs containing up to $\simeq 10^6$ nodes and $\simeq 10^8$ edges. A three-parameter generalized Weibull (also known as a stretched exponential) distribution is a good fit to most of the observed degree distributions. For almost all networks, simple power laws cannot fit the data, but in some cases there is statistical support for power laws with an exponential cutoff. We also calculate the topological (graph) dimension $D$ and the small-world coefficient $\sigma$ of these networks. While $\sigma$ suggests a small-world topology, we found that $D < 4$ showing that long-distance connections provide only a small correction to the topology of the embedding three-dimensional space. | q-bio.NC | q-bio |
The topology of large Open Connectome networks
for the human brain
Michael T. Gastner1,2,* and G´eza ´Odor2
1Yale-NUS College, 16 College Avenue West, #01-220 Singapore 138527
2MTA-EK-MFA, Research Center for Energy, Hungarian Academy of Sciences, P. O. Box 49, H-1525 Budapest,
Hungary
*[email protected]
ABSTRACT
The structural human connectome (i.e. the network of fiber connections in the brain) can be analyzed at ever finer spatial
resolution thanks to advances in neuroimaging. Here we analyze several large data sets for the human brain network made
available by the Open Connectome Project. We apply statistical model selection to characterize the degree distributions of
graphs containing up to ≃ 106 nodes and ≃ 108 edges. A three-parameter generalized Weibull (also known as a stretched
exponential) distribution is a good fit to most of the observed degree distributions. For almost all networks, simple power laws
cannot fit the data, but in some cases there is statistical support for power laws with an exponential cutoff. We also calculate
the topological (graph) dimension D and the small-world coefficient s of these networks. While s suggests a small-world
topology, we found that D < 4 showing that long-distance connections provide only a small correction to the topology of the
embedding three-dimensional space.
Introduction
The neural network of the brain can be considered on structural (wired) as well as on functional connection levels. Precise
structural maps exist only on very small scales. Functional networks, based on fMRI, are available on larger datasets. However
a clear-cut relationship between them is largely unknown. Drawing parallels between neural and socio-technological networks,
neuroscientists have hypothesized that in the brain we have small-world networks, both on a structural1 and functional level.2
Small-world networks are at the same time highly clustered on a local scale, yet possess some long-distance connections that
link different clusters of nodes together. This topology is efficient for signal processing,3, 4 but doubts have remained if the
small-world assumption is generally true for the brain. Despite some evidence that functional networks obtained from spatially
coarse-grained parcellations of the brain are small worlds,5 at a structural cellular level the brain may be a large-world network
after all.4, 6
Like the small-world property, the hypothesis that functional brain networks have scale-free degree distributions became
popular around the turn of the millennium.7, 8 The degree ki of node i is defined as the number of edges adjacent to i. Because
the degree is a basic measure of a node's centrality, the probability Pr(k) that a node has degree k has played a key role in
network science for a long time.9 Especially physicists have popularized power law fits to observed degree distributions.10, 11
When such a fit is statistically justified, the network is called formally "scale-free". Power laws play a crucial role in statistical
physics, where they arise at transitions between an ordered and unordered phase because of the absence of a characteristic
length scale. There are theoretical and empirical arguments that the brain operates near such a critical point.12–15 For this
reason it is plausible to assume that, on a functional level, the connectome's degree distribution is also scale-free. More
sophisticated statistical analyses of the functional connectome justify skepticism about the scale-free hypothesis.16–18 Until
now there are only few results for degree distributions of structural brain networks, and these do not show clear evidence for
power laws.19
In this article we answer whether the structural connectome at an intermediate spatial resolution can be viewed as a scale-
free, small-world network. We analyze large data sets collected by the Open Connectome project (OCP)20 that describe
structural (rather than functional) brain connectivity. The particular data sets chosen by us were processed by members of the
OCP from the raw diffusion tensor imaging data by Landman et al.21 Earlier studies of the structural network have analyzed
much smaller data. For example the network obtained by Sporns et al., using diffusion imaging techniques,22, 23 consists of a
highly coarse-grained mapping of anatomical connections in the human brain, comprising N = 998 brain areas and the fiber
tract densities between them. The entire brain is made up of ∼ 9 × 1010 neurons,24 but current imaging techniques cannot
resolve such microscopic detail. The networks investigated in this article have up to ∼ 106 nodes, which puts them on a scale
1
that is halfway between the earlier coarse-grained view and the complete neural network.
One important measure which could not have been estimated previously because of too coarse-grained data is the topolog-
ical (graph) dimension D. It is defined by
Nr ∼ rD ,
(1)
where Nr is the number of node pairs that are at a topological (also called "chemical") distance r from each other (i.e. a
signal must traverse at least r edges to travel from one node to the other). The topological dimension characterizes how
quickly the whole network can be accessed from any of its nodes: the larger D, the more rapidly the number of r-th nearest
neighbors expands as r increases. Different small-world networks can possess different D, for example due to their distinct
clustering behavior.6 Therefore, the level of small-worldness (quantified for example by the coefficient defined by Humphries
and Gurney25) and the topological dimension contain different information.
Distinguishing between a finite and infinite topological dimension is particularly important theoretically. It has been con-
jectured that heterogeneities can cause strong rare-region effects and generic, slow dynamics in a so-called Griffiths phase,26
provided D is finite.27 Criticality28 or even a discontinuous phase transition is smeared over an extended parameter space.
As a consequence, a signalling network can exhibit behavior akin to criticality, although it does not operate precisely on a
unique critical point that sharply divides an ordered from a disordered phase. This phenomenon is pronounced for the Con-
tact Process,29 a common model for the spread of activity in a network. Subsequent studies found numerical evidence for
Griffiths effects in more general spreading models in complex networks, although the scaling region shrinks and disappears
in the thermodynamic limit if D → ¥
.30–33 Recently Griffiths phases were also reported in synthetic brain networks34–36 with
large-world topologies and with modular organization, which enhances the capability to form localized rare-regions. Real
connectomes have finite size, so they must possess finite D. If in real connectomes D remains small, these models hint at
an alternative explanation why the brain appears to be in a critical state: instead of the self-tuning mechanisms that have
been frequently postulated,37, 38 the brain may be in a Griffiths phase, where criticality exists without fine-tuning. Models
with self-tuning require two competing timescales: a slow "energy" accumulation on the nodes and a fast redistribution by
avalanches when the energy reaches the firing threshold. It is unclear if such a separation of timescales is realistic. Even if the
brain were in a self-organized critical state with clearly separated timescales, Griffiths effects can play an important role due
to the heterogeneous behavior of the system, frequently overlooked when modelling the brain.
Open Connectome brain network data
The data sets analyzed in this article were generated by members of the OCP with the MIGRAINE method described by
Roncal et al.39 In this section we will briefly summarize their methods. Afterwards we will describe our analysis which was
based on the graphs publicly available from the OCP web site.20 The raw input data used by the OCP consist of both diffusion
and structural magnetic resonance imaging scans with a resolution of ≃ 1 mm3 (i.e. the size of a single voxel). MIGRAINE
combines various pieces of software into a "pipeline" to transform this input to a graph with 105–106 nodes.
As an intermediate step, the processing software first generates a small graph of 70 nodes.40 For this purpose the image
is downsampled into 70 regions taken from the Desikan gyral label atlas.41 During this step the software also identifies the
fibers in the brain with deterministic tractography using the Fiber Assignment by Continuous Tracking (FACT) algorithm.42
As stopping thresholds a gradient direction of 70 degrees and a stopping intensity of 0.2 were used.
These fibers are then reanalyzed in the next step of data processing. The Magnetic Resonance One-Click Pipeline outlined
by Mhembere et al.43 generates a big graph where each voxel corresponds to one node. First a "mask" is defined, for example
the 70 regions included in the small graph. Then all data outside the mask are discarded and an edge is assigned to each
remaining voxel pair that is connected by at least one fiber staying within the boundaries of the mask. This procedure will
naturally produce hierarchical modular graphs with (at least) two quite different scales.
At this point each scan has been turned into a network with ≃ 107 vertices and ≃ 1010 edges. However, due to the image
processing algorithm (especially because the mask is chosen conservatively), many of these voxels will become disconnected
and must be considered as noise. To clean up the data, all vertices outside the largest connected component are removed.
According to Roncal et al.39 the remaining graph "keeps essentially all white matter voxels, consisting of ≈ 105 vertices and
≈ 108 edges."
One important point to note is that two voxels A and C are linked by an edge even if there are other voxels B1, . . . ,Bn
between A and C on the same fiber. For example, if one traverses voxels A, B, C on a fiber, the edges (A,B), (A,C) and
(B,C) are all part of the graph. Furthermore, the edges are undirected so that (B,A), (C,A) and (C,B) are also part of the
graph because the FACT algorithm cannot provide information about the direction of an edge. Note that confounding factors
such as the measurement technique, spatial sampling, measurement errors and the network-construction method can affect the
graph data we downloaded.44 We cannot control them, but tested the robustness of our conclusions by modifying one of the
2/14
networks by neglecting a fraction of edges that might have arisen as a consequence of the transitivity rule. Additionally, we
tested the effect of changing the reference null model from a nonspatial model (the Erdos-R´enyi graph) to a spatial one (the
random geometric graph, see section "Small-world coefficient" below).
To save space the OCP data files store only one of the directions (i.e. the upper triangle of the adjacency matrix) so that
the opposite direction must be inferred from the data and inserted into the graph.
There were 3 different sets of big human brain graphs available from the OCP website20 with the abbreviations KKI
(Kennedy Krieger Institute), MRN (Mind Research Network) and NKI (Nathan Kline Institute). The raw data are described
by Landman et al.,21 Jung et al.45 and Nooner et al.,46 respectively. We analyzed the KKI graphs numbered 10 to 19 in more
detail. Some graph invariants (e.g. degrees, clustering coefficients) were calculated and analyzed by Mhembere et al.,43 but
for the present study we have recalculated all invariants directly from the graph data available from the OCP website.
Degree distribution
Model selection
We want to assess how well different probability distributions fit the degrees of the OCP graphs. A first rough-and-ready
visual attempt supports the hypothesis that the tails might be stretched exponentials, for example for the networks KKI-10 and
KKI-18 in Fig. 1. However, such visual techniques have no inferential power. Even if the fitted parameters come from ordinary
least-squares regression toolboxes, the fitted parameters in general do not converge to the true values even if the number of
data points is large. Moreover, least-squares regression lacks a statistically principled criterion for comparing different models
with each other.
KKI−21_KKI2009−18
KKI−21_KKI2009−10
exp(−k0.32)
exp(−k0.32)
10
)
)
k
(
p
(
n
l
−
100
101
102
k
103
104
Figure 1. Empirical degree distributions of the KKI-10 and KKI-18 graphs (solid lines). Dashed lines show a
rough-and-ready approach: ordinary least-squares fits for k > 1000 for the two graphs (short-dashed line for fit to KKI-18,
long-dashed line for KKI-10) suggest stretched exponential tails. While ordinary least-squares fits are not a sound basis for
model selection, we demonstrate in this article that there is indeed statistical evidence in favor of a generalized
three-parameter Weibull distribution with a stretched exponential tail.
A statistically sound framework for model selection is information theory. Here we adopt the information-theoretic
methodology proposed by Handcock and Jones47 who fitted different functions to degree distributions in sexual contact net-
works. The key idea is that a good model should perform well in terms of two opposing objectives. On one hand the model
should have enough flexibility so that it is able to fit the observed distribution. On the other hand it should have only a minimal
number of parameters. In general, the more parameters we have at our disposal, the better we can fit the observation.
The goodness of fit can be quantified by the likelihood function which, under the assumption of independent observations,
has the form
L (v) =
N(cid:213)
i=1
Pr(ki,v).
(2)
Here v is the set of parameters in the model, N the number of observations, and k1, . . . ,kN are the observed degrees. Alterna-
tively we can write the likelihood as
L (v) =
kmax
i=kmin
[Pr(ki,v)]ni ,
(3)
3/14
(cid:213)
where kmin and kmax are the minimum and maximum observed degrees and ni is the number of times we observe the degree
ki. In the graphs KKI-10 through KKI-19, kmin is always equal to 1; kmax ranges from 5154 to 11 241. In the extreme case
of allowing as many parameters as there are observed degrees we can achieve a maximal likelihood of e−nH, where H =
−(cid:229)
because it fits only the particular connectome used as input and sheds little light on general features that different connectomes
might have in common.
N(cid:1) is the Shannon entropy of the data. However, such a highly parameterized model is no longer informative,
N ln(cid:0) ni
kmax
i=kmin
ni
Statisticians have proposed several "information criteria" to address this problem of over-fitting (e.g. Bayesian, deviance
or Hannan-Quinn information criteria). These are objective functions that rate the quality of a model based on a combination
of the likelihood L and the number of parameters K. In this study we apply the Akaike information criterion with a correction
term for finite sample size,48
AICc = −2 ln(L (v)) + 2K +
2K(K + 1)
N − K − 1
,
(4)
where v = ( v1, . . . , vK) is the set of parameters that maximizes L ; as before, N is the number of observations. The last term
in Eq. 4 is a second-order bias correction which, although not in Akaike's original formula,49 gives more accurate results if
N ≃ K.50
compare different models.48 If we denote the AICc value of model j by AIC( j)
then the difference
While the absolute size of AICc is not interpretable, differences in AICc are informative and allow us to quantitatively
c and the minimum over all models by AIC(min)
,
c
j = AIC( j)
c − AIC(min)
c
(5)
estimates the relative expected Kullback-Leibler distance between model j and the estimated best model.51 As a rule of
thumb, models with D
j > 10 on the other hand are unlikely candidates
to explain the data.48
j . 2 have substantial empirical support; models with D
Burnham and Anderson48 list many theoretical reasons in favor of model selection based on AICc. The Bayesian informa-
tion criterion (BIC), although almost equally popular, has been reported to yield poor results when used to fit power-law tails
in probability distributions52 because it tends to underestimate the number of parameters. The Akaike information criterion pe-
nalizes less severely for additional parameters: in the limit N ≫ K the penalty is asymptotically equal to the term 2K in Eq. 4,
whereas the equivalent term in the BIC grows as K ln(N). We carried out Monte Carlo simulations on synthetically generated
probability distributions of the type described in the next section (see supplementary material). We found that model selection
by AICc came close to the true number of parameters. Although the BIC showed acceptable performance, we confirmed that
it indeed favors a too small number of parameters. We therefore advocate the use of AICc rather than BIC for fitting degree
distributions.
Candidate models
The first step of AICc-based model selection is the definition of several candidate models that might generate the observed
distribution. We denote as before by Pr(k) the probability that a node has degree k. The distinctive feature of different
candidate models is the asymptotic decay of Pr(k) for k ≫ 1. Only in this limit we can hope to find scale-free behavior if it
indeed exists. Of course, all real networks are finite so that, strictly speaking, we cannot take the limit k → ¥
. If we restrict
ourselves to only a few high-degree nodes, we have too few data points for a meaningful fit. On the other hand, if we include
too many low-degree nodes, then we may misjudge the correct asymptotic behavior of Pr(k).
We therefore assume for all candidate models that there is an optimal cutoff point kc that separates nodes with degree ≤ kc
from the region where a hypothesized asymptotic function F(k) can fit the data,47
Pr(degree > k) =
k
i=1 Ai
1− (cid:229)
1−(cid:229) kc
i=1 Ai
F(kc) F(k)
if k ≤ kc
if k > kc.
(6)
Each parameter Ak is chosen so that Pr(k) = Ak for k = 1, . . . ,kc. Different families of candidate models can be defined by
different functions F(k). We list all the functions F investigated in this study in Table 1. The exponential function (EXP) is
the candidate that decays most rapidly in the right tail. The power law (POW) has two parameters: an exponent b > 0 and a
k−b .
constant a > 0 that shifts the function to the left or right. The tail k ≫ a has the conventional power law form F(k) (cid:181)
The discrete log-normal (LGN) and Weibull (WBL) distributions are represented by the usual distribution functions of their
continuous namesakes. We also include two three-parameter models: a truncated power law (TPW) and the generalized
Weibull distribution (GWB). In comparison to POW, TPW includes an additional exponential factor which is often used to
4/14
D
model
exponential (EXP)
power law (POW)
log-normal (LGN)
Weibull (WBL)
truncated power law (TPW)
generalized Weibull (GWB)
b
1
F(k)
e−a k
(k + a )−b
a
2 erf(cid:16) lnk−a
√2b (cid:17)
2 − 1
exp(cid:0)−a kb (cid:1)
(k + a )−b e−g k
b (cid:1)(cid:3)
− (k + g )
b
a
exp(cid:2)a (cid:0)g b
Table 1. Investigated candidate models for the degree distribution.
)
k
e
e
r
g
e
d
(
r
P
100
10-1
10-2
10-3
10-4
10-5
10-6
0
observation
EXP
POW
LGN
WBL
TPW
GWB
1000
2000
3000
k
4000
5000
6000
Figure 2. The maximum-likelihood distributions from each model family for matching the degree distribution of network
KKI-18. In this example the generalized Weibull distribution is the best compromise in the right tail (see Table 2).
mimic finite-size cutoffs in the right tail. GWB is a standard three-parameter generalization of WBL with an additional
parameter g , called location parameter, that shifts the distribution to the right or left.53
We can distinguish the members of each candidate model family by the choice of kc, the values of A1, . . . ,Akc and a , b , g .
Model selection by AICc gives a natural criterion for the optimal parameters matching an observed distribution. It is a simple
exercise to prove that L is maximized if Ak equals the observed relative frequency of nodes with degree k,
Ak =
nk
N
.
(7)
For a fixed value of kc, standard numerical algorithms can optimize the remaining parameters a , b and g
in Table 1 to
maximize L . After calculating these maximum-likelihood estimators for every kc between 0 and kmax, we search for the
value of kc that minimizes AICc of Eq. 4. The number K of parameters that we have to insert into this equation is
• K = kc + 1 for EXP,
• K = kc + 2 for POW, LGN and WBL,
• K = kc + 3 for TPW and GWB.
Results of model selection
For each candidate model and each kc = 0, . . . ,kmax we compute the AICc. We then determine the smallest AIC(min)
from the
entire set and calculate for each model j the difference D
j defined in Eq. 5. In Fig. 2 we show for the example of the network
KKI-18 the best-fitting distribution within each candidate model family. We have chosen a logarithmic scale for the ordinate
to highlight the differences in the right tail. While the exponential and simple Weibull distributions decrease too rapidly, the
power law and log-normal distributions decay too slowly. The truncated power law and generalized Weibull distributions are
better in mimicking the overall shape of the distribution.
c
We find that, broadly speaking, this observation is typical of the ten investigated connectomes. As we see from Table 2, in
j < 10 with the GWB distribution. In 6 out of 10 cases, there is a similar
9 out of the 10 investigated networks we can achieve D
5/14
‡
connectome
j
KKI-. . .
EXP
POW LGN WBL
TPW GWB
10
11
12
13
14
15
16
17
18
19
1317.82
1245.42
992.62
767.95
792.10
1094.40
954.28
736.42
1168.50
1006.80
202.90
17.39
43.28
155.43
117.26
120.99
18.81
195.15
109.51
3.91
204.81
0.00
21.83
72.68
124.21
139.87
29.99
216.05
109.90
47.07
73.70
38.75
174.19
221.50
85.74
26.19
66.59
18.16
185.59
190.19
0.00
1.21
0.00
3.45
0.00
86.06
0.00
49.47
51.81
0.56
3.52
0.43
0.73
0.00
9.57
0.00
17.20
0.00
0.00
0.00
Table 2. Smallest relative Akaike information criterion D
highlight in bold font all D
j = AIC( j)
j < 10. For all other models j there is essentially no empirical support.48
c − AIC(min)
c
for each candidate degree distribution. We
id
10
11
12
13
14
15
16
17
18
19
kc
115
309
119
35
114
83
101
21
94
112
a
0.239
1.619
2.597
0.669
0.411
0.064
45.670
0.057
0.358
8.329
b
0.4486
0.2700
0.2382
0.3519
0.4232
0.6021
0.0507
0.6437
0.4260
0.1645
g
110.22
208.37
270.62
100.98
199.25
-18.51
379.24
-4.60
141.72
502.64
Table 3. Summary of fitted parameters of the GWB model for the investigated KKI graphs.
level of evidence for the TPW model which has been previously hypothesized for functional brain networks.54–56 For all other
candidate distributions there is either none or very sporadic statistical support. When interpreting Table 2, one should bear
in mind that even the best-fitting model is only "best" compared with the other tested candidates. True brain networks are of
course far more complicated than any of our candidate models. Even if there were a "true" model, searching for "the" degree
distribution of the human connectome is not a sensible endeavor because it would certainly be a highly complex model that is
unlikely ever to be discovered and included in the set of candidates. As we discuss in the supplementary information, we can
nevertheless sensibly ask which candidate model comes closest to the truth in the sense that it minimizes the Kullback-Leibler
divergence. With this interpretation, we conclude that GWB is generally the best of our candidates. The fact that not all of
the ten data sets are best fitted by the same model does not call this conclusion into question. Just as in traditional p-value
based hypothesis testing, we also expect in AICc-based model selection that the best general model is sometimes rejected by
random chance for a concrete data sample.
A closer look at the fitted GWB values (Table 3) shows that, with the exception of KKI-16 where GWB is rejected by the
AICc, the b exponents lie within one order of magnitude suggesting a common trend if not even universality. This order of
magnitude also agrees with the least-squares fit in Fig. 1.
Dimension measurements
To measure the dimension of the network57 we first computed the distances from a seed node to all other nodes by running the
breadth-first search algorithm. Iterating over every possible seed, we counted the number of nodes Nr with graph distance r or
less from the seeds and calculated the averages over the trials. As Fig. 3 shows, an initial power law crosses over to saturation
6/14
D
due to the finite network sizes. We determined the dimension of the network, as defined by the scaling law (1), by attempting
a power-law fit to the data N(r) for the initial ascent. This method resulted in dimensions between D = 3 and D = 4.
To see the corrections to scaling we determined the effective exponents of D as the discretized, logarithmic derivative of
(1)
Deff(r + 1/2) =
ln Nr − lnNr+1
ln(r)− ln(r + 1)
.
(8)
As the inset of Fig. 3 shows, Deff(r) tends to values below 4 even in the infinite size limit, but the narrow scaling region breaks
down for r > 5. Furthermore, the extrapolated values for D of the connectomes exhibit an increasing tendency with N as we
will now explain in detail.
KKI2009−18
r3.05(5)
KKI2009−10
r3.37(2)
KKI2009−11
KKI2009−42
f
f
e
D
4
3
2
1
0
106
105
104
103
102
r
N
101
100
0
0.1
0.2
1/r
0.3
101
r
102
Figure 3. Number of nodes within graph distance r in the big KKI graphs. Dashed lines show power-law fits. Inset: local
slopes defined in Eq. 8. Crosses correspond to measurements on a regular 1003 lattice.
For better understanding we have also performed the analysis for other graphs besides KKI, possessing graph sizes in
different ranges of N. The finite size scaling results are summarized in Fig. 4, where we also extrapolate to N → ¥
. As one
can see, the dimension values follow the same trend for KKI-, MRN- and NKI-graphs without any clear sign of saturation.
A power-law fit to the data with the form A + BNC is also shown, suggesting that D diverges for infinite N. It is tempting
to extrapolate with this function to larger sizes or even to the infinite size limit. However, since we can rule out that the
degree distributions are scale-free, we cannot assume that such extrapolated graphs faithfully represent connectomes of finer
resolution. For example, using this power-law extrapolation we would overestimate their maximum degree kmax.
Thus the present data does not permit claiming any particular numeric value for the dimension D of the true (i.e. micro-
scopically resolved) brain connectome.
KKI
MRN
NKI
1.71+0.00028*N0.65
3.8
3.3
D
2.8
300000
500000
700000
N
900000
Figure 4. Topological dimension as a function of network size using different data sets. The line shows a power-law fit for
the combined KKI, MRN, NKI results.
We cannot modify the algorithms that generate the OCP graphs, but tested the robustness by randomly removing 20% of
7/14
the directed graph connections in case of the KKI-18 network. This makes the network partially directed, more similar to a
real connectome. As a result the graph dimension did not change much: D = 3.2(2) instead of 3.05(1). Since the majority of
edges are short, a random removal results in a relative enhancement of long connections, but D increases only slightly.
Small-world coefficient
Small-worldness can be characterized by different definitions. One of them is the so-called small-world coefficient s , which
is defined as the normalized clustering coefficient (C/Cr) divided by the normalized average shortest path length (L/Lr),
s =
C/Cr
L/Lr
,
(9)
where the normalization divides the observed quantity (C or L) by the expectation value (Cr or Lr) for an Erdos-R´enyi (ER)
random graph with the same number of nodes and edges.25
There are two different definitions of a clustering coefficient in the literature. The Watts-Strogatz clustering coefficient58
of a network of N nodes is
CW =
1
N
i
2ni/ki(ki − 1) ,
(10)
where ni denotes the number of direct edges interconnecting the ki nearest neighbors of node i. An alternative is the "global"
clustering coefficient,59 also called "fraction of transitive triplets" in the social networks literature,60
C
=
number of closed triplets
number of connected triplets
.
(11)
Both definitions are in common use, but values for CW and CD can differ substantially because Eq. 10 gives greater weight to
low-degree nodes.
The average shortest path length is61
L =
1
N(N − 1)
j6=i
d(i, j) ,
(12)
where d(i, j) is the graph distance between vertices i and j. L is only properly defined if the network is connected because
otherwise the graph distance between some voxels is infinite.
We have calculated C and L for the largest connected components of several KKI networks directly from the edge lists on
the OCP website.20 The CD values are about half of those for CW (see Table 4). A finite size scaling analysis shows that the
values of L, CD and CW decay by increasing the size N (see Fig. 5). For the average shortest path-length this decay is rather
fast; a least-squares regression with the form a + b(1/N)c results in c ≃ 0.29 and a ≃ 0.016. The constant is near zero within
the precision of data, so in the infinite size limit we see small-world behavior. The clustering coefficients are almost constant;
the power-law fit provides a very small slope: c ≃ 0.046 in agreement with the behavior of modular graphs.62 The decreasing
trends for L, CW and CD as functions of N are statistically significant at the 5% significance level; the p-values for t-tests of
zero slope for the log-transformed data are 0.03, 0.002 and 0.04 respectively. Again, finite size scaling has to be interpreted
with caution given that the topology is not scale-free.
As before, we tested the robustness, this time by deleting 10% randomly chosen undirected edges. We obtained very little
changes: L = 11.37 (previously: 11.30), CW = 0.538 (previously: 0.598) CD
= 0.322 (previously: 0.358).
Due to the transitivity of OCP graphs, one may question the validity of using ER graphs as a null model. Especially
the high clustering can be at least partly attributed to the spatial embedding so that, for example, three-dimensional random
geometric graphs63 are useful as comparison. Random geometric graphs have indeed a much higher clustering coefficient,64
CW = 15/32, a similar value to CW in the OC graphs. To circumvent such problems Telesford et al.65 suggested another
small-world criterion, using a "latticized" version of the graph for reference. However, as they pointed out, the latticization
algorithm is computationally too demanding for large graphs. Their algorithm is in practice feasible only for up to N ≃ 104
nodes, thus the necessary memory and run-times are prohibitive in our case.
in Eq. 9.
We determined the clustering coefficient of the corresponding random networks Cr = hki/N, where hki is the mean degree.
We have computed the average path length of the corresponding Erdos-R´enyi networks with the formula66
We therefore kept the ER graphs as our null model and calculated the corresponding small-world coefficient s
Lr =
ln(Nl)− 0.5772
lnhki
+ 1/2 ,
(13)
8/14
(cid:229)
D
(cid:229)
15
10
5
s
t
n
a
i
r
a
v
n
i
k
r
o
w
t
e
n
1.12(1/N)0.046
0.016+670(1/N)0.29
<L>
<Lr>
<CW>
D >
<C
0
1e−06
1.1e−06
1.2e−06
1/N
1.3e−06
1.4e−06
Figure 5. Network invariants as a function of 1/N for the investigated KKI graphs from Table 4. Lines show power-law fits.
KKI
10
11
12
13
14
15
16
17
18
19
N
9.40× 105
8.63× 105
7.44× 105
8.46× 105
7.70× 105
8.47× 105
7.60× 105
7.87× 105
8.49× 105
7.31× 105
Nedges
8.68× 107
7.07× 107
4.98× 107
5.93× 107
5.36× 107
6.94× 107
5.70× 107
5.20× 107
6.63× 107
4.94× 107
hki
184.71
163.84
133.79
140.17
139.10
163.84
150.11
132.29
156.21
134.99
L
11.38
12.25
13.91
12.96
13.10
12.80
12.03
13.00
11.30
13.17
Lr
3.02
3.07
3.14
3.14
3.13
3.06
3.09
3.16
3.09
3.14
CW
5.94× 10−1
5.99× 10−1
6.02× 10−1
6.02× 10−1
6.01× 10−1
5.99× 10−1
6.02× 10−1
6.02× 10−1
5.98× 10−1
6.02× 10−1
CD
3.20× 10−1
3.24× 10−1
3.58× 10−1
3.56× 10−1
3.62× 10−1
3.32× 10−1
3.38× 10−1
3.73× 10−1
3.58× 10−1
3.59× 10−1
Cr
1.97× 10−4
1.90× 10−4
1.80× 10−4
1.66× 10−4
1.81× 10−4
1.94× 10−4
1.98× 10−4
1.68× 10−4
1.84× 10−4
1.85× 10−4
s W
803.26
789.23
757.12
881.74
794.64
740.79
782.48
869.74
888.09
775.96
s
433.09
427.69
450.43
521.58
478.99
411.13
438.63
529.15
531.35
462.90
Table 4. Summary of small-world properties for the studied KKI graphs. N, Nedges: number of nodes and edges. hki: mean
degree. L: average shortest path length. Lr: expectation value for the average shortest path length in Erdos-R´enyi graphs with
the same N and Nedges. CW , CD : clustering coefficients defined by Eq. 10 and 11, respectively. Cr: mean clustering coefficient
in Erdos-R´enyi graphs. s W , s
D : small-world coefficient defined by Eq. 9, based on either CW or CD .
where Nl is the size of the largest component.
, respectively.
Applying these formulas to the KKI graphs, the last two columns of Table 4 show that s W ,s
≫ 1 for all cases, suggesting
small-world behavior according to this definition. These values do not show any tendency with respect to N: the t-tests for
zero slope have p-values 0.45 and 0.72 for s W and s
Conclusions
Let us return to our introductory question: are the structural connectome graphs from the OCP database scale-free, small-world
networks? As far as the adjective "scale-free" is concerned, the answer is clearly no. We have applied model selection based
on the Akaike information criterion to 10 graphs comparing 6 different degree distribution models. The observed distributions
). Most of the exponents b
are best fitted by the generalized Weibull function with a stretched exponential tail (cid:181)
are between 0.2 and 0.5, which may hint at a universal trend. In some cases a truncated power law is a plausible alternative.
However, the truncation occurs at a degree much smaller than the number of nodes in the network so that one cannot regard
these distributions as scale-free.
exp(−kb
Unlike the term "scale-free", the adjective "small-world" does apply to the OCP connectomes in the sense that the small-
world coefficients are much larger than 1. We have performed a finite size scaling analysis using several graphs and found no
dependence between the number of nodes N and the small-world coefficients. The average path length, however, decreases as
N increases. The resolution to this apparent paradox is that the average degreehki increases with N so that there is an increasing
number of shortcuts through the network. On the other hand, we obtained small topological dimensions characteristic of large-
world networks. The dimensions show a tendency to grow as the sizes of the studied graphs increase. The limit N → ¥ here
is taken by increasing N for a fixed voxel size. The absence of a scale-free degree distribution suggests that this limit may not
9/14
D
D
D
be equivalent to fixing the brain volume and instead resolving the details of the connectome at an infinitely small scale. For
this reason, it is difficult to judge whether Griffiths effects can be found in the brain, but the small topological dimensions that
we have observed warrant further investigation.
Our analysis has been based on unweighted graphs. More realistically, however, the connectome is a weighted, modular
network. Links between modules are known to be much weaker than the intra-module connections. Thus, future studies should
take into account that signals in the brain propagate on a weighted, heterogeneous network, where generic slow dynamics is
a distinct possibility.67 The methods presented here to characterize degree distributions and topological dimensions can be
generalized to the weighted case. We hope that, as more precise and finely resolved connectome data will become available,
future research will be able to assess whether Griffiths phases can indeed occur in the brain.
Acknowledgments
We thank the staff of the Open Connectome project for help and discussions. Support from the Hungarian research fund OTKA
(Grant No. K109577) and the European Commission (project number FP7-PEOPLE-2012-IEF 6-4564/2013) is gratefully
acknowledged.
References
1. Hilgetag, C. C., Burns, G. A., O'Neill, M. A., Scannell, J. W. & Young, M. P. Anatomical connectivity defines the
organization of clusters of cortical areas in the macaque monkey and the cat. Philos. Trans. R. Soc. B 355, 91–110 (2000).
2. Sporns, O. & Honey, C. J. Small worlds inside big brains. Proc. Natl. Acad. Sci. USA 103, 19219–19220 (2006).
3. Lago-Fern´andez, L. F., Huerta, R., Corbacho, F. & Siguenza, J. A. Fast response and temporal coherent oscillations in
small-world network. Phys. Rev. Lett. 84, 2758–2761 (2000).
4. Gallos, L. K., Makse, H. & Sigman, M. A small world of weak ties provides optimal global integration of self-similar
modules in functional brain networks. Proc. Natl. Acad. Sci. USA 109,, 2825–2830 (2012).
5. Salvador, R. et al. Neurophysiological architecture of functional magnetic resonance images of human brain. Cereb.
Cortex 15, 1332–1342 (2005).
6. Hilgetag, C. C. & Goulas, A. Is the brain really a small-world network? Brain Struct. Func. 1–6; doi : 10.1007/s00429-
015-1035-6 (2015).
7. Egu´ıluz, V. M., Chialvo, D. R., Cecchi, G. A., Baliki, M. & Apkarian, A. V. Scale-free brain functional networks. Phys.
Rev. Lett. 94, 018102 (2005).
8. van den Heuvel, M. P., Stam, C. J., Boersma, M. & Hulshoff Pol, H. E. Small-world and scale-free organization of
voxel-based resting-state functional connectivity in the human brain. NeuroImage 43, 528–539 (2008).
9. de Solla Price, D. A general theory of bibliometric and other cumulative advantage processes. J. Am. Soc. Inform. Sci. 27,
292–306 (1976).
10. Barab´asi, A.-L. & Albert, R. Emergence of scaling in random networks. Science 286, 509–512 (1999).
11. Caldarelli, G. Scale-free networks: complex webs in nature and technology, (Oxford University Press, Oxford, 2007)
12. Beggs, J. & Plenz, D. Neuronal avalanches in neocortical circuits. J. Neurosci. 23, 11167-11177 (2003).
13. Expert, P. et al. Self-similar correlation function in brain resting-state functional magnetic resonance imaging. J. R. Soc. In-
terface 8, 472–479 (2011).
14. Shew, W. L. & Plenz, D. The functional benefits of criticality in the cortex. Neuroscientist 19, 88–100 (2013).
15. Haimovici, A., Tagliazucchi, E., Balenzuela, P. & Chialvo, D. R. Brain organization into resting state networks emerges
at criticality on a model of the human connectome. Phys. Rev. Lett. 110 178101 (2013).
16. Le, H. Complex network analysis: applications to human brain functional networks. Master's thesis, Universitat Pompeu
Fabra (2013). Available at:
http://www.upf.edu/csim/_pdf/_BestTheses/HoangLe_2012-13.pdf. Date of access: 22/03/2016.
17. Ferrarini L. et al. Non-parametric model selection for subject-specific topological organization of resting-state functional
connectivity. NeuroImage 56, 1453–1462 (2011).
18. Ruiz Vargas, E., Mitchell, D. G. V., Greening, S. G. & Wahl, L. M. Topology of whole-brain functional MRI networks:
improving the truncated scale-free model. Physica A 405, 151–158 (2014).
10/14
19. Humphries, M. D., Gurney, K. & Prescott, T. J. The brainstem reticular formation is a small-world, not scale-free, network.
Proc. R. Soc. B 273, 503–511 (2006).
20. Open Connectome Project, Available at: http://www.openconnectomeproject.org. Date of access:
25/05/2015.
21. Landman, B. A. et al. Multi-parametric neuroimaging reproducibility: a 3T resource study. NeuroImage 54, 2854–2866
(2011).
22. Hagmann, P. et al. Mapping the structural core of human cerebral cortex. PLoS Biol. 6, e159 (2008).
23. Honey, C. J. et al. Predicting human resting-state functional connectivity from structural connectivity. Proc. Natl. Acad.
Sci. 106, 2035–2040 (2009).
24. Lent, R., Azevedo, F. A. C., Andrade-Moraes, C. H. & Pinto A. V. O. How many neurons do you have? Some dogmas of
quantitative neuroscience under revision. Eur. J. Neurosci. 35, 1–9 (2012).
25. Humphries, M. D. & Gurney, K. Network 'small-world-ness': a quantitative method for determining canonical network
equivalence. PLoS ONE 3, e0002051 (2008).
26. Griffiths, R. B. Nonanalytic behavior above the critical point in a random Ising ferromagnet. Phys. Rev. Lett. 23, 17–19
(1969).
27. Munoz, M. A., Juh´asz, R., Castellano, C. & ´Odor G. Griffiths phases on complex networks. Phys. Rev. Lett. 105, 128701
(2010).
´Odor, G. Universality in nonequilibrium lattice systems, (World Scientific, Singapore 2008).
28.
29. Harris, T. E. Contact interactions on a lattice. Ann. Prob. 2, 969-988 (1974).
30.
31.
´Odor, G. & Pastor-Satorras, R. Slow dynamics and rare-region effects in the contact process on weighted tree networks.
Phys. Rev. E 86, 026117 (2012).
´Odor, G. Rare regions of the susceptible-infected-susceptible model on Barab´asi-Albert networks. Phys. Rev. E 87, 042132
(2013).
´Odor, G. Spectral analysis and slow spreading dynamics on complex networks. Phys. Rev. E 88, 032109 (2013).
32.
33. Cota, W., Ferreira, S. C. and ´Odor, G. Griffiths effects of the susceptible-infected-susceptible epidemic model on random
power-law networks. Phys. Rev. E 93, 032322 (2016).
34. Moretti, P. & Munoz, M. A. Griffiths phases and the stretching of criticality in brain networks. Nat. Commun. 4, 2521
(2013).
35. Villegas, P., Moretti, P. & Munoz, M. A. Frustrated hierarchical synchronization and emergent complexity in the human
36.
connectome network. Sci. Rep. 4, 5990 (2015).
´Odor, G., Dickman, R., & ´Odor G. Griffiths phases and localization in hierarchical modular networks. Sci. Rep. 5, 14451
(2015).
37. Bak, P., Tang, C. & Wiesenfeld W. Self-organized criticality. Phys. Rev. A 38, 364–374 (1988).
38. Hesse, J. & Gross, T. Self-organized criticality as a fundamental property of neural systems. Front. Syst. Neurosci. 8, 166
(2014).
39. Roncal, W. G. et al. MIGRAINE: MRI graph reliability analysis and inference for connectomics, Technical report. (2013)
Available at: http://arxiv.org/abs/1312.4875. Date of access: 22/03/2016.
40. Gray, W. R. et al. Magnetic resonance connectome automated pipeline: an overview. IEEE Pulse 3, 42–48 (2012).
41. Desikan, R. S. et al. An automated labeling system for subdividing the human cerebral cortex on MRI scans into gyral
based regions of interest. NeuroImage 31, 968–980 (2006).
42. Mori, S., Crain, B. J., Chacko, V. P. & van Zijl, P. C. M. Three-dimensional tracking of axonal projections in the brain by
magnetic resonance imaging. Ann. Neurol. 45, 265–269 (1999).
43. Mhembere, D. et al. Computing scalable multivariate glocal invariants of large (brain-) graphs. IEEE GlobalSIP 297–300
(2013) Available at:
http://ieeexplore.ieee.org/xpl/abstractAuthors.jsp?reload=true&arnumber=6736874.
Date of access: 22/03/2016.
44. Bialonski, S., Horstmann, M. T. & Lehnertz, K. From brain to earth and climate systems: small-world interaction networks
or not? Chaos. 20, 013134 (2010).
11/14
45. Jung, R. E. et al. Neuroanatomy of creativity. Hum. Brain. Mapp. 31, 398–409 (2010).
46. Nooner, K. B. et al. The NKI-Rockland sample: a model for accelerating the pace of discovery science in psychiatry.
Front. Neurosci. 6, 152. (2012).
47. Handcock, M. S. & Jones, J. H. Likelihood-based inference for stochastic models of sexual network formation. Theor.
Pop. Biol. 65, 413–422 (2004).
48. Burnham, K. P. & Anderson, D. R. Model selection and multimodel inference, (Springer, New York 1998).
49. Akaike, H. A new look at the statistical model identification. IEEE Trans. Automat. Contr. 19, 716–723 (1974).
50. Hurvich, C. M. & Tsai, C.-L. Regression and time series model selection in small samples. Biometrika 76, 297–307
(1989).
51. Burnham, K. P. & Anderson, D. R. Kullback-Leibler information as a basis for strong inference in ecological studies.
Wildlife Res. 28, 111–119 (2001).
52. Clauset, A., Shalizi, C. R. & Newman, M. E. J. Power-law distributions in empirical data. SIAM Rev. 51, 661–703 (2009).
53. Teimouri, M. & Gupta, A. K. On the three-parameter Weibull distribution shape parameter estimation. J. Data Sci. 11,
403–414 (2013).
54. Achard, A., Salvador, R., Whitcher, B., Suckling, J. & Bullmore, E. A resilient, low-frequency, small-world human brain
functional network with highly connected association cortical hubs. J. Neurosci. 26, 63–72 (2006).
55. Hayasaka, S. & Laurienti, P. J. Comparison of characteristics between region-and voxel-based network analyses in resting-
state fMRI data. NeuroImage 50, 499–508 (2010).
56. Joyce, K. E., Laurienti, P. J., Burdette, J. H. & Hayasaka, S. A new measure of centrality for brain networks. PLoS One 5,
e12200 (2010).
57. Gastner, M. T. & Newman, M. E. J. The spatial structure of networks. Eur. Phys. J. B 49, 247–252 (2006).
58. Watts, D. J. & Strogatz, S. H. Collective dynamics of "small-world" networks. Nature 393, 440–442 (1998).
59. Newman, M. E., Moore, C. & Watts, D. J. Mean-field solution of the small-world network model. Phys. Rev. Lett. 84,
3201-3204 (2000).
60. Jackson, M. O. Social and economic networks, (Princeton University Press, Princeton 2008).
61. van den Heuvel, M. P., Stam, C. J., Kahn, R. S. & Hulshoff Pol, H. E. Efficiency of functional brain networks and
intellectual performance. J. Neurosci. 29, 7619–7624 (2009).
62. Ravasz, E. & Barab´asi, A.-L. Hierarchical organization in complex networks. Phys. Rev. E 67, 026112 (2003).
63. Penrose, M. Random geometric graphs, (Oxford University Press, Oxford, 2003).
64. Dall, J. & Christensen, M. Random geometric graphs. Phys. Rev. E 66, 016121 (2002).
65. Telesford, Q. K., Joyce, K. E., Hayasaka, S., Burdette, J. H. & Laurienti, P. J. The ubiquity of small-world networks.
Brain Connect. 1, 367–375 (2011).
66. Fronczak, A., Fronczak, P. & Hołyst, J. A. Average path length in random networks. Phys. Rev. E 70, 056110 (2004).
67.
´Odor, G. Critical dynamics on a large human Open Connectome network. Preprint: arXiv:1604.02127.
Supplementary Information:
Justification of AICc as model selection criterion
Among the various criteria for model selection that have been proposed in the literature, the Akaike and Bayesian information
criteria (commonly abbreviated as AIC and BIC, respectively) are those that are most frequently employed in the analysis of
empirical data. Both criteria have the general form
−2 ln(L (v)) + aK,
(S1)
where L is the likelihood function, v the model parameters that maximize L , and K is the number of parameters. The
difference between AIC and BIC is the prefactor a in the last term,
a =(2
ln(N)
for AIC,
for BIC,
(S2)
12/14
AIC
BIC
c
k
d
e
t
t
i
f
0
5
1
0
0
1
0
5
0
0
50
100
150
true k c
Figure S6. The true crossover point kc in Eq. (S3) versus values fitted by selecting among the candidate models of Eq. (S4).
The selected models either minimize AIC (black filled circles) or BIC (red open squares). The diagonal line indicates perfect
prediction. Standard errors are smaller than symbol sizes.
where N is the number of data points. For N ≥ 8, the BIC gives thus a larger penalty than the AIC for every additional
parameter in the model.
BIC-based model selection is known to be consistent in the sense that it almost surely selects the true model if it is
among the candidates and N is infinitely large.1 AIC-based model selection is generally not consistent, but unlike the BIC
it is, in statistical parlance, efficient: in the limit N → ¥
the AIC selects a candidate model with the smallest Kullback-
Leibler divergence from the true model, even if the true model is not among the candidates.2, 3 The true model is in most
biological applications indeed unlikely to be included in the candidate set because it usually contains a large number of
unknown parameters, which may furthermore increase with N. The consistency of the BIC and the efficiency of the AIC
. While it is insightful to know under which conditions the AIC or BIC are
asymptotically optimal, these properties are ultimately of theoretical rather than practical relevance for finite data sets. The
best guidance which criterion to choose comes from simulation studies involving finite data.
are asymptotic statements valid for N → ¥
With a random number generator we have created finite data that pose a similar challenge to the model selection algorithm
as the observed degree distributions. The data are N = 106 random numbers, independently sampled from the distribution
Pr(kobs > k) =
(exp(cid:2)−b (kc + a )−1k(cid:3)
C(kc,a ,b )× (k + a )−b
if 0 < k ≤ kc,
if k > kc,
(S3)
where kobs is the generated random integer and the prefactor on the second line is C(kc,a ,b ) = (kc +a )
The tail of this distribution is a power law. If k were real-valued rather than an integer, the right-hand side of Eq. (S3) would
be differentiable everywhere. We have chosen this distribution for two reasons. First, we want to test whether it is possible
to determine with information criteria that the true model is a power-law (POW in Table 1). Second, we want to find out
which criterion is better at determining the precise crossover point kc in the presence of random fluctuations. The motivation
is that, if the brain were indeed scale-free, we would correctly identify the power-law model and its parameterization. We
fixed a = 10, b = 2.5 and allowed kc to vary.
b exp(cid:2)−b (kc + a )−1kc(cid:3).
For our first exploratory Monte Carlo simulations, we compared whether AIC- or BIC-based model selection is more
successful at determining the true kc from the POW candidate models
Prcandidate(kobs > k) =
(S4)
1− (cid:229)
(1− (cid:229)
b
if 0 < k ≤ kc,
if k > kc,
.
k
i=1 Ai
kc
i=1 Ai)(cid:16) kc+a
k+a (cid:17)
where a , b and A1, . . . ,Akc are adjustable parameters. The crossover point kc is not itself a parameter in Eq. (S4), but rather
an index for different candidate models.
We summarize the results in Fig. S6. The AIC reconstructs the crossover point almost accurately in the entire range
0 ≤ kc ≤ 150. The BIC underestimates kc, especially when kc is large. The reason lies in the BIC's steeper penalty for
13/14
additional parameters in Eq. (S2). Another obstacle for the BIC is that the true model (i.e. the distribution of Eq. S3 with the
three parameters a , b and kc) is generally not one of the candidates of Eq. (S4) because those contain more parameters than
necessary, in particular one parameter Ai for every 0 < i ≤ kc. The AIC, by contrast, correctly chooses a candidate that still
(k + a )−b
describes the input best among the wrong (i.e. overparameterized) models: for k > kc the distribution decays (cid:181)
whereas there is no power law in the region 0 < k ≤ kc.
Having found that the AIC is better suited for model selection in the present context, we investigate next whether the
AIC reliably identifies the correct model among those listed in Table 1. We generate 10 independent sets of N = 106 random
numbers each. These are drawn from the distribution of Eq. (S3) with a = 10, b = 2.5, and now we also fix kc = 100. For
8 out of these 10 sets, the power law had the smallest AIC. In the remaining two, the power law's AIC differed from the
minimum only by D POW = 0.05 in one case and D POW = 0.30 in the other case. As we have explained after Eq. (5), such small
D POW are still interpreted as substantial empirical support for the power law. We conclude that, if the connectome's degree
distribution were scale-free, the AIC would have correctly identified a power law as a plausible candidate model.
observed network is not the full scale-free network, but only a random sample of 20% of the edges. Such subsampling
of the true network might mimic that the data are collected at a finite resolution and therefore only a fraction of the fiber
tracts might be detected. For a fixed power-law degree distribution (a = 10, b = 2.5, N = 2· 106, kc = 0), we first construct
a corresponding network with the configuration model.4 We then randomly select 20% of the edges. The degrees of this
subgraph have the distribution5
P(kobs = j) =
i= j(cid:18)i
j(cid:19)0.2 j 0.8i− jP(ktrue = i).
(S5)
Here ktrue is the true degree of a node (i.e. in our model a power-law distribution POW as in Table 1 of the main text) and kobs
is the degree observed in the sampled network. We removed all nodes with degree zero from the network to be consistent with
the procedure used by the OCP.
Applying AIC-based model selection with the candidates of Table 1, we find that the differences between the heavy-tailed
distributions POW, LGN, TPW and GWB are smaller for the sampled network than for the full network. In ten test runs, we
could always rule out EXP, WBL and in one case LGN because D > 10, but otherwise D was below this threshold. Still, POW
had the smallest AIC in eight out of ten cases and came in a close second in the remaining two. In summary, AIC-based model
selection would have correctly included the power law in the set of plausible candidates even if the observed network had been
a small sample of a network with a scale-free degree distribution.
There is one fine detail to note when the dimension K of the model approaches the sample size N. In such cases the AIC is
a negatively biased estimate of the Kullback-Leibler information.50 An approximate correction for this bias is the additional
penalty term 2K(K+1)
N−K−1 in Eq. (4). This modified version of the AIC is commonly abbreviated as AICc. Although the difference
is small in the present context, it has been suggested that AICc should routinely be used instead of the AIC.50, 51 We have
followed this advice throughout this article.
References
1. Kim, Y., Kwon, S. & Choi, H. Consistent Model Selection Criteria on High Dimensions. Journal of Machine Learning
Research 13, 1037–1057 (2012).
2. Vrieze, S. I. Model Selection and Psychological Theory: A Discussion of the Differences Between the Akaike Information
Criterion (AIC) and the Bayesian Information Criterion (BIC). Psychological Methods 17, 228–243 (2012).
3. Aho, K., Derryberry, D. & Peterson, T. Model selection for ecologists: the worldviews of AIC and BIC. Ecology 95,
631–636 (2014).
4. Molloy, M. & Reed, B. A critical point for random graphs with a given degree sequence. Random Structures & Algorithms
6, 161–179 (1995).
5. Stumpf, M. P. H., Wiuf, C. & May, R. M. Subnets of scale-free networks are not scale-free: sampling properties of
networks. Proc. Nat. Acad. Sci. 102, 4221–4224 (2005).
14/14
¥
(cid:229)
|
1901.03990 | 1 | 1901 | 2019-01-13T14:22:37 | Formation of three-dimensional auditory space | [
"q-bio.NC"
] | Human listeners need to permanently interact with their three-dimensional (3-D) environment. To this end, they require efficient perceptual mechanisms to form a sufficiently accurate 3-D auditory space. In this chapter, we discuss the formation of the 3-D auditory space from various perspectives. The aim is to show the link between cognition, acoustics, neurophysiology, and psychophysics, when it comes to spatial hearing. First, we present recent cognitive concepts for creating internal models of the complex auditory environment. Second, we describe the acoustic signals available at our ears and discuss the spatial information they convey. Third, we look into neurophysiology, seeking for the neural substrates of the 3-D auditory space. Finally, we elaborate on psychophysical spatial tasks and percepts that are possible just because of the formation of the auditory space. | q-bio.NC | q-bio |
Formation of three-dimensional auditory space
Piotr Majdak∗, Robert Baumgartner∗, Claudia Jenny
Acoustics Research Institute, Austrian Academy of Sciences, Vienna, Austria
Abstract
Human listeners need to permanently interact with their three-
dimensional (3-D) environment. To this end, they require efficient per-
ceptual mechanisms to form a sufficiently accurate 3-D auditory space.
In this chapter, we discuss the formation of the 3-D auditory space from
various perspectives. The aim is to show the link between cognition,
acoustics, neurophysiology, and psychophysics, when it comes to spa-
tial hearing. First, we present recent cognitive concepts for creating
internal models of the complex auditory environment. Second, we de-
scribe the acoustic signals available at our ears and discuss the spatial
information they convey. Third, we look into neurophysiology, seeking
for the neural substrates of the 3-D auditory space. Finally, we elab-
orate on psychophysical spatial tasks and percepts that are possible
just because of the formation of the auditory space.
∗contributed equally; [email protected]; [email protected]
1
1
Introduction
A loud roar and you turn around. Fight or flight? An archaic situation, typ-
ical for our ancestors and animals still living in their original environment.
This concept still applies when transferred to the modern human. "Hi!" and
you turn around - a good friend has just recognized you on a street. Before
you change your walking direction, a car honk lets you look back - you've
just missed that car crossing your path.
Such situations make it obvious: in the jungle and on the street, human
and non-human animals both need a good understanding of the 3-D world
by means of auditory perception. Hearing allows us to create a map of
our environment in order to react in a proper way. Towards this goal, the
auditory system has to answer the question "what is where?".
In this chapter, we review recent advances in understanding how human
listeners form and use the auditory space - from the cognitive, acoustic,
neurophysiological, and psychophysical perspectives. To this end, in Sec. 2,
we describe the problem and elaborate on the potential solutions given by
researchers from cognitive psychology. As the perceptual outcome depends
on the quality with which the acoustic spatial information is conveyed, in
Sec. 3, we describe how this information is encoded within binaural signals.
The solution to the problem should parallel the neural processing actually
happening in our auditory system, and thus, in Sec. 4, we briefly describe the
neural processing of the acoustic signals while focusing on the extraction of
spatial information. Finally, the result of that processing reflects the variety
of spatially oriented tasks human listeners can complete.
In Sec. 5, we
2
describe various psychophysical spatial tasks demonstrating our abilities to
utilize our understanding of the 3-D auditory world.
2 Cognition: representing the world
From the cognitive perspective, the auditory system not only needs to ad-
dress the question "what is where?" in order to form a mental representation
of our environment because sound is ephemeral, i.e., is happening and short-
lived; it is an effect of events, providing information on what happens right
now instead of a long-lasting description of objects' properties. Thus, the
information carried by the sound not only needs to be stored for its pro-
cessing, it also requires consideration of various time scales. On a short
time scale, a crack might mean someone stepping on the floor. Many sim-
ilar cracks, however, would indicate someone opening a door. Thus, the
ephemeral property of sound requires the auditory system to address the
question "what is happening?". Answering that question can be only a ba-
sis for understanding a static environment. Our world, however, is dynamic
and we continuously interact with it. In order to decide which actions have
to be done, our auditory system has to provide a basis for the prediction of
"what will happen next?".
2.1 The ill-posed problem
Auditory scenes consist of objects producing sounds (see Fig. 1a). Percep-
tion, as a process of transforming sensory information to higher levels of
representation, needs to mentally represent these objects and their proper-
3
ties. Here, an "auditory object" can be thought of as a perceptual construct
linking a sound with a corresponding source (Griffiths and Warren, 2004).
Sitting in a park, hearing a honk, a word, and a chirp would let us identify
auditory objects of a car, a human, and a bird. Auditory objects can be as-
signed to a spatial position, among other non-spatial properties. Now, when
the human starts to speak, and the bird starts to sing, these two objects
become sources of acoustic streams. Their mixture arrives at our ears and
the job of the auditory system is to separate them into two auditory streams,
which can be defined as a series of coherent events that can be grouped and
attributed to a single auditory object. Hence, the formation of the auditory
space depends on the capabilities of the listener to form auditory objects
and estimate their spatial properties.
As most of the sound sources are usually located outside of our body,
the formation of the auditory space can be seen as a perceptual task aim-
ing at the reconstruction of the external (or distal) state of affairs (Epstein
and Rogers, 1995). Unfortunately, this is an ill-posed problem emerging
from the fact that not only does the head-centered binaural signal need to
be transformed to a world-centered representation, but also that the infor-
mation available at our ears is insufficient to exactly reconstruct a unique
state. This is because a given binaural signal arriving at our ear drums can
originate from an infinite number of sound-producing events, all of which
having produced exactly that signal. This ill-posed problem results, as a
consequence, in ambiguity in creating auditory objects.
A famous example showing the difficulty of mapping acoustics to a
sound-producing event is the estimation of a 2-D surface given a sound
4
Figure 1: a) Perception as a mental representation of the environment based
on sensation. The objects (left) producing sounds that are perceived by a
listener (center) are represented in an internal model (right). b) Active
inference. Sensations (top) of signals from the environment (left) are used
to update the internal model beliefs (right) on the hidden states of the
environment. The model predictions drive actions (bottom) which allow to
interact with the environment.
wave, posed as "can we hear the shape of a drum?". Indeed, it has been for-
mally shown that for many shapes, one cannot completely differentiate the
5
a)ReflexesProprioceptionMOTOREyesHeadLimbs …SENSATIONSAuditoryVisualSomatic …ENVIRONEMENTHidden statesINTERNAL MODELInternal states(Beliefs)b) shape of the drum because an exact solution of the plane geometry from the
waveform is impossible (Gordon et al., 1992). However, additional or even
redundant information can help and, fortunately, our acoustic environment
is full of redundancy. Recent developments in mathematical tools show the
variety of tricks the auditory system may potentially use when it comes to
utilizing acoustic redundancy and thus better solving inverse problems. For
example, with a known but monaural only signal, the shape of any convex
room can be estimated from the delays between the early reflections in the
room impulse response (Moore et al., 2013). Given four acoustic sensors, the
room shape can be estimated by relying on the delay of just the first reflec-
tions (Dokmanic et al., 2013), showing the power of better solutions given
redundancy (here, multiple sensors). Thus, it is not surprising that binaural
hearing, by providing interaural information as compared to monaural hear-
ing, allows us, to better retrieve the spatial properties of the environment,
better orient ourselves within the environment, and even improve speech
perception (Bronkhorst, 2015, and also Clapp and Seeber, this volume).
Despite the redundancy in the binaural signal, solving the ill-posed prob-
lem requires certain assumptions in order to reduce the infinite number of
potential solutions (e.g., Friston, 2012). Generally, these assumptions are
driven by the goal of an efficient interaction with the environment sampled
by the sensors. At the end of the cognitive process, the solution needs to
provide a basis for decisions that trigger the execution of appropriate ac-
tions. If the vast amount of sensory information was processed from scratch
each time an action was required, most of the actions would happen too late.
Faster processing can be achieved with predictive coding by introducing an
6
internal model (Francis and Wonham, 1976) that predicts the external state
of affairs and is continuously adapted based on incoming sensory data (for
more details on predictive coding, see Aitchison and Lengyel, 2017). Such
models have been introduced in motor-control theory and robotics to de-
scribe reaching movements, to plan movement trajectories, and to model
imagery (for review and discussion, see Grush, 2004).
In cognition, the
term "perceptual inference" has been coined (e.g., Hinton and Ghahramani,
1997).
In predictive coding, the more realistic the model predictions, the more
efficiently actions can be performed. In other words, the objective of the
internal model is to minimize surprise, which then yields minimal correc-
tions to be applied to the performed actions.
In that sense, the process
of cognition can be considered as forming a generator creating hypotheses
and then testing them against the pre-processed sensory information (Gre-
gory, 1980). The free-energy principle has been proposed to explain how
the cognitive system can efficiently create a model predicting the environ-
ment while restricting itself to a limited number of states (Friston et al.,
2006). Free energy, a concept with a long tradition in thermodynamics
(Helmholtz, 1954), is the difference between the internal energy of a system
and the energy required to describe the actual states of that system. In the
free-energy concept, the lower is the free energy, the more efficiently is the
system described. In terms of cognition, a cognitive model creates plausible
predictions that minimize surprise given the states described by the present
sensory information and the internal states describing the model's beliefs
about the environment (hidden states). The free energy acts as a prediction
7
error and is minimized by choosing that most plausible prediction which
most efficiently drives the motor system. The internal states of the model
are then updated based on the new sensory information about the hidden
states of the environment. This process could be termed "active inference"
(for review, see Friston et al., 2016).
2.2 Rules and limitations
In active inference, a model's beliefs represent rules describing plausible en-
vironments. They can be learned throughout the development (Bhatt and
Quinn, 2011) and limit the potential solutions to the plausible ones only.
This application of limitations has several implications. First, they some-
times fail, yielding unrealistic representations. Illusions, i.e., distortions of
the perceived physical reality, are great examples of the consequences of
plausible but wrong assumptions while solving the ill-posed problem (e.g.,
Carbon, 2014). Understanding their origin can help uncover the underly-
ing processes in auditory perception. Second, these limitations allow us to
reduce the vast amount of sensory information to a smaller number of in-
formational units along the ascending pathways of processing. In our case,
a small and discrete number of auditory objects with a finite number of
properties is created from a continuous binaural signal.
In the process,
the frame of reference undergoes a transformation from the head-centered
binaural information to world-centered information about our environment
(Schechtman et al., 2012).
Interestingly, this whole process can be seen as a nonlinear extension of
compressed sensing, a signal processing technique for efficiently acquiring
8
and reconstructing a signal by finding solutions to underdetermined linear
systems (Donoho, 2006).
In compressed sensing, the constraint of spar-
sity is chosen in order to find a solution to the underdetermined system.
Compressed sensing is widely used in signal processing, but it requires a
linear relation between the observation and solution.
In active inference,
this process is described by means of variational Bayes statistics, which
aims to provide an analytical approximation to the posterior probability of
the unobserved variables in order to obtain statistical inference about these
variables.
At the end, it is all about reducing the amount of sensory information.
A widely accepted concept describing the reduction of auditory information
to discrete informational units is auditory scene analysis (ASA, see van de
Par et al., in this volume and Bregman, 1990). ASA assumes that our
auditory system partitions the acoustic signal into auditory streams that
each refer to an auditory object. While good separability between fore- and
background streams was originally believed to constitute the goal (Bregman,
1990), more recent models consider their predictability as the main motiva-
tion for grouping (Winkler et al., 2009). Grouping mechanisms seem to rely
on auditory features like onset, pitch, spectrum, and interaural disparity,
and can act simultaneously and sequentially. Simultaneous grouping as-
sumes that features are integrated to a foreground property, like harmonics
coming from the same instrument are integrated to a single pitch, and fea-
tures deviating from the expected and learned patterns are segregated and
form a background. Sequential grouping integrates and segregates auditory
objects and streams, depending on their temporal properties. It can even
9
override the result of simultaneous grouping: For example, sounds simulta-
neously presented with a different interaural disparity can be grouped into a
single auditory stream, but the same sounds embedded in an acoustic stream
presented from the spatial location corresponding to one of the sounds can
create two auditory objects appearing at the distinct spatial locations (Best
et al., 2007).
Bregman's ASA further involves the concept ("old-plus-new") of compet-
itive processes determining the auditory stream reaching awareness. These
processes have been originally derived as obeying the laws of the Gestalt
theory, which provides a description of our ability to acquire plausible per-
ceptions from the sensory input (Koffka, 1935). The main assumption of
the Gestalt theory is that our perception is influenced by perceptual units
derived from the signal according to the laws of proximity, similarity, clo-
sure, symmetry, common fate, continuity, good Gestalt, and past experience.
Even though the Gestalt theory has difficulties in providing insights into the
neural processes leading to perception (Schultz and Schultz, 2015), the laws
of the Gestalt theory helped in constraining the ambiguity resulting from
the ill-posed problem of perception. Recent findings show that these neural
processes are based on heuristics acquired through learning and experience
(Shinn-Cunningham, 2008). Further, these processes act "top-down" and
can be modulated by other modalities like vision (yielding ventriloquism
or self-motion; Kondo et al., 2012) or auditory attention (Hill and Miller,
2010).
Depending on the relevance of top-down modulations, neural processes
may be distinguished as being reflexive or reflective. Reflexive processes
10
result in very fast reactions (with latencies below 100 ms), which can usu-
ally not be suppressed (Curtis and D'Esposito, 2003). They involve startle
reflex or orienting reflex (Sokolov, 2001), and do not require, but can be
modulated by attention. They can be used to trigger movements toward
auditory sources or intensify the processing of cues that signal approaching
objects (Baumgartner et al., 2017). Thus, they are vital in protecting hu-
mans from hazardous events. In contrast, reflective processes have longer
latencies, require top-down attention, a controlled bias in the preference for
and processing of the information streams (for review, see Knudsen, 2007).
Attention is usually thought to be a single unidirectional process repre-
senting task-specific goals and expectations (top-down, Awh et al., 2012),
however, it can also be modulated by various components of a stimulus
(bottom-up, Arnal et al., 2015), i.e., salient components. Reflective or at-
tentional processes require working memory to develop and test hypotheses
based on the salience in a stimulus (Carlile and Corkhill, 2015). A similar
distinction between reflexive and reflective processes has also been proposed
for speech category learning (Chandrasekaran et al., 2014) and has been
used to describe social behavior (Strack and Deutsch, 2004).
While the reflexive processes in auditory processing have been widely in-
vestigated, the actual reflective processes behind the ASA are not completely
understood yet. Thus, it is not surprising that there is much debate on the
mechanisms underlying ASA, which has been widely discussed from differ-
ent perspectives (e.g., Bizley and Cohen, 2013; Szab´o et al., 2016; Micheyl
et al., 2007; Nelken et al., 2014; Snyder and Elhilali, 2017).
In summary, the free-energy concept provides a solid statistical frame-
11
work for deriving the external state of affairs, and ASA provides a valid
conceptual framework for the cognitive processing of auditory information
(for more details, see van de Par, this volume). The particular result in
terms of a realistic representation of our world depends on the quality with
which the spatial information about the auditory objects is conveyed by the
binaural signal. Thus, in the following section, we describe the acoustic
spatial information encoded in binaural signals.
3 Acoustics: formation of binaural signals
The sounds arriving at our ear drums are acoustically filtered versions of
the sound actually produced in our environment. The filtering is a result of
the interaction of the sound field with the reverberant space and our body
parts such as the head, torso, pinna, and ear canal.
In acoustics, sound fields are commonly described in Cartesian or spher-
ical coordinates. In psychoacoustics, however, the interaural-polar coordi-
nate system, Fig. 2a, better corresponds with the human perception. This
system is created by rotating the poles of the spherical system by 90◦, such
that the two poles are aligned with the interaural axis connecting the two
ears. In that system, the lateral angle α describes the lateral position of a
source, ranges from -90◦ to +90◦, selects the sagittal plane (i.e., all parallel
planes of the median plane), and for directions in the horizontal frontal half-
plane, corresponds to the azimuth angle. The polar angle β describes the
position of the source along the sagittal plane and ranges from -90◦ (bottom)
via 0◦ (eye-level, front), 90◦ (top), and 180◦ (eye-level, back) to 270◦ (bot-
12
Figure 2: a) Interaural-polar coordinate system with the lateral angle α,
the polar angle β, and distance r; b) magnitude spectra of far-field HRTFs
along the median plane as a function of the polar angle; c) energy-time
curves (ETCs) of far-field HRIRs of a left ear along the lateral angle to
the front (top) and back (bottom) of the horizontal plane; d) magnitude
spectra of near-field HRTFs of a left ear for the direction (α, β): (−90◦, 0◦)
(the most-right direction at the horizontal plane) as a function of distance.
Color bar: magnitude in dB.
tom again). Together with the distance r, we use lateral and polar angles
throughout this chapter to describe the spatial position of sound sources.
3.1 Monaural features
For a sound coming from a particular direction, its filtering can be captured
by the binaural pair of head-related transfer functions (HRTFs). While the
13
filtering of the sound happening in the ear canal does not depend on the in-
cidence angle of the sound, the filtering by the head, torso, and pinnae does,
creating direction-dependent changes of the received sound. The directional
information provided by individual HRTFs is referred to as the monaural
spectral features. Their direction-dependent changes are especially apparent
for a sound moving along the median plane of the listener (see Fig. 2b). The
change in the elevation and front/back dimension can be directly linked with
the spectral changes, that is, changing notches and peaks as a consequence
of cancellation and amplification, respectively, caused by various body parts.
The reflections of the torso create spatial frequency modulations up to 3 kHz
(Algazi et al., 2001). The head shadows frequencies above 1 kHz; however,
above 6 to 8 kHz, the effect of the pinna is most prominent (Blauert, 1997).
The reflections at the pinna create peaks and notches at frequencies above
4 kHz. For example, the directionality of the pinna towards the front causes
an attenuation of high frequencies for sounds coming from behind the lis-
tener. This manifests in spectral peaks above 8 kHz for the frontal sound
positions. Further, an increase in sound elevation changes the varying delay
between the direct sound and its reflection along the pinna concha. This
manifests in an upward shift of the spectral notches usually found between
6 and 10 kHz. A further contribution is that of the head.
In general, the pinna is the reason for a large variation in the HRTFs
among listeners, as the pinna's geometry varies among the healthy human
population (Algazi et al., 2001). While HRTFs are similar across listeners
at frequencies up to 6 kHz, differences as large as 20 dB have been found
at higher frequencies (Møller et al., 1995). Listener-specific HRTFs can be
14
acquired by applying system identification approaches on acoustical mea-
surements (for review, see Majdak et al., 2007). The acoustic measurement
including all positions in 3-D space is a resource-demanding procedure and
takes tens of minutes, even when sophisticated measurement methods are
applied. HRTFs can also be calculated based on a geometric representation
of the listener (Kreuzer et al., 2009); however, the demands on geometric
accuracy and computational power are high (Ziegelwanger et al., 2015). Re-
cent developments in the acquisition of the 3-D geometry from photographs
by means of photogrammetric reconstruction (Reichinger et al., 2013) and
numeric algorithms (Ziegelwanger et al., 2016) seem promising in easing the
acquisition of listener-specific HRTFs in the future. While (for research pur-
poses) HRTFs have been measured for a long time, their exchangeability was
limited because of missing standards for their representation. Recently, the
spatially oriented format for acoustics (SOFA) was created (Majdak et al.,
2013a) as a standard, easing their exchangibility, and promoting their usage
in consumer applications.
HRTFs can also be analyzed in the time domain by applying the in-
verse Fourier transformation on each HRTF yielding head-related impulse
responses (HRIRs, see Fig. 2c). HRIRs usually decay within the first 4 ms
and show the direction-dependent delay between the sound source and the
ear. The temporal position of the first onset in an HRIR can be considered as
the broadband time-of-arrival (TOA), which, based on the approximation of
the head as a sphere, for a listener, can be modeled as a spatially-continuous
3-D representation of the broadband delay requiring only few parameters
(Ziegelwanger and Majdak, 2014). Even though HRTFs show a nonlinear
15
spectral phase depending on the sound direction, the HRTF phase spectrum
can be represented by a combination of the minimum phase derived from the
HRTF amplitude and the linear phase corresponding to the TOA (Kulkarni
et al., 1999).
HRTFs also vary with distance, especially in the near field (see Fig. 2d).
This is due to the contribution of the head shadow and changes of the pinna-
reflection paths (Brungart and Rabinowitz, 1999). The nearer the source,
the less diffraction around the head occurs at lower frequencies and the less
intense are the reflection patterns of the pinna. Low-frequency attenuation
of up to 20 dB is a prominent spatial feature encoding distance for near
sounds.
Note that the sound source itself might be a source of spatial cues because
the HRTF filtering is commutative, i.e., the auditory system has no chance
to distinguish whether the evaluated spectrum originates from the sound
source or is an effect of filtering by an HRTF. For example, when listening
to a frequency sweep of notched noise, the uprising spectral notch may be
the reason for the illusion of an elevating source (Blauert, 1997). Alternately,
sounds with spectral ripples may overlap with the monaural spectral features
of HRTFs and interfere with the derivation of directional information from
the binaural signal (Macpherson and Middlebrooks, 2003).
3.2
Interaural cues
Having two ears allows us to sense the sound field at two different spatial
positions. Thus, we have access not only to monaural features, but also to
the combination of the left- and right-ear features, the so-called interaural
16
Figure 3: a) Interaural cross-correlation coefficients (IACCs) of a single
listener's HRTFs for various directions when looking at the listener from the
front. b) IACCs along the median plane calculated for a listener population.
c) IACCs along the horizontal front (gray) and rear (green) half-planes
calculated for a listener population. The population consisted of 97 listeners
from the ARI database (Majdak et al., 2010). The dots show the individual
IACCs, the line and gray area shows the total average and ±1 standard
deviation, respectively, across the population.
cues. The most prominent examples of the interaural cues are the broadband
interaural time and level differences (ITDs, ILDs). Their importance for
sound localization was recognized very early (Lord Rayleigh or Strutt, 1876).
Later, the general dissimilarity of the signals between the two ears, expressed
as binaural incoherence and spectral ILDs, have been found to be important
(Blauert, 1997).
3.2.1
Interaural signal similarity
The similarity between the signals of the listener's two ears has been de-
scribed by many terms:
interaural coherence, interaural cross-correlation,
binaural incoherence, interaural decorrelation, or even the binaural quality
index. And all of them consider the temporal similarity between the left-
17
and right-ear signals, thus intrinsically describe spectral features. The in-
teraural cross-correlation function rLR of the two signals xL and xR is a
function of the interaural lag τ (Goupell and Hartmann, 2006):
(cid:82) xL(t)xR(t + τ ) dt
(cid:113)(cid:82) x2
(cid:82) x2
L(t1) dt1
R(t2) dt2
rLR(τ ) =
,
(1)
This function can vary between -1 and 1 and typically has a single peak.
The lag of that peak corresponds to the broadband ITD and is mostly de-
termined by the lateral position of a sound source. The height of that
peak is usually known as the interaural cross-correlation coefficient (IACC)
and demonstrates the best interaural similarity of the binaural signal (see
Fig. 3a). Note that IACCs are naturally below 1. Typically, free-field IACCs,
even in the median plane, are around 0.9 (see Fig. 3b), because the HRTFs
are not identical as a consequence of non-identical ears. In the horizontal
plane, the IACC decreases with increasing lateral angle of the sound source,
with typical IACCs around 0.4 for the most-lateral directions (see Fig. 3c).
Note that this is a broadband consideration of the IACC and the interaural
dissimilarities may be different and contribute differently across frequencies.
3.2.2
ITD
When talking about the ITD, people usually refer to the broadband ITD.
However, given the limited interaural coherence and the attendant frequency
dependence, a broadband ITD can only be an approximation of delays ap-
pearing between the two ears. Consequently, various methods have been
proposed for the calculation of the ITD, based on HRTFs represented in the
18
time or frequency domains. Figure 4a shows the ITDs as a function of the
lateral angle obtained from an HRTF set by using various methods. In the
time domain, methods either focused on the ITD between the first onsets
(MAX in Fig. 4a, see Andreopoulou and Katz, 2017) and centroids (CTD),
or the lag of the IC peak of HRIRs compared to their minimum-phase ver-
sions (MCM). In the frequency domain, the ITD can be calculated from the
spectral average of the interaural group delay (AGD). While the correct es-
timation method clearly depends on the application, recently, ITDs between
the -30-dB onsets (relative to the peak) in low-pass filtered HRIRs showed
the best congruence with psychophysical tasks when tested with broadband
sounds in humans (Andreopoulou and Katz, 2017). From the geometrical
perspective, there is a long history of various ITD models based on repre-
sentations of the head as a circle, sphere, ellipsoid, and polynomial function
(Xie, 2013).
Generally, ITDs are frequency dependent and theoretical considerations
show that low-frequency ITDs are 50% larger than those at higher frequen-
cies (Kuhn, 1977) with a transition around 1.5 kHz. The maximal ITD
depends on the listener's head diameter and the calculation method and has
a population average of around 850 µs (Algazi et al., 2001). ITDs in that
range imply that sounds with frequencies below 1.2 kHz undergo an inter-
aural phase shift of less than 180◦ when traveling from one ear to the other,
and the interaural phase difference can unambiguously encode the source
direction. At higher frequencies, sounds with wavelengths smaller than the
head diameter provide interaural phase differences larger than 180◦ yielding
ambiguous ITDs. Hence, for stationary tones, ITDs in higher frequencies do
19
provide unique information about the sound's lateral direction. In the case
of amplitude-modulated and multi-tone sounds, the timing of the envelopes
may also be informative even at higher frequencies, yielding envelope ITDs
as a useful acoustic feature (Henning, 1974).
But even broadband ITDs do not provide much information about the
sound's spatial position beyond its lateral direction. Fig. 4b shows contours
of iso-ITDs derived from an HRTF set of a listener. The contours approx-
imate sagittal planes, demonstrating that ITDs are not able to encode the
sound's distance and elevation, nor enable discrimination between front and
back. This finding is not new, as Lord Rayleigh mentioned already in 1876:
"The possibility of distinguishing a voice in front from a voice behind would
thus appear to depend on the compound character of the sound in the way
that it is not easy to understand, and for which the second ear would be
of no advantage" (Lord Rayleigh or Strutt, 1876). Nowadays, the spatial
ambiguity based on the ITD is called the "cone of confusion", or "torus of
confusion" when distance is involved (Shinn-Cunningham et al., 2000).
3.2.3
ILD
ILDs arise because of two effects. First, the head is an obstacle, creating
a shadow for the contralateral ear, and thus ILDs. Because of that, ILDs
increase with both frequency and lateral angle (see Fig. 5a), defined by the
relation between the sound's wavelength and the listener's head size. While
low-frequency ILDs span a range of ±5 dB and increase smoothly for more
lateral sounds, high-frequency ILDs exhibit a span of ±20 dB, with a more
complex relation to the lateral angle. Second, for near-field sounds, the
20
Figure 4: a) ITDs for frontal directions at the horizontal plane, estimated
by various ITD extraction methods (see text). b) Iso-ITD contours; view
at the left ear along the interaural axis. Color bar: ITD in ms.
sound intensity decreases rapidly with the distance to the source, creating
an ILD even at frequencies for which the head is acoustically transparent,
(see Fig. 5c). Such low-frequency near-field ILDs become significant for
distances below 0.5 m (Brungart and Rabinowitz, 1999) and can reach even
beyond 20 dB.
Similar to the front-back ambiguity of the ITD, ILDs do not vary consis-
tently with the polar angle (see Fig. 5c). Thus, ILDs do not encode source
directions along the sagittal planes well either, further contributing to the
cone of confusion based on interaural features.
3.3 Reverberation
Our considerations so far addressed the binaural signal without taking its
origin into account. A single sound source presented in the free field, that
is, just direct path between the sound and listener without any reflections,
21
Figure 5: a) Frequency-dependent ILDs for sounds along the frontal horizon-
tal half-plane. b) Iso-ILD contours; view at the left ear along the interaural
axis. c) Frequency-dependent ILDs for the most-left direction in the hor-
izontal plane as a function of distance. Color bar: ILD in dB. Frequency
dependence shown by filtering each HRTF by a typical Gammatone filter.
is the simplest case.
However, most realistic binaural signals involve listening in reverber-
ant spaces like rooms. When we consider the simple case of a source and
its reflection from a wall, the binaural signal will contain the direct signal
overlapped with filtered versions of itself. The filtering consists of a broad-
band delay (because of the longer propagation path) and spectral changes
(because of the frequency-dependent absorption of the sound on the wall).
While the addition of the reflection can be considered an echo, in the fre-
quency domain, it yields a comb filter, i.e., a spectral ripple with the spectral
density depending on the delay between the direct and reflected sound.
Note that in realistic situations, upon thousands reflections are created
by just a single source. Beginning with the clearly distinguishable early
reflections, their temporal density increases such that after some time they
become a diffuse field, i.e., a sound field with a statistically constant di-
22
rectional distribution.
In addition to specular reflections, diffraction and
diffusion contribute to the complexity of the reverberant sound field.
The sound reflections from walls and other surfaces temporally overlap-
ping with the direct sound, create instantaneous ITD fluctuations. As a
consequence, the interaural similarity in the binaural signal and thus the in-
teraural coherence decreases. (Hartmann et al., 2005). The high-frequency
envelopes seem to be less susceptible to the reduction of the interaural co-
herence caused by reverberation as compared to the low-frequency phase
differences (Ruggles et al., 2012).
Acoustically, the effect of reverberation can be described by the binaural
room impulse response (BRIR), which is basically the HRIR measured in a
room of interest. While a binaural pair of HRIRs are given for the relative
relation between the source and listener's positions, BRIRs further depend
on the absolute positions of the source and listener in the room. While for
HRIRs, the source position change is equivalent to the orientation change
of the listener, in BRIRs, these parameters cannot be exchanged. Hence, a
BRIR is a function of at least the source position, listener position, listener
orientation, room acoustics, and maybe even source orientation.
The position-dependent effect of reverberation implies that it produces
different signals at the two ears. Accordingly, the IACC decreases, for fre-
quencies above 500 Hz, even approaching zero depending on the position
of the source and listener in a room (Hartmann et al., 2005). Sounds from
various directions following the direct sound cause a fluctuation of interau-
ral cues over the time course of the BRIRs, mostly manifesting as a time-
dependent IACC. Hence, IACCs calculated over various ranges of time (and
23
frequencies) are widely used in room acoustics (Mason et al., 2005).
3.4 Dynamic acoustic situations
Even though a sound is an ongoing temporal fluctuation of pressure, its
spatial properties do not change unless the spatial configuration between
the listener, source, and space changes.
In spatial hearing, this situation
is considered as a static one. A widely investigated case is listening to
a static sound source without any head movement. In this situation, the
HRTFs (or BRIRs in a room) do not change. When the source or listener
changes the spatial position and/or orientation, the listening situation is
called "dynamic". As soon as the listener moves (the body or just the head),
the HRTFs (or BRIRs) change, creating a systematic temporal change in
spatial cues.
In order to generally describe spatial changes in listening situations, six
degrees of freedom need to be considered for each object. For example, the
listener's head can be rotated along the horizontal plane (to the left and
right, changing the yaw), along the median plane (up and down, changing
the pitch) and it can be tilted along the frontal plane (to the left and right,
changing the roll). Further, the listener can move along three spatial di-
mensions. This also applies to any sound source with a non-omnidirectional
directivity, like musical instruments, talkers, or loudspeakers.
Generally, head pitch changes the orientation of the pinnae relative to
the source, causing a change in monaural cues, but not necessarily changing
the position of the ears, yielding only a minor change in binaural cues. In
contrast, head yaw changes the ears position in a diametrically opposed
24
fashion, affecting all interaural cues. Hence, horizontal rotations provide
dynamic acoustic cues to distinguish the position of a source in front from
one in back, that is, to acoustically resolve the cone of confusion (Perrett
and Noble, 1995).
The reasons for moving the head due to the sound are manifold. For ex-
ample, the reflexive orienting response, i.e., gaze shift combined with head
movements (and in some animals combined with pinna movements), allows
the listener to orient to the source for further inspection (Sokolov, 2001).
While this reflexive mechanism has been investigated often in the past, not
much is known about intentional listeners' head movements in acoustic en-
vironments (e.g., Leung et al., 2016). Nevertheless, head movements help in
localizing (McAnally and Martin, 2014) and externalizing (Brimijoin et al.,
2013) sounds as well as tracking auditory targets (Leung et al., 2016) -- tasks
requiring a neural basis for the formation of 3-D space.
4 Neurophysiology: coding auditory space
The formation of 3-D space depends not only on the binaural signals but also
on the neural system processing them. Auditory spatial perception is formed
via neural activities ascending from the auditory nerve to the cortex. This
ascending auditory pathway can be anatomically separated into the primary
lemniscal pathway and the non-primary (both lemniscal and non-lemniscal)
pathways (e.g., Straka et al., 2014). Neurons within the primary pathway
process mainly auditory information and project in a tonotopic organization
to the auditory cortex. The non-lemniscal pathways are not fully understood
25
yet (e.g., Lee, 2015) as they link a wide constellation of midbrain, cortical,
and limbic-related sites, integrating different types of sensory information
and providing information about environmental changes even during sleep.
The contribution of both pathways to the formation of the auditory space
is described here following the ascending neural organization.
4.1 Subcortical pathways: reflexive map of auditory space
Auditory processing begins in the cochlea where inner hair cells produce
neural activity (see Fig. 6). The cochlea is not a simple passive sensor; it
is an active sensor whose properties are actively modified by the outer hair
cells innervated by efferent connections. The inner hair cells transmit the
neural information to spiral ganglion cells whose axons form the auditory
nerve (AN, or cochlear nerve). Each of the bilateral ANs projects to the
dorsal cochlear nucleus (DCN) and the ventral cochlear nucleus (VCN) in
the medulla of the brainstem. Both nuclei are specialized in decoding certain
features of the signal (for an overview, see Dehmel et al., 2008, Fig. 1).
VCN networks receive input from AN fibers tuned to similar (mostly low)
frequencies and comprise bushy and octopus cells, which provide the highest
temporal precision of any neuron in the brain. The VCN cells project to the
contralateral (via the trapezoid body, TB) and ipsilateral superior olivary
complexes (SOCs). Given its superior timing properties, the VCN provides
a basis for the decoding of timing-critical acoustic features like the ITD.
The DCN is organized differently. Its cells receive inputs not only from
the high-frequency AN fibers but also from the VCN and various efferent
types of sensory circuits (somatosensory, reticular, vestibular; Muniak and
26
Figure 6: Ascending neural pathways of human auditory processing relevant
in the formation of spatial representation of the environment (see text for
link to the other contralateral neural side (for more
more details).
: link
details on the binaural processing, see Pecka et al., this volume.);
to the visual system;
link to
systems capable of performing actions. Bold lines: the primary auditory
pathway. Dashed and doted lines in the "where" pathway: efferent and
afferent copies, respectively. Bold text in balloons: correspondence to the
cognitive model, see Fig. 1.
link to the somatosensory system;
:
:
:
27
Ryugo, 2014). They are similar to those found in the cerebellum, providing
evidence for their capability of complex information processing.
Indeed,
they form tonotopically organized networks that perform nonlinear spectral
analysis and separate spectral features according to their expectancy (Singla
et al., 2017). This analysis helps in forming acoustic cues required for sound-
source localization in sagittal planes (May, 2000). Further, it may play a role
in attenuating body-generated sounds such as vocalization or respiration,
and thus may improve the signal-to-noise ratio of the received signal (Shore,
2005). Interestingly, DCN cells have also (mostly inhibitory) inputs from the
contralateral CN, indicating an integration of binaural information already
at this very early level of neural processing. The projections of the DCN
cells are manifold. Many projections are direct to the inferior colliculus (IC),
conveying auditory features including the information required for sound
localization based on monaural spectral cues. Other projections are more
spread and include a direct path to the thalamus, allowing the cortex to
prepare for rapid analysis (Anderson et al., 2006), and a path to the nucleus
reticularis pontis caudalis, critical for triggering the acoustic startle reflex
(Meloni and Davis, 1998).
In the SOC, information is mostly processed in three primary nuclei:
the medial superior olive (MSO), the lateral superior olive (LSO), and the
medial and lateral nuclei of the TB. The MSO and LSO have a tonotopic
organization with bilateral inputs from the VCNs. They form the primary
site for the neural computation of ITDs and ILDs (for more details, see
Grothe et al.
in this volume). Combined with the DCN, at this level of
the neural processing, basic auditory spatial features like the ITD, ILD,
28
and monaural spectral features are partially decoded, forming spatial cues
available for processing in further neural structures relevant to auditory
space.
In the midbrain, the IC is an obligatory relay for most of the ascend-
ing auditory information and combines information from other modalities
(Gruters and Groh, 2012). Acoustic features are prepared for the formation
of auditory objects happening at the next synaptic level. The mammal IC
is divided in at least three parts. The central nucleus of the IC is exclusively
auditory. It is organized in sheets of isofrequency laminae, each of which
receives inputs from multiple different nuclei of the brainstem that permit
the decoding of parallel attributes like amplitude and frequency modulation.
Binaural interactions appear to be quite complex and nonlinear. Neverthe-
less, there is strong evidence for the processing of spatial information in all
three dimensions. Single neurons have been found showing ITD sensitivity
similar to that found in psychophysical experiments. Hence, the lateral angle
of a sound can be encoded by means of a single neural path without the need
for a population code (Skottun et al., 2001). Single neurons have also been
found showing a better temporal coding of reverberant stimuli than that of
anechoic stimuli (Slama and Delgutte, 2015). This de-reverberation of the
signals may not only improve robustness in sound recognition but may also
be used to estimate the distance of a sound source by enabling comparisons
between direct and reverberant sound energy. Finally, single neurons have
been found being sensitive to simple analogs of HRTF-like spectral shapes
(Davis et al., 2003), which also show the presence of cells sensitive to the
spectral cues encoding the polar angle of the sound source. The pericentral
29
and external parts of the IC receive ascending inputs from the central IC as
well as descending inputs from the somatosensory system, auditory cortex,
and other higher brain regions. Many projections are bilateral and thus
create numerous feedback loops enabling the integration of the processed
auditory information with that arriving from other sensory systems. From
the IC, the information is transmitted via the brachium of the IC (BIN)
in parallel to the thalamus and superior colliculus (SC). Spatially sensitive
neurons in the BIN, mainly projecting to the SC, were found to prefer the
natural alignment of spatial cues corresponding to a sound-source position
(Slee and Young, 2014).
The SC integrates information from multiple modalities, in particular,
auditory spatial with visual and somatosensory information. Here, maps of
the visual space, body surface, and auditory space arise from spatially or-
dered projections from the retina, skin, and acoustic features, respectively.
Interestingly, ITDs do not seem to contribute much to that spatial map, in-
stead, spatial tuning seems to mostly rely on spectral ILDs (Slee and Young,
2013). The cooperation of neurons combining multimodal afferents and ef-
ferents already at this early neural level decreases reaction time, increases
stimulus detectability, and enhances perceptual reliability, when information
from two modalities is required in a behavioral task (Kayser and Logothetis,
2007). The SC projects to many motor-related parts of the brain and its
organization can be seen as a dynamic map of motor error (Middlebrooks,
2009, pp. 745) with the receptive fields reflecting the deviation between the
angle of gaze (including head and eye position) and the target defined by
the sensory input (visual, auditory, somatic). As a whole, it plays a criti-
30
cal role in the ability to direct behaviors toward specific objects by helping
orient the head and eyes towards something seen and heard (Klier, 2003).
Interestingly, on the one hand, these neural and behavioral abilities seem to
remain active even in the absence of the cortex (Woods, 1964). On the other
hand, dark-reared animals showed a lack of the SC auditory map while still
being able to perform auditory-based behavioral tasks (Blatt et al., 1998).
Hence, while the SC clearly demonstrates the formation of a neural topo-
graphic map of the auditory space, the SC is most likely not responsible
for the creation of the auditory space we are aware of. It is rather a tool
created to fast and mostly reflexively react to the environment in form of
the orienting response (Peck, 1996). From this, one might conclude that the
pathway from SOC via IC to SC forms a reflexive map of the auditory space
(see also Yao et al., 2015), in parallel to the primary ascending path from
the IC to the thalamus and cortex.
4.2 Thalamus and cortex: cognition of auditory space
The cortex plays an important role in the formation of auditory space be-
cause it groups and segregates both spatial and non-spatial perceptual fea-
tures of an auditory scene into streams that each refer to an auditory object
with certain perceptual properties (for review, see Micheyl et al., 2007). Be-
fore reaching the cortex, the thalamus acts as a hub, relaying information
between different subcortical areas and the cerebral cortex. Also, the thala-
mus has been shown to provide an important integration of other modalities
and the preparation of motor responses. Within the thalamus, the audi-
tory information is processed by the medial geniculate body (MGB) of the
31
ventral thalamus. The MGB is further split in dorsal, medial, and ventral
sections.
The dorsal MGB is non-tonotopically organized, its auditory responses
can be influenced strongly by non-auditory inputs, and its neurons project
mainly to the belt of the auditory cortex (Bartlett, 2013). The medial
MGB receives various inputs from the BIN and projects broadly across many
tonotopic, non-tonotopic, multimodal, and limbic cortical areas, terminating
in the cortex and amygdala (for review, see Lee, 2015). There is also a
direct pathway from the DCN to the medial MGB, actually bypassing the
IC (Anderson et al., 2006). This connection shows lower latencies than
those via the IC, which may be advantageous in the role of the medial
MGB in priming the auditory cortex to prepare it for rapid analysis and
recruiting the amygdala for rapid emotional responses such as fear. This
rapid emotional analysis potentially triggers the startle reflex and auditory
looming bias (Bach et al., 2008).
It is not clear yet how the dorsal and
medial parts of the MGB also contribute to the process of forming auditory
objects.
The ventral MGB has been much better investigated; it processes mostly
auditory information, is driven by projections from the central IC, and forms
a tonotopically organized relay of binaural (most) and contralateral-only
(few) information (Lee and Sherman, 2010). The outputs of the ventral
MGB mostly project to the primary auditory cortex and the rostral auditory
area where the acoustic features are further processed to build auditory
objects.
When finally reaching the cortex, the auditory information is spread
32
among various regions. The most involved regions are the superior tempo-
ral gyrus (STG, including the auditory cortex, AC), inferior frontal cortex
(IFC), inferior parietal lobe (IPL), and premotor cortex (PMC). There is
strong evidence for the existence of two largely segregated processing paths,
forming the dual-pathway model of auditory cortical processing, each of
them subserving the two main functions of hearing: "what", the identifi-
cation of auditory objects (e.g., recognition of speech); and "where", the
processing of motion and spatial properties of objects (e.g., sound localiza-
tion; Rauschecker and Tian, 2000).
The "what" pathway follows the antero-lateral route of the AC, which
includes the primary auditory cortex (A1), a rostral area (R), the lateral
and medial belt (including the caudomedial area, CM), and the parabelt.
Moving along the ascending route within the A1, a transition from the rep-
resentation of acoustic features (e.g., response to pure tones) via perceptual
features (e.g., pitch and timbre) to category representation (e.g., auditory
objects) happens. Beginning in the A1, pure-tone sensitive neurons receive
inputs from the ventral MBG and are mostly tonotopically organized. A to-
pographic representation of the auditory space, like that found in the SC, is
mostly missing in the A1. In contrast, responses of some single neurons vary
according to the sound-source direction and result in a 360◦-like panoramic
representation of the space by firing rate (Middlebrooks, 2009), that is, a
neural code for spatial information. The "what" pathway ascends further
to the IFC via other processing areas within the antero-lateral STG. These
and similar connections allow for further processing of non-spatial proper-
ties of auditory objects. From the IFC, this information is transformed into
33
articulatory or motor representations in the PMC.
In contrast, the "where" pathway is highly involved in processing the spa-
tial information retrieved from the auditory stream (for review, see Rauschecker,
2011). While its exact role is still being debated, its processing involves the
separation, location, trajectory, and temporal context of a sound (Ahveninen
et al., 2014). The active regions of the "where" pathway are quite diverse;
they seem to follow a postero-dorsal non-primary route, with projections
from medial and dorsal MGB via the posterior STG (including the planum
temporale, PT, that is, the superior surface of the STG) to the IPL and
dorsal and ventral areas in the PMC (Rauschecker and Tian, 2000). The
spatial information including ITD and binaural coherence seems to be en-
coded by a population code found in various areas of the posterior STG
(Miller and Recanzone, 2009). This information is further projected to the
IPL, that has been found to be strongly involved in the processing of the
spatial auditory information (Arnott et al., 2004). From the IPL, the spatial
features are integrated with information from other sensory modalities and
further projected to the PMC.
Hence, the PMC is activated by the "what" pathway (via IFC, see bold
lines in Fig. 6) and also modulated by the "where" pathway (via IPL). The
modulation by the IPL corresponds to a feed-forward system consisting of
an internal predictive model of the environment located in the PMC and
updated by the multimodal sensory information ascending from the IPL
(see Fig. 6). This allows a quick update of the motor system to the new
sensory situation.
The PMC reacting to predictions based on sensory information is further
34
underlined by findings showing that the PMC is not only involved during
acoustic stimulation but also during musical imagery (Leaver et al., 2009)
where it is responsible for assembling the motor patterns for the poten-
tial production of sound sequences. Efferent feedback from the PMC about
planned motor actions (dashed line in Fig. 6) together with the fast and tem-
porally precise afferent projections from the posterior STG (i.e., an efferent
copy, dotted line in Fig. 6), allow the IPL to compare the spatial auditory
information with the predictive motor states, to decide about the required
adjustments of the internal model, and to minimize the surprise (compare
Sec. 2.1). Further efferent projections to the STG (dashed line in Fig. 6),
on the other hand, allow for modulating the process of feature extraction in
the STG according to the changing feature demands (for more details, see
Rauschecker, 2011).
These considerations show that many cortical regions are involved in
processing features of the auditory space and there is no clear evidence for a
single region representing the auditory space in the cortex per se. The spatial
information is transmitted via the firing rate of the neural population. This
process is further modulated by vision (see Mendonca et al., in this volume),
proprioception (Genzel et al., 2016), and in particular attention (e.g., Fels
et al., in this volume), indicating its reflective nature in the formation of the
auditory space. This is in contrast to the well-localized but reflexive map of
the auditory space found at the level of the SC.
The summary presented in this section is just a simplification of all the
reflexive and reflective processes involved. The human brain is an ana-
log, high-dimensional, recurrent, nonlinear, stochastic, and dynamic system
35
(Dotov, 2014). At the end of the day, these processes form the perception
that allows humans to complete various spatial tasks. In the following sec-
tion, we describe psychophysical spatial tasks, all of them demonstrating
the ability to utilize our understanding of the 3-D auditory world.
5 Psychophysics: listener's abilities given the per-
ceived auditory space
Spatial auditory cues facilitate both reflexive and reflective behavior. Hu-
man psychoacoustic studies usually imply reflective behavioral tasks, while
reflexive behavior, if not targeted explicitly, is more commonly tested in
populations with limited cognitive abilities such as infants. In this section,
we thus review reflective spatial abilities in humans, organized by increasing
cognitive complexity.
5.1 Sound localization
Sound localization describes the (reflective) ability to estimate the spatial
position of the sound source (for review, see Middlebrooks, 2015). Sound
localization experiments are often conducted in anechoic environments or
free-field simulated virtual space and present target sounds either via loud-
speakers or via headphones in virtual auditory spaces simulated by filtering
with HRTFs. Considering normal-hearing conditions, the major cues are
conceptually different for each dimension of the interaural-polar coordinate
system. Alterations of the acoustic environment may change the informa-
tive character of a certain cue and consequently affect its perceptual weight
36
(Keating et al., 2015). Neural plasticity not only enables such context-
dependent weighting between cues but also enables adapting and extending
the set of spatial cues (for review, see Mendon¸ca, 2014).
Interaural cues facilitate sound localization within the lateral dimension.
The duplex theory describes ITD cues being most effective for low-frequency
sounds and ILD cues for high-frequency sounds. Localization of broad-
band sounds in the lateral direction is dominated by low-frequency ITDs
(Macpherson and Middlebrooks, 2002). Besides an intermediate frequency
range between 2 and 4 kHz, where robust localization cues are lacking, hu-
man listeners are accurate in lateral-angle localization. Minimum audible
angles (MAAs) between two successively presented sounds may be as small
as 1◦ for frontal sources, but degrade with increasing laterality to approx-
imately 10◦ (Perrott and Saberi, 1990). These thresholds were derived in
free-field experiments providing natural combinations of ITD and ILD cues.
Tested in isolation, discrimination thresholds of interaural phase differences
(leading to ITDs) also increase by an order of magnitude with increasing
reference difference (e.g., 2◦ at 0◦ reference increasing to 20◦ at 90◦ refer-
ence for a 900 Hz tone, Yost, 1974), which closely resembles the observed
MAAs. ITD cues result either from the temporal fine structure at frequen-
cies below approximately 1.4 kHz or from temporal envelopes at higher fre-
quencies. Consistent with the small perceptual weight in the duplex theory,
just-noticeable differences (JNDs) for such high-frequency envelope ITDs are
at least twice as large as the JNDs at low frequencies (Bernstein and Trahio-
tis, 2002). ILD JNDs are in the range of 1 dB for high-frequency sounds and
depend on amplitude modulation rates. Effects of temporal modulations on
37
ILD JNDs can be explained based on the interaural difference in neural dis-
charge rates with no need for a particular binaural adaptation mechanisms
(Laback et al., 2017). Computational models of lateral-angle localization
have a long history and diversity. Recently, a large-scale attempt to sys-
tematically investigate these approaches has been initiated (Dietz et al.,
2017). The auditory system is at least partly able to adapt to changes in
ITD and ILD cues according to visual and audiomotor feedback (e.g., Tra-
peau and Schonwiesner, 2015). Adaptation aftereffects were only observed
in short-term but not long-term studies (Trapeau and Schonwiesner, 2015),
indicating that different mechanisms are involved.
While spectral-shape cues may be important for lateral-angle localiza-
tion in monaural but not binaural hearing (Macpherson and Middlebrooks,
2002), they are crucial for sound localization within the sagittal dimension.
Consequently, polar-angle localization requires some bandwidth, especially
at high frequencies, in order to achieve high spatial acuity, as indicated by
MAAs as small as 4◦ (Perrott and Saberi, 1990). Due to redundant spectral
information, localization performance is maintained with small uninforma-
tive parts in the spectrum. While spectral information limited to frequencies
below 16 kHz seems to be sufficient, limitations down to 8 kHz cause marked
degradations (Best et al., 2005). Spectral degradations often result in local-
ization responses biased toward the horizon and led to the concept of "eleva-
tion gain" in polar-angle localization (Hofman et al., 1998). Spectral-shape
cues are arguably processed within monaural pathways and thus are often
referred to as "monaural spectral cues" although information from both
ears is combined following a spatially systematic binaural weighting scheme
38
(Macpherson and Sabin, 2007): The contralateral ear contributes less with
increasing lateral eccentricity where the head shadow will naturally degrade
the energy ratio between the target signal and diffuse background noise.
Due to the monaural processing, localization performance can be affected
by frequency modulations in the stimulus spectrum. Template-based models
are able to explain these interactions and show how monaural spectral cues
extracted on the basis of tonotopic gradients are rather independent of nat-
urally common low-frequency modulations in the source spectrum (Baum-
gartner et al., 2014). Listeners are able to learn new spectral-shape cues
(Majdak et al., 2013b; Hofman et al., 1998) and use them simultaneously
with previously acquired cues (Trapeau et al., 2016). Localization perfor-
mance along sagittal planes is particularly listener-specific, but this variation
is hardly explained by only considering the acoustic factor of listener-specifc
HRTFs, suggesting large inter-individual differences in how efficient the au-
ditory system is able to utilize the acoustic information (Majdak et al., 2014;
Baumgartner et al., 2016).
To estimate the distance of a source, listeners have access to a broad vari-
ety of acoustic cues like sound intensity, reverberation characteristics (often
quantified by the direct-to-reverberant energy ratio), near-field ILDs, the
shape of the stimulus spectrum, and others (for review, see Kolarik et al.,
2016). The relative perceptual relevance of these cues and their underlying
neural codes are the subject of debate and are most probably dependent on
the context (for a review, see Hl´adek et al., 2017). Recent studies suggest
that the amount of temporal ILD fluctuations and amplitude modulation
depth likely represent reverberation-related cues (Catic et al., 2015). More-
39
over, spectral-shape cues can affect distance perception (Baumgartner et al.,
2017) and familiarization to non-individualized spectral-shape cues can im-
prove distance perception (Mendon¸ca et al., 2013).
A special case of distance perception concerns distances closer than phys-
ically plausible, namely, inside the listener's head. Sounds are naturally
perceived outside the head (externalized) whereas sounds reproduced with
headphones or hearing-assistive devices are often perceived to originate from
inside the head (internalized, Noble and Gatehouse, 2006). Sound external-
ization is not directly related to the playback device, as free-field signals can
be internalized as well (Brimijoin et al., 2013). Most psychoacoustic studies
investigated sound externalization either with discrimination tasks between
real and virtual sources and/or distance ratings (Hartmann and Wittenberg,
1996). Although perceived distance may be only one of many cues to dis-
criminate between virtual and real sources, the similarity of findings for both
paradigms suggests that distance perception is a crucial component. One
could think that the only reason for using the term "sound externalization"
instead of "distance" is that the percept of sound internalization is an avail-
able option. At first glance, however, there seem to be some contradictions
between studies focusing on sound externalization and distance perception.
For example, decreasing low-frequency ILDs were associated with increas-
ing distance (Brungart et al., 1999) whereas ILDs gradually removed from
the low-frequency partials of a harmonic complex (Hartmann and Witten-
berg, 1996) or decreased from broadband speech (Brimijoin et al., 2013)
were associated with reduced sound externalization. It seems as if sound
internalization is a default state similar to the concept of elevation gain in
40
the case of missing or implausible cues available to the auditory system be-
cause no plausible model of the environment can be established to create an
external state of affairs.
Head rotation causes dynamic cues particularly effective in resolving
front-back confusions (McAnally and Martin, 2014). Self-motion affects the
actual sound location, but this interaction is successfully compensated by
mechanisms responsible for building an allocentric frame of reference (Yost
et al., 2015). Listeners are able to create such allocentric spatial representa-
tions also without visual information (Viaud-Delmon and Warusfel, 2014).
In contrary to self-motion, listeners are sensitive to source motion and are
able to detect movement angles as small as 2◦ depending on source velocity,
stimulus duration, and bandwidth (for review, see Carlile and Leung, 2016).
5.2 From spatial impression to presence
Auditory perception in reverberant spaces like rooms or concert halls is
multidimensional (Cerd´a et al., 2009). Listeners are still able to localize
the direct sound despite the presence of early reflections, which technically
might be interpreted as separate auditory objects.
In a room, it is not
only that humans are able to integrate the divergent spatial cues into a
single auditory object, but also to a certain level, they are even not able to
perceive the reflected copies of the direct sound. This ability to suppress
early reflections and actually to not perceive them as echoes is referred to
as the precedence effect (for review, see Clapp and Seeber, this volume).
The delay of the reflections relevant for the precedence effect depends on
stimulus; it is around 5 ms for short clicks, and can be up to 30 ms for
41
complex stimuli like speech.
Although we simply perceive a single auditory object even in the pres-
ence of its acoustic reflections, the presence of reverberation introduces other
spatial effects, which have been summarized as spatial impression or spa-
ciousness (Kuhl, 1978). Two main components have been found: apparent
source width (ASW) and listener envelopment (LEV, Bradley and Soulodre,
1995).
ASW describes the spatial compactness of a sound event perceived by
the listener. In headphone experiments, the main cue for the compactness
of the perceived sound is the IACC: If it decreases, the sound is perceived
as a wider image (Blauert and Lindemann, 1986). Interestingly, for narrow-
band sounds, listeners are extremely sensitive to the deviation of a perfectly
coherent signal; they can easily discriminate between signals with an IACC
of 1 and 0.99 (Gabriel and Colburn, 1981). The current explanation for
such a high sensitivity is that even in a slightly incoherent signal, large in-
stantaneous ITD and ILD fluctuations occur, which can easily be detected
by the auditory system. The perceptual consequence of the fluctuations de-
pends on their duration, bandwidth, center frequency, and stimulus sound
level (Goupell and Hartmann, 2006). ASW is, however, not much affected
between IACCs of 1 and 0.99 but instead gradually declines with the IACC
(Whitmer et al., 2013). For the extreme case of interaural decorrelation
(IACC of zero), the perceived auditory image splits into two objects ap-
pearing in the left and the right ears, respectively.
In reverberant environments where multiple reflections overlap the di-
rect sound, the ASW has been found to be determined by the lateral energy
42
fraction and IACC of the early arriving sound field, that is, within the first
80 ms of the BRIR (Okano et al., 1998). Deviations of that IACC from one
contributes to the perceived quality of concert halls and has been termed the
binaural quality index (BQI) of room acoustics. Interestingly, as the BQI
increases, the low-frequency ITDs are more likely to be disturbed. Thus, spa-
tial hearing in reverberant situations seems to rely more on high-frequency
ITD cues (transmitted in the signal envelope) than on low-frequency ITD
cues (Ruggles et al., 2012).
LEV describes how immersed in the sound field the listener feels.
It
depends on the level, direction of arrival, and temporal properties of later
(after 80 ms) arriving reflections (Bradley and Soulodre, 1995). Other stud-
ies show that late sound arriving from the side, overhead, and behind the
listener correlates strongly with LEV, suggesting that the LEV can be distin-
guished from the late sound having non-lateral components (Furuya et al.,
2001). The late sound arriving from behind and above the listener seem to
be important as well (Morimoto et al., 2001), showing the relevance of an
accurate formation of 3-D auditory space for the perception of rooms.
The ASW and LEV can be predicted with a model based on the BRIRs
and just noticable differences of the IACC (Klockgether and van de Par,
2014). The overall perceived quality of concert halls can be quantitatively
explained by additional consideration of the reverberation time (Cerd´a et al.,
2015).
While both, the ASW and LEV seem to be the major parameters de-
scribing the spatial impression of a room, they can both be seen as parts of
a larger concept widely used in the context of virtual environments (VE).
43
Here, immersion, as a measure of the psychological sensation of being sur-
rounded (Begault et al., 1998), integrates the objectively derived LEV and
ASW, and further extends to "a psychological state characterized by per-
ceiving oneself to be enveloped by, included in, and interacting with an
environment that provides a continuous stream of stimuli and experiences"
(Witmer and Singer, 1998). In that context, responses to a given level of
immersion have been defined as presence, a measure of the psychological
sensation of being elsewhere (e.g., Slater, 2003). Both immersion and pres-
ence are important for the quality of experience in VE systems (Moller and
Raake, 2014). For example, in headphone-based VE systems, immersion can
be enhanced with the use of listener-specific HRTFs (Vorlander and Shinn-
Cunningham, 2014), being in line with studies showing that our auditory
system prefers the natural combination of ITDs and ILDs (Salminen et al.,
2015). Here, immersion seems to be more easily conveyed via audio than
with vision because audio operates all around the listener even outside the
listener's field of view and without exploratory head movements. Immer-
sion and presence seem to be very related attributes and the underlying
mechanisms are not fully understood yet (for review, see Gaggioli et al.,
2003).
5.3 Other spatially related tasks
Spatial hearing also improves tasks not directly related to the formation
of the 3-D space. A famous example is the cocktail-party effect which de-
scribes the ability to focus on and thus improve the intelligibility of a par-
ticular talker in a multi-talker environment (Bronkhorst, 2015). In such a
44
task, a perception of the spatial world is not required per se, however, the
benefit of spatial separation of maskers from the target, also called spatial
unmasking or release from masking, is clear and has been considered in mod-
els predicting speech intelligibility from binaural signals in many situations
(e.g., Lavandier et al., 2012). Spatial unmasking can further reduce cogni-
tive load in conditions providing similar speech intelligibility (e.g., And´eol
et al., 2017). Spatial attention, that is, knowing "where" to focus, further
modulates the effect of spatial unmasking on a very listener-specific basis
(Oberfeld and Klockner-Nowotny, 2016).
Note that spatial unmasking is not only limited to spatially separated
targets and/or noise. Improved speech intelligibility has also been shown
in listeners once they have adapted to the acoustics of the listening room
(Brandewie and Zahorik, 2010) indicating that while our auditory system is
able to adapt to reverberant spaces, the "noise source" in spatial unmasking
can be both additional sound sources and acoustic reflections of the same
source.
Spatial hearing contributes to other, less-known non-spatial tasks. For
example, spatial impression can increase the emotional impact of orchestra
music by enhancing musical dynamics (Patynen and Lokki, 2016). Looming
bias, that is, the phenomenon that approaching sounds are more salient
than receding sounds, can be significantly mediated by sound externalization
created by the acoustic spatial pinna features alone (Baumgartner et al.,
2017). These and similar findings underline the relevance of the formation
of 3-D auditory space in our everyday life.
45
6 Conclusions
The formation of the auditory space is one of the cognitive processes required
to understand and interact with our environment -- by the sensation of sound.
In that process, the auditory system has to cope with ephemeral acoustic
information about the auditory objects around us. The spatial information,
conveyed by the binaural signals, is encoded by interaural and monaural
features, along various temporal ranges. Our neural auditory system then
creates two representations of the auditory space: a topographically struc-
tured neural network in the superior colliculus, capable of triggering quick
reflexive reactions; and a reflective cortical representation, encoded by neu-
ral populations, capable of modulating other cognitive processes by means
of attention. The reflective representation allows us to consciously perceive
the auditory space and perform spatial tasks.
Many concepts have been proposed for cognitive processes involved in
the formation of the auditory space. Our interaction with the environment
can be seen as a feedforward system with the internal model anticipating
the external (or distal) state of affairs. Feedback coming from hearing and
other senses allows us to compensate for any deviations to the predictions.
Given the ambiguity in the estimation of the external state of affairs from
the limited binaural information, the free-energy principle, also known as
the active inference, seems to be a promising approach to explain how cog-
nition restricts itself to a limited number of states. Further progress in the
development of mathematical methods for solving ill-posed problems and of
experimental methods combining psychophysics with neurophysiology will
46
help to improve our understanding of the formation of the auditory space
in the future. This will lead to advances in many technical applications like
hearing aids driven by spatial attention, listener-specific virtual acoustics,
and dynamic sound reproduction systems.
Acknowledgement
We would like to thank S. Clapp and B. Seeber for their valuable comments
and suggestions. Supported by the Austrian Science Fund (FWF, J 3803-
N30).
References
Ahveninen, J., Kopco, N., and Jaaskelainen, I. P. (2014). "Psychophysics
and neuronal bases of sound localization in humans," Hearing Research
307, 86 -- 97.
Aitchison, L., and Lengyel, M. (2017). "With or without you: predictive
coding and Bayesian inference in the brain," Curr Opin Neurobiol 46,
219 -- 227.
Algazi, V. R., Avendano, C., and Duda, R. O. (2001). "Elevation local-
ization and head-related transfer function analysis at low frequencies," J
Acoust Soc Am 109(3), 1110 -- 22.
And´eol, G., Suied, C., Scannella, S., and Dehais, F. (2017). "The Spatial
Release of Cognitive Load in Cocktail Party Is Determined by the Relative
Levels of the Talkers," J Assoc Res Otolaryngol 1 -- 8.
47
Anderson, L. A., Malmierca, M. S., Wallace, M. N., and Palmer, A. R.
(2006). "Evidence for a direct, short latency projection from the dor-
sal cochlear nucleus to the auditory thalamus in the guinea pig," Eur J
Neurosci 24(2), 491 -- 498.
Andreopoulou, A., and Katz, B. F. G. (2017). "Identification of perceptually
relevant methods of inter-aural time difference estimation," J Acoust Soc
Am 142(2), 588 -- 598.
Arnal, L., Flinker, A., Kleinschmidt, A., Giraud, A.-L., and Poeppel, D.
(2015). "Human Screams Occupy a Privileged Niche in the Communica-
tion Soundscape," Current Biology 25(15), 2051 -- 2056.
Arnott, S. R., Binns, M. A., Grady, C. L., and Alain, C. (2004). "Assessing
the auditory dual-pathway model in humans," NeuroImage 22(1), 401 --
408.
Awh, E., Belopolsky, A. V., and Theeuwes, J. (2012). "Top-down versus
bottom-up attentional control: a failed theoretical dichotomy," Trends
Cogn Sci 16(8), 437 -- 443.
Bach, D. R., Schachinger, H., Neuhoff, J. G., Esposito, F., Di Salle, F.,
Lehmann, C., Herdener, M., Scheffler, K., and Seifritz, E. (2008). "Rising
sound intensity: an intrinsic warning cue activating the amygdala," Cereb
Cortex 18(1), 145 -- 150.
Bartlett, E. L. (2013). "The organization and physiology of the auditory
thalamus and its role in processing acoustic features important for speech
perception," Brain and language 126(1), 29 -- 48.
48
Baumgartner, R., Majdak, P., and Laback, B. (2014). "Modeling sound-
source localization in sagittal planes for human listeners," J Acoust Soc
Am 136(2), 791 -- 802.
Baumgartner, R., Majdak, P., and Laback, B. (2016). "Modeling the Ef-
fects of Sensorineural Hearing Loss on Sound Localization in the Median
Plane," Trends Hear 20, 2331216516662003.
Baumgartner, R., Reed, D. K., T´oth, B., Best, V., Majdak, P., Colburn,
H. S., and Shinn-Cunningham, B. (2017). "Asymmetries in behavioral
and neural responses to spectral cues demonstrate the generality of audi-
tory looming bias," Proc Natl Acad Sci 114(36), 9743 -- 9748.
Begault, D. R., Ellis, S. R., and Wenzel, E. M. (1998). "Headphone and
Head-Mounted Visual Displays for Virtual Environments," J Audio Eng
Soc, Vol. 49, pp. 904 -- 916.
Bernstein, L. R., and Trahiotis, C. (2002). "Enhancing sensitivity to inter-
aural delays at high frequencies by using "transposed stimuli"," J Acoust
Soc Am 112(3), 1026 -- 1036.
Best, V., Carlile, S., Jin, C., and van Schaik, A. (2005). "The role of high
frequencies in speech localization," J Acoust Soc Am 118(1), 353 -- 63.
Best, V., Gallun, F. J., Carlile, S., and Shinn-Cunningham, B. G. (2007).
"Binaural interference and auditory grouping," J Acoust Soc Am 121(2),
1070 -- 1076.
Bhatt, R. S., and Quinn, P. C. (2011). "How does Learning Impact De-
49
velopment in Infancy? The Case of Perceptual Organization," Infancy
16(1), 2 -- 38.
Bizley, J. K., and Cohen, Y. E. (2013). "The what, where and how of
auditory-object perception," Nat Rev Neurosci 14(10), 693 -- 707.
Blatt, B., von Linstow Roloff, E., Withington, D. J., Macphail, E. M., and
Riedel, G. (1998). "Analysis of the superior colliculus auditory space map
function in guinea pig behavior," Neurosci Res Commun 23(1), 23 -- 40.
Blauert, J. (1997). Spatial hearing. The Psychophysics of Human Sound
Localization, revised ed. (The MIT Press, Cambridge, MA).
Blauert, J., and Lindemann, W. (1986). "Spatial mapping of intracranial
auditory events for various degrees of interaural coherence," J Acoust Soc
Am 79(3), 806 -- 813.
Bradley, J. S., and Soulodre, G. A. (1995). "Objective measures of listener
envelopment," J Acoust Soc Am 98(5), 2590 -- 2597.
Brandewie, E., and Zahorik, P. (2010). "Prior listening in rooms improves
speech intelligibility," J Acoust Soc Am 128(1), 291 -- 299.
Bregman, A. S. (1990). Auditory scene analysis, 10 (MIT Press, Cambridge,
MA).
Brimijoin, W. O., Boyd, A. W., and Akeroyd, M. A. (2013). "The Contri-
bution of Head Movement to the Externalization and Internalization of
Sounds," PloS One 8(12), e83068.
50
Bronkhorst, A. W. (2015). "The cocktail-party problem revisited: early pro-
cessing and selection of multi-talker speech," Atten Percept Psychophys
77(5), 1465 -- 1487.
Brungart, D. S., Durlach, N. I., and Rabinowitz, W. M. (1999). "Auditory
localization of nearby sources. II. Localization of a broadband source," J
Acoust Soc Am 106(4), 1956 -- 1968.
Brungart, D. S., and Rabinowitz, W. M. (1999). "Auditory localization
of nearby sources. Head-related transfer functions," J Acoust Soc Am
106(3), 1465 -- 1479.
Carbon, C.-C. (2014). "Understanding human perception by human-made
illusions," Front Hum Neurosci 8, 566.
Carlile, S., and Corkhill, C. (2015). "Selective spatial attention modulates
bottom-up informational masking of speech," Sci Rep 5(1), 8662.
Carlile, S., and Leung, J. (2016). "The Perception of Auditory Motion,"
Trends Hear 20, 1 -- 19.
Catic, J., Santurette, S., and Dau, T. (2015). "The role of reverberation-
related binaural cues in the externalization of speech," J Acoust Soc Am
138(2), 1154 -- 1167.
Cerd´a, S., Gim´enez, A., Cibri´an, R., Gir´on, S., and Zamarreno, T. (2015).
"Subjective ranking of concert halls substantiated through orthogonal ob-
jective parameters," J Acoust Soc Am 137(2), 580 -- 584.
51
Cerd´a, S., Gim´enez, A., Romero, J., Cibri´an, R., and Miralles, J. (2009).
"Room acoustical parameters: A factor analysis approach," Appl Acoust
70(1), 97 -- 109.
Chandrasekaran, B., Koslov, S. R., and Maddox, W. T. (2014). "Toward a
dual-learning systems model of speech category learning," Front Psychol
5, 825.
Curtis, C. E., and D'Esposito, M. (2003). "Success and Failure Suppressing
Reflexive Behavior," J Cogn Neurosci 15(3), 409 -- 418.
Davis, K. A., Ramachandran, R., and May, B. J. (2003). "Auditory Pro-
cessing of Spectral Cues for Sound Localization in the Inferior Colliculus,"
J Assoc Res Otolaryngol 4(2), 148 -- 163.
Dehmel, S., Cui, Y. L., and Shore, S. E. (2008). "Cross-Modal Interactions
of Auditory and Somatic Inputs in the Brainstem and Midbrain and Their
Imbalance in Tinnitus and Deafness," Am J Audiol 17(2), S193.
Dietz, M., Lestang, J.-H., Majdak, P., Stern, R. M., Marquardt, T., Ewert,
S. D., Hartmann, W. M., and Goodman, D. F. M. (2017). "A framework
for testing and comparing binaural models," Hear Res 360, 92 -- 106.
Dokmanic, I., Parhizkar, R., Walther, A., M Lu, Y., and Vetterli, M. (2013).
"Acoustic Echoes Reveal Room Shape," Proc Natl Acad Sci 110(30),
12186 -- 12191.
Donoho, D. L. (2006). "For most large underdetermined systems of lin-
52
ear equations the minimal 1-norm solution is also the sparsest solution,"
Commun Pure Appl Math 59(6), 797 -- 829.
Dotov, D. G. (2014). "Putting reins on the brain. How the body and envi-
ronment use it," Front Hum Neurosci 8(795), 1 -- 12.
Epstein, W., and Rogers, S. J., eds. (1995). Handbook of perception and
cognition Perception of space and motion (Academic Press, San Diego).
Francis, B., and Wonham, W. (1976). "The internal model principle of
control theory," Automatica 12(5), 457 -- 465.
Friston, K. (2012). "Embodied inference and spatial cognition," Cognitive
Processing 13(S1), 171 -- 177.
Friston, K., FitzGerald, T., Rigoli, F., Schwartenbeck, P., O'Doherty, J., and
Pezzulo, G. (2016). "Active inference and learning," Neurosci Biobehav
Rev 68, 862 -- 879.
Friston, K., Kilner, J., and Harrison, L. (2006). "A free energy principle for
the brain," J Physiol Paris 100(1-3), 70 -- 87.
Furuya, H., Fujimoto, K., Young Ji, C., and Higa, N. (2001). "Arrival
direction of late sound and listener envelopment," Appl Acoust 62(2),
125 -- 136.
Gabriel, K. J., and Colburn, H. S. (1981). "Interaural correlation discrim-
ination:
i. bandwidth and level dependence," J Acoust Soc Am 69(5),
1394 -- 1401.
53
Gaggioli, A., Bassi, M., and Delle Fave, A. (2003). "Quality of Experience in
Virtual Environments," in Being there: Concepts, effects and measuere-
ment of user presence in syntetic environments, pp. 122 -- 132.
Genzel, D., Firzlaff, U., Wiegrebe, L., and MacNeilage, P. R. (2016). "De-
pendence of auditory spatial updating on vestibular, proprioceptive, and
efference copy signals," J Neurophysiol 116(2), 765 -- 775.
Gordon, C., Webb, D. L., and Wolpert, S. (1992). "One cannot hear the
shape of a drum," Bull Am Math Soc 27(1), 134 -- 138.
Goupell, M. J., and Hartmann, W. M. (2006). "Interaural fluctuations and
the detection of interaural incoherence: Bandwidth effects," J Acoust Soc
Am 119(6), 3971 -- 3986.
Gregory, R. L. (1980). "Perceptions as hypotheses," Philos Trans R Soc
Lond B Biol Sci 290(1038), 181 -- 197.
Griffiths, T. D., and Warren, J. D. (2004). "What is an auditory object?,"
Nat Rev Neurosci 5(11), 887 -- 892.
Grush, R. (2004). "The emulation theory of representation: motor control,
imagery, and perception," Behav Brain Sci 27(3), 377 -- 396.
Gruters, K. G., and Groh, J. M. (2012). "Sounds and beyond: multisensory
and other non-auditory signals in the inferior colliculus," Front Neural
Circuits 6.
Hartmann, W. M., Rakerd, B., and Koller, A. (2005). "Binaural Coherence
in Rooms," Acta Acust united Ac 91(3), 451 -- 462.
54
Hartmann, W. M., and Wittenberg, A. (1996). "On the externalization of
sound images," J Acoust Soc Am 99(6), 3678 -- 3688.
Helmholtz, H. (1954). On the Sensations of Tone (Dover Books on Music,
NY).
Henning, G. B. (1974). "Detectability of interaural delay in high-frequency
complex waveforms," J Acoust Soc Am 55(1), 84 -- 90.
Hill, K. T., and Miller, L. M. (2010). "Auditory Attentional Control and
Selection during Cocktail Party Listening," Cereb Cortex 20(3), 583 -- 590.
Hinton, G. E., and Ghahramani, Z. (1997). "Generative models for discov-
ering sparse distributed representations," Philos Trans R Soc Lond B Biol
Sci 352(1358), 1177 -- 1190.
Hl´adek, f., Tomoriov´a, B., and Kopco, N. (2017). "Temporal characteristics
of contextual effects in sound localization," J Acoust Soc Am 142(5),
3288 -- 3296.
Hofman, P. M., van Riswick, J. G. A., and van Opstal, A. J. (1998). "Re-
learning sound localization with new ears," Nature Neurosci 1(5), 417 --
421.
Kayser, C., and Logothetis, N. K. (2007). "Do early sensory cortices inte-
grate cross-modal information?," Brain Struct Funct 212(2), 121 -- 132.
Keating, P., Dahmen, J. C., and King, A. J. (2015). "Complementary adap-
tive processes contribute to the developmental plasticity of spatial hear-
ing," Nat Neurosci 18(2), 185 -- 187.
55
Klier, E. M. (2003). "Three-Dimensional Eye-Head Coordination Is Imple-
mented Downstream From the Superior Colliculus," J Neurophysiol 89(5),
2839 -- 2853.
Klockgether, S., and van de Par, S. (2014). "A Model for the Prediction of
Room Acoustical Perception Based on the Just Noticeable Differences of
Spatial Perception," Acta Acust united Ac 100, 964 -- 971.
Knudsen, E. I. (2007). "Fundamental Components of Attention," Annu Rev
Neurosci 30(1), 57 -- 78.
Koffka, K. (1935). Principles of Gestalt psychology (Mimesis Int., London).
Kolarik, A. J., Moore, B. C. J., Zahorik, P., Cirstea, S., and Pardhan,
S. (2016). "Auditory distance perception in humans: a review of cues,
development, neuronal bases, and effects of sensory loss," Atten Percept
Psychophys 78(2), 373 -- 395.
Kondo, H. M., Pressnitzer, D., Toshima, I., and Kashino, M. (2012). "Ef-
fects of self-motion on auditory scene analysis," Proc Natl Acad Sci
109(17), 6775 -- 6780.
Kreuzer, W., Majdak, P., and Chen, Z. (2009). "Fast multipole boundary
element method to calculate head-related transfer functions for a wide
frequency range," J Acoust Soc Am 126(3), 1280 -- 1290.
Kuhl, W. (1978). "Spaciousness (spatial impression) as a component of total
room impression," Acustica 40(3), 167 -- 181.
56
Kuhn, G. F. (1977). "Model for the interaural time differences in the az-
imuthal plane," J Acoust Soc Am 62(1), 157 -- 167.
Kulkarni, A., Isabelle, S. K., and Colburn, H. S. (1999). "Sensitvity of
human subjects to head-related transfer-function phase spectra," J Acoust
Soc Am 105 (5), 2821 -- 2840.
Laback, B., Dietz, M., and Joris, P. (2017). "Temporal effects in interaural
and sequential level difference perception," J Acoust Soc Am 142(5),
3267 -- 3283.
Lavandier, M., Jelfs, S., Culling, J. F., Watkins, A. J., Raimond, A. P.,
and Makin, S. J. (2012). "Binaural prediction of speech intelligibility in
reverberant rooms with multiple noise sources," J Acoust Soc Am 131(1),
218 -- 231.
Leaver, A. M., Van Lare, J., Zielinski, B., Halpern, A. R., and Rauschecker,
J. P. (2009). "Brain Activation during Anticipation of Sound Sequences,"
J Neurosci 29(8), 2477 -- 2485.
Lee, C. C. (2015). "Exploring functions for the non-lemniscal auditory tha-
lamus," Front Neural Circuits 9.
Lee, C. C., and Sherman, S. M. (2010). "Topography and physiology of
ascending streams in the auditory tectothalamic pathway," Proc Natl
Acad Sci 107(1), 372 -- 377.
Leung, J., Wei, V., Burgess, M., and Carlile, S. (2016). "Head Tracking of
Auditory, Visual, and Audio-Visual Targets," Front Neurosci 9.
57
Lord Rayleigh or Strutt, F. (1876). "Our perception of the direction of a
source of sound," Proc Musical Association 2 75 -- 84.
Macpherson, E. A., and Middlebrooks, J. C. (2002). "Listener weighting of
cues for lateral angle: The duplex theory of sound localization revisited,"
J Acoust Soc Ama 111(5), 2219 -- 2236.
Macpherson, E. A., and Middlebrooks, J. C. (2003). "Vertical-plane sound
localization probed with ripple-spectrum noise," J Acoust Soc Am 114(1),
430 -- 445.
Macpherson, E. A., and Sabin, A. T. (2007). "Binaural weighting of monau-
ral spectral cues for sound localization," J Acoust Soc Am 121(6), 3677 --
3688.
Majdak, P., Balazs, P., and Laback, B. (2007). "Multiple exponential sweep
method for fast measurement of head-related transfer functions," J Audio
Eng Soc 55, 623 -- 637.
Majdak, P., Baumgartner, R., and Laback, B. (2014). "Acoustic and non-
acoustic factors in modeling listener-specific performance of sagittal-plane
sound localization," Front Psychol 5(319), 1 -- 10.
Majdak, P., Carpentier, T., Nicol, R., Roginska, A., Suzuki, Y., Watanabe,
K., Wierstorf, H., Ziegelwanger, H., and Noisternig, M. (2013a). "Spa-
tially Oriented Format for Acoustics: A Data Exchange Format Repre-
senting Head-Related Transfer Functions," in Proc 134th Conv Audio Eng
Soc, Roma, Italy, p. 8880.
58
Majdak, P., Goupell, M. J., and Laback, B. (2010). "3-D localization of
virtual sound sources: effects of visual environment, pointing method,
and training," Atten Percept Psychophys 72(2), 454 -- 69.
Majdak, P., Walder, T., and Laback, B. (2013b). "Effect of long-term
training on sound localization performance with spectrally warped and
band-limited head-related transfer functions," J Acoust Soc Am 134(3),
2148 -- 2159.
Mason, R., Brookes, T., and Rumsey, F. (2005). "Frequency dependency
of the relationship between perceived auditory source width and the in-
teraural cross-correlation coefficient for time-invariant stimuli," J Acoust
Soc Am 117(3), 1337 -- 1350.
May, B. J. (2000). "Role of the dorsal cochlear nucleus in the sound local-
ization behavior of cats," Hear Res 148(1-2), 74 -- 87.
McAnally, K. I., and Martin, R. L. (2014). "Sound localization with head
movement:
implications for 3-d audio displays," Front Neurosci 8(210),
1 -- 6.
Meloni, E. G., and Davis, M. (1998). "The dorsal cochlear nucleus con-
tributes to a high intensity component of the acoustic startle reflex in
rats," Hear Res 119(1-2), 69 -- 80.
Mendon¸ca, C. (2014). "A review on auditory space adaptations to altered
head-related cues," Auditory Cogn Neurosci 8, 219.
Mendon¸ca, C., Campos, G., Dias, P., and Santos, J. A. (2013). "Learning
59
Auditory Space: Generalization and Long-Term Effects," PloS One 8(10),
e77900.
Micheyl, C., Carlyon, R. P., Gutschalk, A., Melcher, J. R., Oxenham, A. J.,
Rauschecker, J. P., Tian, B., and Courtenay Wilson, E. (2007). "The
role of auditory cortex in the formation of auditory streams," Hear Res
229(1-2), 116 -- 131.
Middlebrooks, J. C. (2009). "Auditory system: central pathways," in En-
cyclopedia of Neuroscience (Academic Press, Oxford), pp. 745 -- 752.
Middlebrooks, J. C. (2015). "Sound localization," Handb Clin Neurol 129,
99 -- 116.
Miller, L. M., and Recanzone, G. H. (2009). "Populations of auditory cor-
tical neurons can accurately encode acoustic space across stimulus inten-
sity," Proc Natl Acad Sci 106(14), 5931 -- 5935.
Møller, H., Sørensen, M. F., Hammershøi, D., and Jensen, C. B. (1995).
"Head-related transfer functions of human subjects," J Audio Eng Soc
43, 300 -- 321.
Moller, S., and Raake, A. (2014). Quality of Experience: Advanced Con-
cepts, Applications and Methods (Springer).
Moore, A. H., Brookes, M., and Naylor, P. A. (2013). "Room geometry
estimation from a single channel acoustic impulse response," in Proc Eur
Signal Process Conf EUSIPCO, pp. 1 -- 5.
60
Morimoto, M., Iida, K., and Sakagami, K. (2001). "The role of reflections
from behind the listener in spatial impression," Appl Acoust 62(2), 109 --
124.
Muniak, M. A., and Ryugo, D. K. (2014). "Tonotopic organization of ver-
tical cells in the dorsal cochlear nucleus of the CBA/J mouse: Tonotopic
organization of vertical cells in the DCN," J Comp Neurol 522(4), 937 --
949.
Nelken, I., Bizley, J., Shamma, S. A., and Wang, X. (2014). "Auditory
cortical processing in real-world listening: the auditory system going real,"
J Neurosci 34(46), 15135 -- 15138.
Noble, W., and Gatehouse, S. (2006). "Effects of bilateral versus unilat-
eral hearing aid fitting on abilities measured by the Speech, Spatial, and
Qualities of Hearing scale (SSQ)," Int J Audiol 45(3), 172 -- 181.
Oberfeld, D., and Klockner-Nowotny, F. (2016). "Individual differences in
selective attention predict speech identification at a cocktail party," eLife
5, 16747.
Okano, T., Beranek, L. L., and Hidaka, T. (1998). "Relations among in-
teraural cross-correlation coefficient (IACCE), lateral fraction (LFE), and
apparent source width (ASW) in concert halls," J Acoust Soc Am 104(1),
255 -- 265.
Patynen, J., and Lokki, T. (2016). "Concert halls with strong and lateral
sound increase the emotional impact of orchestra music," J Acoust Soc
Am 139(3), 1214 -- 1224.
61
Peck, C. K. (1996). "Visual-auditory integration in cat superior colliculus:
implications for neuronal control of the orienting response," Prog Brain
Res 112, 167 -- 177.
Perrett, S., and Noble, W. (1995). "Available response choices affect local-
ization of sound," Percept Psychophys 57, 150 -- 158.
Perrott, D. R., and Saberi, K. (1990). "Minimum audible angle thresholds
for sources varying in both elevation and azimuth," J Acoust Soc Am
87(4), 1728 -- 1731.
Rauschecker, J. P. (2011). "An expanded role for the dorsal auditory path-
way in sensorimotor control and integration," Hear Res 271(1-2), 16 -- 25.
Rauschecker, J. P., and Tian, B. (2000). "Mechanisms and streams for
processing of "what" and "where" in auditory cortex," Proc Natl Acad
Sci 97(22), 11800 -- 11806.
Reichinger, A., Majdak, P., Sablatnig, R., and Maierhofer, S. (2013). "Eval-
uation of Methods for Optical 3-D Scanning of Human Pinnas," in Int
Conference on 3D Vision, pp. 390 -- 397.
Ruggles, D., Bharadwaj, H., and Shinn-Cunningham, B. G. (2012). "Why
Middle-Aged Listeners Have Trouble Hearing in Everyday Settings," Curr
Biol 22(15), 1417 -- 1422.
Salminen, N. H., Takanen, M., Santala, O., Lamminsalo, J., Alto`e, A., and
Pulkki, V. (2015). "Integrated processing of spatial cues in human audi-
tory cortex," Hear Res 327, 143 -- 152.
62
Schechtman, E., Shrem, T., and Deouell, L. Y. (2012). "Spatial Localization
of Auditory Stimuli in Human Auditory Cortex is Based on Both Head-
Independent and Head-Centered Coordinate Systems," J Neurosci 32(39),
13501 -- 13509.
Schultz, D. P., and Schultz, S. E. (2015). A History of Modern Psychology,
11 ed. (Cengage Learning, Boston, MA).
Shinn-Cunningham, B. G. (2008). "Object-based auditory and visual at-
tention," Trends Cogn Sci 12(5), 182 -- 186.
Shinn-Cunningham, B. G., Santarelli, S., and Kopco, N. (2000). "Tori of
confusion: binaural localization cues for sources within reach of a listener,"
J Acoust Soc Am 107(3), 1627 -- 36.
Shore, S. E. (2005). "Multisensory integration in the dorsal cochlear nucleus:
unit responses to acoustic and trigeminal ganglion stimulation," Eur J
Neurosci 21(12), 3334 -- 3348.
Singla, S., Dempsey, C., Warren, R., Enikolopov, A. G., and Sawtell, N. B.
(2017). "A cerebellum-like circuit in the auditory system cancels re-
sponses to self-generated sounds," Nat Neurosci 20(7), 943 -- 950.
Skottun, B. C., Shackleton, T. M., Arnott, R. H., and Palmer, A. R. (2001).
"The ability of inferior colliculus neurons to signal differences in interaural
delay," Proc Natl Acad Sci 98(24), 14050 -- 14054.
Slama, M. C. C., and Delgutte, B. (2015). "Neural Coding of Sound Enve-
lope in Reverberant Environments," J Neurosci 35(10), 4452 -- 4468.
63
Slater, M. (2003). "A Note on Presence Terminology," Presence Connect
3.
Slee, S. J., and Young, E. D. (2013). "Linear Processing of Interaural Level
Difference Underlies Spatial Tuning in the Nucleus of the Brachium of the
Inferior Colliculus," J Neurosci 33(9), 3891 -- 3904.
Slee, S. J., and Young, E. D. (2014). "Alignment of sound localization cues
in the nucleus of the brachium of the inferior colliculus," J Neurophysiol
111(12), 2624 -- 2633.
Snyder, J. S., and Elhilali, M. (2017). "Recent advances in exploring the
neural underpinnings of auditory scene perception," Ann NY Acad Sci
1396(1), 39 -- 55.
Sokolov, E. (2001). "Orienting Response," in International Encyclopedia of
the Social & Behavioral Sciences (Elsevier, Pergamon), pp. 10978 -- 10981.
Strack, F., and Deutsch, R. (2004). "Reflective and Impulsive Determinants
of Social Behavior," Pers Soc Psychol Rev 8(3), 220 -- 247.
Straka, M. M., Schmitz, S., and Lim, H. H. (2014). "Response features
across the auditory midbrain reveal an organization consistent with a
dual lemniscal pathway," J Neurophysiol 112(4), 981 -- 998.
Szab´o, B. T., Denham, S. L., and Winkler, I. (2016). "Computational Mod-
els of Auditory Scene Analysis: A Review," Front Neurosci 10.
Trapeau, R., Aubrais, V., and Schonwiesner, M. (2016). "Fast and per-
64
sistent adaptation to new spectral cues for sound localization suggests a
many-to-one mapping mechanism," J Acoust Soc Am 140(2), 879 -- 890.
Trapeau, R., and Schonwiesner, M. (2015). "Adaptation to shifted interau-
ral time differences changes encoding of sound location in human auditory
cortex," NeuroImage 118, 26 -- 38.
Viaud-Delmon, I., and Warusfel, O. (2014). "From ear to body:
the
auditory-motor loop in spatial cognition," Auditory Cogn Neurosci 8,
283.
Vorlander, M., and Shinn-Cunningham, B. (2014). "Virtual auditory dis-
plays," in Handbook of virtual environment technology, edited by K. S.
Hale and K. M. Stanney, 2 ed. (CRC Press, Boca Raton), pp. 87 -- 114.
Whitmer, W. M., Seeber, B. U., and Akeroyd, M. A. (2013). "Measuring
the apparent width of auditory sources in normal and impaired hearing,"
Advances in Experimental Medicine and Biology 787, 303 -- 310.
Winkler, I., Denham, S. L., and Nelken, I. (2009). "Modeling the audi-
tory scene: predictive regularity representations and perceptual objects,"
Trends Cogn Sci 13(12), 532 -- 540.
Witmer, B. G., and Singer, M. J. (1998). "Measuring presence in virtual
environments: A presence questionnaire," Presence 7(3), 225 -- 240.
Woods, J. W. (1964). "Behavior of chronic decerebrate rats," J Neurophys-
iol 27, 635 -- 644.
65
Xie, B. (2013). Head-related transfer function and virtual auditory display
(J. Ross Publishing, Plantatation, FL).
Yao, J. D., Bremen, P., and Middlebrooks, J. C. (2015). "Transformation of
spatial sensitivity along the ascending auditory pathway," J Neurophysiol
113(9), 3098 -- 3111.
Yost, W. A. (1974). "Discriminations of interaural phase differences," J
Acoust Soc Am 55(6), 1299 -- 1303.
Yost, W. A., Zhong, X., and Najam, A. (2015). "Judging sound rotation
when listeners and sounds rotate: Sound source localization is a multisys-
tem process," J Acoust Soc Am 138(5), 3293 -- 3310.
Ziegelwanger, H., Kreuzer, W., and Majdak, P. (2016). "A priori mesh grad-
ing for the numerical calculation of the head-related transfer functions,"
Appl Acoust 114, 99 -- 110.
Ziegelwanger, H., and Majdak, P. (2014). "Modeling the direction-
continuous time-of-arrival in head-related transfer functions," J Acoust
Soc Am 135(3), 1278 -- 1293.
Ziegelwanger, H., Majdak, P., and Kreuzer, W. (2015). "Numerical calcu-
lation of listener-specific head-related transfer functions and sound local-
ization: Microphone model and mesh discretization," J Acoust Soc Am
138(1), 208 -- 222.
66
|
1512.03784 | 2 | 1512 | 2016-01-29T11:06:34 | A Chessboard Model of Human Brain and One Application on Memory Capacity | [
"q-bio.NC"
] | The famous claim that we only use about 10% of the brain capacity has recently been challenged. Researchers argue that we are likely to use the whole brain, against the 10% claim. Some evidence and results from relevant studies and experiments related to memory in the field of neuroscience leads to the conclusion that if the rest 90% of the brain is not used, then many neural pathways would degenerate. What is memory? How does the brain function? What would be the limit of memory capacity? This article provides a model established upon the physiological and neurological characteristics of the human brain, which could give some theoretical support and scientific explanation to explain some phenomena. It may not only have theoretically significance in neuroscience, but could also be practically useful to fill in the gap between the natural and machine intelligence. | q-bio.NC | q-bio | A Chessboard Model of Human Brain
and One Application on Memory
Capacity
Chenxia Gu#, Shaotong Wang#, Hao Yu*
(Affiliation): Department of Mathematical Sciences, Xi'an Jiaotong-Liverpool University,
Suzhou, 215123, China
#These authors contributed equally to this work.
*Corresponding author Email: [email protected]
Abstract
The famous claim that we only use about 10% of the brain capacity has recently
been challenged. Researchers argue that we are likely to use the whole brain,
against the 10% claim. Some evidence and results from relevant studies and
experiments related to memory in the field of neuroscience leads to the conclusion
that if the rest 90% of the brain is not used, then many neural pathways would
degenerate. What is memory? How does the brain function? What would be the limit
of memory capacity? This article provides a model established upon the
physiological and neurological characteristics of the human brain, which could give
some theoretical support and scientific explanation to explain some phenomena. It
may not only have theoretically significance in neuroscience, but could also be
practically useful to fill in the gap between the natural and machine intelligence.
Keywords
memory capacity, excitation transmission, neuroscience, machine
intelligence
1. Introduction
The memory of a human being is far more complex than storing and retrieving stored
information in a computer. It includes the association and inference [1, 2] among information,
e.g., the association between two pieces of information such as the appearance of an apple and
the human-language name, apple. To store the association, neural pathways will be generated,
i.e., connections, will be made and more neuron cells must be involved than to simply
memorize the two pieces of information [3]. The nerve cells can perform two states excitatory
and inhibitory. The transformation of a nerve cell between the two states depends on the
irritation obtained from the external environment [4] [5]. The memory forms and stores in
hippocampus which consists of a large number of nerve cells. Every pair of nerve cells can
generate a connection named synapse which is a channel for transmitting the nervous impulse
from one cell to another. In addition, synapse can only conduct impulse in one direction, which
means impulse cannot transmit back to the original cell through the same synapse [6]. In terms
of the formation delivered by synapse, the excitatory nerve cells have a probability to establish
synapse with adjacent cells. The excitement tends to propagate from the cells connected with
excitatory sensor cells to the target neuron cells. Whenever the neuron cell state changes, the
appeared synapse can always be used as a channel for conducting excitement. The pattern of
connections among cells is regarded as the memory of creatures [7].
Inspired by biological neural networks, artificial neural networks (ANNs) are a family of
statistical learning models used to estimate or approximate unknown functions that can depend
on a large number of inputs, which are presented as systems of interconnected "neurons" which
exchange messages between each other [1]. Inheriting the structure of ANNs and its application
on object recognition, a model is established to explain the functioning of the brain, based on
the physiological and neurological characteristics of the human brain. Binary number will be
used to describe the state of one neuron cell where 1 means excitatory and 0 means inhibitory.
The two states of a cell can change when the condition is just right [8]. Then the bridges of each
pair of nerve cells represent the synapse. Finally, the outermost layer of the processor matrix
represents the target cells which are the end of excitement propagation [9].
2. Modelling
The model consists of a sensor-matrix layer and a processor-matrices layer, each of which
contains a certain number of sensor-matrices and processor-matrices respectively. The reason
why we can work within two-dimension matrix is that the exact positions of neuron cells are
not so important, since essentially speaking, memory depends on not the neurons themselves,
but the connections between them. Moreover, due to the complexity of the real process in the
human brain which could be far beyond our imagination, not even within three dimensions
could it be fully illustrated [10-12].
Sensors are used to generate perception which responds to a stimulation from the outside
world. For instance, our eyes, as a visual sensor, can sense the visual features of the
stimulation; ears can sense vocal information and hands can sense tactual information. All the
information will then be processed and transmitted to the processors, where memory-making
process takes place (hippocampus) [13]. After the processors received the information from
the sensors, the different sorts of information will be associated together in some ways and
forms an initial acquisition of the memory subject. The whole memory process involves three
steps, Encoding, Consolidation and Storage, and Retrieval [14] [15]. The transition process
from sensors to processors is called encoding step. The association of all sorts of information
that the processors received from sensors is regarded as the step for consolidation and storage.
For instance, if the visual features of an apple and the literal information of an apple are
associated in a memory process, then next time, one could recall the name, apple, as soon as
he or she see an apple. The recalling is viewed related to the retrieval step.
Although this model does not require a particular memory representation, it is useful to
illustrate the realization of all the three steps with a simplified representation.
2.1. Encoding
To simulate transition process, the system with sensors and processors is considered. Each
state of the sensors or processors can be stored in a matrix, where each entry represents the
state of each neuron cell. The state of a neuron cell is assumed to be either excitatory or
inhibitory, which is indicated by binary numbers, that is, 1 and 0, respectively. An example of
the encoding process from the sensors to the processors is illustrated in Figure 1.
Figure 1 n to m correspondence: an excitatory
cell is denoted by the number 1.
Every cell in the sensor matrix could have its own multiple corresponding cells in the
processor matrix. If one cell in the sensor matrix is set to be 1, then each of the corresponding
cells in the processor matrix will have the value 1 as well [16]. In this model, the
correspondence is set to be one-to-one and is randomly generated. The randomness of
correspondence is of computational significance, since it disperses the distribution of the cells
with value of 1 in the processor which maximizes the utilization of the matrix and maximizes
the distinctions of the information in the sensor. Imagine if the cells with value 1 in the
processor are all concentrating in the center, the results from the operation of the model will
be hard to distinguish.
2.2. Consolidation and Storage
Now constrain our attention to one of the sensors and its corresponding processor, say visual
sensor and its corresponding processor, and the correspondence between them is assumed to
be one-to-one. Therefore, the visual processor, which has a spatial structure in the brain, could
be simply represented by another n by n matrix. Because the message delivered fast in human
brain, the all duration of the transmission from first cell to another cell are regarded equally
[17]. At first, the processor received an initial state of excitation from the sensor. Then, the
excitation transmission process starts. (It is assumed that the excitation could only conduct
from one cell to another once in one stage of the process and it could only be conducted from
the newly excited cells. In another word, the cells that have successfully conducted excitation
will not be able to conduct excitation any more. After the transmission from first cell to
another, a new state of this matrix will be generated.) Before the conduction of excitation, the
path which used to transmit the excitation between a pair of cell is generated firstly. When the
conduction process is finished, the path will be remained there, which forms the memory. The
term bridge is used to indicate whether there is path from one cell to another, and noticed that
every bridge is represented by a directed line segment. If there is a bridge from cell A to cell
B, then it means excitation can be conducted from cell A to cell B. A pair of cells could have
at most two bridges between them, opposite in direction. For an illustration, the cues of the n
by n processor matrix are assumed to lie on the four edges excluding the points at the corners,
as is shown in Figure 2. The inner matrix represents the processor with some initial state and
the red cells on the peripheral edges around this matrix represent the cues.
Figure 2 . Initial configuration and one possible pathway
The conduction process in a processor is regarded as finished once all the cues received the
excitation transmitted by bridge and series of excitation cell. The transmission of the
excitation is under some conduction rules with neurological significance and it is obvious that
the pathway of the excitation will be governed by the cues. Once the initial configuration and
the cues are determined, the excitation will be under conduction between cells in the processor
and ends up with its reach to the cues. One possible pathway is also shown in Figure 2 where
the bridges are indicated by the blue line segments; red cells represent cues.
2.3. Conduction Rules
To deduct the possible pathway, conduction rules are established to ensure that all the cues
could be reached and the establishment of the bridge is optimized. For better illustration,
several rules have been declared. Firstly, the conduction only occurs when bridge has been
established between the two cells. Secondly, design that each cell can build up bridges to its
four adjacent cells, with directions left, right, up and down respectively. The circumstance
with more connections will be considered in later analysis. Thirdly, the transmission process is
assumed to be inherently probabilistic. The distance to a cue determines the probability
whether the bridge will be established between the cell and the next cell which is nearer to that
cue.
Let O and I denote the sets of cues and original excited cell respectively. Suppose that there
are totally m cues and n original excited cells. That is to say, 𝑂 = 𝑂1, 𝑂2, 𝑂3, … , 𝑂𝑚 and 𝐼 =
𝐼1, 𝐼2, 𝐼3, … , 𝐼𝑛. Then theoretically, the conditional 𝑓𝑂𝑖𝐼3 = 𝑝𝑟𝑜𝑏(𝑂𝑖𝐼𝑗) is determined by
𝑓𝑂𝑖𝐼𝑗 =
𝑃𝑂𝑖𝐼𝑗
∑
𝑘∈𝑂
𝑃𝑂𝑘𝐼𝑗
(1)
Where 𝑃 denotes 𝑝𝑟𝑜𝑏(𝑂𝑟⋃𝐼𝑡), the probability that events 𝑂𝑟 and 𝐼𝑡 happen simultaneously.
Obviously, by assumption, the instantaneous set of unreached cues governs the conduction of
the next moment. Moreover, the cues could have accumulative effects on the probability of the
bridge establishment. After the transmission process, the bridges are kept and stored in the
processor, which is regarded as memory.
In the second time, when the figure of apple irritates the sensor again, the excitement will be
transmitted automatically through the path established in the first time and all cues will be
irritated again which indicates the name, apple.
2.4. Illustration
The model is aimed to simulate the memory process and explore the memory capacity. In
another word, it is aimed to find out the limit beyond which the brain would be overloaded and
the association of different types of information could be unreachable. For better illustrate, the
model is now applied to a specific example with a 30*30 processor matrix. The initial
configuration is initialized as shown in Figure 3.
indicating the cells in the processor and red dots
Figure 3. Initial configuration with black dots
The black dots inside represents the initial state of the processor, which could be an image of
an apple. If the red dots outermost refer to the name, apple, then after application of the model,
bridges will be established through which the excitation will be conducted until the reach of
excitation to all the cues eventually. One possible bridges is shown in Figure 4.
indicating the cues.
Figure 4. One possible bridges.
After that, the bridges will be kept as a memory which associates the vision of the apple and
its name. Then next time, if the processor receives the information of the apple again, the
excitation will be transmitted along the bridges stored automatically and finally reach the cues
which indicates the name of the apple. If the processor receives a different piece of
information, the excitation will still be transmitted but may not be associated with the same
cues, the name apple. For instance, if an image and its name (cues represented by red cells) are
defined as shown in Figure 5 (a), another different image indicated in Figure 5 (b) input in a
short time. Then the name output in Figure 5 (a) is different from it in Figure 5 (b), so this is
a successful case. However in Figure 5 (c), another different image obtained by processor in a
short time, the name output is totally the same as it in Figure 5 (a), which means the images
are not be distinguished. Therefore this is a chaotic case [18].
(a)
(b)
(c)
Figure 5 (a)The reference image and its name. (b)Successfully-Distinguished Case: the
processor successfully distinguishes a different image. (c)Chaotic Case: the associated cues are
exactly the same as shown in (a), while the information that the processor rece
3. Results and Analysis
Whether the system is chaotic depends on the state of the cells [19]. In order to describe the
chaotic degree, the concept similarity is introduced here. Two pieces of different image
information are delivered to the same processor in a short period, which means when the second
image message is obtained, the bridges established for the first image have not degraded. Two
cues, one is for the first message and the other is for the second message. Some parts in the two
cue-layers generated respectively are supposed to be different, and the percentage of the same
part in the two cue-layers is indicated by similarity. However if the two cue-layers are exactly the
same, the case is regarded as a chaotic case. First of all, assume that the number of cells with
value of 1 in inner matrix represents the number of information cells included in the image. The
relationship between whether the processor is chaotic and the number of information cells
involved is investigated by the following method. First, change the percentage of the number of
cell with the value of 1 among total number of cells both in image matrix and cue layer from 1%
to 90%. Then, repeat each test 20 times for each percentage and record the similarity of each
different test. The similarity (S) of the results can be calculated by:
𝐺𝑖𝑗 = {
0 𝑖𝑓 𝑖 ≠ 𝑗
1 𝑖𝑓 𝑖 = 𝑗
, 𝑆 =
∑
𝑛
𝑡=1
𝐺𝑎𝑡 𝑏𝑡
𝑛
× 100%
(2)
Where n is the total number of bits in cue layer; 𝑎𝑡 and 𝑏𝑡 represent the 𝑡𝑡ℎ bit in the test sample
and the original sample, respectively.
When S equals 1, which means the model cannot tell the difference between the two different
images, which is called chaotic. If G is 0, that means the two relative cues are not exactly the
same, the difference of the two different images is told by the model. In order to find an optimal
condition to distinguish the different image information, first of all, assume that the number of
cells with value of 1 in image matrix represents the amount of information stored in the image
matrix. The relationship between the whether the processor is chaotic and the number of cells is
investigated by using the following method. Adjust the percentage of the cells storing
information and located in image matrix and cue layer from 1% to 90%. Repeat 50 times in
every percentage and record the times where the similarity is not equal to 1. Then calculate the
successful rate under each condition, where the successful rate is the percentage of the
experiment with S=1 in the whole 50 times experiment. Therefore, Table 1 shows the results.
Table 1 Successful Rate
Based on Table 1, a fitting surface is plotted to find the optimal condition, shown in Figure 6.
The top point of the curve in Figure 6 with the highest success rate (98%) is the highest point in
the whole figure. The condition where the percentage of cells storing information in cue-layer is
10% and the one in image matrix is also 10% is the optimal condition. Therefore, the test result is
record which is in condition 10% density image matrix and 10% cue fulfillment. That means
when the 10% of the brain space is used to store information, the effect of remembering
association is best. That is why in human beings brain, 10% of the storing space is occupied.
However, other cells which not hold information are not useless and they are used to transfer
excitation. In the course of human evolution, the people with a poor memory function are
eliminated and the people with brain in the optimal condition are reserved.
Figure 6. Fitting surface of successful rate.
Image what would happen if much more storage space is used or unbalanced storage space is
used. If 70% of storage space is used both of character part and image part, the successful rate
only 70%. That means if you told the patients that image represent an apple, in the next time he
may call the other totally different image an apple [20]. Therefore the patients will be diagnosed
for a psychopath whose memory is overloaded.
Moreover, in another condition where the two percentages of storage space of character part and
image part are different, the function of brain is also influenced. For instance when the character
capacity is 10% and the image part is 70%, the success rate is only 58%. That means if the man
who has great talent in memorizing images but his capacity of character memorizing is not
balanced with it of image memorizing, the man is also has a tendency to be insane. In other
words, the genius who has gift on one aspect is nearly lunatic.
Generally, this model can be regarded as a mini-brain. The ability of study depends on the
number of cells in use, which means that the length of the period for study decreases with the
size of the matrix [21]. The biggest differences between this model and the real brain are the
number of cells and the number of connections. Therefore, in this section, matrices with different
dimensions and different numbers of connections are adjusted and then the results with different
parameters are compared.
Firstly, the dimension of the matrix is expanded to 50×50, 80×80 and 100×100, respectively.
Thus, correspondingly, the number of neural cells involved increases to 2500, 6400 and 10000,
respectively. The condition with 10% density image matrix and 10% cue fulfillment (the optimal
condition) are set to examine the effect of the size of the matrix on the successful rate. After that
the case with dimension 50×50 has been repeated 50 times and a new successful rate 98% was
obtained which is also comparatively high. The cases with dimension 80×80 and 100×100 were
tested and the successful rates are 100% and 98%, respectively. Therefore, we can speculate that
the successful rate may be independent from the size of matrix.
Secondly, in order to investigate whether the number of connections can affect the results of the
experiment, we double the number of connections and observe, i.e., not only the up, below, left
and right adjacent cell are connected to the central cell but also the up-right, up-left, bottom-left
and bottom-right adjacent cell can be associated. Two pieces of different image information are
given under the optimal condition. The relative cue-layers are different so the two pieces of
image are successfully distinguished.
After that, the experiment has been repeated 50 times and the successful rate obtained is 100%.
Therefore we can speculate that the number of connections may not influence the successful rate
either. According to the statistics, the dimension and the number of connections may have no
effect on the successful rate. Therefore, we can conclude that the percentage of optimal occupied
space in the brain is about 10%.
Finally, to simplify the computation and programming, the number of connections that a neuron
cell could have with the adjacent cells is up to eight in this model. However, in fact, thousands of
connections could be made by one single cell [21].
Hence, further study could enlarge the matrix. Since in the model the matrix used to represent the
processor has comparatively small dimensions, the processor is considered as a whole. When the
matrix becomes larger and larger, partition of the processor could be taken into consideration.
In this model, only two states (excitatory and inhibitory) of a cell are considered. More values
could be used to represent different states to increase the complexity of the situation. Moreover,
the transmission is assumed to happen so quickly that forgetting can be neglected. In some other
cases, forgetting may have significant effects on the memory process.
4. Conclusion
The model has been established to simulate the memory process of the human brain. Through
conducting thousands of experiment by simulation, the statistical results indicate that the
optimal percentage of occupation in the brain is 10%. However, the other cells which are not
occupied are necessary for transmitting the excitation rather than unexploited. Furthermore, it
can be speculated that the successful rate may be independent from the size of the matrix. And
the dimension and number of connections may have no effect on the successful either.
Acknowledgements
This work was supported by grants from the National Natural Science Foundation of China
(No.11204245), and the Natural Science Foundation of Jiangsu Province (No.BK2012637).
References
[1] Grossi, E and Buscema, M. (2007) Introduction to Artificial Neural Networks, European Journal
of Gastroenterology and Hepatology, 19, 1046-1054.
[2] Mastin, L. (2015) Long term memory. http://www.human-memory.net/types\_long.html
[3] Ma, C. and Zhang, N.(2015) Transforming growth factor beta signaling is constantly shaping
memory T-cell population, Proc Natl Acad Sci USA, 112 (35) 11013-11017.
[4] Vogels,T. P., Sprekeler H., Zenke F., Clopath C., and Gerstner W (2011) Inhibitory plasticity
balances excitation and inhibition in sensory pathways and memory networks. Science 16, 1569-
1573.
[5] Hartmann K, Bruehl C, Golovko T, Draguhn A (2008) Fast homeostatic plasticity of inhibition via
activity-dependent vesicularlling. Plos One, doi: 10.1371/journal.pone.0002979.
[6] Huxley, A. F. (1964) Excitation and conduction in nerve: quantitative analysis. Science,145, 1154-
1159.
[7] Okun, M., Lampl, I. (2008) Instantaneous correlation of excitation and inhibition during ongoing
and sensory-evoked activities. Nat Neurosci 11, 535-537.
[8] Luksys G. et al. (2015) Computational dissection of human episodic memory reveals mental
process-specific genetic profiles. Proc Natl Acad Sci USA, 112, 4939-4948.
[9] Kawato, M., Gomi, H. (1992) The cerebellum and VOR/OKR learning models. Trends Neurosci
15(11): 445-453.
[10] Machens, C. K. , Romo, R. , Brody, C.D. (2005) Flexible control of mutual inhibition: A neural
model of two-interval discrimination. Science 307, 1121-1124.
[11] Nenert, R., Allendorfer, J. B., Szaarski, J. P. (2014) A model for visual memory encoding. Plos
One, doi: 10.1371/journal.pone.0107761.
[12] Flores CE, et al. (2015) Activity-dependent inhibitory synapse remodeling through gephyrin
phosphorylation. Proc Natl Acad Sci USA, 112, E65-E72.
[13] Nielsona, D. M., Smitha, T. A., Sreekumara, V., Dennisa, S., Sederberga, P. B. (2015) Human
hippocampus represents space and time during retrieval of real-world memories. Proc Natl Acad
Sci USA, 112(35): 11078-11083.
[14] Mastin, L. (2010). Declarative Memory (Explicit Memory) and Procedural Memory (Implicit
Memory)-Types of Memory-The Human Memory. The Human Memory.
[15] Nomura, H., et al. (2015) Memory formation and retrieval of neuronal silencing in the auditory
cortex. Proc Natl Acad Sci USA, 112(31): 9740-9744.
[16] Markowitza, D. A., Curtisa, C. E., Pesarana, B. (2015) Multiple component networks support
working memory in prefrontal cortex. Proc Natl Acad Sci USA, 112(35): 11084 -11089.
[17] Wada, N., Funabiki, K., Nakanishi, S. (2014) Role of granule-cell transmission in memory trace of
cerebellum-dependent optokinetic motor learning. Proc Natl Acad Sci USA, 111(14): 5373-5378.
[18] Vreeswijk, C. V., Sompolinsky, H. (1996) Chaos in neuronal networks with balanced excitatory
and inhibitory activity. Science, 274(5293): 1724-1726.
[19] McTighe, S. M., Cowell, R. A., Winters, B. D., Bussey, T. J., Saksida, L. M. (2010) Paradoxical
false memory for objects after brain damage. Science, 330, 1408-1410.
[20] Yger, P., Stimberg, M., Brette, R. (2015) Fast earning with wleak synaptic plasticity. J Neurosci ,
35(39): 13351-13362.
[21] Gerstner, W., Naud, R. (2009) How good are neuron models. Science, 326(5951): 37
|
1502.01438 | 1 | 1502 | 2015-02-05T06:39:38 | Neuronal micro-culture engineering by microchannel devices of cellular scale dimensions | [
"q-bio.NC",
"physics.bio-ph"
] | Purpose: The purpose of the current study was to investigate the effect of microchannel geometry on neuronal cultures and to maintain these cultures for long period of time (over several weeks) inside the closed microchannels of cellular scale dimensions.
Methods: The primary hippocampal neurons from E-18 rat were cultured inside the closed polydimethylsiloxane (PDMS) microchannels of varying sizes. The effect of the channel geometry on the spatial and the temporal variations in the neural microenvironment was investigated by studying neural maturation and variation in the media osmolality respectively. The cultures were maintained for longer time spans by PDMS device pretreatment, control on media evaporation (by using hydrophobic ethylene propylene membrane) and an effective culture maintenance protocol. Further, the devices were integrated with the planar microelectrode arrays (MEA) to record spontaneous electrical activity.
Results: A direct influence of channel geometry on neuron maturation was observed with cells in smaller channels maturing faster. The temporal variation in the microenvironment was caused by several fold increase in osmolality within 2-3 days due to rapid media evaporation. With our culture methodology, neurons were maintained in the closed channels as small as 50 microns in height and width for over 1 month in serum free media condition and the time varying spontaneous electrical activity was measured for up to 5 weeks using the MEA.
Conclusions: The understanding of the effect of the culture scale on cellular microenvironment and such long-term culture maintenance will be helpful in studying neuronal tissue development; therapeutic drug screening; and for network level neuronal analysis. | q-bio.NC | q-bio | The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Neuronal micro-culture engineering by microchannel devices
of cellular scale dimensions
Gaurav Goyal and Yoonkey Nam
Department of Bio and Brain Engineering, KAIST, Daejeon 305-701, South Korea
Abstract
Purpose
The purpose of the current study was to investigate the effect of microchannel geometry on neuronal
cultures and to maintain these cultures for long period of time (over several weeks) inside the closed
microchannels of cellular scale dimensions.
Methods
The primary hippocampal neurons from E-18 rat were cultured inside the closed polydimethylsiloxane
(PDMS) microchannels of varying sizes. The effect of the channel geometry on the spatial and the
temporal variations in the neural microenvironment was investigated by studying neural maturation and
variation in the media osmolality respectively. The cultures were maintained for longer time spans by
PDMS device pretreatment, control on media evaporation (by using hydrophobic ethylene propylene
membrane) and an effective culture maintenance protocol. Further, the devices were integrated with the
planar microelectrode arrays (MEA) to record spontaneous electrical activity.
Results
A direct influence of channel geometry on neuron maturation was observed with cells in smaller
channels maturing faster. The temporal variation in the microenvironment was caused by several fold
increase in osmolality within 2-3 days due to rapid media evaporation. With our culture methodology,
neurons were maintained in the closed channels as small as 50 µm in height and width for over 1 month
in serum free media condition and the time varying spontaneous electrical activity was measured for up
to 5 weeks using the MEA.
Conclusions
The understanding of the effect of the culture scale on cellular microenvironment and such long-term
culture maintenance will be helpful in studying neuronal tissue development; therapeutic drug
screening; and for network level neuronal analysis.
Key words:
Hippocampal Neurons, Neuron maturation, PDMS Microfluidic channels, Long term Culture,
Electrophysiology
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Introduction
The usefulness and reliability of the in vitro cell culture models depend on their ability to capture
the characteristic tissue physiology in terms of cell-cell interaction, cell extracellular matrix interaction
and cellular chemical and mechanical microenvironment. However, the conventional cell culture
models currently in use are limited in their ability to mimic the in vivo cellular microenvironment owing
to the large culture scale relative to the cell size, the low cell volume to the extracellular fluid volume
ratio and the convective media flow [1-3]. The escalating understanding of cellular microenvironment
in vitro and the advancements in microtechnology have led to the development of microscale cell culture
models over the past few years [4, 5]. The miniaturized architecture of these microdevices renders the
cell culture scale to be proportional to the spatial resolution of cell-to-cell communication and cell
volume to the extracellular fluid volume ratio greater than one [1], thereby establishing sustained
chemical gradients and simulating the cellular microenvironment analogous to in vivo.
Such miniaturized cell culture devices can also be used to grow nerve cells in confined spaces with
tens of nano-liters of culture media. There have been some attempts to culture mammalian neurons in
closed polydimethylsiloxane (PDMS) microfluidic channels for various neurobiological applications.
Taylor et al. have reported a microfluidic neuron culture device composed of two large fluidic
compartments (width: 1.5 mm, height: 100 µm, length: 8 mm) connected by microchannels. They
selectively guided axons from one compartment to the other and used the device for axonal injury and
regeneration studies [6, 7]. The similar culture device has also been used by some other groups for
studying injury and regeneration of individual axons [8] , culturing neurons derived from embryonic
stem cells [9], for purification of the axonal mRNA [10], for electrophysiological measurements [11],
and to study intracellular pH regulation in the neurites and the soma [12]. Recently, low-density
neuronal cultures inside the narrow PDMS microchannels (channel width: 85 µm, height: 45 µm, length:
variable) have also been reported [13]. Millet et al. proposed a series of chemical treatment of the PDMS
to enhance neuronal survival in the close-channel devices; however, their neuronal cultures inside the
closed microchannels could be maintained only up to 7 days in the static bath devices and up to 11 days
in the perfusion devices made from native, autoclaved or the chemically treated PDMS.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
All of the above discussed investigations have presented a new approach for neuronal analysis in
vitro; however, there is only a limited knowledge available about the effect of microchannel geometry
on shaping up the cellular microenvironment, and its influence on the neuronal cultures inside the
channels. Furthermore, the reported works have not been very successful in maintaining neurons in the
narrow microchannels for long period of time, thereby precluding their use for longitudinal neuronal
studies in vitro. It is difficult to understand the complex cellular microenvironment simulated inside the
microchannels because of its inaccessibility; nevertheless, this knowledge is vital in order to design the
effective and reliable microscale neuron culture models and to engineer the neural microenvironment
in vitro. The other important prerequisite for the usefulness of such culture models is the ability to
culture neurons for a ‘long time’ so that these models can also be used for a variety of in vitro neural
analysis which is currently possible only with the conventional culture systems (using culture flasks
and Petri dishes).
Here we report the investigation of micro-scale neuronal cultures engineered in microchannels of
cellular dimensions. PDMS microchannel devices were fabricated by soft-lithography and E18 rat
hippocampal neurons were cultured up to one month in close channel devices. Early neuronal
development and long-term electrophysiological properties were characterized. The size of the culture
channels was found to be significant parameter and the narrower channels resulted in the faster neuron
maturation. By adequate pretreatment of the PDMS devices, efficient control over media evaporation
and an effective culture maintenance protocol, we maintained the micro-cultures for over one month in
as small as 50 µm high, 50 µm wide and 7 mm long closed microchannels. Further, we successfully
integrated the microchannel cultures with multi electrode arrays (MEAs) and recorded the spontaneous
extracellular activity over a span of 5 weeks. Figure 1 shows the experimental scheme for this work.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Figure 1 Schematic for microdevice integration with microelectrode arrays. Microchannel devices
were assembled on poly-D-lysine coated substrates followed by cell plating in the device and membrane
installation. After the network maturation, MEA could be used to record signals from neurons inside
the microchannels.
Methods
Microchannel device fabrication, preparation and assembly
The masters for the microchannel devices were fabricated by two methods and were divided
into two groups (Table 1).
Table 1 Microchannel dimensions used for maturation study.
Group 1
Group 2
Width
(µm)
1000
1000
200
100
50
Height
(µm)
150
50
50
50
50
Length
(mm)
7
7
7
7
7
The group 1 consisted of 1 mm wide single channel devices fabricated by a clean-room free
process as illustrated in Figure 2. The channel features were made by taping 3M Scotch transparent tape
(MMM600341296, 3M, St. Paul, MN) on clean glass slides. A single layer of tape resulted in 50 µm
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
high features and multiple layers were used to fabricate higher features. The second group consisted of
the devices with 10 parallel microchannels spaced by 200 µm. For each channel, the height was 50 µm
and the width was 50, 100 or 200 µm. The masters were fabricated using standard photolithography
process. Briefly, a 4-inch silicon wafer was coated with the negative photoresist SU-8 2025
(MicroChem Corp., Newton, MA), soft baked and exposed through a film mask using Q4000 Infrared
Mask Aligner (Quintel Corp., Morgan Hill, CA). A positive relief pattern corresponding to the
microchannel geometry was obtained after the post-exposure bake and development.
Figure 2 Clean-room free fabrication process. 3M Scotch transparent tape was pasted on a cleaned
glass slide and was assembled on a normal office paper with straight channel pattern printed on it. The
tape on the glass slide was cut using a surgical knife following the channel pattern visible under the
glass slide. The background tape was removed to obtain positive relief pattern on glass. PDMS was
cast against the tape master and holes were punched in the molded PDMS to obtain the desired device.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
After the masters were fabricated, PDMS (Sylgard 184, Dow Corning Corp., MI) was cast
against the masters to create desired microchannel devices. Input and output ports were made by
punching holes (3 mm or 7 mm diameter) in the fabricated PDMS devices. The devices were then
dipped overnight in methanol to remove any uncured PDMS, followed by thorough washing in
deionized water (4-5 hours dipping) and autoclaving. The inner walls of the PDMS channels were made
selective hydrophilic by assembling the devices on a plain blank glass with the feature side down and
treating with air plasma (Harrick Expanded plasma cleaner, 3 min at 30 W, ~200 mTorr). The devices
were then peeled off from the glass and were reversibly bonded on the poly-D-lysine (100 µg/ml,
Sigma-Aldrich, St. Louis, MO, US) coated cell culture substrates (glass cover-slip or microelectrode
array). The channels were immediately filled with the cell plating media and the assemblies were housed
in the Petri dishes and were placed in a humidified incubator until cell plating.
Cell Culture
For cell culture, E-18 hippocampal neurons from Sprague-Dawley (SD) rat (Koatech Animal
Services, Pyeongtaek-si, Gyeonggi-do, Korea) were suspended in cell plating media consisting of
Neurobasal media (Invitrogen, Gaithersburg, MD, US) supplemented with B-27 (Invitrogen), 2 mM
GlutaMAX (Invitrogen), 12.5 μM L-glutamic acid (Sigma-Aldrich, St. Louis, MO, US) and Penicillin-
Streptomycin (1:100 , Invitrogen). For cell loading, the devices kept in the incubator were taken out
and the excess media from the reservoirs was aspirated. The cell suspension of 3 x 106 cells/ml was
prepared and was loaded through one of the wells to achieve a cell density of 250-300 cells/mm2 inside
the channels. High density cultures were prepared for the electrophysiological experiments to ensure
signal recording. The devices were then again kept inside the humidified incubator maintained at 37°C
and 5% CO2 level. Additional culture media was added to the devices after 20 minutes of cell settling.
Maturation assessment
During early maturation of the neurons in vitro, 5 different stages have been reported [14]. Over
the period of one week after cell plating, neurons undergo several morphological changes before they
attain the functionally mature state and become electrically active. We observed that all the cells do not
grow at the same pace but they are distributed from maturation stage 1 to stage 3 at 1 day in vitro. To
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
explore the effect of culture vessel geometry on neuron maturation, we calculated the percentage of
neurons in different maturation stages growing inside channels of different dimensions. To calculate
the maturation index, the bright-field phase contrast images of the neurons were taken at 1 day in vitro
(DIV) and the cells were classified into different stages following their morphological features. At stage
1, the soma was visible with lamellipodial projections and no neuritees; at stage 2, the minor processes
projected from the soma; stage 3 was marked by one major neurite which would later become the axon
of the cell. The cells were considered in stage 3 if their major neurite was more than twice as long as
the other neurites. The classification of neurons was performed by an individual blinded to the
experimental design to avoid the bias in data collection. The images of the neurons were taken across
the complete length of the channel to avoid any bias originating from location of the neurons inside the
channels. Only the neurons which could easily be classified in the 3 maturation stages were considered
and any cell clusters were neglected from the analysis.
Cell Culture Maintenance
The rapid evaporation of media from the devices is a significant problem for microscale cell
culture. We hypothesized that the continuous loss of aqueous phase from the devices was a major
perturbation in cellular microenvironment in our cultures. To investigate this temporal variation in
microenvironment due to media evaporation, the osmolality of the media in the devices was regularly
measured using the freezing point depression osmometer (K-7400, Knauer, Germany). Further, to
address the problem of evaporation, the devices were installed with a hydrophobic fluorinated ethylene
propylene (FEP) membrane [15] (ALA Scientific Instruments, New York, US) after cell plating. The
FEP membrane was introduced by Potter et al. to prevent media evaporation from their cultures with 1
ml media and is generally used in the labs working with MEA. We expanded the use of this hydrophobic
membrane to microdevices in order to minimize the loss of culture media, thereby maintaining a
favorable chemical microenvironment for healthy cultures. For the membrane installation, FEP
membrane was carefully placed on the PDMS devices and was gently pressed onto the PDMS surface
with a tweezers. The membrane could easily stick to the hydrophobic PDMS surface and make a seal
which prevented the media from evaporation. The membrane could be easily peeled off and reinstalled
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
while changing the media in the devices. For the culture nourishment, media in the devices was replaced
with the maintenance media (Neurobasal media supplemented with B27, Glutamax, and Penicillin-
Streptomycin) after 2 days of cell plating and then half of the media was changed every 3-4 days to
replenish the depleted nutrients and to remove the accumulated wastes [16, 17].
Electrophysiology
To measure electrical activity from the neurons growing inside the channels, the PDMS
microchannel devices were integrated with the planar microelectrode array (Multichannel systems,
Reutlingen, Germany) and MEA 1060 System (Gain 1200, Bandwidth: 10 – 5000 Hz, Multichannel
Systems) was used to collect multichannel extracellular neural signals (‘spikes’). The spike detection
threshold was 5 standard deviations of the background noise and well isolated unit spikes were sorted
out by Offline sorter (Plexon Inc., Dallas, TX, USA). The spike frequency and the burst frequency were
calculated using NeuroExplorer (Nex Technologies, Littleton, MA) following the surprise algorithm
for the burst detection (minimum surprise = 2).
Results
Effect of microchannel geometry on neuron maturation
In order to explore the effect of microchannel geometry on maturation of neurons, the cells
were grown in conventional Petri dish (35 mm diameter with 2 ml culture media) and group 1 & 2
microchannel devices. Figure 3a shows the 5 maturation stages of neurons in the culture as described
by Dotti et al [14]. These stages were used to classify the neuron population in our devices. Figure 3b
shows the neurons at 1DIV distributed into stages 1-3. The neurons in Figure 3b are marked as S1, S2
and S3 for Stage 1, Stage 2 and Stage 3 respectively. The maturation index was calculated for all of our
culture devices and is shown in the Figure 3c.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Figure 3 Maturation of neurons in microchannels. (a) 5 different developmental stages of neurons in
culture, used to classify the neuron population. The image was redrawn from Ref 14. (b) Images
showing neural population distributed from stage 1-2 at 1DIV. The neurons in stage 1, 2 and 3 are
marked as S1, S2 and S3 respectively. (c) Distribution of cells in different maturation stages in different
culture formats. The chi-square statistic was calculated for the data and there is significant evidence
with *p<0.01 that the distribution of cells in different stages is different in all the culture formats.
For the conventional Petri dish, out of 410 cells analyzed 53.2 % neurons were classified in
stage 1 and only 37.1 % neurons were classified in stage 2 and rest in the stage 3. For the microdevices
with 1 mm wide and 150 µm high channels, we observed a slight increase in the stage 1 population as
compared to Petri dish culture. We observed 68.8 % cells in stage 1 and 28.5 % cells in stage 2 and a
very few cells in stage 3 out of the total 414 cells analyzed for these devices. But when the dimensions
of the channels were brought further down, the stage 2 neurons became most predominant in the culture.
It suggested the role of miniaturized dimensions of the culture vessels in accelerated neuron maturation.
For the devices with 1 mm wide and 50 µm high channels, out of 332 neurons 36.1 % were classified
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
in stage 1, 54.8 % in stage 2 and the rest in the stage 3. This switch from stage 1 to stage 2 was seen as
an effect of lowering the channel height with the constant channel width. When the width of the
channels was also reduced to 200 and 100 µm, there was further rise in stage-2 population and a rise in
stage-3 population of the neurons. For the devices with 200 µm wide and 50 µm high channels, out of
434 neurons only 10.1 % cells were in stage 1, and 60.6% and 29.3 % neurons were in stage 2 and 3
respectively. The devices with 100 µm wide and 50 µm high channels had 9.3 % cells in stage 1, 69.7 %
cells in stage 2 and 21.0 % neurons in stage 3 of the total 442 cells analyzed. This trend of accelerated
maturation upon reducing the channel dimensions was not maintained when the width of the channels
was reduced to 50 µm and we observed increase in stage 1 population compared to the devices with
100 µm width. Out of 639 neurons inside channels with 50 µm height and width, 19.7 % were classified
in stage 1, 70. % in stage 2 and the rest in stage 3. These results suggested that the height of
microchannels was a critical design parameter which significantly altered the spatial chemical
microenvironment of the neurons. This observation is in agreement with previous report by Yu et al.
[2] in which they indicated the height of microchannel as a governing parameter in diffusion based
microenvironment simulated in the channels. Our results also suggested that the optimum width of the
channels should be in the range of 100 - 200 µm for maintaining a favorable microenvironment for
neuron growth. We calculated chi-square statistic for all the culture formats and it reflected strong
statistical evidence with p-value <0.001 that the 6 culture devices were different. We also did a series
of follow up chi-square tests by taking one of the devices out at a time from the analysis to see if the
variability was due to any one device. The results suggested that all the culture devices were different
and variability was not due to any one of them.
Osmolality maintenance and neuron survival
The designed devices consisted of microchannels of cellular scale dimensions (down to 50 µm
height and width) which significantly reduced the media required for the culture. The total amount of
media contained in the devices was 50 or 200 µl (corresponding to 3 mm or 7 mm wide media
reservoirs); and routinely the 50 µl media evaporated in a single day from the devices kept inside the
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
humidified incubator which resulted in early cell death. The devices with 200 µl media endured
complete evaporation for 3 - 4 days and in these devices media was replenished before drying out, but
the neurons still could not survive for over 2 weeks of duration (data not shown).
Figure 4 Variation of media osmolality in microdevices and its effect on cell health. Media evaporation
from microdevices caused the osmolality to rise to three fold of the initial value within 4 days but after
membrane installation osmolality was well maintained (a). In hyperosmolar media conditions cells
appeared flattened and the cell nucleus condensed and became distinctively visible (marked by white
arrows) by 2 DIV. (b) & (c) show degrading cells in 100 and 50 µm wide channels respectively. After
membrane installation, media osmolality was maintained and cells appeared round and healthy (d).
Cells shown are at 2 days in vitro.
In order to prevent the loss of aqueous phase from the culture media, the devices were installed
with a water impermeable FEP membrane. After the membrane installation, the original media volume
was maintained in the devices and it helped to preserve the osmolality of the media. Figure 4 shows the
variation in media osmolality in the microchannel devices with and without the membrane and its effect
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
on the morphology of the growing neurons. Without the membrane, the osmolality increased to 3 folds
(657.5±74.5 mOsmol/kg) of the initial value (219±1 mOsmol/kg) in 4 days. Such a large change in
osmolality had significant effects on the health of the neurons inside the microchannels; the neurons
became more flattened, lost their membrane integrity and the cell nucleus shrank and became
conspicuously visible by 2 days after the cell plating (Figure 4b and 4c). Such cell morphology can be
attributed to the loss of cytosol by the neurons due to the hyperosmolar media condition [18-20]. When
the FEP membrane was installed, the osmolality could be successfully maintained within physiological
range (269.5±25 mOsmol/kg) even after 4 days. In the devices with FEP membrane, neurons exhibited
healthy morphology and maintained the spherical cell shape (Figure 4d) and they could be maintained
for several weeks in the microchannels. These observations implied that merely changing the media
frequently was not sufficient to sustain the neural culture in healthy state; rather, an effective means to
prevent media loss was needed to maintain the favorable physiological conditions.
Neuronal characteristics and long-term culture
The cells were labeled with the antibodies against the neuronal marker class III β-tubulin to
observe neuron development and the positive staining was obtained (Figure 5a, cells in a 50 µm wide
channel at 3DIV) at different time points confirming that the cells normally developed to express the
neuronal markers and maintained their neuronal characteristics (data not shown). With the control on
cellular microenvironment, the low density neuronal cultures could be stably and reproducibly
maintained for several weeks inside the closed channels. Figures 5b & 5c show single neurons inside a
50 µm wide (and 50 µm high) channel at 6 DIV and 28 DIV respectively. Figure 5d shows a 6 DIV old
neuron in a 100 µm wide (50 µm high) channel. Figure 5e is the representative image for our long-term
culture, showing a 33 DIV old neuron in a 100 µm wide and 50 µm high channel. The characteristic
morphology of hippocampal neuron was observed, marked by triangular shaped neurons, along with
regular spread and branching of neurites. The neurites were confined to the microchannels and did not
grow under the PDMS constraints.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Figure 5 Neuron development in microchannels and long term culture. Neurons in microchannels
showed characteristic growth of hippocampal pyramidal neurons. (a) Neurons in 50 µm wide channels
stained for neuronal marker class III β- tubulin at 3 DIV. (b) & (c) Single neurons inside 50 µm wide
channels at 6 and 28 DIV respectively. (d) & (e) 6 DIV and 33 DIV old neurons inside the 100 µm wide
channels.
Multichannel neural recordings for the development of neural networks
In order to demonstrate the feasibility of long-term network-level physiology experiments with
the present technique, we integrated our microchannel devices with the planar multi-electrode arrays
that are commonly used for multichannel neural recordings of in vitro neural tissues. For coupling with
the MEA, the microchannel devices were manually aligned with MEA electrodes using an inverted
microscope and were bonded reversibly. A total of 8 devices (with 50 or 100 µm wide channels) were
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
coupled with the MEA. Out of the total 8 microdevices, spontaneous activity from 4 devices could be
recorded up to 5 weeks in vitro. The activity from other 2 devices was recorded for 3-4 weeks and the
rest 2 devices did not show any activity. Figure 6 presents the representative images and recording data
from our electrophysiology experiment. Figure 6a shows the confined neural networks growing inside
8 different microchannels (100 m in width) at 17 DIV. Figure 6b shows a 100 m-wide microchannel
aligned to 30 m-TiN electrodes and dense neurites or somata growing over the electrodes owing to the
confined neural adhesion and growth. Figure 6c represents multichannel neural recordings from three
neighboring surface-embedded microelectrodes. The typical noise level estimated from the electrodes
was 18 20 µVrms and the detected spikes were mainly ‘negative’ spikes with their peak values ranging
from – 100 µV to – 700 µV. For example, the biggest spikes in Figure 6c (marked with the dots) had
negative peak values of – 374.2 µV (7 DIV), – 561.7 µV (17 DIV), and – 682.5 µV (28 DIV). Although
the measured noise level was much higher than the level without channels (2 – 3 µVrms), relatively large
spikes resulted in the high signal-to-noise ratio for reliable spike detection.
The capability of long-term culture allowed us to observe the maturation of the network activity
through the electrophysiological recordings. As the confined neuronal networks matured, the number
of spikes generated from the individual neurons (‘spike rates’) and the degree of temporal clustering of
the spikes (‘burst’) also increased. The estimated values are summarized in table 2. There was a several
fold increase in the network activity from 7 DIV to 28 DIV. The increase in the network activity after
2 weeks is a typical characteristic of the cultured neural networks and can be attributed to synapse
maturation and increase in the synapse density with time [21]. These results demonstrate the possibility
to perform the same level of electrophysiological experiments with neuronal cultures inside closed
microchannel devices as it is possible with macro-scale cultures.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Figure 6 Microchannel culture integration with MEAs. (a) The MEA electrodes overlaid with a 100
µm wide channels and cultured neural networks at 17 DIV. Scale bar 100 µm. (b) Magnified view of
three neighboring electrodes inside the channel marked with dashed lines in (a). Scale bar 100 µm. (c)
Electrical activity recorded from the three electrodes (marked as 1, 2 and 3 in part (a)) at 17, 17, and
28 DIV. Left: 2 sec traces. Right: zoomed traces (100 ms) of the marked segments in left traces.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Table 2 Estimated spike rates and bursts rates in Figure 6c at three different culture ages
Spike rate [spikes/sec]
7 DIV
17 DIV
28 DIV
0.355
1.668
7.085
0.127
1.171
8.638
-
0.085
2.573
Electrode
1
2
3
Burst rate [bursts/min]
7 DIV
17 DIV
28 DIV
2.16
4.56
17.46
0.96
3.90
23.22
-
0.42
8.04
Discussion
Neuron maturation study in microchannels
The maturation of neurons in vitro is influenced by several factors, out of which cell-cell contact
and the presence of neurotrophic factors are the major contributing elements. Neurons with a high
concentration of growth factors and extracellular matrix components around them exhibit accelerated
maturation in culture [22]. The microchannel culture model has been proposed to maintain a high
concentration of
intrinsic growth factors around the cells and to preserve their cellular
microenvironment as compared to the Petri dish culture in which the convective media flow and the
large media volume dilute the trophic factors emanated from the cells and take them away from the
cell’s immediate vicinity [2, 3]. Our rationale was that the spatial distribution of the neurotrophic factors
around the cells depends on the dimensions of the culture vessels. The narrower the culture channel,
the more the neurotrophic factors accumulated around the cells. So we started with the 1 mm wide and
150 µm high channels and then kept reducing the channel dimensions to investigate the effect this
variation in channel geometry may have on neuronal maturation. Our results reflected that the trend was
not linear for all the channel dimensions. When bulk culture was compared to the cultures inside 1 mm
wide and 150 µm high channels, there were more cells in stage 1 in the channel culture than in the bulk
culture. This can be explained by the hypothesis that the concentration of the neurotrophic factors is not
significantly high (inside these channels, when compared to the bulk culture) to accelerate the cell
maturation; rather, the negative factors such as waste accumulation and depleted nutrient condition
come into play to pull the maturation rate down. When the channel dimensions are further reduced, we
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
observed increase in the proportion of cells in stage 2 and stage 3 with each subsequent reduction in
size of the channels. This observation of accelerated growth is in agreement with our rationale of
designing this study. We believe for these three channel sizes, the increasing concentration of the
neurotrophic factors (with the decreasing size of the channels) outweighs the negative factors of
microchannel culture mentioned above and this makes the neurons to mature faster in these channels
These results suggest that the design the microscale culture vessels for neuronal culture should be based
on the optimum dimensions of the channels which can help to emulate the in vivo condition but have
the minimal technical limitations of microscale culture system (high concentration of uncured polymer,
cellular wastes and depleted nutrients; all arise due to small media volume).
The method we have used to investigate the effect of channel geometry on cellular
microenvironment served as a good semi quantitative measure. The advantage of this means of
quantification of neuron growth was that it was quick (because the neurons only needed to be classified
into different growth stages based on the morphological features) and required only bright-field images.
But it is a subjective scoring technique and is susceptible to bias if the experiment in not properly
designed. The group 1 devices used for maturation study were fabricated with an unconventional clean
room free fabrication method. Although this crude method of making the microchannels with tape
masters can suffer from the drawbacks like batch to batch variations, limitation of resolution and the
design complexity, it is however quick and easy. It provides a rapid prototype for microchannel system
(within limited size range) which can be easily adopted by any conventional biology lab to explore a
new paradigm of experimentation without substantial time and cost investments. Although the
usefulness of the devices in this size range depends highly on the experiment design. The fabrication
by this method does not require expensive set up or highly skilled personal. Tape masters can generally
be prepared in a span of few minutes and are stable enough for multiple PDMS moldings. We have
personally used our tape masters more than 20 times to mold the micro channels.
Temporal maintenance of cellular microenvironment
The chemical microenvironment generated around the cells inside the microchannels does not
remain the same but changes with time. This temporal variation can be caused by media evaporation
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
from the culture devices [6], leaching of cytotoxins from PDMS [1,13], waste accumulation and
depleted nutrient and oxygen supply inside the channels [6,7,13]. We addressed the problems of the
cytotoxin leaching from the PDMS by pretreatment and autoclaving the PDMS devices before cell
culture. The other significant issue we encountered was the evaporation of media from the devices
which increased the media osmolality to several fold of the initial value in a span of 3-4 days. Because
the amount of media contained in the micro-devices is very small, the rise in media osmolality due to
evaporation is much higher as compared to large scale cultures. This media loss problem is generally
addressed by frequently adding extra media in the devices (or frequently changing the media in the
devices) which replenishes the lost media volume [6]. But we believe when the media in the
microdevices becomes hyperosmolar, frequently adding more media is not the optimum solution and
can severely affect the cell health. When the cells growing in the hyperosmolar media are exposed to
the fresh media (when media is changed/replenished), they are subjected to an osmotic shock. This
osmotic shock is experienced by the cells every time the media is changed and it exposes the cellular
membrane to osmotic stress. Such a media change protocol worsens the already unhealthy state of the
cells caused by the hyperosmolar cellular environment and expedites cell death. In our opinion,
effective control of media evaporation from the devices is an essential measure to maintain the favorable
microenvironment around the cells for an extended duration. We installed the hydrophobic FEP
membrane on our devices to prevent the evaporation and maintain the media osmolality in the
physiological range. This temporal control over the cellular microenvironment was essential to maintain
healthy neural cultures in the microchannels for long period of time.
Need for the long term culture
The long-term neuronal culture using the conventional culture model has been widely used to
explore a variety of topics in neuroscience including synaptogenesis [23], neural aging and death [24],
synaptic plasticity [25], and neural information processing [26]. For example, the expressions of
glycolytic enzyme GAPDH, synaptic proteins (synaptophysin, α- and β-synuclein) and the level of
glutamate receptors in the cultured neurons were shown to reach to the maximum expression level from
15 to 30 days after cell plating [24]. Moreover, β-synucleins which are implicated in neurodegenerative
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
diseases such as Parkinson’s disease and Alzheimer’s disease were shown to be differentially localized
in dendrites and axons in the immature (3DIV) and mature (20 DIV) neurons, respectively [24]. It has
also been shown that neural network in vitro matures with culture age and starts to exhibit electrical
activity after 2 weeks of cell plating [21,27], indicating 2 weeks of culture age to be minimum time-point
for any meaningful network level electrophysiology experiment. It was therefore important to develop a
reproducible methodology for culturing mammalian neurons in the microchannels for long period of time
so that the microscale culture model can also be used for the similar kind of in vitro experimentation.
Such a culture platform can enable the study of neural maturation, neural aging and death as it can
facilitate definitive and long-term analysis of the chemical and bio-molecular signatures of neurons in
the microchannels. It can also provide temporally rich information for drug experiments revealing the
long term effects of drug exposure to micro scale cultures.
Microdevice integration with the MEA
We coupled our microchannel devices with the planar microelectrode arrays (MEA), to
demonstrate an integrated device that may be used for the long-term noninvasive monitoring of neural
activity and may serve as a means to study the synergistic effect of chemical and electrical stimulation
on neuronal networks in the microchannels. Efforts have also been made previously to integrate
commercial or custom-made MEAs with the microchannel neuron cultures [28-31]. Recently, Dworak
and Wheeler also reported an MEA platform with the PDMS micro-tunnels to selectively record signals
from the isolated axons [32]. In most of these previous works neurons were cultured in open wells with
an easy access to the bulk media and only the neurites were exposed to the controlled microenvironment.
The development of an integrated device with the closed microchannels has been pursued by many
groups because it is an attractive platform for high-throughput drug screening and to study
neurobiology. The electrodes underlying the neural circuits in microchannels can record action
potentials and can be used to investigate the functional changes in the neural activity in a controlled
microenvironment. So far, the success in implementing such a hybrid device had been limited by reliable
long-term neuronal cultures in the microchannels to obtain functionally mature networks. Our
microchannel cultures protocol has enabled us to culture the mammalian neurons inside the “closed
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
microchannels” and to be able to record electrophysiological signals from these neurons over the period
of 5 weeks. Moreover, we have integrated our devices with the commercially available MEA
(Multichannel systems, Reutlingen, Germany) which allows other researchers to easily replicate, use
and extend our work. Many diverse microfluidic techniques can also be integrated with MEA for novel
neuroscience or cell based biosensor applications.
Conclusion
We cultured primary mammalian neurons inside the closed microchannels and investigated the
effect of variation in the cellular microenvironment on our neural cultures. By effectively controlling
the changes in neural microenvironment over time, we maintained the cultures for over 1 month in the
serum free media. The long-term culture further enabled us to record spontaneous electrical activity
from neural cultures when the microdevices were coupled with the microelectrode arrays. We believe
this understanding of the cellular microenvironment and the methodology to maintain healthy long-
term neuronal cultures will empower researchers to design more effective and reliable cell culture
models for in vitro neuronal analysis.
Acknowledgements
This work was supported by Korea Research Foundation Grant (KRF-2007-211-D00123)
and the Brain Research Center of the 21st Century Frontier Research Program from Ministry of
Education, Science and Technology in Korea. Authors also thank Chung Moon Soul Center for
Bioinforrmation and Bioelectronics at KAIST for the financial support.
Reference
1.
2.
Walker, G. M., Zeringue, H. C. and Beebe, D.J. Microenvironment design considerations for
cellular scale studies. Lab Chip 2004; 4(2): 91-97.
Yu, H., Meyvantsson, I., Shkel, I.A. and Beebe, D.J. Diffusion dependent cell behavior in
microenvironments. Lab Chip 2005; 5(10): 1089-1095.
3.
4.
5.
6.
7.
8.
9.
10.
11.
12.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
Yu, H., Alexander, C. M. and Beebe, D.J. Understanding microchannel culture: parameters
involved in soluble factor signaling. Lab Chip 2007; 7(6): 726-730.
Kim, L., Toh, Y. C., Voldman, J. and Yu, H. A practical guide to microfluidic perfusion culture
of adherent mammalian cells. Lab Chip 2007; 7(6): 681-694.
Meyvantsson, I. and Beebe, D. J. Cell Culture Models in Microfluidic Systems. Annu Rev Anal
Chem 2008; 1: 423-449.
Taylor, A. M., Rhee, S. W., Tu, C.H., Cribbs, D. H., Cotman, C. W. and Jeon, N. L. et al.
Microfluidic multicompartment device for neuroscience research. Langmuir 2003; 19(5): 1551-
1556.
Taylor, A. M., Blurton-Jones, M., Rhee, S. W., Cribbs, D. H., Cotman, C. W. and Jeon, N. L.
A microfluidic culture platform for CNS axonal injury, regeneration and transport. Nat Methods
2005; 2(8): 599-605.
Kim, Y., Karthikeyan, K., Chirvi, S. and Davé, D.P. Neuro-optical microfluidic platform to
study injury and regeneration of single axons. Lab Chip 2009; 9(17): 2576-2581.
Shin, H. S., Kim, H. J., Min, S.K., Kim, S.H., Lee, B.M. and Jeon, N.L. Compartmental culture
of embryonic stem cell-derived neurons in microfluidic devices for use in axonal biology.
Biotechnol Lett. 2010; 32(8):1063-70.
Taylor, A., Berchtold, N., Perreau, V.M., Tu, C.H., Li Jeon, N. and Cotman, C.W. Axonal
mRNA in uninjured and regenerating cortical mammalian axons. J Neurosci 2009; 29(15):
4697.
Kanagasabapathi, T., Wang, K.,Mellace, M., Ramakers, G.J. and Decre, M.M. Dual
compartment neurofluidic system for electrophysiological measurements in physically isolated
neuronal cell cultures. Conf Proc IEEE Eng Med Biol Soc. 2009; 2009:1655-1658.
Vitzthum, L., Chen, X., Kintner, D.B., Huang, Y., Chiu, S.Y., Williams, J. and Sun, D. Study
of Na+/H+ exchange-mediated pHi regulations
in
compartmentalized microfluidic devices. Integr Biol (Camb) 2010; 2(1): 58-64.
in neuronal soma and neurites
14.
13. Millet, L. J., Stewart, M. E., Sweedler, J. V., Nuzzo, R. G. and Gillette, M. U. Microfluidic
devices for culturing primary mammalian neurons at low densities. Lab Chip 2007; 7(8): 987-
994.
Dotti, C. G., Sullivan, C. A. and Banker, G. A. The establishment of polarity by hippocampal
neurons in culture. J Neurosci 1998; 8(4): 1454-1468.
Potter, S. M. and DeMarse, T. B. A new approach to neural cell culture for long-term studies.
J Neurosci Methods 2001; 110(1-2): 17-24.
Goslin, K. and Banker, G. Rat hippocampal neurons in low-density culture. In: Banker G,
Goslin K, editors. Culturing nerve cells. Cambridge, Mass., 1991 MIT Press: 251-281.
Kaech, S. and Banker, G. Culturing hippocampal neurons. Nat Protoc 2006; 1(5): 2406-2415.
Russell, J. W., Sullivan, K. A., Windebank, A. J., Herrmann, D. N. and Feldman, E. L. Neurons
undergo apoptosis in animal and cell culture models of diabetes. Neurobiol Dis 1999; 6(5): 347-
363.
17.
18.
15.
16.
19. McNally, H. A. and Borgens, R. B. Three-dimensional imaging of living and dying neurons
20.
21.
22.
23.
24.
25.
with atomic force microscopy. J Neurocytol 2004; 33(2): 251-258.
Van den Eijnde, S. M., Lips, J., Boshart, L., Vermeij-Keers, C., Marani, E., Reutelingsperger,
C. P. and De Zeeuw, C. I. Spatiotemporal distribution of dying neurons during early mouse
development. Eur J Neurosci 1999; 11(2): 712-724.
Jimbo, Y., Kasai, N., Torimitsu, K and Tateno, T. MEA-Based Spike Recording in Cultured
Neuronal Networks. In: Xing, W., Cheng, J., editors. Frontiers in Biochip Technology. 2006
Springer; 88-98.
Parnas, D. and Linial, M. Acceleration of neuronal maturation of P19 cells by increasing culture
density. Dev Brain Res. 1997; 101, 115-124.
Valor, L. M., Charlesworth, P., Humphreys, L., Anderson, C.N. and Grant, S.G. Network
activity-independent coordinated gene expression program for synapse assembly. PNAS 2007;
104(11): 4658.
Lesuisse, C. and Martin, L. J. Long-term culture of mouse cortical neurons as a model for
neuronal development, aging, and death. Dev Neurobiol 2002; 51(1): 9-23.
Fitzsimonds, R. M., Song, H. J. and Poo, M.M. Propagation of activity-dependent synaptic
depression in simple neural networks. Nature 1997; 388(6641): 439-448.
The final publication is available at Springer via http://dx.doi.org/ 10.1007/s13534-011-0014-y
26.
Eytan, D. and Marom, S. Dynamics and effective topology underlying synchronization in
networks of cortical neurons. J Neurosci 2006; 26(33): 8465.
27. Wagenaar, D. A., Pine, J. and Potter, S.M. An extremely rich repertoire of bursting patterns
28.
29.
30.
during the development of cortical cultures. BMC Neurosci 2006; 7(1): 11.
Claverol-Tinture, E., Ghirardi, M., Fiumara, F., Rosell, X. and Cabestany, J. Multielectrode
arrays with elastomeric microstructured overlays for extracellular recordings from patterned
neurons. J Neural Eng 2005; 2(2): L1-7.
Claverol-Tinture, E., Cabestany, J. and Rosell, X. Multisite recording of extracellular potentials
produced by microchannel-confined neurons in-vitro. IEEE Trans Biomed Eng 2007; 54(2):
331-335.
Pearce, T. M., Wilson, J. A., Oakes, S.G., Chiu, S.Y. and Williams, J.C. Integrated
microelectrode array and microfluidics for temperature clamp of sensory neurons in culture.
Lab Chip 2005; 5(1): 97-101.
31. Morin, F., Nishimura, N., Griscom, L., Lepioufle, B., Fujita, H., Takamura, Y. and Tamiya, E.
Constraining the connectivity of neuronal networks cultured on microelectrode arrays with
microfluidic techniques: a step towards neuron-based functional chips. Biosens Bioelectron
2006; 21(7): 1093-1100.
Dworak, B. J. and Wheeler, B. C. Novel MEA platform with PDMS microtunnels enables the
detection of action potential propagation from isolated axons in culture. Lab Chip 2009; 9(3):
404-410.
32.
|
1301.7187 | 1 | 1301 | 2013-01-30T10:29:40 | Emergence of Connectivity Motifs in Networks of Model Neurons with Short- and Long-term Plastic Synapses | [
"q-bio.NC"
] | Recent evidence in rodent cerebral cortex and olfactory bulb suggests that short-term dynamics of excitatory synaptic transmission is correlated to stereotypical connectivity motifs. It was observed that neurons with short-term facilitating synapses form predominantly reciprocal pairwise connections, while neurons with short-term depressing synapses form unidirectional pairwise connections. The cause of these structural differences in synaptic microcircuits is unknown. We propose that these connectivity motifs emerge from the interactions between short-term synaptic dynamics (SD) and long-term spike-timing dependent plasticity (STDP). While the impact of STDP on SD was shown in vitro, the mutual interactions between STDP and SD in large networks are still the subject of intense research. We formulate a computational model by combining SD and STDP, which captures faithfully short- and long-term dependence on both spike times and frequency. As a proof of concept, we simulate recurrent networks of spiking neurons with random initial connection efficacies and where synapses are either all short-term facilitating or all depressing. For identical background inputs, and as a direct consequence of internally generated activity, we find that networks with depressing synapses evolve unidirectional connectivity motifs, while networks with facilitating synapses evolve reciprocal connectivity motifs. This holds for heterogeneous networks including both facilitating and depressing synapses. Our study highlights the conditions under which SD-STDP might the correlation between facilitation and reciprocal connectivity motifs, as well as between depression and unidirectional motifs. We further suggest experiments for the validation of the proposed mechanism. | q-bio.NC | q-bio |
Connectivity Motifs by Synaptic Plasticity
1
Emergence of Connectivity Motifs in Networks of Model
Neurons with Short- and Long-term Plastic Synapses
Eleni Vasilaki1,2, Michele Giugliano1,2,3,∗,
1 Dept. Computer Science, University of Sheffield, S1 4DP Sheffield, UK.
2 Theoretical Neurobiology and Neuroengineering Laboratory, Dept. Biomedical
Sciences, University of Antwerp, B-2610 Wilrijk, Belgium.
3 Laboratory of Neural Microcircuitry, Brain Mind Institute, EPFL, CH-1015
Lausanne, Switzerland.
E-mail: ∗ [email protected]
Abstract
Recent evidence in rodent cerebral cortex and olfactory bulb suggests that short-term dynamics
of excitatory synaptic transmission is correlated to the occurrence of stereotypical connectivity
motifs.
In particular, it was observed that neurons with short-term facilitating synapses form
predominantly reciprocal pairwise connections, while neurons with short-term depressing synapses
form unidirectional pairwise connections. The cause of these structural differences in excitatory
synaptic microcircuits is unknown.
We propose that these connectivity motifs emerge from the interactions between short-term
synaptic dynamics (SD) and long-term spike-timing dependent plasticity (STDP). While the im-
pact of STDP on SD was shown in simultaneous neuronal pair recordings in vitro, the mutual
interactions between STDP and SD in large networks is still the subject of intense research. Our
approach combines a SD phenomenological model with a STDP model that captures faithfully
long-term plasticity dependence on both spike times and frequency. As a proof of concept, we
explore in silico the impact of SD-STDP in recurrent networks of spiking neurons with random ini-
tial connection efficacies and where synapses are either all short-term facilitating or all depressing.
For identical background inputs, and as a direct consequence of internally generated activity, we
find that networks with depressing synapses evolve unidirectional connectivity motifs, while net-
works with facilitating synapses evolve reciprocal connectivity motifs. The same results hold for
heterogeneous networks including both facilitating and depressing synapses. Our study highlights
the conditions under which SD-STDP explains the correlation between facilitation and reciprocal
connectivity motifs, as well as between depression and unidirectional motifs. These conditions may
Connectivity Motifs by Synaptic Plasticity
2
lead to the design of experiments for the validation of the proposed mechanism.
Author Summary
Understanding the brain requires unveiling its synaptic wiring diagram, which encodes memo-
ries and experiences. Synapses are, however, more than mere plugs connecting neurons; they are
communication channels implemented by biophysical systems that display history-dependent prop-
erties. During prolonged repeated activation, synapses may functionally undergo "fatigue" or may
"warm up". These effects, known as short-term plasticity, alter temporarily the communication
properties of a synapse, i.e., over tens to hundreds of milliseconds. At the same time, synapses
undergo structural changes that persist from hours to days, known as long-term plasticity.
Recent progress in electrophysiology enabled simultaneous access to the synaptic wiring dia-
gram in microcircuits, and to its short-term features, i.e., the "fatigue" or "warm-up" properties.
Non-random connectivity patterns were reported, as i) co-occurrence of "fatigue" features and
unidirectional connections between two neurons, or ii) co-occurrence of "warm-up" features and
reciprocal connections. We present a biologically plausible explanation for those patterns, based on
the interaction of short- and long-term synaptic plasticity. We further formulate a computational
model and provide an interpretation of the simulation results by means of its statistical description.
Introduction
Among the most exciting challenges in Neuroscience, the investigation of the brain wiring dia-
gram known as connectomics made spectacular progresses and generated excitement for its per-
spectives (1). Novel discoveries in molecular biology (2; 3; 4), neuroanatomical methods, (5;
6), electrophysiology (7; 8; 9), and imaging (10; 11; 12) have pushed forward the technological
limits, for an ultimate access to neuronal connectivity. This is in fact the most important level
of organization of the brain (13), pivotal to understanding the richness of its high-level cognitive,
computational, and adaptive properties, as well as its dysfunctions.
At the microcircuit level (14; 15; 16; 17; 18), the non-random features of cortical connec-
tivity have recently raised a lot of interest (7; 9). The occurrence of stereotypical connectivity
motifs was experimentally demonstrated and, in some cases, accompanied by physiological infor-
mation on neuronal and synaptic properties (7; 19; 20; 9), on activity-dependent short-term (21;
Connectivity Motifs by Synaptic Plasticity
3
22) and long-term plasticity (23; 24) and rewiring (25; 26). These physiological details are com-
plementary to anatomical connectivity mapping, and are known to underlie structural, dynamical,
and computational network properties (27). In fact, information transmission between neurons in
the brain takes place by means of more than mere "connectors". For instance, short-term dynam-
ics (SD) of synaptic efficacy is very common and has been quantified as transient and reversible
facilitation or depression of postsynaptic responses, upon repeated presynaptic activation (28).
Relaying information to downstream neurons in a microcircuit is thus dependent on past-history,
determined by presynaptic activation frequencies, and shaped by the SD profile at each synapse
(21).
We address recent experimental findings obtained in rodent cortices (19), where short-term facil-
itation and depression were found to correlate to specific, pairwise, connectivity motifs: neurons,
connected by synapses exhibiting short-term facilitation, form predominantly reciprocal motifs;
neurons, connected by synapses exhibiting short-term depression, form unidirectional motifs. In-
terestingly, the same over expression of connectivity motifs has been observed in another brain
area, i.e. the excitatory microcircuitry of the olfactory bulb (29). The cause of these structural
differences are largely unknown.
Inspired by the theory by Clopath et al. (2010) on the relationship between neural code and
cortical connectivity (30), we hypothesise that interactions between short-term and long-term
synaptic plasticity could explain the observed pairwise connectivity motifs. In (30), spike-timing
dependent long-term synaptic plasticity (STDP) was shown to account in silico for the emergence of
non-random connectivity motifs in networks of model neurons (30). Adopting the same framework,
we add to it a standard phenomenological description of SD (31) , and study SD-STDP interactions
in recurrent networks of Integrate-and-Fire neurons (32).
Extensive computer simulations of the evolution of synaptic efficacies under both external and
self-generated activity have revealed an over-expression of reciprocal motifs on facilitating synapses,
and of unidirectional motifs on depressing synapses. The departure from the initial random con-
nection efficacies results from the interplay between neuronal activity and synaptic plasticity mech-
anisms, over long (STDP) and short time scales (SD). Supported by a mean-field analysis (33; 34;
35; 36; 37), we conclude that the SD-STDP interplay is controlled by the distribution of firing
rates in the network, through the switch of STDP from a correlational "pre-post" temporal mode
at low firing rates, to a "Hebbian" rate mode at high firing rates, earlier reported experimentally
Connectivity Motifs by Synaptic Plasticity
4
by Sjostrom et al. (38). Short-term facilitation provides a positive feedback in recurrent microcir-
cuits and it boosts firing, while depression acts as a negative feedback and it effectively attenuates
high firing rates and discouraging their reverberations.
In the "Hebbian" mode, STDP mostly
gives rise to long-term potentiation (LTP) of synaptic efficacy, which becomes stronger resulting in
reciprocal motifs. This leads to increasing firing rates of bidirectionally coupled neurons, with LTP
acting as a long-term positive feedback, until the connection efficacies are ultimately stabilized by
saturation.
In the "pre-post" (temporal) mode instead, it is the spike timing that drives LTP,
which strongly favors the emergence of spatially asymmetric, i.e., unidirectional, motifs (30).
Materials and Methods
We study the properties of networks of excitatory spiking neurons (32), connected via plastic,
current-based, synapses(21; 31; 36; 35; 30; 37) via a set of simulations. We introduce a convention
for distinguishing between "strong" and "weak" connections consistent with (30), and a simple
measure to quantify the occurrence of pair-wise motifs in our simulations. We finally provide a
Wilson-Cowan firing rate model that is helpful for the interpretation of the numerical results. The
numerical values of all parameters are indicated in Table 1.
Neuron model
The network is composed of identical adaptive exponential Integrate-and-Fire neurons (32), each
described by a membrane potential Vm(t) and by a spike-frequency adaptation variable x(t)(39).
Below a threshold Vθ, Vm(t) satisfies the charge-balance equation
cm Vm = gleak (Eleak − Vm) + gleak ∆T e(Vm − VT ) / ∆T − x + Isyn + Iext
(1)
where Isyn is the synaptic input from other neurons and Iext the external (background) input
currents. When a spike occurs, i.e., Vm(t) crosses Vθ, Vm is reset to a Ereset.
The spike-frequency adaptation variable x(t) evolves as
τx x = a (Vm − Eleak) − x
(2)
When a spike occurs, x evolves as x → x + ∆x. The numerical integration of Eq. 1 is suspended
Connectivity Motifs by Synaptic Plasticity
5
for a period of time τarp following each spike, to mimic absolute refractoriness, during which Vm
remains "clamped" at Ereset.
The model details are not essential to our conclusions (see Supplemental Information, Figure
S7).
External (background) inputs
Each neuron is identified by an index i = 1, 2, 3 . . . and it receives a time variant input Iext i
according to the following protocols.
Toy Network: ( e.g., Fig. 1B, 10 neurons) Iext i consists of a 0.5 nA constant current, as well
as periodic 1 nA square pulses. Pulses occur as in a traveling wave of activity, which moves every
5 msec from one unit, e.g., the i-th neuron, to its next index neighbour, e.g., the (i + 1)-th neuron,
see also (30). Each pulse is delivered in turn to all neuronal indexes, as an extremely narrow bell-
shaped profile with unitary amplitude and standard-deviation of 0.5, resulting in neighbouring
neurons being only slightly stimulated. Each pulse is of sufficient amplitude to elicit neuronal
firing in the unit where the bell-shaped profile is centred on, e.g., the i-th unit (unless refractory).
Large Network: (e.g., Fig. 5, 1000 neurons) Iext is as in the toy network with the addition of
an uncorrelated gaussian noisy current (40), with mean µ, standard deviation σ = 200 pA, and
autocorrelation time length τsyn = 5 msec. Parameter µ is drawn randomly before launching the
simulation and for each neuron of the network, from a gaussian distribution with mean 200 pA
and unitary coefficient of variation. The noisy current mimics asynchronous synaptic inputs from
(not explicitly modelled) background populations (41; 42).
Internal (synaptic) inputs
Neurons connect to each other according to a fixed anatomical wiring matrix [Cij ], which indicates
whether neuron j-th projects to neuron i-th, i.e., Cij = 1, or not, i.e., Cij = 0. The matrix [Cij ] is
obtained from an all-to-all connectivity without autapses (i.e., Cii = 0), upon randomly pruning
≈ 20% of its elements (see e.g. Fig. 1B), to introduce variability in neuronal firing. A more
substantial reduction of the structural connectivity does not affect qualitatively our conclusion,
although it downscales the number of plastic synapses available for statistical analysis. This would
require larger networks to be simulated in order to obtain the same statistical confidence over our
results, and would require upscaling Gij (e.g., through increasing Wmax, see the next subsections)
Connectivity Motifs by Synaptic Plasticity
to obtain the same network firing rates.
The i-th neuron receives at any time a synaptic current Isyn i, described as
Isyn i = −Isyn i / τsyn +
N
∞
Xj=1
Xf
Cij Gij δ(t − tf
j )
6
(3)
where tf
j represents the occurrence time of the f -th spike emitted by the j-th presynaptic neuron,
and Gij the amplitude of the postsynaptic current (PSC), corresponding to the activation of the
synapse by the presynaptic j-th neuron. The Dirac's delta function δ(t) represents the occurrence
of a presynaptic action potential. Eq. 3 models incoming individual PSCs as waveforms with
instantaneous rise time and slower decay time (43), imposes a linear postsynaptic superposition of
effects, and implies that in the lack of any presynaptic spiking activity, Isyn i decays exponentially
to zero with a time-constant τsyn, and when a presynaptic spike is fired it evolves as Isyn i →
Isyn i + Cij Gij .
Frequency-dependent short-term synaptic dynamics (SD)
Gij defines the amplitude of the PSC from presynaptic neuron j-th to postsynaptic neuron i-th.
On short timescales, its value changes as a function of the activation history of the presynaptic
neuron, leading to transient and reversible depression or facilitation of synaptic efficacy (21). This
is referred here as homosynaptic plasticity and implemented by having Gij proportional to the
amount of used resources for neurotransmission uij rij and their maximal availability Aij, i.e.,
Gij = Aij uij rij .
Frequency-dependent short-term synaptic dynamics is described by the following equations,
with a different set of parameter values to capture depressing or facilitating synapses (21):
rij = (1 − rij )/τrec − P∞
kj
uij rij δ(t − tkj )
uij = −uij/τf acil + P∞
kj
U (1 − uij) δ(t − tkj )
(4)
For the sake of notation, indexes have been dropped from parameter U as well as from the time
constants τrec and τf acil in Eqs. 4, although in general each synapse has its own parameters (see
Table 1). Eqs. 4 reduce to the following update rules: i) when no spike is fired by the presynaptic
Connectivity Motifs by Synaptic Plasticity
7
neuron j, uij and rij recover exponentially to their resting values, U and 1, respectively;
ii)
as a presynaptic spike occurs, rij is reduced as rij → (1 − uij) rij , while uij is increased as
uij → uij + Uij (1 − uij). The impact of short-term plasticity of PSCs amplitude is exemplified
in Fig. 1C,F.
Spike-timing dependent long-term plasticity (STDP)
We further extend the description of PSCs (Eqs. 3-4), by an additional scaling factor Wij , to
incorporate STDP (24), see also (44):
Gij = Wij Aij uij rij .
(5)
Wij changes on timescales longer than τrec and τf acil and according to the correlated activity of
both pre- and postsynaptic neurons, as in (35), capturing spike-triplets effects (35; 30; 37). This
is referred here as heterosynaptic plasticity. Each neuron is complemented by four variables, i.e.,
q1, q2, o1, o2, which act as running estimates of the neuron firing rate, averaged over distinct time
scales, i.e., τq1 , τq2 , τo1, τo2 . In the lack of any firing activity of the j-th neuron, those variables
exponentially relax to zero:
τq1 q1j = −q1j
τq2 q2j = −q2j
τo1 o1j = −o1j
τo2 o2j = −o2j
(6)
while each time the neuron fires, these variables are increased by a unit:
q1j → q1j + 1
q2j → q2j + 1
o1j → o1j + 1
o2j → o2j + 1.
(7)
This enables a compact formulation of STDP: when the j-th neuron spikes, the following updates
are performed for all the indexes i:
Wij → Wij − η o1i (t) (cid:2)A−
Wji → Wji + η q1i (t) (cid:2)A+
3 q2j (t − ǫ)(cid:3)
3 o2j (t − ǫ)(cid:3)
2 + A−
2 + A+
(8)
as the j-th neuron is presynaptic to all connected i-th neurons, and postsynaptic to all connected
i-th neurons, respectively. Numerically, the evaluation of q2j and o2j is performed just before the
Connectivity Motifs by Synaptic Plasticity
8
neuron j-th spikes, as indicated by the infinitesimal time-advance notation of ǫ. When no spike
occurs, Wij maintains indefinitely its value. The instantaneous value of each Wij is bounded in the
range [0 ; Wmax]. Unless otherwise stated, Wij is initialized from a uniform random distribution
between 0 and Wmax, prior to the start of each simulation.
We remark that our choice for combining STDP and SD together aims at a phenomenological
description of their interactions. It is assumed that all STDP variables are local as well as that
STDP and SD models act independently. This allows us to easily analyse network interactions and
avoid the paradox of pre-synaptic firing detection under strong short-term depression of synaptic
efficacy. In fact, we assume for simplicity that the STDP coincidence-detector is located at each
synapse and that such a detector is able to access timing information of presynaptic spikes, re-
gardless on whether they result in large or small PSCs. Similarly, the same coincidence-detector
will also sense all the spikes emitted by the post-synaptic neuron. This is the only information
required to calculate the long-lasting synaptic modification.
Convention on "strong" and "weak" connections and network symmetry index
In this study, we focus on the appearance or disappearance of a strong connection between two
neurons, only for units that are anatomically connected, i.e., Cij = 1. For the sake of comparison,
we adopted the framework of Clopath et al. (2010), where the activity-dependent appearance or
disappearance of a connection conventionally occurs in terms of a competition among the "strong"
links in a "sea" of weak synapses (7). As in their paper, we adopt the convention of identifying as
"strong" those connections whose corresponding STDP synaptic factor Wij is above the 2/3 of its
upper bound Wmax.
With such a definition, we measure the symmetry in the network connectivity by counting
reciprocal or unidirectional motifs as (but see e.g., (45) for alternative definitions)
s(W ) = 1 − (0.5 N (N − 1) − M )−1
N
N
Xi=1
Xj=i+1
W ∗
ij − W ∗
ji
(9)
where N is the size of the matrix W , as well as the size of the network. The symmetry index s(W )
takes values in the range [0 ; 1] and depends on the average similarity between elements of W that
are on symmetric positions with respect to the diagonal. Following our previous convention, Wij
and Wji are first normalized and then zero-clipped: W ∗
ij = Wij / Wmax if Wij > 2/3 Wmax,
Connectivity Motifs by Synaptic Plasticity
9
and otherwise W ∗
ij = 0.
In Eq. S46, M represents the number of null pairs {W ∗
ij , W ∗
ji} =
{0, 0} that occur as a consequence of clipping or by initialization. Evaluating s(W ) on networks
with a majority of unidirectional connections results in values close to 0 (e.g., Fig. 2A), while
its evaluation on networks with a majority of reciprocal connections results in values close to 1
(e.g., Fig. 2B). For uniform random matrices W , the full statistics of s(W ) can be calculated
(see Supplemental Information) and employed for deriving a significance measure for s as a p-
value, representing the probability that the value of s observed in simulations could result by
chance. This symmetry measure is useful only in silico, where the entire connectivity matrix
as well as its maximal synaptic efficacy are known. The same measure cannot be considered
as an operational tool for quantifying experimentally-extracted microcircuit connectivity, where
statistical counting of connectivity motifs and their comparison to chance-level are relevant (7).
Approaching a qualitative comparison to the experiments of Wang et al. (2006), such a statistical
counting has been used in the large simulations of Fig. 5.
Mean-field Network Description
We approximate the firing rate of the network of Integrate-and-Fire units through its mean-field
dynamical description (46; 47; 34), closely following earlier work (31; 36). We consider the sim-
plifying hypotheses that i) the network consists of one or more non overlapping subpopulations of
excitatory neurons (see Figs. 4A-B and Fig. 6A) and that ii) neurons within each subpopulation
share identical synaptic coupling, connectivity, and short-term synaptic plasticity properties, i.e.,
all depressing or all facilitating, as in Fig. 6A; see Fig. 7A for an exception. We also assume that iii)
for each of the (sub)populations, the individual neuronal firing follows a Poisson distribution, with
instantaneous mean firing rate E(h), which depends monotonically on the corresponding overall
input currents h. Under these hypotheses, neurons can be only distinguished by the subpopulation
they belong to, i.e., depressing D or facilitating F, and their firing can be identified as ED(hD)
and EF (hF ). For the case of two subpopulations (Fig. 6A), hD and hF evolve over a characteristic
time scale τ as
τ hD = −hD + JDD ED + JDF EF + Iext
τ hF = −hF + JF D ED + JF F EF + Iext
(10)
Connectivity Motifs by Synaptic Plasticity
10
where Iext represents the external input, JDD and JF F the average synaptic efficacies of recurrent
connections within each subpopulation, and JDF and JF D the average synaptic efficacies of inter-
subpopulations connections. On a first approximation, J can be considered as the ensemble average
of Gij . Firing rates ED and EF are computed from hD and hF as threshold-linear frequency-
current response functions: ED = [hD − θ]+ and EF = [hF − θ]+, with [x]+ = max{x; 0}
(for alternatives see (48; 49)). JDD, JF F , JDF , JF D undergo plastic changes, on both the short
and long time scales:
indicating by Jab the coupling between the presynaptic population b and
the postsynaptic population a, then Jab = Wab Aab ub xb with a, b ∈ {D, F }. The mean-field
variables ub and xb depend only on the presynaptic firing rate Eb, and capture the short-term
homosynaptic plasticity in the mean-field approximation of Eqs. 4 (31), as:
uD = (UD − uD) /τf acil D + UD (1 − uD) ED
xD = (1 − xD) /τrec D + uD xD ED
uF = (UF − uF ) /τf acil F + UF (1 − uD) ED
(11)
xF = (1 − xF ) /τrec F + uF xF EF
On longer timescales, the mean-field approximation of STDP given in (35) is adopted here by
altering the factor Wab, which evolves as a function of both presynaptic Eb and postsynaptic Ea
firing rates:
Wab = −A−
2 τo1 Eb Ea − A−
3 τo1 τq2 E2
b Ea + A+
2 τq1 Eb Ea + A+
3 τq1 τo2 Eb E2
a . (12)
1
η
Results
We hypothesize that synaptic connectivity reflects the interaction of short-term plasticity, such
as homosynaptic depression and facilitation (21), with long-term plasticity. Inspired by a recent
theory, which links connectivity to spiking activity and to the neural code (30), we ask the question:
could known relationships between short-term plasticity and spiking activity in recurrent networks
(31; 33; 36) explain the occurrence of connectivity motifs (19; 29) observed experimentally? In
Connectivity Motifs by Synaptic Plasticity
11
support to our hypothesis, we present simulations of networks of excitatory neuron models with
activity-dependent plastic connections. We then study the necessary conditions for the emergence
of these specific connectivity motifs by standard firing rate models.
A toy microcircuit model
In order to demonstrate and test the key features of our hypothesis, we consider a simplified rep-
resentation of excitatory neuronal microcircuits, as a network of adaptive exponential Integrate-
and-Fire units (32). The validity of our results does not depend on the specific details of the model
neuron and of its spike-frequency adaptation parameters chosen here (Supplemental Information,
Figure S7). Neurons are connected to each other through excitatory synapses (Fig. 1B), whose
efficacy undergoes either short- or long-term plasticity. Using a standard model for short-term
synaptic dynamics (SD) (21), we simulate homosynaptic, use-dependent modifications of the post-
synaptic potentials (PSPs) amplitude. We also employ a recently introduced phenomenological
description of heterosynaptic long-lasting potentiation or depression of PSPs, able to capture with
great accuracy both spike-timing and frequency effects (35). We refer to this long-term associative
plasticity as spike-timing dependent (STDP). Both SD and STDP occur simultaneously with neu-
ronal dynamics, although across distinct timescales. We note that STDP interacts with short-term
plasticity as a frequency-independent scaling factor of PSPs amplitude, rather than contributing
to a redistribution of synaptic efficacy in the sense of (50; 51) (see Discussion). In addition to
excitatory PSPs evoked by trains of presynaptic spikes, neurons receive an external current input,
deterministically played back over and over, as a traveling wave of activity (Fig. 1B). Such an exter-
nal stimulation recreates the temporal coding of Clopath et al. (2010), which imposes deterministic
spike-timing correlations among evoked spikes (see Methods). The external input can be regarded
an oversimplified generic e.g., thalamic, input with known temporally correlated structure.
We define two microcircuits, identical for all aspects of neuronal properties, maximal synaptic
efficacy, anatomical connectivity, and external inputs, with the exception of the SD properties.
Specifically, one microcircuit includes exclusively short-term facilitating synapses (Fig. 1C), while
the other includes only depressing synapses (Fig. 1F). In order to describe changes in the micro-
circuit connectivity, we adopt the same framework of Clopath et al. (2010) where connections
form or disappear via competition among the "strong" links in a "sea" of weak synapses (7) (see
Methods). The connectivity, which is randomly initialized, slowly evolves into largely non-random
Connectivity Motifs by Synaptic Plasticity
12
configurations during the simulation (Fig. 1D,G). At the steady-state, these configurations match
the experimentally observed correlations: reciprocal motifs emerge in cell pairs more often than
unidirectional motifs, when synapses are facilitating; the opposite occurs when synapses are de-
pressing. This is revealed both by direct inspection of the synaptic efficacy matrix [Wij ] (e.g., see
Fig. 8) and by quantification of its symmetry index s (see Methods). Although other statistical
measures can be defined, as for instance the percentage of weak elements in [Wij ] or the sparseness
of the emerging connectivity, here we focus on the reciprocal versus unidirectional features of its
strong elements, according to the same convention of Clopath et al. (2010). When s takes values
close to 1, almost all of the existing pairwise anatomical connections are reciprocal and [Wij ] is
almost a symmetric matrix. On the other hand, for values of s close to 0, unidirectional or very
weak connection motifs prevail.
We underline that the prevalence of external inputs over recurrent synaptic inputs (or vice
versa) is not imposed a priori but it depends on the interaction between SD and STDP. For
facilitating microcircuits, the overall recurrent (synaptic) input that a neuron receives becomes
progressively larger than the external input, as soon as more and more reciprocal connectivity
motifs are established by STDP. For depressing microcircuits, the opposite holds, as more and
more connections are weakened by STDP.
These results, summarized in Fig. 1 for a sample microcircuit composed by ten neurons,
have been confirmed over 2000 repeated simulations, analysed in Fig. 2. Over a slow timescale,
STDP results in a very small degree of symmetry (s = 0.01 ± 0.01, mean ± stdev), with
high significance (p < 10−4) for each of the repeated simulations involving depressing synap-
tic connections (Fig. 2C). Under identical external inputs to the microcircuit, STDP leads in-
stead to a large degree of symmetry (s = 0.61 ± 0.10, mean ± stdev), with high signif-
icance (p < 10−4) in about 75% of the simulations involving facilitating synaptic connec-
tions (Fig. 2D). In the remaining 25% of the cases, the resulting symmetry value was in the
range [0.25; 0.55], suggesting that to some extent, the variability observed in experiments (19;
29) was replicated in the simulations, where not 100% of the times the motifs correlations emerged.
In this simplified example, the variability is attributed mostly to the sparse anatomical connectiv-
ity and the irregular firing regimes this imposes. Doubling the simulation time (not shown) led to
a very minor reduction of the variability on s, by less than 3% and only for the facilitating synaptic
connections, suggesting that stationarity had been already reached.
Connectivity Motifs by Synaptic Plasticity
13
A simple mechanism for the emergence of motifs
In the microcircuits considered above, the statistics of the anatomical connectivity, the external
inputs, as well as the single-neuron properties are identical. Any difference in long-term modifica-
tion of connections could only arise i) from differences in the collective neuronal activity and take
place ii) through the spike-timing and firing-rate dependent mechanisms underlying connection
motifs formation, as first described in (30). In the microcircuits considered here, neuronal activ-
ity is known to depend on the connectivity and the short-term changes in PSPs amplitude (31).
Indeed, during the 2000 simulations of Fig. 2, the microcircuits including short-term depressing
synapses collectively fired at 20 ± 0 Hz (see also Fig. 1H), lower than the recurrent microcir-
cuits employing short-term facilitating synapses, which fired at 59 ± 4 Hz (see also Fig. 1E).
Facilitating synapses charge up with ongoing presynaptic activity and they recruit more connected
neurons, as in a positive feedback. On the contrary, depressing synapses get soon fatigued, re-
sulting into a weaker subsequent recruitment of postsynaptic neurons, whose firing would further
depress synaptic efficacy, as in a negative feedback.
We identify such an asymmetry as the cause underlying motifs emergence, and the firing-rate
dependence of STDP (24) as its mechanism. In particular, it is the functional switch of STDP from
a correlational "pre-post" temporal mode at low firing rates, to a "Hebbian" rate mode at high
firing rates, that entirely accounts for the asymmetry in the connectivity motifs. In addition to the
spike-timing information (Fig. 2A), the STDP model employed here faithfully captures the strong
dependency on the neuronal firing rates (38) (Fig. 2B). For the sake of illustration, we isolate the
impact on synaptic efficacy of Eqs. 6,7,8, in a two-neuron system, with one neuron projecting to the
other via a single synapse. We simulate the long-term change in PSPs amplitude at that synapse
(see Fig. 2B and (35)), imposing 75 pre-post spike pairing events, evoked at different uniform firing
frequencies. We study two cases: (i) each presynaptic spike precedes the postsynaptic spike by
10 msec, i.e., tpre < tpost, or (ii) vice versa, i.e., tpre > tpost. In agreement with the experiments
(38), above 30 − 40 Hz long-term potentiation (LTP) prevails on long-term depression (LTD), even
in those cases when spike-timing per se would promote LTD, i.e., tpre > tpost. Below that critical
frequency, LTP or LTD reflects causal or anti-causal relationships between pre- and post-synaptic
firing times, respectively (24).
It follows that facilitating microcircuits would preferentially develop reciprocal connectivity
motifs, irrespectively of the spike-timing, due to their higher firing rates that promote LTP. The
Connectivity Motifs by Synaptic Plasticity
14
consequent increased recurrence in the connections would further increase the overall population
firing rate, above the critical frequency. This increment in the firing rates leads at first place to
LTP. Depressing microcircuits instead do not develop preferentially reciprocal motifs, but reflect
the asymmetric temporal structure of the input. On a first approximation, their connections
would rather not promote persistent high activity regimes; instead they mostly respond to external
stimuli. These stimuli are imposed here at low rates and give rise to feed-forward links, due to the
repeating traveling-wave features of the stimulation, as in (30).
These specific connectivity and activity configurations are stable when imposing reflecting
boundaries, as in the numerical implementation of STDP (52; 51; 53). These boundaries limit
the synaptic efficacy to a minimal and a maximal values (see Methods). We also remark that the
firing rate dependence of STDP, shown in Fig. 2B, is not captured by all the models proposed in
the literature. In our case, the minimal set of STDP model features sufficient for motifs emergence
must include the reversal of LTD into LTP at high firing rates.
As a conclusive illustration of the simple mechanism for motifs emergence, we performed addi-
tional negative and positive control simulations (Fig. 3), studying the impact of alternative formu-
lations of STDP. By setting A+
3 = A−
3 = 0, and A+
2 = 4.5 10−3 in Eq. 8 and slightly modifying
Eq. 7 (see the Supplemental Information) one can reproduce the pair-based STDP model (52;
54). When parameters are adjusted so that this model has an identical spike-timing dependency
(i.e., compare Fig. 3A to Fig. 2A), the resulting frequency dependency is completely different (i.e.,
compare Fig. 3C to Fig. 2B), showing no reversal of LTD into LTP at high frequencies. Vice
versa, by inverting signs and swapping the values A+
2 , A+
3 , A−
2 , and A−
3 in the same equations,
it is possible to ad-hoc reverse the temporal dependency of STDP, as described experimentally
for anti-STDP (aSTDP) (55), while leaving the frequency dependence roughly intact (i.e., com-
pare Fig. 3B,D to Fig. 2A,B) and featuring the reversal of LTD into LTP at high frequencies. Of
course, this modified "triplet rule" model should not be generalized as novel accurate description
of aSTDP, since more data would be anyway needed to access its firing rate dependence. Panels
E-H of Fig. 3 repeat the simulations of Fig. 2, and demonstrate that only when the frequency-
dependence of STDP is realistic (38), the heterogeneity in the network firing rates leads to the
emergence of asymmetric connectivity motifs (compare Fig. 3E,G or Fig. 3F,H to Fig. 2C,D).
Connectivity Motifs by Synaptic Plasticity
15
Large homogeneous microcircuits
Due to the simplicity of the mechanism disclosed above, we can analyse larger populations of
neurons interacting via homogeneous and heterogeneous short-term plastic synaptic connections.
Prior to running numerical simulations, we qualitatively investigate the generality of our results
and the conditions for the asymmetry in the emerging firing rates. We carry out this analysis
by a standard Wilson-Cowan firing rate description of neuronal networks dynamics (see Eq. 10)
(47). This approach has been already extended to the case of short-term synaptic depression
and facilitation (see Eq. 11) (33; 56; 36), long-term Hebbian plasticity (57), or both (58). This
technique can be used as long as the characteristic times of long-lasting plasticity are much longer
than those of neural and synaptic short-term dynamics, as in our case.
For the sake of simplicity, we initially omit STDP (Eq. 11), and consider recurrent networks
of neurons connected by homogeneous synapses, facilitating (Fig. 4A) or depressing (Fig. 4B). We
sketch these networks as oriented graphs, where arrows represent connections with average efficacy
JF F or JDD, correspondingly. We do not explicitly include inhibition, which, nevertheless, does
not qualitatively alter the validity of our results, as examined in the Supplemental Information.
Instead, we distinguish whether the average external input to a generic neuron is zero, i.e., referred
to as balanced inputs (Iext = 0 in Eq. 10), or is set to a positive value, i.e., referred to as
unbalanced inputs, (Iext = 5 in Eq. 10). The population firing rates can be studied via standard
methods as in dynamical system theory (59), i.e., analyzing the system of Eq. 10 and evaluating
their steady-state solutions.
For instance, given a specific combination of external inputs and recurrent average synaptic
efficacies, i.e., JF F = JDD = 4 for unbalanced inputs and JF F = JDD = 10 for balanced
inputs, the steady-state solutions of Eq. 10 provides the equilibrium firing rates. These can be
then compared to the critical frequency of STDP (see Fig. 2B), invoking the same mechanisms for
motifs emergence proposed by Clopath et al., (2010):
if above the critical frequency, then LTP
prevails and STDP will reinforce reciprocal synaptic efficacy; if below the critical frequency, then
LTP or LTD will be determined by spike-timing information. Panels C and D in Fig. 4 illustrate
graphically in the plane E, h, the actual location of the steady-state solutions of Eq. 10. These
are derived by means of geometrical analysis techniques, similarly to the nullclines intersection
methods (59). In fact, at the equilibrium Eq. 10 can be rearranged so that its roots correspond
to the intersections between the unitary slope line (dashed) and a given curve (continuous lines;
Connectivity Motifs by Synaptic Plasticity
16
see Supplemental Information). The shape and bending of such a curve depend on the external
input Iext and on the SD parameters, so that the firing rates emerging in the recurrent networks
of Fig. 4A-B can be compared to each other (33; 56; 36). The dynamical stability of each steady-
state solution is displayed by a different marker symbols: circles for stable, squares for unstable
equilibrium points. Similarly to Fig. 2B, the approximate location of the STDP critical frequency
has been indicated by a gray shading. Although we did not chose the parameters of the rate models
to quantitatively match the Integrate-and-Fire simulations (40; 57), we conclude that homogeneous
facilitating networks generally fire at higher firing rates than depressing networks, for the same set
of average synaptic coupling and external inputs, and when engaged in reverberating activity (40).
Mathematically the two networks share the same mean-field description, Eqs. 10-11, although
their numerical parameters are considerably different (see Table 1) (19; 29). Specifically, short-
term depressing synapses have a much larger recovery time constant from depression (τrec) than
facilitating synapses. Facilitating synapses have instead a much larger recovery time constant from
facilitation (τf acil) than depressing synapses.
Along these lines and by studying the asymptotic limits of the mean-field equations, we found
heuristically that the dominating parameter is τ −1
rec. This sets an upper limit to the location of any
possible firing rate of each network. Then, a network with low values of τ −1
rec, i.e., where the time
scale of recovery from depression is very long, would fire slower than a network with comparatively
higher values of τ −1
rec, i.e., where the time scale of recovery from depression is very fast or negligible.
These two cases correspond to the depressing and facilitating networks, respectively (Figs. 4A-B).
These general considerations have been validated and confirmed in networks of 1000 neurons,
see Methods. In these simulations, neurons are connected with maximal synaptic efficacy A = 6 pA.
In order to increase the biological realism, in these large scale simulations we introduce fluctuating
random inputs to each neuron, mimicking background network activity (42). Therefore, each neu-
ron receives an uncorrelated noisy current, as well as a periodic wave-like stimuli (see Methods).
Figs. 5A-B display the count of the occurrence of unidirectional versus reciprocal connectivity
motifs. Similar to the small toy model, analysed in Fig. 2, unidirectional depressing connec-
tions significantly outnumber the reciprocal depressing connections, while facilitating reciprocal
connections prevail on unidirectional facilitating connections. As indicated by Figs. 5D-E, the
distributions of the firing rates of the two networks feature the same heterogeneous distributions
of firing rates: networks of homogeneous depressing short-term plastic synapses fire at generally
Connectivity Motifs by Synaptic Plasticity
17
low rates, while networks of homogeneous facilitating synapses fire at higher rates. Finally, the
symmetry index s, computed after a very long simulation run, results in a value of 0.28 for the
depressing network and of 0.99 for the facilitatory network.
Heterogeneous microcircuits
We further study the more general case of a heterogeneous network (Fig. 6A) with two subpopu-
lations: one with facilitating synapses, and the other with depressing synapses. Arrows represent
synaptic connections, with average efficacies JF F , JDD, JF D, and JDF . Synapses established
within neuronal pairs that belong to the same subpopulation, share, by definition, the same SD
properties, i.e., short-term depressing or short-term facilitating, but not both simultaneously. On
the contrary, synapses established within neuronal pairs that belong to distinct subpopulations
have, by definition, heterogeneous SD properties. Thus a total of five categories of connectivity
motifs are possible in this network:
facilitatory reciprocal motifs, depressing reciprocal motifs,
facilitatory unidirectional motifs, depressing unidirectional motifs, and reciprocal motifs with both
facilitation and depression. For the first four categories, experiments support strong non-random
occurrences (19; 29). For the case of reciprocal motifs with both facilitation and depression, no
extensive experimental information has been published. Our results suggest that non-random oc-
currences of the first four categories arises from SD-STDP interactions, and predict that the last
category should be largely underexpressed, compared to chance level.
We first examine the impact of STDP in a simplified two-neuron system, with one neuron
projecting to the other via a single synapse. Figs. 6B,D, show the long-term change in PSPs
amplitude as a function of the pre- and postsynaptic firing rates, at that single synapse. Since in
a heterogeneous network pre- and postsynaptic firing frequencies may differ, we swept the firing
frequencies of the two neurons throughout all the possible combinations, within a realistic range.
We study two cases: each presynaptic spike precedes the postsynaptic spike by 10 msec, i.e., tpre <
tpost, or vice versa, i.e., tpre > tpost. We emphasise that only synapses established within neuronal
pairs that belong to distinct subpopulations can experience heterogeneous pre- and postsynaptic
firing rates. In this case however, the impact of spike timing information becomes negligible as
soon as pre- and postsynaptic neurons fire at different frequencies. In the small minority of cases
where this is not true, pre- and postsynaptic frequencies are integer (sub)multiples of each other,
and a transient synchronization of spike times occurs periodically.
In these circumstances, the
Connectivity Motifs by Synaptic Plasticity
18
timing information has a specific impact, as revealed graphically by bright or dark lines in the
plots of Figs. 6B,D. In all other cases, overall plasticity profiles reflect the conventional associative
Hebbian LTP/LTD and its consequences (60; 61; 57).
Intuitively, the heterogeneous population of Fig. 6A can lead to the emergence of connectivity
motifs. To illustrate this point, we first ignore that subpopulations might interfere with each other's
firing rate. We assume that the facilitating and depressing subnetworks would be still characterized
by higher or lower firing rates, respectively, as previously presented for homogeneous networks.
Then, the LTP/LTD maps of Figs. 6B,D suggest that such an initial asymmetry of emerging
firing rates can be maintained indefinitely. The connections JDF are increasingly weakened and
the connections JF D strengthen. The resulting configuration, sketched in Fig. 6C, is stable. We
tested and confirmed this statement under the mean-field hypothesis, by studying the dynamics of
Eqs. 10, 11, and 12, in addition to computing their equilibria. Fig. 6E-F display the mean firing
rates of each subnetwork, and the initial time course of synaptic efficacies, with JDF progressively
becoming weaker. This is true only when inter-population efficacies, i.e., JF D and JDF , are
initialized to slightly weaker values than the intra-population efficacies, i.e., JF F and JDD. Such
an initial difference between inter-population and intra-population couplings could however emerge
from fully homogeneous couplings, i.e., JF F = JDD = JDF = JF D = 1, as demonstrated in
the Supplemental Information.
We further confirmed that the synaptic configuration, indicated by the mean field models, is
present in large microscopic numerical simulations, as in Fig. 5C (see Methods). These simulations
involve 1000 identical Integrate-and-Fire units, subdivided in two subpopulations of equal size, with
their anatomical connectivity set to 80% of all possible connections, similar to the homogenous
case. The maximum synaptic efficacy is set to A = 12 pA and the initial synaptic connections Wij
are randomly initialized. As in the mean-field model, the inter-population terms Wij are initialized
to weaker values than the intra-population terms (but see Supplemental Information for alterna-
tives). Each neuron receives an uncorrelated background noisy current as well as periodic wave-like
stimuli, similar to the homogeneous case. As indicated by Fig. 5F, the distributions of the spiking
frequencies of individual neurons of the two subnetworks feature the same heterogeneous distribu-
tion of firing rates as in the previously studied homogeneous networks: subnetworks of depressing
short-term plastic synapses fire at generally low rates, while subnetworks of facilitating synapses
fire at higher rates. The location of the critical firing frequency for the STDP is represented again
Connectivity Motifs by Synaptic Plasticity
19
as a grey shaded area.
Results in Fig. 5C show all the possible synaptic combinations. As in the data of Wang et
al. (2006), reciprocal motifs are significantly co-expressed with facilitatory synapses and unidirec-
tional motifs with depressing synapses. The actual motif count is compared to the null-hypothesis
of having statistical independence between the connection occurrence within a pair of neurons,
estimated at a 95% confidence interval upon the same hypothesis of Bernoulli repeated, inde-
pendent, elementary events. The frequency Q of observing a connection between two neurons,
regardless of its SD properties, is first estimated by direct inspection of the connectivity matrix
[Wij ]. Then the conditional occurrence frequencies of a facilitatory synapse QF and of a depress-
ing synapse QD = (1 − QF ) are computed, given that a connection exist between two neurons.
The null hypothesis for each possible combination is given by standard probability calculus, under
the hypothesis of independence of the identical events. For instance, the occurrence frequency
of observing by chance no connections within a neuronal pair is (1 − Q)2, while the occurrence
of observing by chance a reciprocal motifs with mixed depressing and facilitating properties is
2 (Q2 QF QD). Finally, in this example, the symmetry index s, computed after a very long simu-
lation run, resulted in a value of 0.18 for the depressing subnetwork and of 0.66 for the facilitation
subnetwork.
Microcircuits with overlapping SD properties
In the heterogeneous network of Fig. 6, as well as in the homogeneous networks, we make the
assumption that the SD profile is determined primarily by the identity of the projecting neuron.
This has been experimentally found in the olfactory, visual, and somatosensory cortices as well as
in other brain areas (62; 63; 64; 65). Nonetheless, the assumption on the projection-cell specificity
can be removed in order to theoretically explore the impact of SD heterogeneity across distinct
synaptic connections established by the same presynaptic neuron (66).
We assumed that a generic neuron had a certain probability pD of establishing a short-term
depressing synapse with a target neuron, and a probability 1 − pD of establishing a short-term
facilitating synapse with another one. In this case, individual neurons are still indistinguishable
and their mean-field synaptic input can be described mathematically Fig. 7A (see Supplemental
Information).
For small values of pD, the emerging firing rates approximate those of a network of facilitating
Connectivity Motifs by Synaptic Plasticity
20
synapses, while for large values of pD the firing rates behave as for a network of depressing synapses.
In other words, the mixed networks behave dynamically as an intermediate case between two
extremes. This result is quantified in figure Fig. 7B, where the location of the stable equilibrium
points has been analysed under the mean-field hypotheses and plotted as a function of pD, for
different external inputs regimes. The qualitative location of the critical firing frequency for the
STDP is represented as a grey shaded area. In this situation, and as an explicit consequence of
the lack of any structure, i.e., compare Fig. 7A with Fig. 6A, STDP fails to discriminate individual
connections within the network, but rather shapes them as reciprocal or as unidirectional motifs,
depending on the particular choice of pD.
Connectivity Motifs by Synaptic Plasticity
21
Discussion
Our results indicate that time- and frequency-dependent STDP mechanisms may be responsible,
through internally generated spiking activity in recurrent network architectures, for the observation
that excitatory neurons connected by short-term facilitating synapses are more likely to form
reciprocal connections, while neurons connected by short-term depressing synapses are more likely
to form unidirectional connections. More specifically:
1. the internally generated firing rates in model networks with facilitating connections are higher
than in networks with depressing connections, under identical background inputs;
2. neurons, participating to such an internally generated activity, are likely to form bidirectional
connections with each others when firing at sufficiently high rates, reflecting the "Hebbian"
mode of STDP; neurons firing at low rates are likely to form unidirectional connections,
reflecting the temporally asymmetric "pre-post" temporal mode of STDP;
3. once formed, these connectivity motifs persist and self-sustain themselves through the internal
firing regimes of the network, from which the motifs emerged from;
4. externally generated inputs, strongly depolarizing or strongly hyperpolarizing individual neu-
rons, when prevailing over internally generated activity, may lead to opposite motifs emer-
gence consistent with Clopath et al. (2010).
Relationships to previous work and additional mechanisms
The impact of long-term associative synaptic plasticity in recurrent networks of spiking neurons
has been earlier studied by many investigators, who proposed mechanisms for the unsupervised
formation of stimulus-driven dynamical attractors of network activity, in the context of working-
memory states (57). Within the same aims, the interactions between long-term plasticity and SD
were also already partly explored, both in numerical simulations and in mean-field descriptions
(58). Here, we focused on specific long-term plasticity mechanisms (STDP) (24) to study the
emergence of network structure (67; 68; 69; 70) and connectivity motifs (30; 71). To the best of
our knowledge, this is the first time that SD is viewed as key element for the emergence of network
connectivity, and shown to capture, to a certain extent, the experimental observations.
Our study builds upon assumptions of Clopath et al.
(2010):
in silico activity-dependent
(dis)appearance of a connection occurs in terms of a competition among "strong" links in a "sea"
Connectivity Motifs by Synaptic Plasticity
22
of weak synapses, which are undetectable by cellular electrophysiology. There is still no general
agreement on whether STDP and its variants can entirely account for developmental wiring, fine-
tuning, or remapping of microcircuits connections (72). However, the role of STDP for connectivity
development has been positively discussed for systems where information is present at fast time
scales (73), as it might be relevant for the neural code of cortex and olfactory microcircuitry. In
addition, the STDP "triplet" rule proposed by Pfister and Gerstner (35), which is the central
ingredient for this work, captures the behaviour of developmental plasticity models (37) widely
used in classic studies of connectivity development (74). As a consequence, our results are built
upon well-known phenomenological models, which have been shown to have excellent agreement
with experimental data sets (21; 35; 30). We have, therefore, neglected more accurate biophysical
descriptions of SD (75; 76; 77; 78) and STDP (79; 80), aiming at a minimalist description of neu-
ronal excitability, synaptic transmission, and synaptic plasticity. The advantage of our approach
is that the functional consequences of the interactions between SD and STDP could be analysed
by standard mean-field analysis (34; 36).
Our study concludes that the symmetry of the connections may be influenced by the internally
generated spatio-temporal correlations of neuronal firing (81). Therefore, connectivity motifs might
not emerge exclusively from correlations governed by the external inputs (30). Both internally
generated and externally imposed correlations are likely to play a role and therefore the results of
Clopath et al. (2010) still hold. For instance, in a microcircuit connected by purely depressing
synapses, external activity can still induce heterogeneity in the firing rates. If a subset of neurons
was strongly depolarized by external selective inputs, then units would fire above the critical
frequency, and non-random bidirectional connectivity might equally emerge (see Fig. 8). Similarly,
strong hyper polarization by external selective inputs, would cause neurons to fire below the critical
frequency, even though neurons are connected by recurrent facilitating synapses (not shown)..
Out of many possible features affecting internally generated activity, such as cellular excitability,
architecture of long-range connections, and SD properties, here we focused on the last one. We note
however that the dataset of Wang et al. (2006) contains additional information on cell excitability
and its correlation to motif emergence. Cells displaying accommodating discharge patterns (i.e., ex-
hibit spike frequency adaptation) are found to establish mostly unidirectional motifs. Instead, cells
displaying nonaccommodating discharge patterns establish mostly reciprocal motifs (19). A differ-
ential expression of spike-frequency adaptation currents results in (non)accommodating discharge
Connectivity Motifs by Synaptic Plasticity
23
patterns. This is known to affect the firing rate distributions in recurrent model networks (82; 83;
39), and generally lower the location of steady-states stable equilibria (Figs. 4C-E). For the model
parameters employed here, we did not observe differences when repeating all our simulations with-
out spike-frequency adaptation. Adaptation currents in the model are not strong enough to reduce
substantially the steady-state firing rates of e.g., the facilitatory subnetworks below the critical
firing rate for STDP, and violate our conclusions. On the contrary, as spike-frequency adaption is
reported by Wang et al. in neurons participating to unidirectional motifs, we speculate that it par-
ticipate jointly with depressing SD in maintaining STDP in its correlational "pre-post" mode: once
more this determines the emergence of unidirectional connectivity motifs. Non-accommodating dis-
charge patterns would by definition not interfere with the output firing rates, in networks connected
by facilitating synapses and bound to emerge reciprocal motifs, by the "Hebbian" mode of STDP.
Our study addresses the emergence of motifs in the cortical pyramidal microcircuitry, but as
preliminary data collected in the olfactory bulb become available, our framework could tentatively
model a common principle for synaptic wiring. Long-term and short-term plasticity has been ex-
perimentally found among olfactory mitral cells, (84; 29), and STDP was reported in the rodent
and insect olfactory systems (85; 86), and invoked at the output of mitral cells to explain decor-
relation of sensory information (87). We, therefore, speculate that similar mechanisms for the
emergence of connectivity motifs are common in both the cortical and olfactory microcircuitry.
Our investigation of heterogeneous network architectures revealed that STDP and SD led to
a stable configuration of developing connections, compatible to the experimental observations.
Under the simplified hypotheses that we adopted, the configurations of connectivity motifs and
their relationship to the underlying neuronal collective dynamics are stable. Our results have
been grounded on both numerical simulation results and mean-field analysis. The latter, however,
ignored spike-timing correlations between spike-trains. It could thereby only assess the conditions
for the emergence of strong connections. While this is sufficient to anticipate the emergence of
reciprocal connectivity motifs, its converse does not per se implies the emergence of unidirectional
motifs. In such a case, connections would for instance tend to weaken or to disappear. It is only
through the numerical simulations of the full network of spiking neurons that we could show that
the causal spike-timing information, if present, gives rise to unidirectional connectivity via STDP
mechanisms.
We also note that the major difference in SD properties, which accounts for motifs emergence,
Connectivity Motifs by Synaptic Plasticity
24
is the heterogeneity of the time constant representing the short-term depression recovery τrec. In
this respect, our results and conclusions would be qualitatively unchanged by replacing facilitating
synaptic properties with linear, i.e., non-depressing, properties. Along these lines, we hope that
our results might prompt new experiments on connectivity motifs. We predict that the value of
τrec, in a pair of connected neurons, should be inversely correlated to the occurrence frequency of
reciprocal motifs.
Simplifying assumptions
Among the key simplifying hypotheses of this paper, we note that STDP was assumed to scale
only the SD parameter G, while leaving the parameter U unaltered (88). This choice is consistent
to what has been reported at distinct synapses of the central nervous system, although it might
not be representative of all cortical areas (50). Although the debate on pre- and postsynaptic
expression of both SD and STDP is fierce, our choice of SD and STDP interaction is in part an
arbitrary hypothesis, and it serves as a first solid ground for our conclusions. Enabling STDP
to modify the parameter U would have partly altered in an activity-dependent manner, the SD
profile of a synapse. This would have made isolating SD contribution more complex, and relating
our findings to previous theoretical works (67; 68; 69; 30; 70) less straightforward.
The model itself is purely phenomenological, and does not capture biophysical details, but
rather the interaction of SD and STDP via a set of variables locally known to the synapse. It
does however maintains the desirable compatibility with experimental data. In addition, explicit
and systematic details on STDP at excitatory facilitatory synapses in the ferret prefrontal cortex
are currently scarce (19). While more efforts, both experimental and theoretical, should be un-
doubtedly devoted in these directions, the hypothesis of scaling G and not U remains a simplifying
assumption, in the view of lacking of a systematic understanding on how STDP affects all the
parameters of the SD model (89).
The presence of two classes of excitatory cortical synapses (19), with a distinct set of SD param-
eters such as τrec, τf acil, and U , prompted us to consider in our model the existence of two classes
of excitatory neurons. The consequence of having these classes within the same microcircuitry, as
well as their postsynaptic target preferences for connection, are crucial issues that deserve more
experimental and theoretical efforts, but see however (36; 90). We thus explicitly neglected hetero-
geneity in the target synaptic preference of neurons, when expressing a facilitating or depressing SD
Connectivity Motifs by Synaptic Plasticity
25
profile (66). We instead considered homogeneous and heterogeneous neuronal populations, where
the facilitating or depressing SD profile was determined by the projecting neuron. Although such
a scenario for SD seems realistic for synapses between principal neurons, it might not be accurate
for all the synapses in neocortical microcircuits.
The mixed microcircuitry of Fig. 7 and its preliminary analysis are an example of our attempt to
relax these assumptions. Depending on the probability of establishing a facilitating (or depressing)
connection, fully mixed networks were shown to behave as a continuum between the two extremes
of the homogeneous populations studied earlier, i.e., networks of all facilitating or all depressing
connections. These mean-field predictions were further confirmed in numerical simulations of
large networks (not shown). In these cases, SD-STDP interactions will not necessarily lead to the
desirable heterogeneity for the connectivity motifs, as the heterogeneity in the emerging firing rates
does not occur. Structured local microcircuit architectures, with heterogeneous single-cell type,
number of afferents, and firing rates are the obvious topics for a future study. We believe that
the simple case of the heterogeneous two-subpopulation network studied in this work was a first
necessary step towards the highlighted future directions.
Another point that deserves some discussion is our choice for a structured, foreground periodic
stimulation, employed in all simulations with more or less intense asynchronous background activ-
ity. In the toy microcircuit (Fig. 1), the cyclic stimulus was needed to demonstrate the emergence
of a large number of asymmetric connections. It is, however, neither a strict requirement for our
interpretation of reciprocal and unidirectional motifs formation, nor a necessary condition for their
unbalance. This cyclic, deterministic, stimulus is an oversimplified way to invoke the temporal
coding strategy of (30). It was chosen to unambiguously relate unidirectional pairwise motifs for-
mation to spiking activity. Without such an input, and under the presence of strong uncorrelated
background activity, unidirectional motifs would tend to be very sparse. We underline that our
intention was to expose to the very same conditions networks identical in every aspects but in the
synaptic SD, and validate the heterogeneity conditions of the connectivity matrix.
Limits of our approach
Our proposed mechanism for non-random pattern emergence is based on the sole interactions be-
tween STDP and SD. Obviously, it is unlikely that these mechanisms operate independently of
other synaptic phenomena. Homeostatic plasticity could for instance continuously rescale synap-
Connectivity Motifs by Synaptic Plasticity
26
tic efficacy and make SD heterogeneities less predominant in determining connectivity motifs. In
the lack of a priori experimental information, we chose the maximal synaptic efficacy A to share
the exact same value for all the microcircuits we examined, in order to ensure a fair comparison.
With all its limitations, our proposal may still provide a simple working hypothesis on one com-
ponent underlying the emergence of connectivity, linked to short-term synaptic dynamics, along
the same lines of the theory proposed by Clopath et al. (2010). Its validity could be challenged
by experiments that interfere and probe the emergent firing activity, e.g., in local in vitro cultured
microcircuits with known synaptic properties (91). Our framework might be also useful for inves-
tigating further structure-function relationships at the subcellular level, by altering the synaptic
machinery, or by employing (future) genetically-encoded fluorescent reporters of synaptic efficacy
and dynamics. The use of optogenetics and genetically encoded neuronal voltage- and calcium-
sensors, may lead to experimental validation or falsification of our hypothesis, which might directly
contribute to understand short- and long-term plasticity interactions.
Further, the results of our simulations do not automatically capture the over-expression of
reciprocal connectivity cortical motifs over the unidirectional ones (7). This is violated in the case
of short-term depressing synapses, where the expression of reciprocal motifs is clearly at chance
level Fig. 5A, but it is however not the case for Fig. 5B. Clopath et al. (2010) model these aspects
by employing different neuronal coding strategies, i.e., time- or rate-codes. In our work, in order
to isolate the heterogeneity component of SD properties, we considered a common type of spatio-
temporal activation and ignored, for the sake of simplicity, heterogeneous neural coding strategies.
We are however confident that including more complex network architectures, as well as more
specific spatio-temporal correlations, as in Clopath et al. (2010), additional non-random features
could be explained.
In our results, we use oversimplified network architectures to illustrate that two recurrent
excitatory networks, identical in any aspect apart of their synaptic properties, evolve different
connectivity motifs. We simulated and studied these models analytically, and we identified the
mechanism behind the heterogeneity as frequency-related. It would not be an obvious easy task to
explain simultaneously all the properties of cortical connectivity, as experimentally observed (19).
The choice of our homogeneous architecture, the lack of cell diversity, and the isolated evolution of
neuronal dynamics, considerably differ from the actual cortex as well as from the variety of internal
and external stimuli it receives during development and during lifetime. Nevertheless, the choice
Connectivity Motifs by Synaptic Plasticity
27
of our protocol serves in demonstrating a potential mechanisms behind the motif formation, for
facilitating and depressing synapses.
We would like to emphasise that our theory refers only to one of many possible, perhaps com-
peting, mechanisms that contribute to stereotypical motifs emergence. Alternative explanations
and a causal demonstration of the key ideas we suggest, remain to be provided. It might be of
interest exploring to which extent developmental changes in SD, such as the switch from depression
into facilitation at synapses between layer 5 pyramidal neocortical neurons (64), occurring after
postnatal day (P) 22, are mirrored by changes in motifs statistics. For marginal pair-wise proba-
bility of connection, Song et al. (2005)(7) report no significant dependence on age, but provide no
systematic characterization of motifs statistics beyond P20.
It may be also possible to attempt a chronic manipulation of the firing rates of neuron (sub)populations,
by pharmacologically altering synaptic profiles, e.g., modulating postsynaptic receptor desensitisa-
tion, changing the presynaptic probability release, or interfering with neurotransmitter recycling.
As future directions, more complex heterogeneous anatomical architectures and single-cell prop-
erties should be incorporated within the same computational modeling framework. Very specific,
non-random initial architectures, e.g., small-world and scale-free (92), could be explored, extend-
ing our results towards other aspects that determine reciprocal or unidirectional motifs, possibly
beyond the firing levels and towards, e.g., , the density of hub nodes, ranking orders, heavy tails
in neighbours distribution, etc.
How critical is the STDP "triplet" rule for the validity of our results? Any STDP model that is
able to capture the high frequency effect on plasticity, as revealed by the experiments of Sjostrom
et al. (2001), will reproduce our results if combined with short-term facilitating and depressing
synapses. Not all STDP models are consistent with the data of Sjostrom . For instance, Froemke
and Dan (93) proposed an STDP model capturing pre- and postsynaptic frequency effects. This
model predicts that in high frequencies no reversal of LTD into LTP takes place, an inversion that is
critical for the mechanism we propose for motifs emergence. Interestingly, however, a generalization
of the "triplet" rule that we adopt here is able to capture the experimental data of Froemke and
Dan (2002) (94). We also note that pair-based STDP models may still support the emergence
of reciprocal connectivity motifs and replicate our results (not shown). In fact, by adjusting the
numerical values of its parameters (e.g., A+
2 = 2A−
2 in Eqs. S39 and S42, see Supplemental
Information), LTP can be made prevailing over LTD above a frequency similar to the critical value
Connectivity Motifs by Synaptic Plasticity
28
of the triplet-model (72; 95). Nevertheless, when the pair-based STDP parameters are chosen to
match the 10Hz temporal window of the triplet-STDP model employed here (compare Figs. 2A
and 3A), the frequency-dependent profile observed by Sjostrom et al. (2001) is not reproduced
(compare Figs. 2B and 3C) and SD-STDP interactions do not lead to the same results (compare
Figs. 2C,D and 3E,G). This served here as a negative controlsk for our results. We also remark that
excluding heterogeneity in spike propagation delays made our analysis simpler. It is known that
the distribution of delays considerably affect emerging network states, besides having an obvious
effect on STDP (96). In this respect, not only STDP and SD interactions might be altered by
propagation delays, but the collective dynamical properties of recurrent networks could also vary.
Once more, for the sake of simplicity we chose to consider a minimalist scenario, setting the ground
for future, more extended, investigations.
Finally, we underline the great value of the availability of physiological information accom-
panying anatomical connectivity. These complementary data-set contains precious statistical in-
formation regarding the expression of microcircuit motifs, which we are starting to recognize (7;
9). We believe that computational modeling is in this context a very powerful tool to explore
additional hypotheses and challenge further theories.
Acknowledgments
We are grateful to C. Clopath, J.-P. Pfister, M. Pignatelli, A. Carleton, P. Dayan, A. Loebel for
discussions and comments.
A Supplementary Text
For the sake of illustration of the standard mathematical techniques employed in the main text,
we consider two extreme simplifications of the mean-field description of a recurrent networks with
short-term plastic synapses (Eqs. 9-10; see also (36), and references therein): the predominance
of depressing mechanisms or the predominance of facilitatory mechanisms. We analyze these two
cases and specifically study the stability of dynamical equilibria. These simplified models and their
analysis have been already provided in the literature, with varying degree of details. We further
comment on the analysis of the full model of short-term synaptic plasticity and comment on the
impact of adding recurrent inhibition to our scenario. We then consider and numerically analyse the
Connectivity Motifs by Synaptic Plasticity
29
mean-field description of a mixed population, where a generic neuron may establish simultaneously
depressing and facilitating synapses to its targets. We finally examine and partly relax the necessary
conditions for a heterogeneous network of two interacting subpopulations to display emergence of
connectivity motifs. For the case of a random matrix (i.e., as a null hypothesis), we provide some of
the statistical properties of the symmetry measure we adopted to quantify the connectivity motifs
to derive a confidence measure. In the last section, we provide model details for the pair-based,
STDP as well as for the anti-STDP "triplet" model, discussed in the main text.
A.1 A single population with depressing synapses
The Eqs. 9-10 of the main text can be simplified under the hypothesis of very fast recovery from
facilitation. This is the same of assuming that τf acil is very small and that u = U does not vary in
time, as a consequence. In this case, only short-term depression is modifying the synaptic efficacy,
so that the mean-field firing rate dynamics of a large recurrent network of excitatory neurons, can
be described by a system of two dynamical equations (33):
τ h = −h + A U x E + Iext
x = (1 − x) /τrec − U x E
(S13)
These are coupled non-linear differential equations that can be analyzed by standard methods
of dynamical systems theory (59). We will consider here the derivation of the equilibria for h and
x, and indicate how their stability was assessed. By definition of equilibrium, h = 0 and x = 0
for those points (also called fixed points). Substituting these conditions in Eqs. S13, we obtain two
non-linear algebraic equations,
h = A U x E + Iext
x = 1 / (1 + U τrec E)
(S14)
Using the second equation to replace x as it appears in the first, we get an implicit equation in
the unknown h:
Connectivity Motifs by Synaptic Plasticity
30
h = A U E / (1 + U τrec E) + Iext
(S15)
Eq. S15 can be solved numerically (e.g., by the Newton-Raphson method, (97)), given a specific
set of parameters A, U , τrec, and Iext. Alternatively, its solution(s) can be interpreted graphically
as the intersection(s) between two functions of h, F1(h) and F2(h), in the cartesian plane,
F1(h) = h
F2(h) = A U E / (1 + U τrec E) + Iext
(S16)
with F1(h) is the unitary slope line (see Fig. S1). Retaining the graphical interpretation, it is
possible to appreciate intuitively the existence of the equilibrium points and their dependence
on the parameters, as explored in Fig. S1. To this aim it is useful, prior to plotting F2(h), to
analytically determine some of its mathematical properties, such as the asymptotic limits and
derivatives.
We observe that, by definition of E = [α (h − θ)]+, F2(h) is zero for values of h lower than θ,
h ≤ θ. As h → +∞, F2(h) tends to an horizontal positive asymptote, occurring at (J τ −1
rec + Iext).
For values of h larger than θ, the first derivative of F2(h) is always positive, indicating that the
function is monotonically increasing. In the same range of h, the second derivative is instead always
negative, therefore indicating that the function is convex. Moreover, the tangent line to F2(h) at
h = θ is the steepest of all the tangents to subsequent points (h > θ).
F2(h) = α A U / [1 + α τrec U (h − θ)]2
F2(h) = − 2 α2 τrec A U 2 / [1 + α τrec U (h − θ)]3
(S17)
The value of the slope of the tangent line at F2(h) at h = θ is particularly informative when
drawing F2(h), since at any coordinates h > θ all the other tangent lines are by definition less
steep than it. This maximal slope ( F2(θ) = α A U ) can then be used as a necessary condition
for the intersections between F2(h) and the unitary slope line, at least when Iext ≤ θ.
When Iext ≤ 0, there is always an intersection between F2(h) and F1(h) at h = Iext. This
Connectivity Motifs by Synaptic Plasticity
31
is a stable equilibrium point (i.e., see below for the discussion of the stability). If α A U ≤ 1,
there will be for Iext ≤ 0 no other intersections, since F1(h) = h has unitary slope itself. Given
the necessary condition α A U > 1, there exist a minimal value for those parameters, above
which other two intersections (i.e., one stable and one unstable) with the unitary slope lines occur
(compare Fig. S1A,B, and C, which were obtained for increasing values of A).
Determining analytically such values requires imposing the condition where the unitary slope
line becomes tangent to F2(h). Mathematically this can be expressed by stating that when Iext ≤ θ
there is a specific (intersection) point h0 > θ between F1(h) and F2(h). By definition, this point
lays on the unitary slope line F2(h0) = h0, with F2(h0) = 1, i.e., F2(h) has a unitary first
derivative at h0 (see Fig. S1B).
F2(h0) = α A U (h0 − θ) / (1 + α τrec U (h0 − θ)) + Iext = h0
F2(h0) = α A U / [1 + α τrec U (h0 − θ)]2 = 1
(S18)
Simplifying the algebraic manipulations, we note the apparent similarities between the two
equations above. We express (part of the) numerator and denominator of the right hand sides of
each equations, by denoting the common terms as N0 and D0, respectively, as it follows
F2(h0) = N0 (h0 − θ) /D0 + Iext = h0
F2(h0) = N0 / D2
0 = 1
(S19)
Thus, from the second equation we derive N0 / D0 = D0, and we substitute it in the first,
obtaining
h0 = θ + p(θ − Iext) / (α U τrec)
(S20)
The above expression is of course only defined when the argument of the square root is positive,
which is consistent with our previous hypothesis (i.e., Iext < θ). One can now replace h0 in the
second equation of Eqs. S19 and obtain the corresponding critical value of A, associated to the
existence of such a (double) intersection (Fig. S1B):
Connectivity Motifs by Synaptic Plasticity
A0 = (cid:16)1 + pα U τrec (θ − Iext)(cid:17)2
/ (α U )
32
(S21)
In summary, when Iext ≤ θ there is always one (stable) equilibrium at h = Iext and of possibly
other two intersections (one stable and one unstable; compare Fig. S1A,B, and C), depending on
the strength of A with respect to A0. For Iext > θ, the situations changes and the scenario
simplifies considerably, with only one (stable) intersection for any other choice of the parameters
(see Fig. S1D).
The analysis of the stability of the equilibrium points conclude our discussion. Since the
dynamical system described Eqs. S13 is non-linear, stability of equilibrium points must be related
to the linearized system. The linearization is obtained for each equilibrium point by first-order
Taylor expansion of Eqs. S13 around that point. Let's compactly rewrite Eqs. S13 as
h = G1 (h, x) = (−h + A U x E + Iext) /τ
x = G2 (h, x) = (1 − x) /τrec + U x E
The linearized system, around a generic equilibrium point (h0, x0), is then
(S22)
(S23)
h ≈ G1 (h0, x0) + ∂G1/∂h(h0,x0) (h − h0) + ∂G1/∂x(h0,x0) (x − x0)
x ≈ G2 (h0, x0) + ∂G2/∂h(h0,x0) (h − h0) + ∂G2/∂x(h0,x0) (x − x0)
Assessing the stability of the above system reduces to studying the Jacobian matrix M :
M (h0, x0) =
∂G1/∂h ∂G1/∂x
∂G2/∂h ∂G2/∂x
(h0,x0)
(S24)
By definition, it is possible to make explicit the Jacobian matrix as
Connectivity Motifs by Synaptic Plasticity
M (h0, x0) =
(−1 + α A U x0)/τ
(α A U (h0 − θ))/τ
−α U x0
−τ −1
rec − α U (h0 − θ)
33
(S25)
In particular, the real part of the two eigenvalues associated to M (h0, x0) has been analyzed for
each equilibrium point (h0, x0). The eigenvalues λ1,2 were computed as the roots of the following
algebraic second order equation
det (I − λ M (h0, x0)) = 0
(S26)
where I indicated the 2 × 2 identity matrix and det() indicates the computation of the determinant
of a square matrix. When at least one of the eigenvalue had positive real part, the equilibrium
point was classified as unstable. When both eigenvalues had negative real parts, the equilibrium
point was classified as stable. Nothing can be however concluded on the stability of the non-linear
system, in the cases in which one or both eigenvalues have zero real part (and the other has negative
real part). Stable and unstable equilibrium points have been graphically represented as circles and
squares in Fig. S1, respectively, as well as in Figure 4C-E.
A.2 A single population with non-depressing facilitating synapses
The Eqs. 9-10 of the main text can be again simplified under the hypothesis of very fast recov-
ery from depression τrec. The mean-field firing rate dynamics of a single neuronal population,
recurrently connected by short-term plastic synapses, can be rewritten as:
τ h = −h + A u E + Iext
u = (U − u) /τf acil + U (1 − u) E
(S27)
While this hypothesis is not as realistic as the one of the previous section, it is preparatory for
the analysis of the full model. As for the previous case, we consider the derivation of the equilibrium
points for h and u as well as the assessment of their stability. Substituting the definitions of
equilibrium, h = 0 and u = 0, into Eqs. S27, we get
Connectivity Motifs by Synaptic Plasticity
34
(S28)
h = A u E + Iext
u = U (1 + E τf acil) / (1 + U τf acil E)
Using the second equation to replace u as it appears in the first, one obtains an implicit equation
in h:
h = A U (1 + E τf acil) E / (1 + U τf acil E) + Iext
(S29)
As for Eq. S15, numerical methods can be used for solving Eq. S29, looking for values of h that
satisfy the equivalence given a specific set of parameters A, U , τf acil, and Iext. The solution(s)
of Eq. S29 can be also graphically interpreted as the intersection(s) in the cartesian plane of two
functions of h, F1(h) and F3(h) as defined below
F1(h) = h
F3(h) = A U (1 + E τf acil) E / (1 + U τf acil E) + Iext
(S30)
We observe that by definition of E = [α (h − θ)]+, F3(h) is zero when h ≤ θ. When
h → +∞, the function diverges to infinity, but it can also be approximated by the straight line
F3(h) ≈ α A h. We also note that for values of h larger than θ, the first derivative of F3(h)
is positive, indicating that the function is monotonically increasing. In the same range of h, the
second derivative is also positive, therefore indicating that the function is concave.
F3(h) =
α A U [1 + 2 α τf acil (h − θ) + α2 τ 2
[1 + α τf acil U (h − θ)]2
f acil U (h − θ)2]
F3(h) = 2 α2 τf acil A U (1 − U ) / [1 + α τf acil U (h − θ)]3
(S31)
As opposed to the previous case, the value of the slope of the tangent line at F3(h) at h = θ
is not particularly relevant when drawing F3(h), since at any coordinates h > θ all the other
tangent lines are by definition steeper than it. The minimal slope is
F3(θ) = α A U can then
Connectivity Motifs by Synaptic Plasticity
35
used in combination with the asymptotic approximation F3(h) ≈ α A h (i.e., the maximal slope
is α A). It is then clear that, for 0 ≤ Iext ≤ θ, a sufficient and necessary conditions for having
always two equilibrium points (i.e., one stable at the value h = Iext when 0 ≤ Iext ≤ θ, and
the other unstable) is represented by α A > 1 (Fig. S2A,B). All the considerations on how to
assess the dynamical stability of these equilibrium points hold, and the expression of the Jacobian
matrix M , whose eigenvalues determine the stability, is given below:
M (h0, u0) =
(−1 + α A u0)/τ
(α A U (h0 − θ))/τ
−α U (1 − u0) −τ −1
f acil − α U (h0 − θ)
(S32)
A.3 Single population with short-term plastic synapses
In the general case, the mean-field equations of a single neuronal population, recurrently connected
by short-term plastic synapses, are given by
τ h = −h + A u x E + Iext
x = (1 − x) /τrec − u x E
(S33)
u = (U − u) /τf acil + U (1 − u) E
with the neuronal gain function chosen as a threshold-linear relationship between input (mean)
current h and output firing rate E = [α (h − θ)]+ (see also Eqs. 9-10 of the main text). The
analysis of this system, including its equilibrium points, has been already given elsewhere (36).
Supporting the necessary condition on the symmetry breaking by long-term plasticities mentioned
in the Results section of the main text, here we derive a simple observation on the analytical
properties of these equilibrium points. According to the definition, we substitute h = 0, x = 0,
and u = 0 into Eqs. S33, and get
Connectivity Motifs by Synaptic Plasticity
36
h = A u x E + Iext
x = 1 / (1 + u τrec E)
(S34)
u = U (1 + E τf acil) / (1 + U τf acil E)
By appropriate substitutions of x and of u into the first equation, it is possible to express it as
h = F4(E(h)), an implicit equation in h:
F4(E) =
E−2 + E−1U τf acil + E−1U τrec + U τf acilτrec
AU (cid:0)E−1 + τf acil(cid:1)
+ Iext.
(S35)
We observe that for h → +∞, E(h) → +∞ and F4(E) → A τ −1
rec + Iext, implying the
existence of an horizontal asymptote. This intuitively suggests that for any choice of the other
parameters compatible with the existence of multiple equilibrium points (i.e., intersections between
F4(h) and the unitary slope line), the uppermost equilibrium point (i.e., always stable) will change
its location proportionally to A and to τ −1
rec, for the same choice of Iext. Hence, a high value of
τrec, as in depressing synapses, will give a lower asymptote versus a low value, as in facilitating
synapses.
Let's now consider two independent populations of excitatory neurons, one recurrently con-
nected by short-term depressing synapses (i.e., τrec > τf acil) and one by short-term facilitating
synapses (i.e., τf acil > τrec) and both receiving identical non-zero external inputs Iext. As for
the previous considerations on the horizontal asymptote of F4(E), for an appropriate choice of A
(i.e., large enough to have multiple equilibrium points) or for any value of Iext > θ, the firing
rate uppermost equilibrium point of the facilitating population will always be larger than the firing
rate uppermost equilibrium point of the depressing population. Together with the specific firing
rate dependence of STDP, arising from the triplet-interactions, this consideration rules out that
reciprocal motifs of short-term depressing synapses will outnumber unidirectional motifs of facil-
itating synapses. The stability analysis for the depressing and facilitating populations (with the
parameters used in our simulations) is provided in the main text (see Fig. 3D-F).
Assessing the stability of the above system reduces to linearization of the system around the
fixed points by the use of a Taylor expansion and the study of the so called Jacobian matrix
Connectivity Motifs by Synaptic Plasticity
M (h0, x0, u0) of the system, which for the sake of completeness, we report below:
(−1 + αAu0x0)/τ
(αAu0(h0 − θ))/τ
(αAx0(h0 − θ))/τ
−αu0x0
−τ −1
rec − αu0(h0 − θ)
−αx0(h0 − θ)
αU (1 − u0)
0
−τ −1
f acil − αU (h0 − θ)
37
(S36)
A.4 The impact of recurrent inhibition
We extend the description of the system given in Eqs. S13, to the case where recurrent inhibition
is explicitly accounted for. The mean-field firing rate dynamics of two neuronal populations, one
excitatory and one inhibitory, recurrently connected by short-term excitatory plastic synapses and
by frequency-independent inhibitory synapses (as in Fig. S3), can be rewritten as:
τe he = −he + Aee U x E − Aei I + Iext
τi
hi = −hi + Aie U x E
(S37)
x = (1 − x) /τrec + U x E
with E = [αe (he − θe)]+ and I = [αi (hi − θi)]+. We consider here the derivation of the
equilibrium points for he, hi, and x, and we indicate how their stability was assessed. Substituting
the conditions he = 0,
hi = 0, and x = 0 in Eqs. S37, we obtain three non-linear algebraic
equations
he = Aee U x E − Aei I + Iext
hi = Aie U x E
(S38)
x = 1 / (1 + U τrec E)
Using the second and third equations to replace x and I in the first, we get an implicit equation
in the unknown he:
Connectivity Motifs by Synaptic Plasticity
38
he =
Aee U E
1 + U τrec E
− Aei (cid:20)αi (cid:18) Aie U E
1 + U τrec E
− θi(cid:19)(cid:21)+
+ Iext
(S39)
We make the assumption that the synaptic coupling from the excitatory to the inhibitory pop-
ulation is sufficiently strong Aie > τrec θi, so that short-term depression of that pathway does not
prevent steady recruitment of inhibition at higher firing rates of the excitatory population. We also
assume for simplicity that recurrent excitation is also sufficiently strong so that Aee > αi Aei Aie.
Under these hypotheses, when E is below a certain value, i.e., E < θi / [U (Aie − τrec θi)],
the above implicit equation coincides with Eq. S15 and it can be written as if inhibition was not
present:
he = Aee U E / (1 + U τrec E) + Iext
(S40)
Instead, above that critical value for E (i.e., and therefore for he), Eq. S39 does not change
formally, apart from its coefficients:
he = Aee U E / (1 + U τrec E) + Iext
(S41)
with Aee = Aee − αi Aei Aie and Iext = Iext + αi θi Aei.
It is easy to verify that
0 < Aee < Aee and Iext > Iext.
The critical value for the recurrent inhibitory inputs to affect the excitatory population can
be translated into a condition on he, i.e., he > θe + θi / [αe U (Aie − τrec θi)]. Under
our previous hypothesis, such a critical value for he is always larger than the threshold θe for
the activation of the excitatory neurons. There will exist a range of activation for he above θe,
where the impact of inhibition is negligible. For larger activation he, inhibition kicks by step-wise
decreasing the parameter Aee.
All in all, the presence of recurrent inhibition in the system does not alter qualitatively the
conclusions on the existence of equilibrium points of the mean field description. The statement
on the separation of the uppermost equilibrium points, associated respectively to the short-term
depressing and the short-term facilitating networks, remains true, since the horizontal asymptote
shares the same indirect proportionality relationship with the time constant τrec of recovery from
depression.
Connectivity Motifs by Synaptic Plasticity
39
A.5 Single population with mixed synapses
We consider the special case of a homogeneous network of neurons, whose connections to distinct
target postsynaptic neurons can be simultaneously short-term depressing and short-term facilitat-
ing. For the sake of simplicity and for distinguishing this case from the mixed populations studied
in the main text (see Fig. 4), neurons are assumed to be indistinguishable from each other. How-
ever, every neuron has a certain probability pD to establish a short-term depressing connection
with its postsynaptic target. The same neuron has probability 1 − pD to establish a short-term
facilitating connection to another postsynaptic neurons. Under these simplifying hypotheses, and
by definition of conditional expected value (98), the mean-field firing rate dynamics of the neuronal
population, recurrently connected by short-term plastic synapses, is
τ h = −h + A [pD uD xD + (1 − pD) uF xF ] E + Iext
xD = (1 − xD) /τrecD − uD xD E
uD = (UD − uD) /τf acilD + UD (1 − uD) E
(S42)
xF = (1 − xF ) /τrecF − uF xF E
uF = (UF − uF ) /τf acilF + UF (1 − uF ) E
The cases pD = 0 and pD = 1 have been already examined in the previous sections. For
intermediate values of pD, an evaluation of the equilibria of Eqs. S42 has been carried out numer-
ically, resulting in a qualitatively similar behavior to the extreme cases, with the location of the
(stable) equilibrium points to be intermediate between those of a short-term depressing neuronal
network and those of a short-term facilitation neuronal network. As expected, for increasing values
of pD the location of all equilibrium points of E(h) (if any) decreases monotonically.
Connectivity Motifs by Synaptic Plasticity
40
A.6 Emergence of motifs when initial intra- and inter-population cou-
pling are identical
A specific initial configuration for the intra- (i.e., JF F , JDD) and inter-population synaptic effica-
cies (i.e., JF D, JDF ) has been indicated in the main text as a necessary condition for subsequent
"symmetry-breaking" of the spontaneously emerging firing rates, in the depressing and the fa-
cilitating subpopulations (Fig. 4A,C). In this section, we relax such a condition and show how,
by an appropriate external stimulation protocol, the same symmetry breaking could occur. As a
consequence, the emerge of connectivity motifs is generally not impaired when synaptic couplings
are set to identical values (i.e., JF F = JDD = JF D = JDF ). Other protocols and conditions might
lead to the same configuration and here we focus on the simplest.
We assume that the two subpopulations receive a common external input and an alternating
pulsed stimulus component. As in a recurring traveling wave of external activity, each subpop-
ulation is alternatively exposed to an pulsating input component, so that both facilitatory and
depressing subpopulations are activated but never at same time. Due to intrinsic subpopulation
properties, determined by short-term synaptic plasticities and reviewed in the previous sections,
and as a direct consequence of the associative character of long-term plasticity (Fig. 4B,D) dis-
cussed in the main text, this stimulation protocol leads to stronger intra-population synaptic
coupling and weaker inter-population synaptic coupling (see Fig. S5). As discussed in the results
of the main text, such a configuration is retained indefinitely, even in the absence of external
alternating stimulation.
In order to understand why such a stimulation protocol succeeds in developing inter- and intra-
population coupling asymmetries, we must examine the STDP parameters (e.g., ) in its mean
field formulation (see Eq. 11 of the man text). Let's assume that the firing rate ED of the
depressing subpopulation is larger than the firing rate EF of the facilitating subpopulation (i.e.,
say, ED = k EF , with k > 1). This configuration of firing rates is forced by the external input
component, which alternates in time and across the subpopulations. Under these conditions, the
STDP would modify intra-population synaptic coupling JDD as follows:
∆JDD = −A−
2 τo1 E2
D + A+
3 τr1 τo2 E3
D
(S43)
while for the intra-population synaptic couplings, the STDP results into:
Connectivity Motifs by Synaptic Plasticity
∆JDF = −
1
k
∆JF D = −
1
k
A−
2 τo1 E2
D +
1
k
A+
3 τr1 τo2 E3
D
A−
2 τo1 E2
D +
1
k2 A+
3
τr1 τo2 E3
D
41
(S44)
(S45)
It is now easy to prove that ∆JDD > ∆JDF > ∆JF D, hence identical initial values
for JDD, JDF , and JF D would slowly modified heterogeneously and leading to JDD > JDF
and to JDD > JF D. Similar considerations can be repeated when the firing rate EF of the
facilitating subpopulation is larger than the firing rate ED of the depressing subpopulation (i.e.,
say, EF = k ED, with k > 1), concluding that all in all that the stimulation protocol would
shape synaptic efficacies as JDD > JDF , JDD > JF D, JF F > JDF , and JF F > JF D, therefore
privileging intra-population coupling to inter-population ones, as shown in Fig. S5B.
A.7 Statistics of the symmetry index
In order to quantify and describe concisely the symmetries of the emerging network connectivity
matrix [Aij] of size N × N , we defined the following quantity (but see e.g., (45), for alternative
definitions):
s = 1 −
1
(0.5 N (N − 1) − M )
N
N
Xi=1
Xj=i+1
A∗
ij − A∗
ji
(S46)
s intuitively represents the mean absolute difference between elements that are on symmetric
positions, with respect to the diagonal of the matrix. By definition, the elements A∗
ij are obtained
from Aij upon first normalizing its numerical values to the maximal allowed Amax and then
clipping them to a lower fraction z. For instance, choosing z = 2/3, if Aij > 2/3 Amax then
A∗
ij = Aij / Amax, and otherwise A∗
ij = 0. In the Eq. S46, M represents the number of null
pairs {A∗
ij, A∗
ji} = {0, 0} as a consequence of clipping. Then, s can be rewritten in terms of an
arithmetic average of a set of K observations of a random variable q:
s = 1 −
1
K
K
Xi=1
qi
(S47)
Assuming for simplicity that each element of [Aij ] is independently drawn from a uniform
distribution (i.e., between 0 and Amax), the probability density distribution of q, fq(Q), can be
Connectivity Motifs by Synaptic Plasticity
42
derived analytically (98). Because the arithmetic average is an unbiased estimator of the expected
value of the random variable it samples (i.e., in this case q) (98), most of the statistical properties
of s can be immediately derived from the distribution of q. First of all, the expected value of s is
given by the expression below
hsi = Amax (cid:18)1 −
1 − z
1 − z2 (cid:20) (1 − z)2
3
+ z (1 + z)(cid:21)(cid:19)
(S48)
Deriving the expression for the variance is less straightforward and requires estimating the
expected value of 1/K. In fact K coincides with the number of terms in the double sum of Eq. S46
and it is by definition not a fixed quantity, but a realization of a binomial random variable. An
approximated expression for the variance of s is then
V ar{s} ≈ A2
max
2
N (N − 1) (1 − z2) (cid:18)1 +
2 z2
N (N − 1) (1 − z2)(cid:19) V ar{q}
(S49)
where
V ar{q} =
1
1 − z2 (cid:18) (1 − z)4
6
+
2 z (1 − z3)
3
(cid:19) − (1 − hsi)2
(S50)
The validity of all the above expressions have been tested and validated numerically, directly
estimating the average and variance of s across thousands of uniform random matrices [Aij ], for
several values of z in the range [0.1; 0.9], finding an excellent agreement.
Finally, from the Central Limit Theorem (98), we can expect the density distribution of s to be
approximately Gaussian, at least for small values of z. By the above statistical expressions, when
studying the impact of short- and long-term synaptic plasticity in shaping microcircuit connectivity
motifs (see Figs. 1-2,5), we expressed the significance of the observed values of s as the chance
level, i.e., the (Gauss-distributed) probability that the observed value of s could be obtained by
chance from a random uniform matrix.
Connectivity Motifs by Synaptic Plasticity
43
A.8 Alternative STDP models
A.8.1 Pair-based STDP model
By appropriately choosing A+
2 , A−
2 , and setting to zero both A+
3 and A−
3 , earlier phenomenological
models of pair-based STDP can be rephrased as a special case of the triplet-based model. For the
pair-based STDP, each neuron of the network needs only two indicator variables, i.e., q1 and o1,
instead of four. In the lack of any firing activity of the j-th neuron, those variables exponentially
relax to zero:
τq1 q1j = −q1j
τo1 o1j = −o1j
(S51)
As the j-th neuron fires, the variables must be instantaneously updated. For such update rule,
there are two distinct scenarios determining how successive pre-post or post-pre events interact
and affect synaptic efficacy: i) all-to-all spike pairs interactions,
q1j → q1j + 1
o1j → o1j + 1
(S52)
and ii) nearest-spike interactions,
q1j → 1
o1j → 1.
(S53)
where the update rules do not allow accumulation of effects. Finally, when the j-th neuron spikes,
the following updates are performed over all the indexes i:
Wij → Wij − η A−
2 o1i (t)
Wji → Wji + η A+
2 q1i (t)
(S54)
Following the considerations by Pfister and Gerstner (54), pair-based models of STDP with
all-to-all interactions must be excluded, as they do not reproduce realistic (i.e., BCM) features
of synaptic plasticity. We then consider the triplet-based STDP model and alter some of its
parameters as it follows: A+
3 = A−
3 = 0, A+
2 = 4.5 10−3, A−
2 = 7.1 10−3. We also replaced
the all-to-all spike pairs interactions by a nearest spike interaction, by modifying the update rule
for q1 and r1 (and for q2 and r2, although those state variables are anyway irrelevant, upon setting
A+
3 = A−
3 = 0). By doing so, we obtain the pair-based STDP plasticity rule matching the
Connectivity Motifs by Synaptic Plasticity
44
exact same temporal window of the STDP triplet model employed here. In order to prove that
there is correspondence in terms of the temporal window, but altered frequency-dependence, we
subjected both the triplet-based and the pair-based models to 75 pairing events, at low frequency
10Hz. The STDP temporal windows at such a frequency are undistinguishable from each other
(compare Fig. 2A with Fig. 3A). The frequency-dependence is computed across the same number
of pairing events, imposing a pre-post or post-pre delay of 10msec. Note that for sufficiently
large frequency of the pre-post pairs, the curves corresponding to pre-post and to post-pre become
symmetric with respect to a horizontal line (see Fig. 3B; i.e., around 75 − 80%, corresponding to
long-term depression). Such a symmetry implies that in the case of random occurrence of pre-post
or post-pre timing, the LTP and the LTD components would cancel each other on the average.
A.8.2 Triplet-based anti-STDP model
This model is obtained from the triplet-based, by setting A+
3 = 7.1 10−3, A−
3 = 6.1 10−3,
A+
2 = 0, and A−
2 = 3.5 10−3, by leaving unchanged the update rules for q1, q2, o1, and o2 as
in the original triplet model, and by modifying the actual weight update equations as it follows:
when the j-th neuron spikes, the following updates are performed for all the indexes i:
Wij → Wij + η o1i (t) (cid:2)A+
Wji → Wji − η q1i (t) (cid:2)A−
3 q2j (t − ǫ)(cid:3)
3 o2j (t − ǫ)(cid:3)
2 + A+
2 + A−
(S55)
Note that this is only a tentative proposal for an anti-STDP rule, since experimental of data
is not yet available for all induction protocols earlier employed for STDP. In particular, we ignore
the frequency-dependency of the anti-STDP and by the new parameter set we roughly leave it
untouched (compare Fig. 2B with Fig. 3D).
References
1. Seung H (2009) Reading the book of memory: Sparse sampling reading the book of memory:
Sparse sampling versus dense mapping of connectomes. Neuron 62: 17-29.
2. Wickersham I, Lyon E DC Barnard, Mori T, Finke S, Conzelmann K, et al. (2007) Monosy-
naptic restriction of transsynaptic tracing from single, genetically targeted neurons. Neuron
Connectivity Motifs by Synaptic Plasticity
45
53: 639-47.
3. Zhang F, Aravanis A, Adamantidis A, de Lecea L, Deisseroth K (2007) Circuit-breakers:
optical technologies for probing neural signals and systems. Nat Rev Neurosci 8: 577-81.
4. Lichtman J, Livet J, Sanes J (2008) A technicolour approach to the connectome. Nat Rev
Neurosci 9: 417-22.
5. Denk W, Horstmann H (2004) Serial block-face scanning electron microscopy to reconstruct
three-dimensional tissue nanostructure. PLoS Biol 2: e329.
6. Chklovskii D, Vitaladevuni S, Scheffer L (2010) Semi-automated reconstruction of neural
circuits using electron microscopy. Current Opinion in Neurobiology 20: 667 -- 75.
7. Song S, Sjostrom P, Reigl M, Nelson S, Chklovskii D (2005) Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS Biol 3: e68.
8. Hai A, Shappir J, Spira M (2010) In-cell recordings by extracellular microelectrodes. Nat
Methods 7: 200-2.
9. Perin R, Berger T, Markram H (2011) A synaptic organizing principle for cortical neuronal
groups. Proc Nat Acad Sci USA 108: 5419-24.
10. Friston K (2011) Functional and effective connectivity: A review. Brain Connectivity 1:
13-36.
11. Minderer M, Liu W, Sumanovski L, Kugler S, Helmchen F, et al. (2012) Chronic imaging
of cortical sensory map dynamics using a genetically encoded calcium indicator. J Physiol
590: 99-107.
12. Wedeen V, Rosene D, Wang R, Dai G, Mortazavi F, et al. (2012) The geometric structure
of the brain fiber pathways. Science 335: 1628-34.
13. Kandell E, Schwartz J, Jessel A (2008) Principles of Neural Science. McGraw-Hill.
14. Silberberg G, Grillner S, LeBeau F, Maex R, Markram H (2005) Synaptic pathways in neural
microcircuits. Trends in Neuroscience 28: 541-51.
15. Grillner S, Markram H, De Schutter E, Silberberg G, LeBeau F (2005) Microcircuits in
action - from cpgs to neocortex. Trends in Neuroscience 28: 525-33.
Connectivity Motifs by Synaptic Plasticity
46
16. Douglas R, Martin K (2007) Recurrent neuronal circuits in the neocortex. Curr Biology 17:
R496-500.
17. Douglas R, Martin K (2007) Mapping the matrix: the ways of neocortex. Neuron 56: 226-38.
18. Binzegger T, Douglas K RJ Martin (2004) A quantitative map of the circuit of cat primary
visual cortex. J Neurosci 24: 8441-53.
19. Wang Y, Markram H, Goodman P, Berger T, Ma J, et al. (2006) Heterogeneity in the
pyramidal network of the medial prefrontal cortex. Nat Neurosci 9: 534-42.
20. Silberberg G, Markram H (2007) Disynaptic inhibition between neocortical pyramidal cells
mediated by martinotti cells. Neuron 53: 735-46.
21. Tsodyks M, Markram H (1997) The neural code between neocortical pyramidal neurons
depends on neurotransmitter release probability. Proc Nat Acad Sci USA 94: 719-23.
22. Varela J, Sen K, Gibson J, Fost J, Abbott L, et al. (1997) A quantitative description of
short-term plasticity at excitatory synapses in layer 2/3 of rat primary visual cortex. J
Neurosci 17: 7926-40.
23. Buonomano DV, Merzenich M (1998) Cortical plasticity: From synapses to maps. Annu
Rev Neurosci 21: 149-86.
24. Markram H, Gerstner W, Sjostrom P (2011) A history of spike-timing-dependent plasticity.
Front Synaptic Neurosci 3: 1-24.
25. Chklovskii D, Mel B, Svoboda K (2004) Cortical rewiring and information storage. Nature
431: 782-8.
26. Le Be' V, Markram H (2006) Spontaneous and evoked synaptic rewiring in the neonatal
neocortex. P Natl Acad Sci USA 103: 13214-9.
27. Haeusler S, Schuch K, Maass W (2009) Motif distribution, dynamical properties, and com-
putational performance of two data-based cortical microcircuit templates. J Physiol (Paris)
103: 73-87.
28. Cowan W, Sudhof T, Stevens C (2003) Synapses. The Johns Hopkins Univ. Press.
Connectivity Motifs by Synaptic Plasticity
47
29. Pignatelli M (2009) Structure and Function of the Olfactory Bulb Microcircuit. Ph.D. thesis,
´Ecole Polytechnique F´ed´erale de Lausanne, http://library.epfl.ch/en/theses/?nr=4275.
30. Clopath C, Buesing L, Vasilaki E, Gerstner W (2010) Connectivity reflects coding: a model
of voltage-based stdp with homeostasis. Nat Neurosci 13: 344-52.
31. Tsodyks M, Pawelzik K, Markram H (1998) Neural networks with dynamic synapses. Neural
Comp 10: 821-35.
32. Brette R, Gerstner W (2005) Adaptive exponential integrate-and-fire model as an effective
description of neuronal activity. J Neurophysiol 94: 3637-42.
33. Tsodyks M (2005) Methods and models in neurophysics, Elsevier, chapter Synaptic Dynam-
ics. pp. 245-66.
34. Renart A, Brunel N, Wang X (2004) Mean-Field Theory of Irregularly Spiking Neuronal
Populations and Working Memory in Recurrent Cortical Networks, volume Computational
Neuroscience: A Comprehensive Approach. CRC Press.
35. Pfister JP, Gerstner W (2006) Triplets of spikes in a model of spike timing -- dependent plas-
ticity. J Neurosci 26: 9673-82.
36. Barak O, Tsodyks M (2007) Persistent activity in neural networks with dynamic synapses.
PLoS Comput Biol 3: e35.
37. Gjorgjieva J, Clopath C, Audet A, Pfister JP (2011) A triplet spike-timing -- dependent plas-
ticity model generalizes the bienenstock -- cooper -- munro rule to higher-order spatiotemporal
correlations. Proc Nat Acad Sci USA 108: 19383 -- 8.
38. Sjostrom P, Turrigiano G, Nelson S (2001) Rate, timing, and cooperativity jointly determine
cortical synaptic plasticity. Neuron 32: 1149-1164.
39. Liu Y, Wang X (2001) Spike-frequency adaptation of a generalized leaky integrate-and-fire
model neuron. J Comput Neurosci 10: 25-45.
40. Brunel N, Amit D (1997) Model of global spontaneous activity and local structured activity
during delay periods in the cerebral cortex. Cereb Cortex 7: 237-52.
Connectivity Motifs by Synaptic Plasticity
48
41. Roxin A, Brunel N, Hansel D, Mongillo G, Van Vreeswijk C (2011) On the distribution of
firing rates in networks of cortical neurons. J Neurosci 31: 16217-26.
42. Tuckwell H (1989) Stochastic Processes in the Neurosciences. Society for Industrial and
Applied Mathematics.
43. Sterrat D, Graham B, Gillies A, Willshaw D (2011) Principles of Computational Modelling
in Neuroscience. Cambridge University Press.
44. Morrison A, Diesmann M, Gerstner W (2008) Phenomenological models of synaptic plasticity
based on spike timing. Biological Cybernetics 98: 459-78.
45. Huber L, Baker F (1979) Evaluating the symmetry of a proximity matrix. Quality and
Quantity 13: 77-84.
46. Wilson H, Cowan J (1972) Excitatory and inhibitory interactions in localized populations
of model neurons. Biophys J 12: 1-24.
47. Dayan P, Abbott L (2001) Theoretical neuroscience: computational and mathematical mod-
eling of neural systems. The MIT Press: Cambridge, Massachusetts.
48. La Camera G, Giugliano M, Senn W, Fusi S (2008) The response of cortical neurons to
in vivo-like input current: theory and experiment i. noisy inputs with stationary statistics.
Biological Cybernetics 99: 279-301.
49. Giugliano M, La Camera G, Fusi S, Senn W (2008) The response of cortical neurons to
in vivo-like input current: theory and experiment ii. time-varying and spatially distributed
inputs. Biological Cybernetics 99: 303-18.
50. Markram H, Tsodyks M (1996) Redistribution of synaptic efficacy between neocortical pyra-
midal neurones. Nature 382: 807-9.
51. Abbott L, Nelson S (2000) Synaptic plasticity: taming the beast. Nat Neurosci 3: 1178-83.
52. Song S, Miller K, Abbott L (2000) Competitive hebbian learning through spike-timing-
dependent synaptic plasticity. Nat Neurosci 3: 919-26.
53. Rubin J, Lee D, Sompolinsky H (2001) Equilibrium properties of temporally asymmetric
hebbian plasticity. Phys Rev Lett 86: 364-7.
Connectivity Motifs by Synaptic Plasticity
49
54. Pfister JP, Gerstner W (2006) Beyond pair-based stdp: A phenomenological rule for spike
triplet and frequency effects. MIT Press, volume 17 of Advances in Neural Inf. Proc. Syst.,
pp. 1409-16.
55. Bi G, Rubin J (2005) Timing in synaptic plasticity: from detection to integration. Trends
in Neuroscience 28: 222-8.
56. Holcman D, Tsodyks M (2006) The emergence of up and down states in cortical networks.
PLoS Comput Biol 2: e23.
57. Del Giudice P, Fusi S, Mattia M (2003) Modelling the formation of working memory with
networks of integrate-and-fire neurons connected by plastic synapses. J Physiol (Paris) 97:
659-681.
58. Del Giudice P, Mattia M (2001) Long and short-term synaptic plasticity and the formation
of working memory: A case study. Neurocomputing 38-40: 1175-1180.
59. Strogatz S (1994) Nonlinear dynamics and chaos. Reading, MA (USA): Addison Wesley.
60. Kempter R, Gerstner W, Van Hemmen J, Wagner H (1996) Temporal coding in the sub-
millisecond range: model of barn owl auditory pathway. Advances in neural information
processing systems : 124 -- 130.
61. Gerstner W, Kempter R, Van Hemmen J, Wagner H (1996) A neuronal learning rule for
sub-millisecond temporal coding. Nature 383: 76 -- 78.
62. Bower J, Haberly L (1986) Facilitating and nonfacilitating synapses on pyramidal cells: a
correlation between physiology and morphology. Proc Nat Acad Sci USA 83: 1115-9.
63. Stratford K, Tarczy-Hornoch K, Martin K, Bannister N, Jack J (1996) Excitatory synaptic
inputs to spiny stellate cells in cat visual cortex. Nature 382: 258-61.
64. Reyes A, Sakmann B (1999) Developmental switch in the short-term modification of unitary
epsps evoked in layer 2/3 and layer 5 pyramidal neurons of rat neocortex. J Neurosci 19:
3827-35.
65. Geracitano R, Kaufmann W, Szabo G, Ferraguti F, Capogna M (2007) Synaptic hetero-
geneity between mouse paracapsular intercalated neurons of the amygdala. J Physiol 585:
117-34.
Connectivity Motifs by Synaptic Plasticity
50
66. Markram H, Wang Y, Tsodyks M (1998) Differential signaling via the same axon of neocor-
tical pyramidal neurons. Proc Nat Acad Sci USA 95.
67. Morrison A, Aerstsen A, Diesmann M (2007) Spike-timing-dependent plasticity in balanced
random networks. Neural Comp 19: 1437-67.
68. Gilson M, Burkitt A, van Hemmen J (2010) Stdp in recurrent neuronal networks. Front
Comput Neurosci 4: 1-15.
69. Gilson M, Burkitt A, Grayden D, Thomas D, van Hemmen J (2009) Emergence of net-
work structure due to spike-timing-dependent plasticity in recurrent neuronal networks. iii:
Partially connected neurons driven by spontaneous activity. Biological Cybernetics 101:
411-26.
70. Kunkel S, Diesmann M, Morrison A (2011) Limits to the development of feed-forward struc-
tures in large recurrent neuronal networks. Front Comput Neurosci 4.
71. Bourjaily M, Miller P (2011) Excitatory, inhibitory, and structural plasticity produce corre-
lated connectivity in random networks trained to solve paired-stimulus tasks. Front Comput
Neurosci 5: 1-24.
72. Song S, Abbott L (2001) Column and map development and cortical re-mapping through
spike-timing dependent plasticity. Neuron 32: 339-50.
73. Butts D, Kanold P (2010) The applicability of spike time dependent plasticity to develop-
ment. Front Synaptic Neurosci 2: 1-9.
74. Cooper L, Intrator N, Blais B, Shouval H (2004) Theory of Cortical Plasticity. World
Scientific, Singapore.
75. Fuhrmann G, Segev I, Markram H, Tsodyks M (2002) Coding of temporal information by
activity-dependent synapses. J Neurophysiol 87: 140-8.
76. Fuhrmann G, Cowan A, Segev I, Tsodyks M, Stricker C (2004) Multiple mechanisms govern
the dynamics of depression at neocortical synapses of young rats. J Physiol 557: 415-38.
77. Loebel A, Silberberg G, Helbig D, Markram H, Tsodyks M, et al. (2009) Multiquantal release
underlies the distribution of synaptic efficacies in the neocortex. Front Comput Neurosci 3.
Connectivity Motifs by Synaptic Plasticity
51
78. Pan B, Zucker R (2009) A general model of synaptic transmission and short-term plasticity.
Neuron 62: 539-54.
79. Rubin J, Gerkin R, Bi G, Chow C (2005) Calcium time course as a signal for spike-timing-
dependent plasticity. J Neurophysiol 93: 2600-13.
80. Graupner M, Brunel N (2012) Calcium-based plasticity model explains sensitivity of synaptic
changes to spike pattern, rate, and dendritic location. Proc Nat Acad Sci USA 109: 3991-6.
81. Vogels T, Rajan K, Abbott L (2005) Neural network dynamics. Annu Rev Neurosci 28:
357-76.
82. Gigante G, Del Giudice P, Mattia M (2007) Frequency-dependent response properties of
adapting spiking neurons. Mathematical Biosciences 2: 336-51.
83. Rauch A, La Camera G, Luescher H, Senn W, Fusi S (2003) Neocortical pyramidal cells
respond as integrate-and-fire neurons to in vivo-like input currents. J Neurophysiol 90:
1598-612.
84. Pimentel D, Margrie T (2008) Glutamatergic transmission and plasticity between olfactory
bulb mitral cells. J Physiol 586.8: 2107-119.
85. Cassenaer S, Laurent G (2012) Conditional modulation of spike-timing dependent plasticity
for olfactory learning. Nature doi:10.1038/nature10776.
86. Gao Y, Strowbridge B (2009) Long-term plasticity of excitatory inputs to granule cells in
the rat olfactory bulb. Nat Neurosci 12: 731-3.
87. Linster C, Cleland T (2010) Decorrelation of odor representations via spike timing-dependent
plasticity. Front Comput Neurosci 4: 1-11.
88. Buonomano DV (1999) Distinct functional types of associative long-term potentiation in
neocortical and hippocampal pyramidal neurons. J Neurosci 19: 6748-54.
89. Markram H, Pikus D, Gupta A, Tsodyks M (1998) Potential for multiple mechanisms,
phenomena and algorithms for synaptic plasticity at single synapses. Neuropharmacology
37: 489-500.
Connectivity Motifs by Synaptic Plasticity
52
90. Mongillo G, Barak O, Tsodyks M (2008) Synaptic theory of working memory. Science 319:
1543-6.
91. Bi G, Poo M (1998) Synaptic modifications in cultured hippocampal neurons: dependence
on spike timing, synaptic strength, and postsynaptic cell type. J Neurosci 18: 10464 -- 72.
92. Barabasi A (2005) Taming complexity. Nature Physics 1: 68-70.
93. Froemke R, Dan Y (2002) Spike-timing-dependent synaptic modification induced by natural
spike trains. Nature 416: 433-438.
94. Clopath C, Gerstner W (2010) Voltage and spike timing interact in stdp - a unified model.
Front Synaptic Neurosci 2.
95. Kepecs A, van Rossum MC, Song S, Tegner J (2002) Spike-timing-dependent plasticity:
common themes and divergent vistas. Biological Cybernetics 87: 446-58.
96. Mattia M, Del Giudice P (2004) Finite-size dynamics of inhibitory and excitatory interacting
spiking neurons. Phys Rev E Stat Nonlin Soft Matter Phys 70: 052903.
97. Press W, Teukolsky S, Vetterling W, Flannery B (2007) Numerical Recipes: The Art of
Scientific Computing. Cambridge University Press.
98. Papoulis A, Pillail SU (2002) Probability, random variables, and stochastic processes.
McGraw-Hill, 4th edition.
Connectivity Motifs by Synaptic Plasticity
53
Figures and Figure Legends
A
X
Y
feedforward (strong)
connection, from X to Y
C
pre
X
Y
reciprocal (strong)
connection, between X and Y
post
4
3
B
5
6
7
8
9
cycling stimuli
sequence
2
10
1
F
pre
post
D
V
m
0
9
/
1
6
5
7
200 msec
n +
c ilit a ti o
D
P
o rt-t e r m f a
S
T
h
s
reciprocal (strong)
facilitating connections
4
3
E
2
1
10
25 msec
8
9
s
h
ort-term d
S
T
e
pre
D
P
ssio
n +
unidirectional (strong)
depressing connections
4
3
G
H
5
7
6
V
m
0
9
/
1
200 msec
2
10
1
5
4
3
2
1
5
4
3
2
1
8
9
25 msec
Figure 1. Emergence of connectivity motifs in a toy model network. Unidirectional
(reciprocal) strong excitatory connections are indicated (A) as dashed (continuous) line
segments, representing the topology of the network (B). Each model neuron receives periodic
spatially alternating depolarizing current pulses, strong enough to make it fire a single action
potential. Synapses among connected neurons display (C) short-term facilitation of postsynaptic
potential amplitudes. Spike-Timing Dependent Plasticity (STDP) leads to strengthened
connections and result into a largely reciprocal topology (D). Modifying the short-term plasticity
profile into depressing (F), leads to a largely unidirectional topology shaped by STDP(G).
Distinct motifs of strong connections arise from short- and long-term plasticities, due to distinct
firing patterns (compare E and H), under identical external stimulation and initial connections.
Parameters: Aij = 400pA.
Connectivity Motifs by Synaptic Plasticity
54
A
]
%
[
e
g
n
a
h
c
P
D
T
S
140
120
100
80
60
40
B
]
%
[
e
g
n
a
h
c
P
D
T
S
250
200
150
100
50
P
T
L
D
T
L
40
10 Hz
−20
0
20
t - t [ms]
post pre
t
> t
pre
post
t
post
< t
pre
P
T
L
D
T
L
10
20
30
40
50
Pairing frequency [Hz]
C
D
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
y
r
t
e
m
m
y
s
x
i
r
t
a
m
c
i
t
p
a
n
y
S
y
r
t
e
m
m
y
s
x
i
r
t
a
m
c
i
t
p
a
n
y
S
0
0
short-term depressing synapses + STDP
t = 0
t = 200
mean
68%
20 40 60 80 100 120 140 160 180
time [a.u]
0
100
50
#
short-term facilitating synapses + STDP
t = 0
t = 200
mean
68%
20 40 60 80 100 120 140 160 180
time [a.u]
0
100
50
#
Figure 2. Statistics of motifs emergence in a toy model network. When decoupled from
recurrent interactions, an isolated model synapse undergoes long-term changes depending on pre-
and postsynaptic spike timing (A) and pairing frequency (B). Above a critical frequency (gray
shading), spike timing no longer matters and long-term potentiation of synaptic efficacy (LTP)
prevails on long-term depression (LTD). Panels C, D: The simulations of Fig. 1 were repeated
2000 times, each time starting from a random initial topology. STDP progressively induced a
persisting non-random reconfiguration of strong connections, quantified across time by a
symmetry index (see Methods). Neurons connected only by short-term depressing synapses
evolved strong unidirectional connections, corresponding to a low symmetry index. This is
displayed in panel C, as an average across the 2000 simulations (left panel). Initial and final
distributions of symmetry index values are also shown (right panel, gray and black histogram
respectively). Neurons connected only by short-term facilitating synapses evolved instead strong
bidirectional connections with high symmetry indexes (D).
Connectivity Motifs by Synaptic Plasticity
55
A
C
E
G
y
r
t
e
m
m
y
s
x
i
r
t
a
m
c
i
t
p
a
n
y
S
y
r
t
e
m
m
y
s
x
i
r
t
a
m
c
i
t
p
a
n
y
S
B
D
F
Spike-triplets aSTDP model
]
%
[
e
g
n
a
h
c
P
D
T
S
140
120
100
80
60
40
P
T
L
D
T
L
40
−20
0
20
t - t [ms]
post pre
Spike-triplets aSTDP model
]
%
[
e
g
n
a
h
c
P
D
T
S
250
200
150
100
50
pre
t
> t
post
t
post
< t
pre
P
T
L
D
T
L
10
20
30
40
50
Pairing frequency [Hz]
y
r
t
e
m
m
y
s
x
i
r
t
a
m
c
i
t
p
a
n
y
S
y
r
t
e
m
m
y
s
x
i
r
t
a
m
c
i
t
p
a
n
y
S
short-term depressing synapses +
spike-triplet anti-STDP
t = 0
t = 200
mean
68%
1
0.8
0.6
0.4
0.2
0
40
120
80
time [a.u]
160
0
50 100
#
1
0.8
0.6
0.4
0.2
short-term facilitating synapses +
spike-triplet anti-STDP
t = 0
t = 200
mean
68%
0
40
120
80
time [a.u]
160
0
50 100
#
Spike-pairs STDP
−20
0
20
t - t [ms]
post pre
Spike-pairs STDP
t
> t
pre
post
t
post
< t
pre
P
T
L
D
T
L
40
P
T
L
D
T
L
]
%
[
e
g
n
a
h
c
P
D
T
S
140
120
100
80
60
40
]
%
[
e
g
n
a
h
c
P
D
T
S
120
100
80
60
10
20
30
40
50
Pairing frequency [Hz]
short-term depressing synapses +
spike-pairs STDP
t = 0
t = 200
mean
68%
1
0.8
0.6
0.4
0.2
0
40
120
80
time [a.u]
160
0
50 100
#
H
1
0.8
0.6
0.4
0.2
short-term facilitating synapses +
spike-pairs STDP
t = 0
t = 200
mean
68%
0
40
120
80
time [a.u]
160
0
50 100
#
Figure 3. STDP key-features for motif emergence. The pair-based STDP model, with
temporal window shown in panel A and matching exactly Fig. 2A, exhibits a different
frequency-dependency (panel C) than the triplet-based STDP model (Fig. 2B). Modifying the
triplet-based STDP parameters to ad-hoc invert its temporal window (e.g., as in anti-STDP,
panel B, compare to panel A), yet leaves its frequency-dependency and the LTD-reversal (gray
shading) unchanged (panel D). Repeating the study of Fig. 2 with these two modified models, we
find that i) the pair-based STDP fails to account for motifs emergence (panels E,G, compare to
Fig. 2), while anti-STDP succeeds (panels F,H, compare to Fig. 2). Parameters: see the
Supplemental Information.
Connectivity Motifs by Synaptic Plasticity
56
A
JFF
EF
B
JDD
ED
I
I
C
40
30
)
h
(
E
20
10
0
0
unbalanced ext. inputs
Facilitating
Depressing
Stable equilibria
D
100
)
h
(
E
80
60
40
20
balanced ext. inputs
Facilitating
Depressing
Stable equil.
Unstable equil.
10
20
h
30
40
0
20
40
60
80
100
h
E
balanced ext. inputs (zoom)
20
15
)
h
(
E
10
5
0
Depressing
Facilitating
Stable equil.
Unstable equil.
15
20
5
10
h
Figure 4. Mean-field analysis of firing rate equilibria, in homogeneous networks
without long-term plasticity. The firing rate of homogeneous recurrent networks, including
short-term facilitating synapses (A) or depressing excitatory synapses (B), was studied by
standard mean-field analysis. Average synaptic efficacies are indicated by JF F or by JDD,
correspondingly. Excitatory and inhibitory external inputs are modeled by a single term, taking
positive, zero or negative values. A zero value corresponds to balanced excitatory/inhibitory
inputs, while a non-zero value corresponds to unbalanced excitatory/inhibitory inputs. The
steady-state firing rate (i.e., E, in a.u.) are the roots of the equation E(h) = h (see Eqs. 10,11
and the Supplemental Information), whose graphical solution is provided (C-E), for facilitating
(gray) or depressing (black) synapses, without long-term plasticity. Panel E is a zoomed view of
D. (Un)stable firing rate equilibria are indicated by filled circles (squares). Networks with
facilitating synapses fire at higher rates than networks with depressing synapses, as emphasized
by the gray shading.
Connectivity Motifs by Synaptic Plasticity
57
B
C
Actual
Null hp
-,-
F,- D,-
F,F D,F D,D
F
Depressing subopulation
Faciliating subpopulation
A
Actual
Null hp
-,-
D,-
D,D
D
0.8
0.6
0.4
0.2
0
0.07
0.05
0.03
0.01
0.6
0.4
0.2
0
0.35
0.25
0.15
0.05
s
n
o
i
t
c
a
r
f
s
f
i
t
i
o
M
g
n
g
r
e
m
E
s
n
o
i
t
i
u
b
i
r
t
s
d
s
e
a
r
g
n
i
r
i
F
t
-,-
F,-
F,F
E
Actual
Null hp
0.4
0.3
0.2
0.1
0
0.14
0.1
0.06
0.02
0
20
40
60
80
100
0
20
Firing Rate [Hz]
40
80
Firing Rate [Hz]
60
100
0
20
40
60
80
100
Firing Rate [Hz]
Figure 5. Results from numerical simulations of large recurrent networks of model
neurons with short- and long-term plasticity. Homogeneous and heterogeneous recurrent
networks made of 1000 Integrate-and-Fire model neurons were numerically simulated, under
identical conditions. Panel A shows the comparison of the emergence of weak or no connectivity
pairs (indicated as "-,-"), of unidirectional strong connectivity pairs ("→", "D,-"), and of
reciprocal strong connectivity pairs ("↔", "D,D") for a homogeneous network of neurons
connected by depressing synapses: strong unidirectional depressing connections significantly
outnumber reciprocal depressing ones. The fractions of emerged motifs (black) is significantly
different than the null-hypothesis (white) of random motifs occurrence. Panel B repeats this
quantification for a homogeneous network with facilitating synapses: strong connections are only
found on reciprocal connectivity pairs ("↔", "F,F") and all emerging motifs are non-random.
Panel C repeats the same quantification for a heterogeneous network with both short-term
facilitating and depressing synapses. Emerging motifs display highly non-random features and
confirm that reciprocal facilitatory motifs ("↔", "F,F") outnumber unidirectional facilitatory
motifs ("→", "F,-"), and that unidirectional depressing motifs ("→", "D,-") outnumber reciprocal
depressing motifs ("↔", "D,D"). Panels D-F display the steady-state firing rate distributions,
corresponding to homogeneous depressing, homogeneous facilitating, and heterogeneous networks
respectively. The plots confirm that heterogeneity in connectivity motifs is accompanied by
bimodal firing rates above and below the critical frequency, represented here by a grey shading.
Connectivity Motifs by Synaptic Plasticity
58
A
JFF
EF
C
JFF
EF
E
]
.
u
.
a
[
e
t
a
r
50
40
30
20
10
0
JDD
ED
JDD
ED
JFD
JDF
I
JFD
JDF
I
B
]
z
H
[
t
s
o
p
f
70
40
10
D
]
z
H
[
t
s
o
p
f
70
40
10
t > t
post pre
LTP
LTD
10
70
40
[Hz]
f
pre
t < t
post pre
LTP
LTD
10
70
40
[Hz]
f
pre
+20%
+10%
-10%
P
T
L
D
T
L
+20%
+10%
-10%
P
T
L
D
T
L
F
Depressing
Facilitating
80
40
time [a.u.]
120
]
.
u
a
.
[
W
JFD
JFF
5
4
3
2
JDF
1
JDD
0
80
40
time [a.u.]
120
Figure 6. Mean-field simulation of a heterogeneous network with short- and
long-term plasticity. The firing rate evolution of a heterogenous recurrent network, including
both short-term facilitating and depressing synapses (A), was estimated by numerically solving
the corresponding mean-field equations. The average synaptic efficacies among and across
populations, indicated by JF F , JDD, JF D, and JDF , undergo long-term modification. Panels
B,D show the long-term changes of an isolated synapse (decoupled from recurrent interactions)
depending on pre- and postsynaptic spike timing (i.e., tpre, tpost) and frequencies (i.e., fpre,
fpost). When fpre and fpost are varied independently, long-term potentiation (LTP) and
depression (LTD) emerges as in associative Hebbian plasticity. This suggests that JF F and JF D
will become significantly stronger than JF D and JDD (C) and that such a configuration will be
retained indefinitely. This was confirmed by simulations (E-F) plotting the temporal evolution of
the firing rates EF (black trace) and ED (gray trace), and of the synaptic efficacies (F). The
heterogeneity occurs by separation of emerging firing rates (Fig. 4), as emphasized by the grey
shading. Parameters: Iext = 5, τ = 10 msec, with initial conditions JF F = JDD = 3, and
JDF = JF D = 1 (see Supplemental Information for alternatives).
Connectivity Motifs by Synaptic Plasticity
59
A
JFF
1-p
D
E
p
D
JDD
I
B
100
balanced ext. inputs
unbalanced ext. inputs
80
60
40
.
l
i
l
u
q
e
e
b
a
t
s
-
E
20
0
0.2
0.4
0.6
0.8
1
p
D
Figure 7. Mean-field analysis of firing rate equilibria, in homogeneous networks with
overlapping short-term synaptic properties and no long-term plasticity. Panel A
represents the sketch of a recurrent network where a clear segregation between subpopulations of
depressing- or facilitating-only synapses does not occur. A neuron has a probability pD of
connecting to its postsynaptic target by a depressing synapse, and 1 − pD of connecting to its
target by a facilitating synapse. Panel B plots the location of the equilibria of the firing rate E,
under distinct external inputs conditions and for increasing values of pD. For the same
parameters of Fig. 4, stable equilibria move as a function of pD, taking intermediate values
between the two extreme cases, i.e., pD = 0 and pD = 1; compare to panels D-F of Figs. 4.
Connectivity Motifs by Synaptic Plasticity
60
A
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
1
2
3
4
5
6
7
8
9
10
C
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
E
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
1
2
3
4
5
6
7
8
9
10
1
2
3
4
5
6
7
8
9
10
s = 0.38, p = 0.3
B
4
3
9
8
7
6
5
4
3
2
1
10
9
8
7
6
5
4
3
2
1
10
9
8
7
6
5
4
3
2
1
6
6
6
D
F
5
7
5
7
5
7
8
9
4
3
8
4
9
3
2
1
10
2
1
10
2
1
10
X
X
X
X
X
9
X
X
10
X
X
X
X
X
9
X
X
10
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
10
X
X
X
X
9
1
2
3
4
5
6
7
8
Presynaptic neuron
s = 0.12, p = 0.002
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
1
2
3
4
5
6
7
8
Presynaptic neuron
s = 0.58, p = 0.00002
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
X
1
2
3
4
5
6
7
8
Presynaptic neuron
8
9
Figure 8. External activity may still induce reciprocal motifs emergence in
depressing networks. As in Figure 1, the synaptic connectivity matrix of a network of ten
neurons was randomly initialized and pruned (A,B, i.e., pruning is indicated by the "X"
symbols). Internally generated activity, as in Figures 1-2, contributes to the emergence of
non-random unidirectional motifs, resulting in an asymmetric matrix W (C,D). However, if five
units of the same network (gray circles) are externally stimulated above the STDP critical
frequency (see the Results of the main text), a non-random connectivity emerges, featuring
reciprocal motifs and a symmetric connectivity submatrix (E,F; upper left corner, dashed
rectangle). The values indicated above panels A,C,E represent the symmetry index and its
significance.
Connectivity Motifs by Synaptic Plasticity
Symbol Description
Forward Euler method integration time step
Number of simulated neurons
Membrane capacitance
Membrane leak conductance
Resting membrane potential
After-spike reset potential
Spike steepness of the exponential IF model
Spike emission threshold of the exponential IF model
Threshold voltage parameter of the exponential IF model
Absolute refractory period
Voltage dependence coefficient of the spike frequency adaptation
Spike-timing dependence parameter of the spike frequency adaptation
0.0805 nA
61
Value
0.1 msec
10 − 1000
281 pF
30 nS
−70.6 mV
−70.6 mV
2 mV
20 mV
−50.4 mV
2 msec
4 nS
144 msec
5 msec
0.8
0.1
900 msec
100 msec
100 msec
900 msec
7.1 10−3
0
0
6.5 10−3
16.8 msec
101 msec
33.7 msec
114 msec
6 − 12 pA
5
3
1
dt
N
cm
gleak
Eleak
Ereset
∆T
Vθ
VT
τarp
a
∆x
τx
τsyn
UD
UF
τrec D
τrec F
Time constant of the spike frequency adaptation
Excitatory postsynaptic currents decay time constant
Release probability, for depressing synapses
Release probability, for facilitatory synapses
Time constant of recovery from depression, for depressing synapses
Time constant of recovery from depression, for facilitating synapses
τf acil D Time constant of recovery from facilitation, for depressing synapses
τf acil F Time constant of recovery from facilitation, for facilitating synapses
A−
2
A−
3
A+
2
A+
3
τq1
τq2
τo1
τo2
STDP model LTD amplitude for post-pre event
STDP model LTD amplitude for post-pre event (triplet-term)
STDP model LTP amplitude for pre-post event
STDP model LTP amplitude for pre-post event (triplet-term)
STDP model decay time of presynaptic indicator q1
STDP model decay time of presynaptic indicator q2
STDP model decay time of postsynaptic indicator o1
STDP model decay time of postsynaptic indicator o2
Ai j
Maximal synaptic efficacy
Wmax
Upper boundary for STDP dimensionless scaling factor Wij
θ
η
Threshold of the frequency-current response curve for mean-field models
STDP plasticity rate
Table 1. Parameters employed in the simulations: STDP parameters are as in the minimal
all-to-all triplet model described in Pfister and Gerstner (2006); short-term depression and
facilitation parameters as in (Wang et al., 2006); neuron parameters are as in (Clopath et al.,
2010).
Connectivity Motifs by Synaptic Plasticity
62
20
18
16
14
12
10
8
6
4
2
0
20
18
16
14
12
10
8
6
4
2
0
A
0
5
C
0
5
10
h
10
h
15
20
B
h0
0
5
10
h
15
20
D
20
18
16
14
12
10
8
6
4
2
0
20
18
16
14
12
10
8
6
4
2
0
15
20
0
5
15
20
10
h
Figure S1. Graphical representation of the function F2(h), for different values of A
and Iext. The function F2(h) has been plotted for different values of the maximal synaptic
efficacy A and external input Iext, resulting in one intersection only (A - A = 3, Iext = 0) ,
three intersections with two of them coincident with each other (B - A = 5.488, Iext = 0),
three distinct intersections (C - A = 7, Iext = 0), and finally one intersection for larger
external input (D - A = 3, Iext = 4). The remaining parameters were: U = 0.8,
τrec = 500 msec, θ = 3, α = 1. The scripts to generate these plots and to carry out
asymptotic analysis on the stability of the equilibrium points, see the text, are available online
from the ModelDB (accession number 143082).
Connectivity Motifs by Synaptic Plasticity
63
A
20
18
16
14
12
10
8
6
4
2
0
B
20
18
16
14
12
10
8
6
4
2
0
0
5
10
h
15
20
0
5
15
20
10
h
Figure S2. Graphical representation of the function F3(h), for different values of A.
The function F3(h) has been plotted for different values of the maximal synaptic efficacy A and
external input Iext, resulting in one (A - A = 0.5, Iext = 0) or two intersections (B - A = 3.5,
Iext = 0). The remaining parameters were: U = 0.1, τf acil = 500 msec, θ = 3, α = 1. The
scripts to generate these plots and to carry out asymptotic analysis on the stability of the
equilibrium points, see the text, are available online at the ModelDB (accession number 143082).
Aee
E
Iext
Aie
Aei
I
Figure S3. Mean field description for two recurrently connected populations of
excitatory and inhibitory neurons. Excitatory neurons are recurrently connected by
short-term plastic synapses and project to a population of inhibitory neurons. Inhibitory neurons
project back to the excitatory population with non-plastic, linear synapses.
Connectivity Motifs by Synaptic Plasticity
64
A
20
18
16
14
12
10
8
6
4
2
0
B
20
18
16
14
12
10
8
6
4
2
0
0
5
10
h
15
20
0
5
15
20
10
h
Figure S4. Graphical representation of the equivalent of functions F2(h) and F3(h), in
the presence of inhibition. Increasing the coupling between the excitatory and the inhibitory
population substantially bends down the curves previously analyzed as F2(h) and F4(h), lowering
the firing rate of the equilibrium points (Aei = Aie / 10; A - Iext = 0, B - Iext = 5).
]
.
u
a
.
[
t
e
a
r
A
70
60
50
40
30
20
10
0
Depressing
Facilitating
50
100
150
200
time [a.u.]
]
.
.
u
a
[
W
5
4
3
2
1
0
B
JFD
JFF
JDF
JDD
50
100
150
200
time [a.u.]
Figure S5. Formation of asymmetric intra and extra population synapses. The
subopulations D and F of Fig. 4A receive a common input current (IC = 12.5) and an
alternating stimulation component (IED, IEF ), oscillating periodically between two amplitude
levels (i.e., 2 and 0) every 1 [a.u.] of time: whenever IED = 0, IEF = 2, and vice versa. The
temporal evolution of the firing rate of each population is shown in A, while the corresponding
long-lasting plastic changes of the synaptic coupling (i.e., JF F , JDD, JF D, JDF ) is plotted in B,
upon initialization to the same value (i.e., 1). The same final configuration is generally obtained
even by randomly initializing JF F , JDD, JF D, and JDF by a Gaussian distribution with mean 1
and standard deviation 0.001, for 90 out of 100 simulation runs, demonstrating a degree of
robustness.
Connectivity Motifs by Synaptic Plasticity
65
A
Actual
Null hp
-,-
D,-
D,D
D
Actual
Null hp
-,-
D,-
D,D
G
Actual
Null hp
-,-
D,-
D,D
J
Actual
Null hp
-,-
D,-
D,D
0.8
0.6
0.4
0.2
0
0.8
0.6
0.4
0.2
0
0.8
0.6
0.4
0.2
0
0.8
0.6
0.4
0.2
0
s
n
o
i
t
c
a
r
f
s
f
i
t
i
o
M
g
n
g
r
e
m
E
0.6
0.4
0.2
0
s
n
o
i
t
c
a
r
f
s
f
i
t
i
o
M
g
n
g
r
e
m
E
s
n
o
i
t
c
a
r
f
s
f
i
t
i
o
M
g
n
g
r
e
m
E
s
n
o
i
t
c
a
r
f
s
f
i
t
i
o
M
g
n
g
r
e
m
E
0.6
0.4
0.2
0
0.6
0.4
0.2
0
0.6
0.4
0.2
0
Actual
Null hp
-,-
F,-
F,F
E
Actual
Null hp
-,-
F,-
F,F
H
Actual
Null hp
-,-
F,-
F,F
K
Actual
Null hp
-,-
F,-
F,F
0.4
0.3
0.2
0.1
0
0.4
0.3
0.2
0.1
0
0.4
0.3
0.2
0.1
0
0.4
0.3
0.2
0.1
0
B
C
Actual
Null hp
-,-
F,- D,-
F,F D,F D,D
F
Actual
Null hp
-,-
F,- D,-
F,F D,F D,D
I
Actual
Null hp
-,-
F,- D,-
F,F D,F D,D
L
Actual
Null hp
-,-
F,- D,-
F,F D,F D,D
Figure S6. External stimulation protocols and models details do not affect motifs
emergence. The figure examines the fraction of motifs, emerging when neurons receive periodic
input wave-like stimulation (A,B,C, exactly as in Figure 6A-C, for ease of comparison). However,
when the periodic stimulation is omitted (D,E,F), as well as when each neuron receives instead a
ten percent shared random background inputs with each other (G,H,I), very similar results
emerge. The simulations of panels D,E,F, when repeated without spike-frequency adaptation
mechanisms from each unit of the network(s) (J,K,L), still give rise to the same results.
Connectivity Motifs by Synaptic Plasticity
66
N = 10
]
s
[
E
0.6
0.5
0.4
0.3
0.2
0.1
0
Unitary matrix
Random
Triangular unitary matrix
0
0.2
0.4
h
0.6
0.8
Figure S7. Impact of the clipping threshold parameter h on the symmetry measure
s. The continuous trace and × markers show the expectation of the symmetry index for a
random connectivity matrix with 20% of its elements randomly pruned, over 10000 samples of
10 × 10 matrices, with error bars indicating the standard deviation. The simulations were
repeated for a unitary matrix (i.e., leading to the maximum possible value for s) and for a upper
triangular unitary matrix (i.e., leading to the minimum possible value for s). The value of
h = 2/3 used in this work, chosen for consistence to earlier works, leads to a middle point
between the two extremes considered and thus provide a good discriminating condition when
using the statistics of a random matrix as a null hypothesis.
|
1804.05766 | 1 | 1804 | 2018-04-16T16:10:00 | Universal Features in Phonological Neighbor Networks | [
"q-bio.NC"
] | Human speech perception involves transforming a countinous acoustic signal into discrete linguistically meaningful units, such as phonemes, while simultaneously causing a listener to activate words that are similar to the spoken utterance and to each other. The Neighborhood Activation Model (NAM~\cite{Luce:1986,Luce:1998}) posits that phonological neighbors (two forms [words] that differ by one phoneme) compete significantly for recognition as a spoken word is heard. This definition of phonological similarity can be extended to an entire corpus of forms to produce a phonological neighbor network~\cite{Vitevitch:2008} (PNN). We study PNNs for five languages: English, Spanish, French, Dutch, and German. Consistent with previous work, we find that the PNNs share a consistent set of topological features. Using an approach that generates random lexicons with increasing levels of phonological realism, we show that even random forms with minimal relationship to any real language, combined with only the empirical distribution of language-specific phonological form lengths, are sufficient to produce the topological properties observed in the real language PNNs. The resulting pseudo-PNNs are insensitive to the level of lingusitic realism in the random lexicons but quite sensitive to the shape of the form length distribution. We therefore conclude that "universal" features seen across multiple languages are really string universals, not language universals, and arise primarily due to limitations in the kinds of networks generated by the one-step neighbor definition. Taken together, our results indicate that caution is warranted when linking the dynamics of human spoken word recognition to the topological properties of PNNs, and that the investigation of alternative similarity metrics for phonological forms should be a priority. | q-bio.NC | q-bio | Universal Features in Phonological Neighbor Networks
Kevin S. Brown,1, 2, 3, 4 Paul D. Allopenna,5 William R. Hunt,1 Rachael
Steiner,5 Elliot Saltzman,6 Ken McRae,7, 8 and James S. Magnuson4, 5
1Department of Biomedical Engineering,
University of Connecticut, Storrs, CT, USA
2Department of Physics, University of Connecticut, Storrs, CT, USA
3Institute for Systems Genomics, University of Connecticut, Storrs, CT, USA
4Connecticut Institute for the Brain & Cognitive Sciences,
University of Connecticut, Storrs, CT, USA∗
5Department of Psychology, University of Connecticut, Storrs, CT, USA
6Department of Physical Therapy and Athletic Training,
Boston University, Boston, MA, USA
7Department of Psychology, University of Western Ontario, London, Ontario, CAN
8Brain & Mind Institute, University of Western Ontario, London, Ontario, CAN
(Dated: July 4, 2018)
8
1
0
2
r
p
A
6
1
]
.
C
N
o
i
b
-
q
[
1
v
6
6
7
5
0
.
4
0
8
1
:
v
i
X
r
a
1
Abstract
Human speech perception involves transforming a countinous acoustic signal into discrete lin-
guistically meaningful units, such as phonemes, while simultaneously causing a listener to activate
words that are similar to the spoken utterance and to each other. The Neighborhood Activa-
tion Model (NAM [1, 2]) posits that phonological neighbors (two forms [words] that differ by one
phoneme) compete significantly for recognition as a spoken word is heard. This definition of phono-
logical similarity can be extended to an entire corpus of forms to produce a phonological neighbor
network [3] (PNN). We study PNNs for five languages: English, Spanish, French, Dutch, and Ger-
man. Consistent with previous work, we find that the PNNs share a consistent set of topological
features. Using an approach that generates random lexicons with increasing levels of phonological
realism, we show that even random forms with minimal relationship to any real language, combined
with only the empirical distribution of language-specific phonological form lengths, are sufficient
to produce the topological properties observed in the real language PNNs. The resulting pseudo-
PNNs are insensitive to the level of lingusitic realism in the random lexicons but quite sensitive
to the shape of the form length distribution. We therefore conclude that "universal" features seen
across multiple languages are really string universals, not language universals, and arise primarily
due to limitations in the kinds of networks generated by the one-step neighbor definition. Taken
together, our results indicate that caution is warranted when linking the dynamics of human spoken
word recognition to the topological properties of PNNs, and that the investigation of alternative
similarity metrics for phonological forms should be a priority.
∗ [email protected]; http://kbrown.research.uconn.edu
2
I.
INTRODUCTION
The preception and recognition of acoustic speech, known in psycholinguistics as spo-
ken word recognition (SWR), requires that human listeners rapidly map highly variable
acoustic signals onto stable linguistically relevant categories (in this case, phonemes, i.e. the
consonants and vowels that comprise a language's basic sound inventory) and then piece
together sequences of phonemes into words, all without robust cues to either phoneme or
word boundaries (see here [4, 5] for reviews). Decades of research on human spoken word
recognition has led to a consensus on three broad principles: (1) SWR occurs in a continuous
and incremental fashion as a spoken target word unfolds over time, (2) words in memory
are activated proportionally to their similarity with the acoustic signal as well their prior
probability (computed as a function of their frequency of occurrence) in the language, and
(3) activated words compete for recognition. A key difference between theories is how to
characterize signal-to-word and word-to-word similarity. Most theories incorporate some set
some sort of similarity threshold, and pairs of words meeting that threshold are predicted
to strongly activate each other and compete. Perhaps the most influential definition for
the phonological similarity of spoken words is the concept of phonological neighbors posited
under the Neighborhood Activation Model (NAM) by Luce and colleagues [1, 2]. NAM
includes a gradient similarity metric and a threshold metric, although only the latter is
widely used (and we focus on it here). The threshold metric defines neighbors based on the
Deletion-Addition-Substitution (DAS) string metric, which states that two words are neigh-
bors (i.e., they are sufficiently similar to strongly activate one another and compete) if they
differ by no more than the deletion, addition, or substitution of a single phoneme. Thus,
cat has the deletion neighbor at, addition neighbors scat and cast, and many substitution
neighbors, such as bad, cot, and can. NAM predicts that a target word's recognizability is
determined according to a simple frequency-weighted neighborhood probability rule which is
defined by the ratio of the target word's prior probability to the summed prior probability of
all its DAS-linked neighbors. The NAM rule predicts a greater proportion of the variance in
spoken word recognition latencies (10-27%, depending on task [lexical decision, naming, or
identification in noise] and conditions [signal-to-noise ratio] [1]) than any other measure that
has been tested (e.g., log word frequency alone accounted for 5-10% of variance in Luce's
studies).
3
The focus of the NAM approach has typically been used to characterize the recognizability
of single words according to the sizes (densities) of their locally defined neighborhoods. More
recently, it has been realized that viewing the structure of the phonological lexicaon globally
as a complex network enables the probing of connections between both large and small scale
network topology and human spoken word recognition. Thus, rather than considering a word
and its neighbors in isolation, the set of neighbor relationships for an entire lexicon can be
represented as an unweighted, undirected graph [3] in which words (phonological forms) are
represented by nodes and two words are joined by an edge if they meet the standard NAM
DAS threshold. The NAM approach can be translated to the network context to mean
that (frequency-weighted) node degree is important for predicting latencies in spoken word
recognition. There are also prior indications that other topological properties (e.g. node
clustering coefficient [6, 7], closeness centrality [8], and second neighbor density [9]) may
also explain some aspects of SWR that the frequency-weighted neighborhood probability it
is based upon does not.
Previous studies have shown that what we will call the phonological neighbor network, or
PNN, for English has some features of both Watts-Strogatz [10] and Barabasi-Albert [11]
graphs.
It has a relatively short mean geodesic path length and high clustering coeffi-
cient, but also has a degree distribution that is at least partially power law [3]. Subsequent
analyses of additional languages (English, Spanish, Hawaiian, Basque, and Mandarin) have
shown these characteristics to be broadly shared across languages when PNN graphs are
constructed using NAM's DAS rule [12]. On the basis of these results, Vitevich and col-
leagues have assigned importance to these language "universals" and argued that many of
these properties are sensible if not essential (e.g. , high degree assortativity, which measures
the tendency of nodes to be connected to other nodes of similar degree, can buffer against
network damage) [12].
However, making claims about SWR on the basis of the properties of PNNs alone is
potentially fraught for at least two reasons. First, PNNs are static representations of lexical
structure, whereas spoken words are processed incrementally over time. Second, different
measures of word similarity will result in radically different PNNs. NAM's DAS rule is
based on a relatively simple string distance metric that provides a local measure of inter-
word similarity that is insensitive to the sequence of phonemes in a word. Thus, while
NAM's DAS metric accounts for substantial variance using a regression-based approach
4
(predicting response latencies for many words), there is substantial evidence from studies
examining competition between specific pairs of words with different patterns of position-
dependent phonological overlap that words whose onsets overlap compete more strongly
than words that are matched in DAS similarity but whose onsets are mismatched (e.g.,
battle would compete more strongly with batter than with cattle [13]). Marslen-Wilson
and colleagues [14, 15] proposed a threshold metric that gives primacy to onset similarity.
They focused on the notion (consistent with many priming and gating studies [15]) that the
"cohort" of words activated by a spoken word is restricted to words overlapping in their
first two phonemes. Thus, the cohort competitors of cat include not just DAS neighbors
overlapping at onset (can, cab, cast) but also longer words that would not be DAS neighbors
(cattle, castle, cabinet). In addition, the cohort metric predicts that rhyme (i.e. a word's
vowel and following consonants) neighbors (cat-bat, cattle-battle) do not compete because
they mismatch at onset, despite high DAS similarity. A PNN based on a simple onset
cohort rule (connect words that overlap in the first two phonemes) would obviously have
very different structure than a DAS-based PNN. When using PNNs to compare lexical
structure between languages, we must consider the potential role of the similarity metric
itself in determining the network's structure and topology. This possibility calls into question
any universal (language-independent) claims about SWR based on DAS networks. Prior
work has demonstrated that this is likely true at least in English, as PNNs constructed
from a random lexicon with the same phonological constraints as English are basically
indistinguishable from the real language network [16, 17].
Here, we explore this possibility further by extending DAS-based PNNs to four languages
in addition to English: Spanish, French, German, and Dutch. We show that PNNs for these
languages have degree distributions and topological properties similar to PNNs previously
constructed for English, Spanish, Hawaiian, Mandarin, and Basque [12]. We then show, by
separating words by number of syllables, that all five language networks consist of aggrega-
tions of at least two very different networks, as has been previously suggested for English [16].
We also note for the first time the effects of homophones such as bare and bear on PNN
structure. Finally, using a set of models that generate random lexicons with varying levels
of phonological realism, we show that even extremely simple random lexicons, along with
language-specific phoneme inventories and distributions of phonological form length (that is,
the frequency distributions of words of different lengths, ignoring all other linguistic details),
5
can create pseudo-PNNs that share all the properties of real PNNs. In fact, adding phono-
logical constraints (e.g., phonotactic constraints on phoneme sequences) does very little to
improve pseudo-PNN match to language-based PNNs. While these pseudo-PNNs are quite
insensitive to the level of realism in the lexicon, they are extremely sensitive to the empirical
form length distribution, which we show drives all of the observed differences among English,
French, German, Dutch, and Spanish. Our results suggest that the primary determinant of
the observed topology of PNNs is the neighbor definition itself, which dramatically limits
the network structures possible in PNNs. In addition, our work strongly motivates the con-
sideration of alternate phonological similarity metrics and suggests that it is important to
try to understand the formative dynamics underlying the observed phonological form length
distributions.
II. DATA
We used the freely available online CLEARPOND [18] database to construct DAS-based
PNNs for five languages: English, Dutch, German, French, and Spanish. CLEARPOND
is described in detail elsewhere [18], but in brief, it includes phonological transcriptions of
orthographic forms and frequency information for over 27, 000 words from each language.
Frequency information for English [19], Dutch [20], German[21], and Spanish [22] is derived
from the SUBTLEX database which counts word occurrences in television and movie subti-
tles. French frequency information is derived from Lexique [23], a fusion of an older French
language database (Frantext) with word occurrence information derived from webpages. For
all five languages we constructed PNNs based on the DAS rule described above: two words
were neighbors and therefore linked with a bidirectional, unweighted edge, if they differed by
no more than a single phoneme deletion, addition, or substitution. After PNN construction,
we found that, in each language, a significant percentage of the words had no phonological
neighbors, ranging from 24% (French) to 45% (Dutch). All singleton words were excluded
from any further analysis, since their topological properties are either trivial (e.g. they are
all degree zero) or undefined (e.g. clustering coefficient).
In all five languages, the mean
length of the neighborless words is larger than that of the words with neighbors, but this
difference is not statistically significant (permutation test).
6
III. EMPIRICAL ANALYSIS OF PHONOLOGICAL NEIGHBOR NETWORKS
A. Degree Distributions and Topology
Figure 1 shows the degree distributions for the five PNNs constructed from the CLEAR-
POND data (compare also to Figure 10 in the original CLEARPOND paper [18]), and Table I
gives a summary of some of the common topological measures employed in the empirical
analysis of networks, all of which have been specifically highlighted in prior PNN research.
All five language degree distributions are best fit (via maximum likelihood) by a truncated
power law, as tested via likelihood ratio [24]. In addition, we observe that all PNNs have: (i)
relatively high clustering, (ii) short mean geodesic paths, (iii) extraordinarily high values of
degree assortativity, and (iv ) relatively small giant connected components (the largest con-
nected subgraph in the network). Thus, all five PNNs have similar degree distributions and
topological characteristics, and they combine some features of Watts-Strogatz [10] graphs
(high clustering) with Barabasi-Albert graphs [11] (power law degree distribution). High
degree assortativity and small giant component sizes are features of the PNNs that are not
displayed by either WS or BA graphs. These features are all consistent with previous studies
on English alone [3, 16] and other languages not studied here [12].
The grouping of languages in Figure 1 is rather surprising. Essentially, Spanish is by
itself, Dutch and German have quite similar degree distributions, and English and French
are grouped together. One might expect different clustering based on language typology;
for example, with the two Romance languages (French and Spanish) grouped together. We
will show that the observed clustering can be explained without any reference to the specific
history of words. Instead, the structure of the phonological form length distribution, along
with target language phoneme frequencies, are all that is required.
B.
Islands and Frequency Assortativity
Given the relatively modest size of the giant connected component in all five languages
(see Table I), it is worth examining the connected component size ("island size") distribu-
tion Pc for each of the five PNNs. Power law distributions for the sizes of the connected
components Pc have been previously observed in PNNs for both English and Spanish [25].
Figure 2 shows that this power law distribution of component sizes is broadly shared over
7
N
m
¯k
GC size
EN
18983
76092
8.01
0.66
NL
15360
36158
4.71
0.56
DE
17227
41970
4.87
0.58
ES
FR
20728
21177
36111
145426
3.48
0.43
13.7
0.74
C (all/GC)
0.23/0.28
0.16/0.23
0.21/0.24
0.18/0.20
0.24/0.25
l
α
6.68
1.0∗
8.48
1.84∗
8.73
1.2∗
9.41
2.1∗
6.85
1.04∗
r (all/GC)
0.73/0.70
0.74/0.69
0.75/0.70
0.71/0.62
0.71/0.68
TABLE I. Topological measures for the five PNNs. Language names are abbreviated using their
two-letter ISO language codes: EN (English), NL (Dutch), DE (German), ES (Spanish), and FR
(French). N is the number of nodes, m is the number of edges, ¯k is mean degree, GC size is the
fraction of network nodes that are in the giant connected component, C is clustering coefficient, l
is mean geodesic path length, α is the power law exponent of the degree distribution, and r is the
degree assortativity coefficient. Two values occurring in the table with a forward slash denote that
quantity computed for the entire graph and only the giant component. Fits to degree distributions
were performed via maximum likelihood [24] starting at k = 2, except for French which began at
k = 10. Asterisks denote that the best fitting distribution is not strictly power law but rather
truncated power law, as determined via a likelihood ratio test [24].
all five languages. In fact, the island size distribution is more robustly power law than the
PNN degree distribution itself, albeit over a relatively modest range (less than a factor of
100).
We now remark on a previously unobserved feature of PNNs, again present in all five lan-
guages. All five languages show a weak but statistically significant degree of word-frequency
based assortativity. Simply, words of similar usage frequency tend to be connected to each
other in the PNN. We computed frequency assortativity by dividing the continuous word
frequency data into ten equal-mass bins and then computing an assortativity coefficient and
jackknife standard deviation using the definitions in Newman [26]. The values ranged from
0.1 in English to 0.24 in Spanish, which correspond to between 26 and 47 Jackknife standard
8
deviations. This is weak relative to degree assortativity in these networks (see Table I), but
not insignificant on the scale of assortativity coefficients found in other social, biological,
and technological networks [26].
C. DAS Graphs as Mixtures
There is a deep physical basis for observing power laws in thermodynamics. Diverging
length scales at critical points mean that there are correlations at all scales in the system.
Critical point behavior cannot depend on any quantity (like a force) with an associated
length scale, but rather only on scale-free quantities like symmetries and conservation laws.
Critical point phenomena then become universal, in the sense that the same behavior (critical
exponents) is observed in systems that may have radically different forces but the same set
of symmetries.
The converse is not true. Observation of power laws does not necessarily indicate any
deep phenomena at work. Power laws in empirical data can arise from a wide variety of
reasons, many of them mundane. One of the simplest is Simon's famous demonstration [27]
that multiplicative (rather than additive) random noise can yield heavy-tailed distributions.
Another way to obtain power laws is via mixture distributions; in this case apparent scale-free
behavior arises by simply mixing several distributions, each with well-defined but different
scales.
Indications that the degree distribution of the PNN for English results from a mixture
of distributions of different scales have been advanced by others [16]. Degree distributions
for English PNNs separately constructed from short and long (in phonemes) words showed
different shapes and, at least for short words, displayed markedly less power-law behavior.
In Figure 3 we show that this result also holds for the CLEARPOND English corpus, as
well as for Dutch, German, Spanish, and French. We divided all words in each corpus into
two classes: monosyllabic and polysyllabic.
Figure 3 clearly shows that connectivity among only monosyllabic words differs from
polysyllabic word connectivity. The monosyllabic degree distributions look less like power
laws than do the polysyllabic degree distributions, and monosyllabic words are in general
more densely connected than are polysyllabic words. This raises the possibility that the
PNN degree distribution may arise as a mixture of distributions.
In all five languages,
9
networks formed from polysyllabic words have degree distributions that are much closer to
(truncated) power laws than are the monosyllabic word networks. In addition, note that
(with the exception of French) the polysyllabic degree distributions are much more similar
across the five languages than the monosyllabic graph degree distributions or those of the
full graphs (see Figure 1)
In Appendix A, we look more closely at phonological neighbor graphs formed exclusively
from monosyllabic or polysyllabic words, and compare them to graphs containing all words
in each corpus (see Table VI). We found that some of the full PNN topological properties are
present in both the monosyllabic and polysyllabic networks (e.g., degree assortativity and
clustering coefficient). However, others are markedly different or disappear. The component
or "island" size distribution Pc is driven entirely by the polysyllabic words; the monosyllabic
words are almost completely connected (an unsurprising outcome of the DAS rule; shorter
words, such as cat, are much more likely to have DAS neighbors than long words like cat-
apult). The full PNN graphs have short (∼ 7) average path lengths primarily because the
monosyllabic graphs have extremely short average path lengths (∼ 5) and the polysyllabic
graphs have long (∼ 10) ones. When we compare the local properties of the monosyllabic
words in both the monosyllabic and full graphs, numbers of neighbors and second neigh-
bors are highly correlated. However, clustering is more weakly correlated, indicating that
explanations of latencies in SWR that appeal to node clustering [6] coefficient as a predictor
may be quite sensitive to whether or not polysyllabic words were included as items in the
experiment.
At least three questions remain. First, do constraints imposed by the one-step neighbor
DAS similarity measure explain the apparently universal topological features seen across all
five languages? If so, what explains the observed differences in the degree distributions in
Figure 1? Finally, how much lexical structure is required to generate PNNs that resemble
those of real languages? In what follows, we address these three questions in detail.
IV. PSEUDOLEXICONS
Figure 3 and additional results that we present in the Appendix A suggest that the
truncated power law behavior observed in the five PNNs might be the result of mixing
subgraphs with different connectivity properties. The left panel of Figure 4 again shows
10
the degree distributions for the five languages, this time with all homophones removed. We
discuss homophones in detail in Appendix B; in brief, we remove homophones because our
random lexicon models produce phonological forms (rather than written words) directly
and cannot properly account for homophones. The right panel shows the distribution Pl of
words of length l phonemes. The Pl distributions are underdispersed relative to Poisson (not
shown); note also that they are all zero-truncated, as there are no words in any language that
consist of zero phonemes. A particularly intriguing feature of the five language Pl is that they
cluster similarly to the degree distributions shown in Figure 1. English and French together,
then German and Dutch, and Spanish by itself. While this could be entirely coincidental or
a result of previously undetected cross-linguistic similarities, below we will show that it is
not.
A. Models
To determine which topological features of the PNNs arise due to specific features of real
languages and which are driven purely by the DAS connection rule, we adopt and extend
an approach inspired by previous work on the English PNNs [16]. We generate corpora of
random phonological forms using generative rules that include varying amounts of real lin-
guistic detail. We denote such a corpus of random strings of phonemes a pseudolexicon. Each
pseudolexicon is paired with a target language, since all the models use some information
from the real language for construction. Specifically, pseudolexicons are created from the
phonemic inventory of each language (the set of all phonemes that occur in the language),
with lexicon size constrained to be approximately the same as the real-language lexicon for
the target language (for example, about 22,000 unique words [i.e., excluding homophones]
for English CLEARPOND), and with the same form length distribution as the target lan-
guage. To match the length distribution, the length of each random string is first specified
by drawing a random integer from a form length distribution Pl defined on the positive
integers excluding zero. In all cases, the pseudolexicon has a form length distribution which
we specify. Specifically, we consider the following six models for pseudolexicons. We have
named the models using terminology taken from the Potts [28] and Ising [29] models. Each
includes progressively greater language-specific detail relating to phonological structure. We
expected that we would get better a successively better match to a given target-language
11
PNN as we included more detail.
• Infinite Temperature (INFT). Each phoneme in the string is drawn uniformly
from the target language's phoneme inventory.
• Noninteracting, Uniform Field (UNI). Each phoneme in the string is drawn
randomly using its observed frequency in the real language's lexicon.
• Noninteracting, Consonant/Vowel Uniform Field (CVUNI). Each position in
the random string is either a consonant or a vowel drawn randomly using observed
positional consonant/vowel frequencies in the real lexicon. Specifically, we use the real
language's corpus to compute the position-dependent probability that position l is a
consonant or a vowel. The particular consonant or vowel placed at that position is
drawn uniformly from the respective set of items (lists of consonants and vowels).
• Noninteracting, Consonant/Vowel Field (CV). Positions are selected to be con-
sonants or vowels exactly as in CVUNI. The particular consonant or vowel placed at
each position is selected using observed frequencies of consonants and vowels from the
real lexicon.
• Noninteracting, Spatially Varying Field (SP). Each phoneme is drawn randomly
from real positional frequencies in the target lexicon. For example, if a language has
an inventory of twenty phonemes, we use the real lexicon to compute a πl,x that gives
the probability that phoneme x occurs at position l, and then use this table to assign
a phoneme to each position of the random string.
• Nearest Neighbor Interactions (PAIR). The first phoneme in each string is drawn
using a positional probability. Subsequent phonemes are drawn via the following rule.
If the phoneme at position k is x, then the phoneme at position k + 1 is drawn using
the empirical probability (from the real lexicon) that phoneme x(cid:48) follows phoneme x.
We have listed the models in rough order of complexity; INFT uses the least amount of
information about the real language's structure and PAIR the most. We note that while
it is possible to generate real words (particularly short ones) from the models above, the
vast majority of the strings produced bear no resemblance to real words in any of the five
12
languages. The only model that avoids unpronounceable diphones is PAIR; in the other
models unpronounceable diphones occur frequently.
For each pseudolexicon, we discarded any duplicate items. This is why we removed
homophones from the real languages; we did not generate orthographic tags for the random
phonological forms, so duplicated forms in the pseudolexicon all represent a single node. We
then formed a pseudo-PNN by using the DAS rule to connect items in the pseudolexicon to
one another. As with the real PNNs, before any analysis we discarded nodes in the pseudo-
PNNs with degree zero. Figure 5 shows the degree distribution of the Francis & Kucera 1982
English corpus (FK) [30] and its six corresponding pseudolexicons. We first show the fit to
FK, rather than CLEARPOND English, due to our ability to better control the contents of
the FK corpus (see Appendix B) for details). Each of the six pseudolexicons had as its input
Pl the empirical English Pl (e.g. Figure 4, right panel). We note that, while the sizes of the
pseudolexicons were fixed to the real-language target lexicon, once the pseudonetworks are
formed, they may have fewer nodes than this, since many pseudowords may be neighborless
and hence not appear in the graph.
B. English Networks
Figure 5 shows that even minimal levels of linguistic realism yield a pseudo-PNN with
a stikingly similar degree distribution to the real English PNN. Even UNI, which includes
nothing beyond overall phoneme frequencies and the empirical Pl, looks quite similar to FK.
Table II, which lists the same topological properties we previously showed in Table I tells an
even more compelling story. First, the putatively lingustically relevant topological properties
discussed earier - high clustering, short mean path length, high degree assortativity, and (to
some extent) small giant components - are present in all of the pseudolexicons whose degree
distributions match that of FK. Giant component size is the least well-matched property
in all of the models, though it is still smaller than observed in many real-world networks.
Furthermore, even INFT, in which degree distribution (and hence mean degree) is a poor
match to FK, has high clustering, short mean path length, and high degree assortativity.
INFT includes almost nothing about the target language except the form length distribution
and the phoneme inventory. We also note the noisiness in the degree distribution of INFT.
While one might hypothesize that this is a result of its relatively small size, the degree
13
distribution of INFT does not become smooth even for larger (10,000 node) graphs (not
shown).
FK INFT UNI CVUNI CV SP PAIR
N
m
¯k
7861
1891
2922
22745
2841
7501
5.79
3.0
1947
3532
3.63
0.80
0.22
6.65
1.0∗
0.64
3022
3139
4346
8687
8811
12319
5.74
5.61
5.67
0.85
0.85
0.87
0.27
0.25
0.25
5.30
1.0∗
5.40
1.0∗
5.63
1.0∗
0.48
0.49
0.45
5.13
0.85
0.25
5.34
1.0∗
0.45
GC size
0.69
C
l
α
r
0.21
6.38
1.0∗
0.67
0.77
0.19
7.26
1.0
0.60
TABLE II. Topological measures for the FK English corpus and six pseudolexicons matched to it.
All rows of the table are as described in Table I.
Figure 6 and Table III shows the same information for the CLEARPOND English
database and pseudolexicons matched to it. We note first that all the conclusions that held
for FK hold for CLEARPOND English. Again, even a model as naive as UNI has a very
similar degree distribution to the real English PNN and very similar topological character-
istics. INFT, again despite having a degree distribution that is an extremely poor match to
English CLEARPOND, has high clustering coefficient and high degree assortativity. Com-
pared to FK, some differences are evident. Chiefly among them is that all the models now
have too low of a mean degree, arising because the model degree distributions have large-k
tails that are too short. However, given the analysis and discussion in Appendix B, this is
to be expected. As discussed there, our models do not include analogs to inflected forms
(e.g., WALK, WALKS, WALKED). We also have not attempted to model homophones
(which have been removed in our pseudo-lexicon PNNs) or proper nouns. All three of these
item types preferentially affect the tail shape of Pk. We also note that the CLEARPOND-
matched pseudolexicons tend (except for SP) to undershoot the English giant component
size, though they still match the fundamental observation that the GC is a relatively small
portion of the full network.
14
EN INFT UNI CVUNI CV
SP
PAIR
18252
3192
5911
59965
3748
13281
N
m
¯k
6.6
GC size
0.65
C
l
α
r
0.21
6.81
1.0∗
0.70
2.35
0.36
0.19
16.7
1.0∗
0.83
4.49
0.41
0.25
6.53
1.0∗
0.74
3942
6857
3.48
0.47
0.22
9.43
1.0∗
0.85
6219
7098
8705
16373
18821
20922
5.27
0.38
0.27
5.77
1.0∗
0.68
5.31
0.63
0.25
10.7
1.0∗
0.71
4.81
0.35
0.25
9.35
1.0∗
0.72
TABLE III. Topological measures for the CLEARPOND English corpus (EN) and six pseudolex-
icons matched to it. All rows of the table are as described in Table I.
C. Five Language Pseudonetworks
We now compare pseudo-PNNs to real PNNs for all five languages: English, Spanish,
Dutch, German, French. For this comparison, we used only the UNI model, since it has
a very similar degree distribution to the English PNN Pk despite containing almost no
information about real language phonology and constraints. In each case, the pseudo-PNN
is matched in total corpus size and form length distribution to its target language. The left
panel of Figure 7 shows the true degree distributions for the five language PNNs (shown
also in Figure 4) and the right panel of Figure 7 shows the pseudo-PNNs using the UNI
model. Furthermore, Table IV shows topological parameters for Spanish, French, German,
and Dutch and their matched UNI pseudo-PNNs. We omit English in Table IV because
that information is contained in Table III.
Figure 7 and Table IV together show that, as in English, the UNI model is able to come
remarkably close in shape and topological properties to the real phonological neighbor net-
works, despite not resembling the real language's phonology in any way. The clustering of
the five language degree distributions for the pseudo-PNNs mimics that seen in the real
PNNs, particularly in the manner in which Spanish is separated from the other languages.
Given the way the UNI pseudo-PNNs were constructed, this grouping must be driven en-
tirely by the form length distribution. Table IV shows that the pseudo-PNNs match their
15
target languages quite well overall, with some properties extremely similar, e.g. clustering
coefficients and degree assortativity.
FR pFR
ES
pES DE pDE NL
pNL
12164
7854
20018
2198
16787
4141
14943
3938
32753
21577
31812
2852
35402
8749
31697
9408
5.49
0.36
0.27
7.49
1.0∗
0.75
3.16
2.60
4.17
4.23
4.24
4.79
0.43
0.32
0.57
0.33
0.55
0.57
0.19
0.19
0.21
0.25
0.16
0.27
9.49
1.0∗
9.99
1.0∗
8.88
1.0∗
5.34
1.0∗
8.5
1.0∗
11.59
1.0∗
0.70
0.65
0.71
0.67
0.73
0.66
N
m
¯k
5.38
GC size
0.72
C
l
α
r
0.28
7.13
1.0∗
0.59
TABLE IV. Topological measures for four phonological neighbor networks (FR, ES, DE, NL)
and matched UNI pseudo-PNNs (pFR, pES, pDE, pNL). All rows of the table are as described in
Table I.
In Figure 2 we showed that the component size distributions for all five language PNNs
follow a power law, even moreso than the degree distributions for the PNNs themselves.
This has previously only been observed in English and Spanish [25]. However, even these
component size distributions do not arise out of any fundamental or universal phonological
properties. In the left panel of Figure 8 we reprint Figure 2 to allow easy comparisons. In
the right panel we show component size distributions for the five pseudo-PNNs. While the
span of Pc is somewhat reduced in the pseudo-networks, all the pseudographs clearly have
power law size distributions with exponents similar to their target languages. Thus, even
the island size distribution is essentially an artifact of the neighbor definition.
D. Sensitivity to the Form Length Distribution
The previous section demonstrates that the topological properties of phonological neigh-
bor networks constructed using the one-step DAS rule are driven not by any real linguistic
feature but by the connection rule itself. While the resulting PNNs are remarkably insensi-
tive to the degree to which real phonological constraints are used in their construction, we
16
have also shown that the PNNs are sensitive to the shape of the form length distribution
Pl. In this section, we investigate that sensitivity further. We do that by generating four
more UNI lexicons with different input form length distributions and compare the resulting
pseudo-PNNs to the FK English database. The four form length distributions are as follows.
• EMP. Pl is the empirical form length distribution for FK, exactly as in Figure 5 and
Table II.
• ZTP(1x). Pl is a zero-truncated Poisson (ZTP) model fit to the empirical distribution.
The ZTP distribution has the form
PZTP(k; λ) =
λk
(eλ − 1)k!
,
(1)
which assuming independence among the empirical length values xi leads to a model
N(cid:89)
λli
i=1
(eλ − 1)li!
likelihood
L(λ) =
(2)
in which λ can be determined via numerical maximization L(λ).
• ZTP(1.5x) This model is a zero-truncated Poisson model for Pl with a mean equal
to 1.5 times the mean of the ML λ of ZTP-1X.
• GEO. Here Pl follows a geometric distribution
PGEO(k; p) = p(1 − p)k−1
(3)
for which the parameter p is chosen to make the mean of GEO equal to the mean of
the empirical English Pl.
We chose ZTP(1x) as a simple but relatively poor approximation to the form length
distributions of the real languages; the real form length distributions are all underdispersed
relative to Poisson. ZTP(1.5x) as compared to ZTP(1x) is similar to the difference between
the Spanish Pl and those of English and French (see Figure 4). GEO is chosen to have an
identical average length to English phoneme strings, but otherwise has a shape radically
different than any Pl we observe. Figure 9 shows degree distributions for pseudo-PNN
constructed using UNI pseudolexicons, with each of these four choices for Pl, along with
the degree distribution of FK for comparison. The inset in Figure 9 shows the form length
17
FK EMP ZTP(1x) ZTP(1.5x) GEO
7861
2959
2753
22745
7954
13127
N
m
¯k
5.79
GC size
0.69
C
l
α
r
0.21
6.38
1.0∗
0.67
5.38
0.85
0.24
5.19
1.0∗
0.44
9.54
0.85
0.28
4.48
1.0∗
0.53
211
304
2.88
0.64
0.19
4.71
1.74
0.46
3592
35938
20.0
0.95
0.35
3.73
1.0∗
0.49
TABLE V. Topological measures for four UNI pseudo-PNNs (EMP, ZTP-1X, ZTP-1.5X, GEO)
and the real FK phonological neighbor network. All rows of the table are as described in Table I.
distribution used to produce the pseudolexicon yielding the PNN in the main panel. Table V
compares the topological properties of those four pseudo-PNNs to FK and each other.
It is clear from Figure 9 that the shape of the PNN degree distribution is extremely
sensitive to the form length distribution. Even the relatively small differences in the shape
of EMP and ZTP(1x) lead to large changes in the tail mass of the degree distribution.
The difference between the degree distribution of ZTP(1x) and ZTP(1.5x) is similar to the
difference between the degree distributions of English or French and Spanish (see Figure 4).
In addition, Table V shows that the PNN made from ZTP(1.5x) is much smaller (fewer nodes
and edges) than any of the other models. This is expected given the reduction in probability
of short phonological forms in ZTP(1.5x) when compared to EMP, ZTP(1x), or GEO; the
probability that two strings from the UNI pseudolexicon that differ in length by one unit or
less are neighbors decays exponentially with string length. Note also from Table V that no
matter what effect Pl has on the degree distribution of the resulting PNN, all graphs show
high clustering coefficients, short mean free paths, and high degree assortativity.
V. CONCLUSION
We have shown that observed "universal" topological features of phonological neighbor
networks [12] - truncated exponential degree distributions, high clustering coefficients,
18
short mean free paths, high degree assortativity and small giant components - are string
rather than language universals. That is, inferences from networks based on similarity re-
garding language ontogeny or phylogeny are suspect, in light of our analyses demonstrating
that similar network structures emerge from nearly content-free parameters. One might
object to this strong interpretation. The DAS rule obviously captures important relations
that predict significant variance in lexical processing due to similarity of phonological forms
in the lexicon. Networks based on DAS are able to extend DAS's reach, as was previously
demonstrated with clustering coefficient [6, 7]. Note, though, that clustering coefficient re-
lates to familiar concepts in word recognition that have not been deeply explored in the
spoken domain: the notion of neighbors that are friends or enemies at specific positions, dis-
cussed by McClelland and Rumelhart in their seminal work on visual word recognition [31].
Consider a written word like make, with neighbors such as take, mike, and mate. Take is an
enemy of the first letter position in make, but a friend at all other letter positions, where
it has the same letters. A written word with a clustering coefficient approaching 1.0 would
have many neighbors that all mismatch at the same position (thus making them neighbors
of each other). A word with a similar number of neighbors but a low clustering coefficient
(approaching N/L, that is, N neighbors evenly distributed of L [length] positions) would
have more evenly distributed neighbors. For spoken word recognition, the results of Chan
and Vitevitch [6] suggest that a high clustering coefficient exacerbates competition because
it is heavily loaded on a subset of phoneme positions, creating high uncertainty.
In our
view, this reveals important details about phonological competition, but not ontogeny or
phylogeny of English, or other specifically linguistic structure. Indeed, given the similarity
in the distribution of clustering coefficients (among other parameters) in English and in our
abstract PNNs, we interpret instances of (e.g.) high clustering coefficient as string universals
rather than language universals.
While phonological neighbor network topology is largely insensitive to the degree of real
phonological structure in the lexicon used to construct the neighbor network, we found some
amount of sensitivity to the input form length distribution Pl. Even relatively subtle changes
in Pl can lead to observable changes in the degree distributions of the resulting neighbor
networks, and differences among the five languages we studied here can be almost wholly
attributed to differences in form length distributions among the five languages. However,
even this sensitivity is only partial. Form length distributions that look nothing like any
19
of the languages we consider here (GEO, although GEO may partially resemble the Pl of a
language like Chinese), that generate network degree distributions that we do not observe,
still yield high clustering coefficients, short mean free paths, and high degree assortativity.
The question of what leads to a given language's Pl is a question about language evolution
that will be much more difficult to explain, though some parallels might be drawn with work
that seeks to understand the evolution of orthography [32–34].
At an even deeper level, it may be perilous to attach too much meaning to the topology
of any similarity network of phonological forms, at least with respect to human performance
in psycholinguistic tasks. This is because these networks do not "do" anything; they have no
function. They are not connectionist networks that attempt to model phoneme perception,
like TRACE [35] or TISK [36]. No matter how they are constructed, they are basically static
summaries of the structure of the speech lexicon; they do not perform a processing function.
Insofar as the similarity measure aligns with latency data from human spoken words tasks
(e.g. picture naming [7], lexical decision [6], etc.), network properties may encode some
features of human performance. While there is evidence that some aspects of human task
performance may be predicted from features of neighbor networks [3, 6–9], it is clear from our
study that care must be taken in interpreting the results of studies of phonological networks.
If the static structure of the lexicon were to be paired with a dynamics that represents mental
processing, it would be possible to test the utility of phonological similarity networks for
explaining human performance in psycholinguistic tasks.
ACKNOWLEDGEMENTS
We acknowledge the seed grant program from the Connecticut Institute for Brain and
Cognitive Sciences (CT IBaCS) for supporting this work.
[1] P. A. Luce, Research on Speech Perception, Technical Report No. 6: Neigborhoods of words
in the mental lexicon, Tech. Rep. (Speech Research Laboratory, Department of Psychology,
Indiana University, 1986).
[2] P. A. Luce and D. B. Pisoni, Ear Hear. 19, 1 (1998).
[3] M. S. Vitevitch, J. Speech Lang. Hear. Res. 51, 408 (2008).
20
[4] C. A. Fowler and J. S. Magnuson, in The Cambridge Handbook of Psycholingustics, edited by
M. Spivey, K. McRae, and M. Joanisse (Cambridge University Press, 2012) pp. 3–25.
[5] J. S. Magnuson, D. Mirman, and E. Myers, in The Oxford Handbook of Cognitive Psychology,
edited by D. Reisberg (Oxford University Press, 2013) pp. 412–441.
[6] K. Y. Chan and M. S. Vitevitch, J. Exp. Psychol. Hum. Percept. Perform. 35, 1934 (2009).
[7] K. Y. Chan and M. S. Vitevitch, Cogn. Sci. 34, 685 (2010).
[8] S. R. S. Iyengar, C. E. V. Madhavan, K. A. Zweig, and A. Natarajan, Top. Cogn. Sci. 4, 121
(2012).
[9] C. S. Q. Siew, Psychon. B. Rev. 24, 496 (2017).
[10] D. J. Watts and S. H. Strogatz, Nature 393, 440 (1998).
[11] A. L. Barab´asi and R. Albert, Science 286, 509 (1999).
[12] S. Arbesman, S. H. Strogatz, and M. S. Vitevitch, Int. J. Bifurcation Chaos 20, 679 (2010).
[13] P. D. Allopenna, J. S. Magnuson, and M. K. Tanenhaus, J. Mem. Lang. 38, 419 (1998).
[14] W. D. Marslen-Wilson and A. Welsh, Cognitive Psychol. 10, 29 (1978).
[15] W. D. Marslen-Wilson, in Cognitive Models of Speech Processing: The Second Sperlonga Meet-
ing, edited by G. T. M. Altmann and R. Shillock (Erlbaum, 1993) pp. 187–210.
[16] T. M. Gruenenfelder and D. B. Pisoni, 52, 596 (2009).
[17] M. Stella and M. Brede, J. Stat. Mech.: Theory Exp. 5, P05006 (2015).
[18] V. Marian, J. Bartolotti, S. Chabal, and A. Shook, PLOS ONE 7, e43230 (2012).
[19] M. Brysbaert and B. New, Behav. Res. Methods 41, 977 (2009).
[20] E. Keuleers, M. Brysbaert, and B. New, Behav. Res. Methods 42, 643 (2010).
[21] M. Brysbaert, M. Buchmeier, M. Conrad, A. M. Jacobs, J. Bolte, and A. Bohl, Exp. Psychol.
58, 412 (2011).
[22] F. Cuetos, M. Glez-Nosti, A. Barb´on, and M. Brysbaert, Psicol´ogica 32, 133 (2011).
[23] B. New, C. Pallier, M. Brysbaert, and L. Ferrand, Behav. Res. Methods Instrum. Comput.
36, 516 (2004).
[24] A. Clauset, C. R. Shalizi, and M. E. J. Newman, SIAM Rev. 51, 661 (2009).
[25] S. Arbesman, S. H. Strogatz, and M. S. Vitevitch, Entropy 12, 327 (2010).
[26] M. E. J. Newman, Phys. Rev. E 67, 026126 (2003).
[27] H. A. Simon, Biometrika 42, 425 (1955).
[28] R. B. Potts, Math. Proc. Cambridge Philos. Soc. 48, 106 (1952).
21
[29] E. Ising, Z. Phys. 31, 253 (1925).
[30] W. N. Francis and H. Kucera, Frequency analysis of English usage: Lexicon and Grammar
(Houghton Mifflin, 1982).
[31] J. L. McClelland and D. E. Rumelhart, Psychol. Rev. 88, 375 (1981).
[32] M. H. Christiansen and S. Kirby, Trends Cogn. Sci. 7, 300 (2003).
[33] J. B. Plotkin and M. A. Nowak, J. Theor. Biol. 205, 147 (2000).
[34] R. F. i Cancho and R. V. Sol´e, Proc. Natl. Acad. Sci. USA 100, 788 (2003).
[35] J. L. McClelland and J. L. Elman, Cognitive Psychol. 18, 1 (1986).
[36] T. Hannagan, J. S. Magnuson, and J. Grainger, Front. Psychol. 4, 563 (2013).
Appendix A: Syllable Level Analysis
To deterimine the number of syllables, we count vowels and dipthongs in the phonological
transcription of each word. In addition, we correct for words that end in a phonological 'l'
with no vowel preceeding the final phoneme. For example, the English word able has only
a single vowel but is a two-syllable word. We note here that syllable boundaries are much
harder to determine, but we do not need to decompose the word into its constituent syllables
for any of our analysis.
For each language we built two additional graphs: one for the monosyllabic (MS) words
only and one for the polysyllabic (PS) words only. Just as in the full PNN, these two new
graphs used the DAS rule to determine if two words should be connected by an edge. We
show results for English and Dutch are in Table VI. Table VI shows that the topological
properties of the PNNs arise by mixing two very different kinds of graphs. For quantities like
the clustering coefficient and degree assortativity, this mixing is very mild. The MS graphs
tend to cluster more strongly than the PS graphs, and vice versa for degree assortativity,
but the differences are not extreme. This is not the case for the rest of the topological
measures. Despite having far fewer nodes, the MS graphs have tenfold greater edge density.
The MS graphs are almost completely connected; all "islands" in the English and Dutch
PNNs are induced by the structure of PS graphs. Mean geodesic paths are quite short in
the MS graphs and long in the PS graphs. The MS graphs do not have power law degree
distributions at all; that arises due to mixing with the PS graphs (all power-law or truncated
power law) in the full graph.
22
EN MS+PS EN MS
EN PS NL MS+PS NL MS
NL PS
N
m
d
¯k
GC size
18983
76092
0.0004
8.01
0.66
5979
50232
0.003
16.8
0.98
13004
19808
0.0002
3.0
0.46
15630
36158
0.0003
4.71
0.31
2808
16785
0.004
11.96
0.97
12552
18396
0.0002
2.93
0.43
C
l
α
r
rf
0.23/0.28
0.3/0.3
0.19/0.26
0.16/0.23
0.31/0.30
0.13/0.20
6.68
1.0∗
4.63
-
10.3
1.04∗
4.62
1.84∗
11.8
-
8.73
1.72
0.73/0.70
0.65/0.65
0.74/0.66
0.74/0.69
0.59/0.59
0.74/0.65
0.104(4) [26σ] 0.068(4) [15σ] 0.089(7) [13σ] 0.126(5) [25σ] 0.055(8) [7σ] 0.100(7) [14σ]
TABLE VI. Topological measures for graphs produced from the CLEARPOND English and Dutch
corpora. MS+PS is the full PNN (see also Table I), MS is a graph formed from only the monosyl-
labic words, and PS a graph formed from only the polysyllabic words. With the exception of edge
density d and frequency assortativity coefficient rf , all symbols in this table are the same as those
in Table I, and the quantities in the tabhle separated by forward slashes have the same meaning
as in Table I. Edge density is defined as 2m/N (N − 1), where m is the number of edges and N the
number of nodes in the graph.
We also compared node-level topology for the MS words in the MS only graph and the full
PNN (MS+PS). Most quantities are almost perfectly correlated for these two: these include
number of neighbors (degree), number of second neighbors, and eigenvector centrality. All
of these quantites are highly correlated with R2 ≥ 0.95. Node clustering coefficient for the
MS words in the two English graphs is more weakly similar (R2 = 0.8), with large outliers
(see Figure 10). It would be interesting to revisit the proposed relationship between node
clustering and spoken word recognition [6] facility in light of these findings.
When we performed the same syllable-level calculations for the other three languages in
the CLEARPOND database, we find a consistent story (results not shown). In all cases,
MS giant component sizes are much larger than PS GC sizes, MS edge densities are close to
tenfold larger, and MS mean geodesic path lenghts are much shorter. PS degree distributions
23
are well-fit by truncated power laws and have much more consistent power law exponents
than we see for the full PNNs in Table I. All five languages except Dutch have power law
exponents in the range 1.0 − 1.04. Dutch is better fit by a non-truncated power law with
exponent 1.72 (see Table VI). Furthermore, clustering coefficient and degree assortativity
are similar in the MS and PS graphs, just as in English and Dutch. As in English and
Dutch, clustering coefficients are larger for the MS graphs in German, Spanish, and French.
Appendix B: Lexical Issues
In this section we discuss some lexical issues that are relevant to the pseudolexicon models
we construct in Section IV. Table VII shows the thirteen most highly connected words in
English CLEARPOND. What should be clear from Table VII is that proper nouns (Lowe)
and homophones (see,sea) are overrepresented. Our pseudolexicon models directly generate
phonological forms with no orthographic tags; we thus cannot represent homophones in our
models and must remove them from the PNN graphs for comparison. Another category
of words not represented in Table VII that we cannot easily model is inflected forms, for
example word-final 's' for plurals in English. Lemmas and their inflected forms occur much
more frequently than expected at random, particular for longer (multisyllabic) words. We
discuss the effects of these three categories of words (homophones, inflected forms, and
proper nouns) on the resulting PNN degree distribution, and explain our protocol for their
removal.
We began with the Francis & Kucera 1982 English corpus [30] (hereafter FK), which
consists of the Brown corpus of English words, along with prononuciation (phonological
transcription) for every item and an indication as to whether the item is a lemma or an
inflected form. We used FK rather than CLEARPOND here because, as shown below,
we lack some of this word-level information in CLEARPOND and cannot remove all three
categories of words in the CLEARPOND PNNs. We successively removed inflected forms,
proper nouns, and homophones from FK as follows.
• Proper Nouns. Any word whose orthographic (written) form begins with a capital
letter is assumed to be a proper noun. This rule applies equally well to the PNNs
for FK and CLEARPOND. However, we emphasize here that because of the rules for
capitalization in German (all nouns are capitalized), we cannot systematically remove
24
word
degree
Lea
Lee
Lew
loo
lieu
Lai
lye
lie
Lowe
low
male
see
sea
68
68
66
66
66
63
63
63
62
60
60
60
60
TABLE VII. The thirteen words in English CLEARPOND with the highest degree. Note the
prevalence in this list of (i ) proper nouns and (ii ) homophones (e.g. ,see,sea).
proper nouns for all five languages in CLEARPOND.
• Inflected Forms. FK includes lemma numbers for all the words, so we can simply
remove any words that are not lemmas. We do not have this information for any words
in CLEARPOND and thus cannot remove them. To try to remove inflected forms in
CLEARPOND we could, for example, remove all words with word-final phonological
'z'. This would remove English plurals but also improperly remove some lemmas
(size). Even if this were desirable, we would need different rules for all five languages.
Therefore we are forced to keep all inflected forms in the CLEARPOND PNNs.
• Homophones Homophones are items with identical phonological transcriptions but
different orthography. These are relatively simple to remove in both FK and CLEAR-
POND English, and the same procedure works in any language. We search the nodes
for sets of items with identical phonological transcriptions. For example, see and sea
would comprise one homophone set in English, and lieu, loo, and Lou another. One
25
of the items from each homophone set, chosen at random, is kept in the PNN and the
nodes corresponding to all other items in the set are deleted.
Figure 11 shows the degree distribution of the FK PNN when inflected forms, proper
nouns, and homphones were successively removed. Two features of this figure deserve men-
tion. First, the main effect of these classes of words is in the tail of the degree distribution.
Secondly, removal of inflected forms causes very little change compared to removal of proper
nouns and homophones. It is relatively easy to understand why the largest changes to the
degree distribution occur at large k, at least for homophones. Consider a single orthographic
form w that is also a homophone with degree d. All of the other orthographic forms in its
homophone set are connected to both w and all of the d neighbors of w. If there are N
words in the homophone set, we end up with N nodes each with degree d + N − 1. Thus,
homophone sets can boost the degree of both their neighbors (since a neighbor of one is a
neighbor of all other words in the set) and the homophones themselves. As an example, a
homophone set of size 10 in which one of the words has 10 neighbors yields 10 nodes with
degree 19. Removing members of the homophone set will therefore tend remove nodes of
large degree and therefore shift the tail of Pk.
Figure 12 compares removal in English CLEARPOND to FK. We first note that, despite
being based on completely different corpora, the unaltered English CLEARPOND and FK
yield similar PNNs. In addition, as in FK, removal of homophones and proper nouns in
CLEARPOND tends to truncate the tail of the degree distribution. As we noted above, the
only class of words that we can consistenly remove from all five CLEARPOND languages is
homophones, and we remove these for all model comparisons. The number of homophone
sets, mean set size, and the number of nodes removed from the graph when the removal
procedure described above is implemented, for each of the five languages in CLEARPOND
is shown in Table VIII. Table VIII indicates that for all languages except French, the majority
of homophone sets are pairs like see and sea.
26
Language
EN
DE
ES
FR
NL
NH
731
440
1059
9013
417
µH
2.09
2.10
2.03
2.63
2.08
Nodes Removed
795
485
1123
14735
449
TABLE VIII. Number of homophone sets NH , mean homophone set size µH and the number of
nodes removed from the CLEARPOND PNNs when homophones are removed. Note the wide
variation in the number of homophones across the five languages.
27
FIG. 1. Log-log plot of the logarithmically binned phonological neighbor network (PNN) degree
distribution Pk for five languages: English (EN), Spanish (ES), French (FR), Dutch (NL), and
German (DE).
28
100101102k10−410−2100PkENESFRNLDE FIG. 2. Size distributions Pc vs. c for the connnected components ("islands") in all five languages.
Languages are abbreviated using their two-letter ISO codes (see Table I). The giant component
has been excluded from this figure for all five languages; it sits far to the right for each language.
The minimum island size is two because we have removed any singleton nodes (loners) from the
PNNs, as discussed in Section II.
29
100101102c10−510−310−1PcENESFRNLDE FIG. 3. Degree distributions Pk vs. k for PNNs formed from exclusively monosyllabic (left panel)
or polysyllabic (right panel) words in each lexicon. Each language is abbreviated by its two letter
ISO code; see the caption to Table I for the key to these codes.
30
100101102k10−310−210−1100PkENESFRNLDE100101102k10−310−210−1100PkENESFRNLDE FIG. 4. The left panel shows the degree distributions Pk versus k for the five CLEARPOND PNNs.
Compare to Figure 1; this figure differs because homophones have been removed from the graphs
as detailed in Appendix B. The right panel shows the distribution Pl of phonological form lengths
in each of the five languages from the CLEARPOND corpora. Note that all these distributions are
only defined for l ≥ 1; length zero words do not exist.
31
100101102k10−410−2100PkENESFRNLDE051015l0.00.10.2PlENESFRNLDE FIG. 5. Degree distributions Pk versus degree k for the Francis and Kucera 1982 corpus (FK)
along with the six pseudolexicons fit to it. See the text for a key to the abbreviations for the
pseudolexicons.
32
100101102k10−310−210−1PkFKINFTUNICVUNICVSPPAIR FIG. 6. Degree distributions Pk versus degree k for the CLEARPOND English corpus (EN)
along with the six pseudolexicons fit to it. See the text for a key to the abbreviations for the
pseudolexicons.
33
100101102k10−310−210−1PkENINFTUNICVUNICVSPPAIR FIG. 7. The left panel shows the degree distributions Pk versus k for the five CLEARPOND PNNs.
This is a reprint of the left panel of Figure 4, so that comparisons may be more easily made. The
right panel shows degree distributions for pseudo-PNNs, each of which is produced using the UNI
model (see Section IV A) and matched to the target language.
34
100101102k10−510−310−1PkDEENNLFRES100101102k10−510−310−1PkpDEpENpNLpFRpES FIG. 8. The left panel shows the component size distribution Pc versus c (compare to Figure 2).
The right panel shows component size distributions for pseudo-PNNs, each of which is produced
using the UNI model (see Section IV A) and matched to the target language.
35
100101102c10−510−310−1PcDEENNLFRES100101102c10−510−310−1PcpDEpENpNLpFRpES FIG. 9. Degree distributions (main panel) for UNI pseudo-PNNs constructed using four different
phonological length distributions: the empirical English form length distribution (EMP), a zero-
truncated Poisson fit to the empirical distribution (ZTP(1x)), a zero-truncated Poisson with shifted
mean (ZTP(1.5x)), and a geometric distribution (GEO) with the same mean as EMP. The real FK
network is shown for comparison, and the inset shows the four different form length distributions.
36
100101102k10 310 210 1PkFKEMPZTP(1x)ZTP(1.5x)GEO051015l0.000.050.100.150.20PlEMPZTP(1X)ZTP(1.5X)GEO FIG. 10. Clustering coefficient for each MS node in the MS graph (x-axis) and in the full English
CLEARPOND PNN (y-axis). The R2 of the correlation between the two sets of values is 0.8.
37
0.00.51.0Ci(MS)0.00.51.0Ci(Full) FIG. 11. Double log plot of the degree distribution Pk against k for the FK phonological neighbor
network (black circles) when inflected forms (magenta), proper nouns (cyan), and homophones
(yellow) are successively removed.
38
100101k10−310−210−1PkFKFK-inflFK-infl-propFK-infl-prop-homo FIG. 12. Double log plot of the degree distribution Pk against k for the CLEARPOND English
PNN and FK when various classes of words are removed. The CLEARPOND English PNN de-
gree distribution is show unaltered (green) and after homophones and proper nouns are removed
(magenta), while FK is show unaltered (cyan) and when homophones, proper nouns, and inflected
forms are all removed (red).
39
100101k10−410−310−210−1PkCPCP-homo-propFKFK-homo-prop-infl |
1807.09194 | 1 | 1807 | 2018-07-24T15:47:11 | Chromatic transitions in the emergence of syntax networks | [
"q-bio.NC",
"cond-mat.dis-nn"
] | The emergence of syntax during childhood is a remarkable example of how complex correlations unfold in nonlinear ways through development. In particular, rapid transitions seem to occur as children reach the age of two, which seems to separate a two-word, tree-like network of syntactic relations among words from a scale-free graphs associated to the adult, complex grammar. Here we explore the evolution of syntax networks through language acquisition using the {\em chromatic number}, which captures the transition and provides a natural link to standard theories on syntactic structures. The data analysis is compared to a null model of network growth dynamics which is shown to display nontrivial and sensible differences. In a more general level, we observe that the chromatic classes define independent regions of the graph, and thus, can be interpreted as the footprints of incompatibility relations, somewhat as opposed to modularity considerations. | q-bio.NC | q-bio | Chromatic transitions in the emergence of syntax networks
Bernat Corominas-Murtra1,2,3, Mart´ı S`anchez Fibla4, Sergi Valverde2,3 and Ricard Sol´e2,3,5
1 (1) Institute of Science and Technology Austria,
Am Campus 1, A-3400, Klosterneuburg, Austria
2Complex Systems Lab, ICREA-Universitat Pompeu Fabra, Dr Aiguader 88, 08003 Barcelona, Spain
3Evolution of Technology Lab, Institut de Biologia Evolutiva (CSIC-UPF),
Passeig Maritim de la Barceloneta, 37-49, 08003 Barcelona, Spain
4SPECS, Technology Department, Universitat Pompeu Fabra, Roc Boronat 138, 08018 Barcelona, Spain
5 Santa Fe Institute; 1399 Hyde Park Road; Santa Fe; NM 87501; USA
The emergence of syntax during childhood is a remarkable example of how complex correlations
unfold in nonlinear ways through development. In particular, rapid transitions seem to occur as
children reach the age of two, which seems to separate a two-word, tree-like network of syntactic
relations among words from a scale-free graphs associated to the adult, complex grammar. Here we
explore the evolution of syntax networks through language acquisition using the chromatic number,
which captures the transition and provides a natural link to standard theories on syntactic structures.
The data analysis is compared to a null model of network growth dynamics which is shown to
display nontrivial and sensible differences. In a more general level, we observe that the chromatic
classes define independent regions of the graph, and thus, can be interpreted as the footprints of
incompatibility relations, somewhat as opposed to modularity considerations.
Keywords: Complex Networks, Graph Colouring, Modularity, Syntax
I.
INTRODUCTION
The origins of human language have been a matter of
intense debate. Language is a milestone in our evolu-
tion as a dominant species and is likely to pervade the
emergence of cooperation and symbolic reasoning [1 -- 4].
Maybe the most defining and defeating trait is its vir-
tually infinite generative potential: words and sentences
can be constructed in recursive ways to generate nested
structures of arbitrary length [3, 5]. Such structures are
the product of a set of rules defining syntax, which are
extracted by human brains through language acquisition
during childhood after a small sample of the whole combi-
natorial universe of sentences has been learned. And yet,
in spite of its complexity, syntax is accurately acquired
by children, who master their mother tongue in a few
years of learning. Indeed, around the age of two, linguis-
tic structures produced by children display a qualitative
shift on their complexity, indicating a deep change on
the rules underlying them [6, 7]. This sudden increase of
grammar complexity is known as the syntactic spurt, and
defines the edge between the two words stage, where only
isolated words or combinations of two words occur, to a
stage where the grammar rules governing this syntax are
close to the one we can find in adult speech -although the
cognitive maturation of kids makes the semantic content
or the pronunciation different from the adult one. How
can we explain or interpret such nonlinear pattern?
Statistical physicists have approached the problem of
language evolution showing for example that nontrivial
patterns are shared between language inventories (collec-
tions of words) and some genetic and ecological neutral
models [8] -- see [9] and references therein. However, most
of these models do not make any assumption about the
role played by actual interactions among words, or, more
generally, linguistic units, which largely define the nature
of linguistic structures. In this context, a promising ap-
proach to its structure and evolution involves considering
language in terms of networks of interconnected units in-
stead of unstructured collections of elements (e.g., words
or syllabes) [10]. In this context, syntactic networks, in
which nodes are words and links the projection of actual
syntactic relations, have been shown to be an interesting
abstraction to grasp general patterns of language produc-
tion [10 -- 12]. Specially valuable has been the quantitative
data obtained from syntax networks obtained along the
process of syntax acquisition [7, 13, 14].
At the fundamental level, syntax can be understood as
a set of symbols associated under a restricted universe
of combinations somewhat similar to chemistry. Atoms
and words would then be linked through compatibility
relations defining what can be combined and what is for-
bidden. The power of this picture is supported by the use
of linguistic methods in the systematic characterization
of chemical structures [15]. Chemical structure diagrams
can thus be seen as some sort of language, with chemical
species and bonds as key ingredients. In a more abstract
fashion, we can say that general rules of combining ele-
ments within a given set of interacting pieces with well
defined functional meaning is at work in both language
and chemistry.
Following the chemical analogy, where abstract classes
of "nodes" can be defined, we will take advantage of
graph colorability theory as a general framework to de-
tect transitions based on qualitative changes of compati-
bilities. Specifically, we suggest that a new combinatorial
approach grounded on a graph colouring may enable a
better understanding of the evolution of networks having
internal relations of compatibility (e.g., some kind of syn-
tactic rules). In this context, we propose the chromatic
number -and associated measures- of the graph [16 -- 18] as
8
1
0
2
l
u
J
4
2
]
.
C
N
o
i
b
-
q
[
1
v
4
9
1
9
0
.
7
0
8
1
:
v
i
X
r
a
2
FIG. 1: Optimal colourings of syntactic networks before and after the syntactic spur. (a) A syntactic network before the
transition (3th corpus) is largely bipartite (this network accepts a 2-coloring). (b) Post-transition network (7th corpus) is
remarkably more complex, which corresponds to high chromatic number χ(G7) = 6. All networks coming from Peter dataset.
Time spent between these two corpora is about two months and a half (see text).
an indicator of network complexity. Within our context,
it will be a surrogate of syntactic complexity. The chro-
matic number is defined as the minimal number of colors
needed to paint all nodes of the graph in a way that no
adjacent nodes have the same color [17]. The q-coloring
problem, i.e., to know whether a graph can be colored
with q different colors is one of the most important N P -
complete problems. From the statistical physics point of
view, an analogous problem is defined within the context
of the Potts model [19].
It is worth to emphasize that transitions in the evolu-
tion of the chromatic number have been widely studied
in models of random graphs [18, 20 -- 22]. Our exploration
over sequences of syntactic graphs mapping child lan-
guage acquisition also displayed transitions in the chro-
matic number (see below). This is, to the best of our
knowledge, the first time that such transitions have been
reported in a real system. Classes of nodes would be de-
fined precisely by the fact that there are no connections
among them, a measure conceptually opposite to graph
modularity.
The remaining of the paper is organized as follows.
Section II is devoted to a brief revision of the so-called
Potts model as the way to introduce the chromatic num-
ber. In section III we apply these theoretical constructs
to our problem and we analyze the obtained data by us-
ing different estimators of relevance, the most prominent
of them being a null model of random sentence genera-
tion. In section IV we discuss the obtained results and
we highlight a number of potential impacts of this kind
of complexity estimators for complex networks.
II. GRAPHS AND COLORING: BASICS
We will work over undirected graphs. An undirected
graph G(V, E) -hereafter, G- is composed by the set of
V = {v1, ..., vn} nodes and a set E = {ej(cid:107)1 ≤ j ≤ m} ⊆
V × V of edges. Each (unordered) pair ej = {vi, vk}
depicts a link between nodes vi and vj. The number of
links k(vi) attaching node vi is the degree of the node
and (cid:104)k(cid:105) is the average degree of the graph G. The degree
distribution P (k) accounts for the probability to select a
node at random having degree k. The identity card of a
graph is the so-called Adjacency matrix, a(G), which is
defined as follows:
(cid:26) 1, iff (∃ek ∈ E) : (ek = {vi, vj})
aij =
0, Otherwise
.
We observe that the adjacency matrix of undirected
graphs is symmetrical, i.e., aij = aji.
The computation of the chromatic number can be for-
mulated as the following combinatorial problem: What is
the minimal number of 'colours' needed to paint all nodes
abitputturngethavedidfixkickpullthrowseeontheretapemoreainrightpieceanbacktoolrecorderoffdaddy'sbabybagaroundmypencilwriteairboxupballhittrainlightcarswheelsbikeboyswant of the graph in such a way that no single node is con-
nected to neighbors having the same color? We can map
this problem into the antiferromagnetic q-dimensional
Potts model at T = 0 [19]. This model is a general-
ization of the classical Ising model for lattices: at every
node of this lattice we place a particle having a spin which
energetically constraints the state of its neighbors. Tra-
ditionally, spins can have only two states, namely ↑(cid:105) and
↓(cid:105). In the Potts model, compatibility relations take into
account an arbitrary number q > 2 of different states.
Let us consider a partition of nodes V containing q dif-
ferent classes, namely, Gq(V ) = {g1, ..., gq} of V , i.e.:
(cid:92)
(cid:91)
Gq = ∅ and
Gq = V
,
(1)
The state σi of node vi indicates the class of Gq(V ) to
which the node belongs to, i.e., σi ∈ gj. Let Fq(V ) be
the ensemble of all partitions of V containing q different
classes. Every element in Fq(V ) has the following energy
penalty1:
H(Gq) = J
aijδ(σi, σj)
,
(2)
(cid:88)
i<j
where J = 1 is the coupling constant and δ is the Kro-
necker symbol:
(cid:26) 1, iff i = j
δ(σi, σj) =
0, Otherwise
.
Intuitively, the higher the presence of pairs of con-
nected nodes belonging to the same state, the higher will
be the energy of the global state of the graph. Given
a fixed q, the configurations displaying minimal energy
may have an amount of non-solvable situations, leading
to the unavoidable presence of connected nodes at the
same state. This phenomenon is called frustration, and
for these configurations, the ground state of the Hamil-
tonian defined in (2) displays positive energy. If there is
no frustration, i.e., ∃Gq ∈ Fq(V ), we can find a partition
that satisfies:
H(Gq) = 0
,
(3)
and we say that the graph is q-colorable, being the q dif-
ferent colors the q different classes or members of Gq.
When the graph is q-colorable, there is at least one par-
tition Gq ∈ Fq(V ) such that, if vi, vj ∈ V belong to the
same class or color of the partition, namely gl ∈ Gq. We
deduce that:
(vi, vj ∈ gl) ⇒ aij = 0 .
(4)
Relation (4) maps color classes onto disjoint sets of graph
elements (adjacent nodes have a different color). Now,
1 In our approach, the energy units of this Hamiltionian are arbi-
trary.
3
the coloring problem consists in finding the minimal num-
ber of classes (or colors) required to properly paint the
graph. This is the so-called Chromatic Number of the
graph G:
χ(G) = min{q : (∃Gq ∈ Fq(V )) : H(Gq) = 0}
(5)
Now suppose network partition(s) G∗q ∈ Fq(V ) having
minimal energy, see equation (2), given a number of
colours q:
.
G∗q = min
Gq∈Fq(V )
{H(Gq)}
.
(6)
In general, the process of search for the chromatic num-
ber yields a decreasing sequence of energies ending at
H(G∗χ(G)) = 0:
H(G∗1) ≤ ... ≤ H(G∗χ(G)) = 0 ,
(7)
In order to assess the statistical significance of chromatic
numbers, we define the relative energy of any q-coloring
as follows:
H(G∗q)
E
,
fq(χ) =
(8)
where E is the number of edges in the graph G. This
quantity 0 ≤ fq(χ) ≤ 1 corresponds to the minimal (rel-
ative) number of frustrated links or violations (i.e., when
adjacent nodes have the same color).
Despite the high complexity of this problem (comput-
ing the chromatic number in an arbitrary graph is a N P -
hard problem) several bounds can be defined. A lower
bound can be defined from the so-called Clique number.
A clique is a subgraph in which every node is connected
to all other nodes in the subgraph. The Clique number
ω(G) is the size of the largest clique in the graph, which
is a natural lower bound for χ(G) [17]:
ω(G) ≤ χ(G)
.
(9)
Alternatively, an upper bound on χ(G) can be defined by
looking at the K-core structure of G. The K(G) core is
the largest subgraph whose nodes display degree higher
or equal to K. Now, lets K∗(G) be the K-core with
largest connectivity that can be found in G:
K∗ = max{K : K(G) (cid:54)= ∅}
.
(10)
Then, it can be shown that K∗ sets an upper bound to
the chromatic number [17]:
χ(G) ≤ K∗ + 1
.
(11)
Finally, let us mention that, for some families of random
graphs the chromatic number has an asymptotic behavior
depending on the average connectivity [18], χ(G) ∼ (cid:104)k(cid:105)
log(cid:104)k(cid:105)
However, the above relationship does not hold for scale-
free networks with exponent 2 < γ < 3. These heteroge-
nous networks cannot have a stable value of the chro-
matic number because their clique number (9) diverges
with the graph size, even at constant (cid:104)k(cid:105) [23].
GP 4
1
GP 5
1
GP 6
1
GP 7
1
GP 8
1
GP 9
1
GP 10
1
GP 11
1
4
1/49 5/105 66/434 131/644 87/589 157/903 104/659 95/717
8/434 31/644 15/589 40/903 20/659 10/717
f1(χ)
f2(χ)
f3(χ)
f4(χ)
f5(χ)
GP 2
1
0
0
0
0
GC2
1
GP 1
1
0
0
0
0
GC1
1
0
0
GP 3
1
0
0
0
0
GC3
1
0
0
f1(χ)
f2(χ) 6/140 5/119 11/156 6/128 10/152 14/199 61/361 65/442 71/439 93/592 131/687
16/592 29/687
f3(χ)
1/592
4/687
f4(χ)
9/361 11/442 8/439
0
0
0
0
0
0
0
0
0
8/644
1/644
GC7
1
0
0
GC8
1
8/903
0
GC9
1
2/659
0
GC10
1
0
0
GC11
1
0
0
0
GC4
1
0
0
0
GC5
1
0
0
GC6
1
0
0
TABLE I: Relative energy values of q-colorings in the Peter (top) and Carl (bottom) datasets (see text).
III. THE EVOLUTION OF χ ALONG SYNTAX
ACQUISITION
of child's utterances obtaining a sequence of 11 syntactic
graphs corresponding to the sequence of Peter conversa-
tions GP 1, ...,GP 11 and Carl's conversations GC1, ...,GC11.
Here we study the evolution of the chromatic num-
ber through language development as captured by syn-
tax graphs. We compare the chromatic number with the
lower and upper bounds provided by the clique number
and the maximal K-core, respectively. We assess the
relevance of computed chromatic numbers with the cor-
responding minimal energy. The combination of these
two measurements enable us to interpret the nature of
the chromatic number. Specifically, we can check wether
changes in this number reflects a global pattern or instead
some anomalous behaviour of a small, localized subgraph.
Finally, we provide further validation of our analysis by
comparing chromatic numbers in empirical and synthetic
networks obtained through a random sentence generator.
A. Building the Networks of Early Syntax
Through the process, networks built upon the aggrega-
tion of syntactic structures from child's productions grow
and change in a smooth fashion until a rapid transition
occurs [7, 14, 24] -- see also [10]. We reconstruct syntac-
tic networks by projecting the raw constituent structure,
i.e., phrase structure of children's utterances, into linear
relations among lexical items [25]. Then, we aggregate all
these productions in a single graph where nodes are lexi-
cal items and links represent syntactic relations between
them [7, 24]. These networks provide a unique window
into the patterns of change occurring in the language ac-
quisition process.
The two cases studied here are obtained from the
CHILDES Database [26, 27] which includes conversations
between children and parents. Specifically, we choose Pe-
ter and Carl's corpora, whose structure has been accu-
rately extracted and curated. For both Peter and Carl's
corpora, we choose 11 different recorded conversations
distributed in approximately uniform time intervals rang-
ing from the age of ∼ 20 months to the age of ∼ 28
months. The chosen interval corresponds to the period
in which the syntactic spurt takes place. From every
recorded conversation, we extract the syntactic network
B. Chromatic transition from bipartite to
multicoloured networks
From our graph collection -see section III A-, we obtain
two sequences of chromatic numbers sP (χ) and sC(χ)
corresponding to the evolution of the chromatic number
in Peter and Carl datasets, respectively:
sP (χ) = χ(GP 1), ..., χ(GP 11)
sC(χ) = χ(GC1), ..., χ(GC11)
.
The above sequences display similar patterns with some
interesting differences, see figure (2) and figure (4). For
example, the middle stages of both datasets show an in-
crease in the chromatic number. At the stage when the
syntactic spur takes place, Peter's dataset sP displays a
sharp transition from a nearly constant, low chromatic
number to a high chromatic number, which is fully con-
sistent with the emergence of complex syntax. First tree
networks in sP accept 2-colorings, i.e., they are bipar-
tite, see figure (1). The grammar at this stage mainly
generates pairs of complementary words, like:
(cid:104)verb, noun(cid:105) or
(cid:104)adjective, noun(cid:105)
.
Typical productions of this stage are, for example, "car
red" or "horsie run". This pre-transition pattern, also-
called 2-word stage, corresponds to a highly restrictive
grammar, e.g., syntactic structures like (cid:104)verb, verb(cid:105) do
not exist.
Instead, relations between lexical items are
strongly constrained by their semantic content. On the
other hand, Carl's sequence sC shows χ ≥ 3 from the
very beginning -- i.e.
these networks are not bipartite.
A detailed inspection of Carl's productions at this stage
shows the presence of functional particles from the very
beginning. This suggests that, in general, high chromatic
numbers relates to high grammar flexibility, being this
flexibility provided by the hinge role that have these par-
ticles in the global functioning of grammar.
5
FIG. 2: Evolution of the chromatic number χ (solid line), maximum clique ω (dashed line with crosses) and the maximum
core K∗ + 1 (dashed line with squares) in (top) real networks and in (bottom) an ensemble of n = 20 simulated networks (see
text). Inset: Comparison between the time evolution of the K-core size (dashed line) and the maximum K∗-core size (solid
line). Left and right panels correspond to Peter and Carl datasets, respectively. Shaded gray areas correspond to standard
deviation in the case of the simulated instances.
Still, the behaviour of χ(G) can be quite sensitive to
the anomalous behaviour of small subgraphs. For ex-
ample, the transition of χ(G2) = 2 to χ(G3) = 3, when
Peter is about 23 months old, is due a single triangle in a
(largely) bipartite network -- see figure (2) left. A combi-
nation of measurements enables us to assess whether the
chromatic number represents the behaviour of a small
number of nodes or it is the natural outcome of global
network features. For example, we can compare χ(G)
with the lower bound given by the clique number (9)
and the upper bound provided by the maximal K-core
connectivity (10). Therefore, each sequence sP (χ), sC(χ)
will be accompanied by two sequences, namely Ω, κ:
ΩP,C = ω(GP 1,C1), ..., ω(GP 11,C11)
κP,C = K∗(GP 1,C1), ..., K∗(GP 11,C11)
.
Since every graph can be associated to a measure of χ rel-
evance -- see eq. (8) -- , we will have a third sequence of f 's
for every Peter's graph and another associated to every
Carl's graph -see table (I). For example, figure (2) -top-
shows a clear trend towards increasing maximum clique
and maximum K-core combined with increased relevance
-- see table I -- , which indicates that the final chromatic
number cannot be longer associated to any trivial clique.
In any case, the relevance of the chromatic number as a
202224262823456789Peter instancesTimeline (Months) K+1 (max k−core + 1)χ (chromatic number)ω (max clique)222426280204060 K+1 (max k−core + 1)size of k−core202224262823456789Carl instancesTimeline (Months)2224262802040602122232425262728024681012Simulated instances from PeterTimeline (Months)**** not optimal (timed out)222426285101520202224262824681012Simulated instances from CarlTimeline (Months)2224262851015202530Figure1:Evolutionofthechromaticnumberχ(solidline),maximumcliqueω(dashedlinewithcrosses)andthemaximumcoreK∗+1(dashedlinewithsquares)in(top)realnetworksandin(bottom)anensembleofn=20simulatednetworks(seetext).Inset:ComparisonbetweenthetimeevolutionoftheK-coresize(dashedline)andthemaximumK∗-coresize(solidline).LeftandrightpanelscorrespondtoPeterandCarldatasets,respectively.1 global complexity estimator is much more feasible after
the transition. Both Peter and Carl sequences show that
the chromatic number is often close to the clique number
-see figure 2 (top). Maximum K∗-core size is generally
more than twice the maximum clique size (see figure 2
(inset)). Then, the whole network structure (or a large
part of it) has enough connectivity to enable the emer-
gence of a non-trivial K-core structure. This is consis-
tent with a manual inspection of grammars that generate
a great amount of combinatorial complexity, i.e., a rich
collection of compatibility relations.
C. Real syntax versus null model
Here, we compare the evolution of the chromatic num-
ber in real and simulated networks. A data-driven,
syntax-free model that generates random child's utter-
ances having the same statistics of word production as
Peter and Carl datasets is used as a null model [7]. This
model definition enables us to assess if the high com-
binatorics displayed by post-transition networks emerge
directly from an increasingly rich vocabulary. We build
our model by extracting the following statistical parame-
ters from the 11 recorded conversations in Peter and Carl
corpora:
1. The number of sentences SP (i), SC(i) in the Peter
and Carl datasets.
2. The probability distribution of structure lengths or
the probability P (s) that any syntactic structure
has s words. We obtain two different distributions,
one for each dataset.
3. We assume that the probability of the i-th most
frequent word is a scaling law:
1
Z
i−β
p(i) =
(12)
with 1 ≤ i ≤ Nw(T ), β ≈ 1 -- i.e., Zipf's law -- and
Z is the normalization constant:
,
Nw(T )(cid:88)
(cid:18) 1
(cid:19)β
i
i=1
Z =
.
(13)
Notice that Z depends on lexicon size, Nw(T ),
which grows slowly at this stage.
We run the above model in the two datasets by gen-
erating SP,C(i) random sentences, each experiment is
repeated 20 times. From the collection of randomly gen-
erated syntactic structures we construct a comparable
sequence of syntax networks following the same method
as in the real datasets -see section III A. Figure (3) shows
that our model generates random syntax networks with
size and connectivity comparable to the ones measured in
real networks. These statistical indicators display a huge
increase during the studied period, being this increase
6
FIG. 3: Evolution of (top) the mean degree and (bottom)
size of the largest connected component in the real (strong
solid lines) and simulated (weak solid lines) syntax networks.
Shaded gray areas correspond to standard deviation in the
case of the simulated instances.
sharper around the age of two, i.e., during the syntactic
spurt [7]. As discussed in section II both the mean con-
nectivity and network size play an important role when
determining the values of ω, χ and K∗.
Now, we compute the sequence of averaged chromatic
numbers, sP (χ), sC(χ), for the simulated Peter and Carl
syntax networks. Similarly, we generate the sequences of
average clique number ΩP,C and the average maximum
K-core κP,C. The most salient property we find when
comparing real networks obtained from both Peter and
Carl's corpora with their randomized counterparts is a
huge increase of χ, ω and K∗ in the simulated networks.
That is, the ensemble of random strings displays higher
complexity parameters than the real corpora. For exam-
ple, at the end of the studied period, the three complexity
estimators are close to 10 in Peter simulations and close
to 9 Carl simulations -see figure (2) bottom.
A very interesting feature is found at the first stages
of the simulated Peter sequence: the random networks
2022242628123456789Mean DegreeTimeline (Months)Mean Degree PeterCarlSimulated PeterSimulated Carl2022242628050100150200250300350Maximum Connected ComponentTimeline (Months)Number of nodes PeterCarlSimulated PeterSimulated CarlFigure2:Evolutionofthe(top)averageconnectivityand(bottom)sizeofthelargestconnectedcomponentinthereal(solidlines)andsimulated(dashedlines)syntaxnetworks.2 7
FIG. 4: Relationship between (right) the average degree and the chromatic number and (right) the size of the largest connected
component and the chromatic number in the Peter (solid circles) and Carl (open squares). Simulation results are at the bottom
while real data is shown at the top.
are no longer bipartite -see section III B. In particular,
the third corpus has an average chromatic number of 4,
which is significantly higher than the observed chromatic
number.
In this case, the two-stage grammar imposes
severe constraints on what is actually plausible in any
pre-transition syntactic structure. This trend is also ob-
served at latter stages of language acquisition. In general,
simulated networks have higher chromatic numbers than
empirical networks, although both two types of networks
have similar connectivities -by definition. In some cases,
the average chromatic number of the graphs belonging
to the random ensemble is twice the real one, see figure
(2). To better understand the nature of these deviations,
we have compared the behaviour of chromatic numbers
against mean connectivity and the size of the largest
connected component. Figure 4 shows a well-defined,
non-trivial deviation between real networks and random
networks. For example, the plot of chromatic number
and network size display a quasi linear relationship in
simulated networks -see figure (4) bottom. These plots
suggest that the chromatic number is capturing essential
combinatorial properties of the underlying system, which
cannot be reproduced with a simple, syntax-free random
generation model.
IV. DISCUSSION
Syntax is a characteristic, complex and defining fea-
ture of language organization.
It pervades its capacity
for unbounded generative power of the linguistic system
[5], allows sentences to be organized in highly structured
ways and is acquired in almost full power by children
after being exposed to a limited repertoire of examples.
Syntax is also one aspect of the whole: semantic and
phonological aspects need to be taken into account, and
they are all embedded in (and run by) a cognitive, brain-
embodied framework [28]. Because of the dominant role
played by how words actually interact with each other,
computational and theoretical approaches dealing with
word inventories or other statistical trends ignoring inter-
actions are likely to be limited. Following previous work
that takes advantage of complex networks approaches
to language organization [10] we have made a step fur-
ther in studying the organization of syntax graphs using
graph coloring. The motivation of this approximation is
twofold. On the one hand, graph colorability allows to
properly detect correlations that are not capture by topo-
logical approaches. On the other hand, it seems a natural
way to substantiate previous claims connecting syntax
with compatibility relations common with other types of
systems, such as chemical structures. Since graph color-
ing naturally defines compatibility through the presence
or absence of a common label to every pair of nodes, it
12345671234567Chromatic NumberAverage degree PeterCarl1234567050100150200250300Chromatic NumberSize GCC2468100246810Chromatic NumberAverage degree Simulated PeterSimulated Carl246810050100150200250300Chromatic NumberSize GCCFigure3:Relationshipbetween(right)theaveragedegreeandthechromaticnumberand(right)thesizeofthelargestconnectedcomponentandthechro-maticnumberinthePeter(solidcircles)andCarl(opensquares).Simulationresultsareatthebottomwhilerealdataisshownatthetop.3 seems the right framework to study the process of net-
work growth in child language.
The behavior of the chromatic number accurately
marks the syntactic spurt in language acquisition, i.e.,
it is a footprint of the generative power of the underly-
ing grammar. There are limitations associated to the
network definition. Syntactic relations are structure-
dependent, not sequence dependent. Because the net-
work is an aggregation of text sequences, it cannot fully
grasp the hierarchical nature associated to syntactic con-
structs. Still, the chromatic number is a global measure-
ment that can detect grammar constraints by analyz-
ing the pattern of network interaction at different scales.
That is, the network representation is an indicator of
global linguistic performance and includes some combi-
natorial signal which can be properly detected with the
chromatic number.
In this context, standard network
measurements like average degree, clustering or degree
distribution are much more limited.
There are other, broader implications of our work. The
chromatic number can be viewed as a reciprocal measure
of standard community detection. Here, the chromatic
number defines a partition of the network in classes of
unlinked nodes. This definition is particularly relevant
in networks where some kind of compatibility relation
8
is at work in the wiring process. In this case, the stan-
dard community structure can be misleading, because el-
ements of the same class cannot be connected. The case
for syntactic graphs is paradigmatic but the partition in-
duced by the chromatic number could shed light into the
behaviour of many other systems. Additionally, we have
proposed to assess the statistical significance of these par-
titions with the sequence of minimal violations -see equa-
tion (8). Future work should explore how the chromatic
number (and related measures) can be exploited to de-
tect forbidden links in the network. Deviations of the
chromatic number (as the ones observed in this paper)
suggest the presence of combinatorial constraints that
must be taken into account, for example, when defining
proper null-models.
Acknowledgments
We thank Complex Systems Lab members for fruitful
conversations. This work was supported by the James
McDonnell Foundation (BCM, SV, RSV) and the EU
program FP7-ICT- 270212 (MSF).
[1] J. Maynard-Smith and E. Szathm`ary, The Major Tran-
sitions in Evolution (University of New York Press. New
York, 1997).
A. Gomila, Adaptive Behavior 20, 427 (2012).
[15] K. R. S. J. Tauber, J. Chem. Doc 11, 139 (1971).
[16] R. l. Brooks and W. T. Tutte, Proceedings of the Cam-
[2] D. Bickerton, Language and Species
(University of
bridge Phylosophical Society 39, 194 (1941).
Chicago Press. Chicago, 1990).
[17] B. Bollobas, Modern Graph Theory (Springer, 1998), cor-
[3] M. D. Hauser, N. Chomsky, and T. W. Fitch, Science
rected ed.
298, 1569 (2002).
[18] B. Bollob´as, Random Graphs (Cambridge University
[4] M. H. Christiansen and S. Kirby, Trends in Cognitive
Press, 2001).
Sciences 7, 300 (2003).
[5] N. Chomsky, Language and problems of knowledge (MIT
Press. Cambridge, Mass, 1988).
[6] A. Radford, Syntactic Theory and the Acquisition of En-
glish Syntax: the nature of early child grammars of En-
glish (Oxford. Blackwell, 1990).
[19] F. Y. Wu, Rev. Mod. Phys. 54, 235 (1982).
[20] B. Bollob´as, comb 8, 49 (1988).
[21] D. Achlioptas and M. Molloy, The Electronic Journal of
Combinatorics 6, R29 (1999).
[22] L. Zdeborov´a and F. Krzakala, Phys. Rev. E 76, 031131
(2007).
[7] B. Corominas-Murtra, S. Valverde, and R. Sol´e, Ad-
[23] G. Bianconi and M. Marsili, EPL (Europhysics Letters)
vances in Complex Systems (ACS) 12, 371 (2009).
74, 740 (2006).
[8] R. A. Blythe and A. J. McKane, Journal of Statisti-
cal Mechanics: Theory and Experiment 2007, P07018
(2007).
[9] R. V. Sol´e, B. Corominas-Murtra, and J. Fortuny, Jour-
nal of The Royal Society Interface 7, 1647 (2010).
[24] B. Corominas-Murtra, arXiv.org:0704.3708 (2007).
[25] F. Hristea and M. Popescu, Dependency Grammar Anno-
tator, Building Awareness in Language Technology (Ed-
itura Universitatii din Bucaresti, 2003).
[26] L. Bloom, L. Hood, and P. Lightbown, Cognitive Psy-
[10] R. V. Sol´e, B. Corominas-Murtra, S. Valverde, and
chology pp. 380 -- 420 (1974).
L. Steels, Complexity 15, 20 (2010), ISSN 1099-0526.
[11] R. Ferrer-i-Cancho, R. Kohler, and R. V. Sol´e, Phys. Rev.
E 69, 051915 (2004).
[12] J. L. Steele, Ph.D. thesis, Doctoral dissertation, Univer-
sity of Queensland (2009), 305 pages.
[13] J. Y. Ke and Y. Yao, Journal of Quantitative Linguistics
[27] L. Bloom, P. Lightbown, and L. Hood, Monographs of
the society for Research in Child Development. Serial 160
(1975).
[28] R. Jackendoff, Foundations of Language: Brain, Mean-
ing, Grammar, Evolution (Oxford University Press,
USA, 2002).
15, 70 (2008).
[14] L. Barcel´o-Coblijn,
B. Corominas-Murtra,
and
|
1902.10070 | 1 | 1902 | 2019-02-06T14:01:34 | Warped phase coherence: an empirical synchronization measure combining phase and amplitude information | [
"q-bio.NC",
"nlin.CD"
] | The entrainment between weakly-coupled nonlinear oscillators, as well as between complex signals such as those representing physiological activity, is frequently assessed in terms of whether a stable relationship is detectable between the instantaneous phases extracted from the measured or simulated time-series via the analytic signal. Here, we demonstrate that adding a possibly complex constant value to this normally null-mean signal has a non-trivial warping effect. Among other consequences, this introduces a level of sensitivity to the amplitude fluctuations and average relative phase. By means of simulations of Roessler systems and experiments on single-transistor oscillator networks, it is shown that the resulting coherence measure may have an empirical value in improving the inference of the structural couplings from the dynamics. When tentatively applied to the electroencephalogram recorded while performing imaginary and real movements, this straightforward modification of the phase locking value substantially improved the classification accuracy. Hence, its possible practical relevance in brain-computer and brain-machine interfaces deserves consideration. | q-bio.NC | q-bio |
Warped phase coherence: an empirical synchronization
measure combining phase and amplitude information
Ludovico Minati,1, 2, 3, a) Natsue Yoshimura,4, 5 Mattia Frasca,6 Stanis law Dro zd z,2, 7
and Yasuharu Koike4
1)Tokyo Tech World Research Hub Initiative - Institute of Innovative Research,
Tokyo Institute of Technology, Yokohama 226-8503, Japan
2)Complex Systems Theory Department, Institute of Nuclear
Physics - Polish Academy of Sciences (IFJ-PAN), 31-342 Krak´ow,
Poland
3)Center for Mind/Brain Sciences (CIMeC), University of Trento, 38123 Trento,
Italy
4)FIRST - Institute of Innovative Research, Tokyo Institute of Technology,
Yokohama 226-8503, Japan
5)PRESTO, JST, 332-0012 Saitama, Japan
6)Department of Electrical Electronic and Computer Engineering (DIEEI),
University of Catania, 95131 Catania, Italy
7)Faculty of Physics, Mathematics and Computer Science,
Cracow University of Technology, 31-155 Krak´ow, Poland
(Dated: 27 February 2019)
The entrainment between weakly-coupled nonlinear oscillators, as well as between
complex signals such as those representing physiological activity, is frequently as-
sessed in terms of whether a stable relationship is detectable between the instan-
taneous phases extracted from the measured or simulated time-series via the ana-
lytic signal. Here, we demonstrate that adding a possibly complex constant value
to this normally null-mean signal has a non-trivial warping effect. Among other
consequences, this introduces a level of sensitivity to the amplitude fluctuations
and average relative phase. By means of simulations of Rossler systems and ex-
periments on single-transistor oscillator networks, it is shown that the resulting
coherence measure may have an empirical value in improving the inference of the
structural couplings from the dynamics. When tentatively applied to the elec-
troencephalogram recorded while performing imaginary and real movements, this
straightforward modification of the phase locking value substantially improved the
classification accuracy. Hence, its possible practical relevance in brain-computer
and brain-machine interfaces deserves consideration.
In between the extremes of complete asynchrony and perfect synchronization
between non-linear dynamical systems, complicated trajectories can show asso-
ciations in some aspects but not others. A frequent observation is the presence
of a relatively stable phase relationship between the positions of two systems
along their respective orbits, while the sizes, i.e. amplitudes, of these orbits
fluctuate more or less independently. We introduce a straightforward algebraic
change to the well-known phase locking value and demonstrate how the result-
ing warping, among other consequences, confers a level of sensitivity to the
amplitude fluctuations. We find that the measure thus obtained is potentially
useful, for example as regards enhancing the ability of classifying imaginary
actions based on short electroencephalogram segments.
a)Author to whom correspondence should be addressed. Electronic addresses: [email protected]
and [email protected]. Tel.: +39 335 486 670. URL: http://www.lminati.it.
2
I.
INTRODUCTION
Networked nonlinear, possibly chaotic oscillators which are nonidentical by virtue of para-
metric heterogeneities or structural nonequivalence pervade the natural and physical world.
In such systems, the absence of an invariant manifold implies a nontrivial effect of the
coupling strength. Entrainment initially ensues in the form of an increasingly stable phase
relationship coexisting with incoherent amplitudes, because the energy transfer rate is in-
sufficient to fully overcome the repulsive forces and lock a common trajectory. As coupling
is further strengthened, amplitude fluctuations become increasingly entrained and finally
a common trajectory is attained, hallmarking complete synchronization.1 -- 3 Though this
behavior is generally well-established, even in simple situations such as coupled Rossler
systems non-trivial routes to synchronization can be observed depending on the topology of
the underlying attractors.4 The measurement of synchronization remains an area of active
investigation, and a variety of approaches are available. On the one hand, measures of phase
locking are particularly sensitive to weak couplings, and are accordingly of wide interest.
For example, they represent an empirical means of establishing functional associations be-
tween brain areas, based on noisy, broadband signals such as the electroencephalogram.
However, they completely disregard amplitude information: particularly in the presence of
noise, this can be an undesirable limitation, since equal weighting is given to the synchrony
of small and large fluctuations.5,6 On the other hand, measures such as cross-correlation,
spectral coherence, mutual information, and generalized synchronization include both phase
and amplitude information. However, they may be poorly sensitive in situations of partial
synchronization where amplitude fluctuations remain largely uncorrelated and may require
longer time-series for reliable estimation.7,8 Accordingly, the phase locking value and its vari-
ants are often the methods of choice for the design of connectivity-based brain-computer
and brain-machine interfaces, also in virtue of their rapid convergence and robustness to
non-stationarity.9,10 A measure primarily indexing phase synchronization but capable of
including a variable amount of amplitude information could have considerable practical
utility.
II. FORMULATION
Given a set of n real-valued time-series sj(t) with j = 1 . . . n, the first step in quantify-
ing the phase locking is obtaining the corresponding analytic signals. These separate the
instantaneous amplitude and phase components rj(t) ∈ [0,∞) and αj(t) ∈ [−π, π],
ψj(t) = sj(t) + isj(t) = rj(t)eiαj (t) ,
where i = √−1 and sj(t) is the Hilbert transform of sj(t)
sj(t) =
1
π
p.v.(cid:20)Z ∞
−∞
sj(τ )
t − τ
dτ(cid:21) ,
(1)
(2)
and where p.v. denotes the Cauchy principal value of the integral.11 The instantaneous
relative phase between a pair of time-series or signals (j, k) can then be obtained with
From this, the degree of phase synchronization, or phase locking value, is defined as
∆αjk(t) = arg(ψj(t)ψk(t)) .
djk = hei∆αjk(t)it = hei[αj (t)−αk(t)]it .
(3)
(4)
While this parameter fully represents the level of coherence in Kuramoto-like networks, it
only provides a partial account of entrainment in systems possessing free amplitude, such as
Stuart-Landau and Rossler oscillators. This encompasses the near-totality of experimental
3
FIG. 1. Geometry on the complex plane of warping the instantaneous phase α into angle θ de-
pending on the amplitude r and constant parameter c (here assumed real non-negative) with Eq.
(5). Red hue denotes radii between the origin and the circle having radius r centered around (c, 0).
scenarios. Furthermore, in complex networks of such systems, amplitude dynamics support
the onset and maintenance of phase synchronization.12
Normally, one has hψj(t)i = 0. The present work is a preliminary exploration concerned
with the notable effects of adding a possibly complex constant parameter c to the analytic
signal, such that
θj(t, c) = arg(ψj (t) + c) .
(5)
By this straightforward operation, the instantaneous phases αj(t) are warped into angles
θj(t, c) which also depend on the amplitudes rj (t). Unlike in a simple weighting, these are
determined by the magnitude c as well as by the warping direction c/c = ±1,±i.
First, let us consider that the underlying geometric relationships determine four possible
situations on the complex plane, hereby illustrated without loss of generality assuming a
uniform distribution α ∈ [−π, π] and c/c = 1 (Fig. 1). If c = 0, there is no warping and
θ = α. For 0 < c < r, one still has θ ∈ [−π, π] but the distribution is gradually drawn
towards a peaked one, eventually leading again to a uniform distribution in θ ∈ [−π/2, π/2]
when c = r. For c > r, the distribution is morphed towards an increasingly narrow centrally-
symmetric one. In the limit c → ∞, one has θ ∈ [− tan−1(r/c), tan−1(r/c)].
In other words, the instantaneous amplitude r influences the effect of the phase α on the
angle θ in an intricate manner, which can be conceptualized as follows. On the one hand,
the surface is "twisted", so that if r ≪ c, δα is shrunk into δθ ≈ 0, whereas if r ≫ c one
has δθ ≈ δα. As discussed below, when θ is used as a basis for determining synchronization,
this emphasizes the phase coherence of large-amplitude fluctuations. On the other hand,
4
FIG. 2. Warping the phase α into the angle θ with Eq. (5), assuming c = 1, visualizing the
different effect depending on whether r < c, r = c or r > c as also captured in Eqs. (6) and
(7).
the surface is also "plucked", so that for r ≈ c a large swing is observed in the vicinity of a
singularity point at α = arg(−c), around which the relationship between θ and α diverges
(Fig. 2).
Importantly, Eq. (5) implies a possibly non-monotonic relationship between the phase α
and the angle θ, namely when c > r (Fig. 3). Hence, after the warping, the angle θ can
no longer be considered as a phase because, assuming a fixed frequency, it may not satisfy
the fundamental requirement of increasing (or decreasing) monotonically with time. When
c > r, the synchronization measure described below and based on this value should then
not be considered as properly representing phase synchronization. At most, it only has an
empirical value.13
Another perspective on the warping effect is obtained by considering, again assuming c/c =
±1,
∂θ
∂α
∂θ
∂r
=
=
r2 + cr cos α
r2 + 2cr cos α + c2 and
c sin α
r2 + 2cr cos α + c2 .
(6)
(7)
5
FIG. 3. Relationship between the phase α and the angle θ observed as c is increased, corresponding
to the four situations in Fig. 1 and assuming r = 1. For c = 2, the relationship becomes non-
monotonic, i.e. dθ/dα crosses zero. Hence, the angle θ no longer represents a phase.
Here, the equations for c/c = ±i are readily obtained by replacing cos α with sin α, sin α
with − cos α, and c with c/i. While the effect of r on ∂θ/∂α is non-trivial and depends on
α, away from α = arg(−c) a positive relationship is evident, whereby phase fluctuations are
magnified when amplitude is large.
Second, a measure of warped phase coherence can be introduced by replacing the phase α
with the angle θ in Eq. (4), as
wjk(c) = hei[θj (t,c)−θk(t,c)]it ,
(8)
where by definition djk = wjk(0). At a given setting of c, wjk(c) reflects the level of
synchrony as visible after the phase warping. However, as the value of c is changed an
undesirable effect is noted: due to the influence of c on the distribution of θ, in c → ∞
one has w → 1. As depicted in Fig. 1, this is purely because θ → 0. When appropriate,
for visualization purposes, this issue can be addressed by normalizing, for example
e.g.
according to
wjk(c) =
where
jk(c)
wjk(c) − w′
jk(c)
1 − w′
,
jk(c) = hei[θj (t,c)−Ω[θk(t,c)]]it .
w′
(9)
(10)
Here, Ω [x] reshuffles xu according to a random sequence of index u. This operation re-
tains the angle distributions p(θj) and p(θk), but alters the distribution corresponding to
p(θj − θk). The underlying rationale is that, due to the reshuffling operation which destroys
any temporal relationship, w′
jk(c) approximates the value of wjk(c) that would be observed
if the signals were completely incoherent, taking into account the angle distributions p(θj)
and p(θk). Consequentially, this allows rescaling the coherence values onto a range closer
to [0, 1]. This operation can have complex consequences depending on the distributions.
However, for the avoidance of doubt, it is herein introduced only as a potentially convenient
form of rescaling. It is not as an essential step in the determination of warped coherence: in
fact, this rescaling has the drawback that the shuffling operation in Eq. (10) can introduce
random error.
In the absence of entrainment, one expects wjk(c) ≈ 0; contrariwise, as
exemplified in Section III, one can have wjk(c) > 0 or even wjk(c) < 0.
6
FIG. 4. Random signals with identical amplitude (left) or phase (right): distributions of relative
warped angles before (∆θ) and after (∆θ′) reshuffling, and resulting normalized warped phase
coherence w(c) (dashed lines denote maximum and minimum). Red color intensity proportional to
the logarithm of probability density.
III. RANDOM SIGNALS
The precise effect of warping depends on the phase and amplitude distributions. Never-
theless, the additional sensitivity of warped phase coherence to amplitude synchronization
can be initially illustrated via two paradigmatic cases involving pairs of random signals
j = 1, 2 having the form ψj = rjeiαj , where rj and αj are random variables. In one case,
the amplitudes are identical but the phases are not, i.e. r1 = r2 and α1 6= α2, whereas in
the other case the converse is true, i.e. r1 6= r2 and α1 = α2.
In both cases, rj and αj are respectively drawn from a Gaussian distribution with µ = 0
and σ = 1, and from a uniform distribution in [−π, π]. Unless otherwise noted, here and
thereafter the signals are rescaled to unitary average amplitude, i.e. hψi = 1.
Let us consider the angle differences for the initial and reshuffled signals as per Eqs. (8)
and (10), namely ∆θ = θ1(t, c) − θ2(t, c) and ∆θ′ = θ1(t, c) − Ω [θ2(t, c)]. While p(∆θ)
7
FIG. 5. Effect of non-zero average relative phase on the normalized warped phase coherence w(c)
between two random signals with fully coherent phases and amplitudes.
and p(∆θ′) both tend towards Dirac delta functions in c → ∞, they do so along different
trajectories: their respective distributions determine the form of w(c). Namely, in the case
of identical amplitude but unrelated phases, as expected one has w(0) = 0, above which the
measured coherence increases until w(1.1) = 0.16 then decreases again. Contrariwise, in
the case of identical phase but unrelated amplitudes, as expected one has w(0) = 1, above
which the measured coherence decreases until w(1.8) = 0.51, then marginally increases
again (Fig. 4).
For c > 0, depending on the strength of amplitude entrainment, w(c) can therefore increase
or decrease with respect to the amplitude-agnostic observation established by the non-
warped value at w(0), effectively corresponding to the phase locking value. It is therefore
evident that the effect of warping is fundamentally different from that of recently-proposed
amplitude-weighted versions of the phase locking value. In these cases, regardless of the
weighting one always has null and unitary expected value when the underlying phases are,
respectively, fully incoherent and fully coherent.5,6
Importantly, in the context of two coupled non-linear dynamical systems, the situation of
entrained amplitudes but incoherent phases is rather unphysical. That is because, even in
presence of generalized synchronization, a functional relation between the instantaneous
states must be present to retain coherent amplitudes. However, such scenario can easily
arise via modulation mechanisms, for example when two signal sources are such that their
instantaneous phases are generated by independent random processes, but their cycle ampli-
tudes are generated by other processes which are entrained. This is empirically observed for
neurophysiological signals such as the electroencephalogram, which represent a convoluted
ensemble average over a large number of partially-coherent non-linear generators. Indeed,
even exotic situations such as preferential phase-amplitude couplings can be observed for
such signals, and may index fundamental functions such as memory formation.14
On another note, phase synchronization clearly does not necessarily imply the identity of
phases, hence the effect of the average relative phase should be considered. The above
case with α1 = α2 is not generally representative, but only an initial demonstration of the
sensitivity to the amplitudes r1, r2.13 Considering the case r1 = r2 and α2 = α1 + β where
β ∈ [−π, π] is a constant parameter, reveals that warped phase coherence is sensitive to
the average relative phase, eventually leading to a change in sign. When β = 0, w(c) = 1
regardless of c; however, for increasing β, w(c) gradually decays and, for sufficiently strong
warping, one observes w(c) < 0 when β > π/2 (Fig. 5). This is reminiscent of linear
correlation and obviously deviates from the canonical definition of phase synchronization,
which is by definition agnostic to relative phase. However, under certain conditions it
can be empirically relevant. The reason is that in the absence of time delays or phase
8
FIG. 6. Warped phase coherence w(c) as a measure of synchronization between pairs of coupled
Rossler systems, showing separately the effects of warping direction (left) and of removing ampli-
tude and/or average relative phase (right; r = 1, arghei∆ αi = 0). a), b) and c) denote different
system parameter settings, see Section IV for description.
frustration, two attractively-coupled systems between which a stable phase relationship is
established tend towards zero relative phase when the energy exchange rate is sufficiently
large.1 -- 3
The examples which follow illustrate the potential advantages of warped phase coherence
as a measure of synchronization across diverse simulated and experimental scenarios.
9
FIG. 7. Warped phase coherence as a means of inferring the structurally-coupled nodes from the
generated time-series in random networks of Rossler systems. a) Accuracy in classifying the coupled
vs. uncoupled node pairs h(c), b) Corresponding average difference in coherence h∆w(c)i.
FIG. 8. Warped phase coherence w(c) as a measure of synchronization (left) in an alternative set
of three pairs of coupled Rossler systems, featuring phase-coherent and funnel attractors (right).
See Section IV for description.
IV. R OSSLER SYSTEMS
To gain further insight beyond a purely stochastic situation, we next considered a rep-
resentative case of synchronization of non-linear dynamics. Namely, two Rossler systems
were diffusively coupled at the variable xj, having the form
10
x1,2 = −y1,2 − z1,2 + u(x2,1 − x1,2)
y1,2 = x1,2 + a1,2y1,2 .
z1,2 = b1,2 + z1,2(x1,2 − c1,2)
(11)
These were integrated by means of an explicit Runge-Kutta (4,5) formula, the Dormand-
Prince pair, from t = 0 to t = 20 × 103 starting from and averaging over 100 sets of
identical random initial conditions with x1(0) = x2(0) ∈ [−1, 1], y1(0) = y2(0) ∈ [−1, 1] and
z1(0) = z2(0) ∈ [−1, 1]. We assessed the level of synchronization in terms of s = x for three
parameter settings.
First, we considered the case of a1 = a2 = 0.3, b1 = 0.2, b2 = 0.22, c1 = 5.7, c2 = 6.3,
u = 0.1, which represents a moderate level of coupling and mismatching, i.e. 10% in param-
eters b and c, within a region featuring well-developed chaoticity with dense folding.15,16
Despite some different fluctuations, regardless of the warping direction the coherence grad-
ually increased from w(0) = 0.42 to w(4) ≈ 0.55 (Fig. 6a).
As shown, unlike the non-warped value w(0), for c > 0 w(c) depends on both the am-
plitude fluctuations and the average relative phase between the initial signals. To better
understand the consequences of warping, it is useful to separate these effects by removing
either or both elements from the initial signals. That is, on the one hand setting r = 1, i.e.
considering ψj(t) = eiαj (t), and on the other hand setting arghei∆ α12i = 0, that is, consider-
ing ψj(t) = rj (t)ei αj (t), where α1 = α1 − arghei∆α12i/2 and α2 = α2 − arghei∆α21i/2. Doing
so revealed that, in this case, the increase in w(c) reflects amplitude synchronization: it is
insensitive to the average relative phase but abolished by removing the amplitude compo-
nent, leaving only a dip around c = 1, possibly related to the "plucking" effect described
previously (Fig. 6a).
Second, we changed a1 = a2 = 0.25 and reduced u = 0.045, yielding an overall comparable
level of entrainment as the previous example but in a region with sparser folding.15,16 In
this case, the trend of w(c) was considerably different, with w(0) ≈ w(4) and, in lieu of the
increase, a dip around c ≈ 1 was observed for all settings. This result serves to illustrate
that this measure is sensitive to the system dynamics, responding to the warping in a dif-
ferent manner depending on the level of folding and the envelope features (Fig. 6b).
Third, we considered an uncoupled case with weak mismatching, setting a1 = a2 = 0.2,
b1 = 0.2, b2 = 0.205, c1 = c2 = 5.7, u = 0. Clearly, in this case E[w(c)] = 0, however one
may observe w(c) > 0 as a consequence of the finite simulation time. Whereas w(0) = 0.23,
for c ≥ 1 we observed w(c) ≈ 0, hallmarking an improved representation of the actual lack
of entrainment in this configuration. In contrast with the first case, here the effect reflected
an increased sensitivity to the non-null average relative phase, considering which enhanced
detection of the actual independence between the two systems (Fig. 6c).
We subsequently considered 300 sparse random networks of n = 30 Rossler systems dif-
fusively coupled at variable xj according to random binary undirected graphs G obtained
from an Erdos-R´enyi G(n, p) model having p = 0.05. Denoting with gjk the entries of the
corresponding adjacency matrix, for each system j = 1 . . . n one has
n
Xk=1
xj = −yj − zj + u
yj = xj + ayj .
zj = bj + zj(xj − cj)
gjk(xk − xj )
(12)
Here, we set a = 0.3 and u = 0.05, and the other parameters were drawn randomly from
bj ∈ [0.19, 0.21] and cj ∈ [5.4, 6.0]. Integration was performed from t = 0 to t = 6 × 103
11
(≈ 850 cycles), in this case starting from non-identical random initial conditions xj(0) ∈
[−1, 1], yj(0) ∈ [−1, 1] and zj(0) ∈ [−1, 1]. These simulations addressed the identification
of asynchrony vs. weak synchronization between the coupled node pairs, among which
hw(0)i = 0.2: this represents one of the scenarios previously evaluated while inferring
connectivity from time-series in Kuramoto and Rossler networks17. Denser networks with
stronger couplings, wherein physical links may need to be distinguished from emergent
forms of preferential entrainment, were not considered here.18,19
For each network, given sj = xj the synchronization matrices wjk(c) were calculated with
Eq. (9) for each setting of c (here, the non-normalized values wjk(c) were not considered).
The corresponding values were then sorted and the indices (j(c)
q ) denoting the q =
1 . . . m most intensely synchronized node pairs were recorded, where m denotes the number
of edges in the graph G. The accuracy h(c) ∈ [0, 1] in inferring a predetermined number
of structurally-coupled node pairs from synchronization of the generated time-series was
thereafter determined as
q , k(c)
h(c) =
1
m
m
Xq=1
gj(c)
q k(c)
q
,
(13)
and averaged over all the networks. The accuracy steadily increased with c from the level
of observed for the non-warped value, h(0) = 0.74, towards h(4) ≈ 0.90. Notably, in this
case, both amplitude and average relative phase information were relevant, and removing
either reduced the accuracy, respectively down to h(4) = 0.79 and h(4) = 0.87. Since the
amplitude fluctuations were slow compared to the time-scale of oscillation cycles, the warp-
ing direction had a limited effect (Fig. 7a). For c >, 1, h(c) settled on a plateau close to
the accuracy level observed for the correlation coefficient computed on the raw signals sj,
namely h = 0.92. By comparison, the discriminatory ability of synchronization of the am-
plitude envelopes rj devoid of phase information was considerably lower, namely h = 0.70
(data not shown). However, even within the plateau region, the average difference in syn-
chronization between the structurally-coupled and the uncoupled nodes h∆w(c)i continued
increasing (Fig. 7b).
The sensitivity of warped phase coherence to the system dynamics can be further illustrated
in a slightly different scenario, derived from a study wherein three qualitatively-different
types of transitions to phase synchronization were observed.4 Namely, let us consider two
coupled Rossler systems
x1,2 = −ω1,2y1,2 − z1,2
y1,2 = ω1,2x1,2 + a1,2y1,2 + u(y2,1 − y1,2) ,
z1,2 = b1,2 + z1,2(x1,2 − c1,2)
(14)
where ω1 = 0.98, ω2 = 1.02, b1 = b2 = 0.1, c1 = c2 = 8.5 and where parameter a controls
the attractor topology. For relatively low values, e.g. a1 = a2 = 0.16, the attractors are
phase-coherent and phase synchronization ensues already for coupling strengths at which
the amplitudes remain only weakly correlated. For intermediate and high values, e.g. a1 =
a2 = 0.22 and 0.28, the attractors become funnel: in the former case, phase synchronization
sets in via an interior crisis, whereas in the latter case, it appears as a manifestation of a
generalized relationship. Setting the coupling strength to values suitable for observing a
similar level of phase locking w(0), respectively u = 0.04, 0.13 and 0.18, revealed a marked
difference in the effect of increasing the warping level (Fig. 8). For the phase-coherent
attractors, a sharp drop was observed from w(0) = 0.98 to w(4) ≈ 0.26, whereas for the two
settings yielding funnel attractors, one has w(0) ≈ w(4) ≈ 0.95. While a detailed analysis
is beyond the scope of this work, this is in line with the previously-established facts. In
the former case, phase diffusion was weak, so phase synchronization ensued quickly but
in the absence of coherent amplitude fluctuations; by contrast, in the latter two cases, it
was the manifestation of a deeper level of generalized entrainment.4 This was confirmed
by considering different coupling strengths, at which the level of entrainment accordingly
12
changed but the qualitative difference in the effect of the warping parameter remained: for
the phase-coherent attractors, a drop is always observed, until the coupling is strong enough
to entrain the amplitudes, whereas for the funnel attractors, the effect of warping remains
always more constrained.
V. SINGLE-TRANSISTOR OSCILLATORS
To evaluate the proposed synchronization measure in an experimental scenario, we next
considered a structured network of coupled single-transistor oscillators. Each node consists
of an "atypical" circuit resembling a Hartley oscillator, whose autonomous dynamics are
controlled via a resistor placed in series to the external DC voltage source. For certain
resistance ranges, transition to chaos occurs, and manifests primarily in the form of cycle
amplitude fluctuations (Fig. 9a).20 A ring comprising 90 such oscillators was studied,
wherein nodes are diffusively coupled to their neighbors via resistors RC connected between
the transistor collectors. A minority of nodes also have an additional resistor RL towards
a distant site, delineating four segments which yield a "toy model" of the hub nodes in a
complex network (Fig. 9b). The system was physically realized with discrete components
on a circuit board (Fig. 9c).21
For the present purpose, we reanalyzed publicly available time-series recorded in a collectively-
chaotic phase (50,000 points, ≈ 750 cycles; example in Fig. 9d).22 We focused on the
fact that the 9 long-distance connections were implemented through stronger couplings
RL = 40 Ω than the 90 short-range links RC = 750 Ω realizing the ring; hence, despite
them being weakened by parasitic effects, one expects stronger entrainment to be detectable.
Accordingly, the non-warped value w(0) revealed that synchronization was elevated within
the collective of interconnected nodes in the distant segments, and it noticeably diffused to
the other nodes along the ring outside these "hub regions". Increasing the magnitude of the
warping parameter, particularly along the c/c = −1 direction and for w(c), had a striking
"focusing" effect: it gradually reduced the intensity of the synchronization diffused outside
these regions and attenuated the clusters spontaneously formed along the ring, delineating
increasingly clearly the distantly-coupled segments (Fig. 10).
Considering the same approach as in Section IV for identifying the nine long-distance
links from the time-series, a non-monotonic effect on accuracy was noted, starting from
h(0) = 0.89, markedly decreasing and then recovering, eventually reaching h(−2) = 1 (Fig.
11a); an analogous difference was observed for the non-normalized values w(c) (data not
shown). Underlying this observation there appeared to be two competing, qualitatively dif-
ferent effects recalling the simulations of Rossler systems in Fig. 6b versus Fig. 6c. Among
the long-distance coupled nodes the situation was one of near-complete synchronization,
and warping had a non-monotonic effect on the measured coherence, delineating a dip
down from hw∈(0)i = 0.98. As a consequence, for c < 2 these links transiently became
less clearly distinguishable from preferential entrainments along the weaker short-distance
connections (Fig. 11b). Differently from the coupled Rossler oscillators and the electroen-
cephalogram, the dip location and depth markedly depended on the warping direction. The
reason is that activity was spike-like, implying that the amplitudes fluctuated considerably
on the temporal scale of individual cycles, hence there was considerable spectral overlap
between sj(t) and rj(t). On the other hand, across the rest of the network the measured
coherence steadily decreased with warping, delineating a sharp drop from hw6∈(0)i = 0.37
to hw6∈(4)i ≈ 0.17 (Fig. 11c), primarily due to increased sensitivity to the non-null average
relative phase (data not shown). As a consequence, the ensemble of directly structurally-
connected nodes eventually became more clearly distinguishable.
The relationship between phase- and amplitude-related effects can be intricate, and a
proper comparison with other synchronization measures is beyond the scope of this initial
study, in that it would require comprehensive consideration of diverse system dynamics,
coupling strengths, time-series lengths etcetera. Nevertheless, for the only purpose of visual
comparison, three other representative synchronization measures2 were computed for these
13
FIG. 9. Single-transistor oscillator network. a) Ring fragment showing three adjacent oscillators,
wherein R ≈ 1180 Ω (adjusted individually), V = 5 V, L1 = 10 µH, L2 = 8.2 µH, C = 30 pF,
and wherein RC = 750 Ω and RL = 40 Ω implement, respectively, short- and long-distance links
(additional 2.2 nF blocking capacitor in series with RL not shown). b) Network topology, with
the ring segments representing "hub regions" in red and corresponding node numbers. c) Physical
realization of an oscillator. d) Representative time-series. See Ref. (21) for a detailed description.
14
FIG. 10. Effect of increasing the warping parameter −c on the spatial pattern of warped phase
coherence hwjk(c)i measured from the experimental network and averaged over six rotations of the
link topology over the physical oscillators, as detailed in Ref. (21). Underlying adjacency matrix
represented as a graph in Fig. 9b.
FIG. 11. Warped phase coherence as a means of inferring the stronger long-distance links from the
experimental time-series. a) Accuracy in classifying the long-distance coupled vs. the other node
pairs h(c), b) and c) Corresponding average synchronization inside hw∈(c)i and outside hw6∈(c)i
the ensemble of long-distance coupled node pairs.
time-series. First, linear correlation, which assumes y(t) ∝ x(t). Second, generalized syn-
chronization, which assumes a possibly more complicated functional relationship between
the trajectories y(t) = Φ(x(t)); this was estimated via a rank-based measure, namely the
L-index, setting τ = 40 ns, w = 100 ns, m = 5 and k = m + 1.23 Third, normalized mutual
information, which assumes a possibly non-linear but scalar relationship y(t) = Φ(x(t));
this was estimated as IXY/√HXHY, where I and H represent, respectively, the mutual
information and the Shannon entropies, calculated over N = 16 bins over a maximum lag
d = ±40 ns.24 Taking into account the different sensitivity of these measures, also to noise,
it appears evident that the resulting synchronization matrices are overall more similar to
those yielded by warped phase coherence w(c) with relatively large values of c than to
that obtained for the non-warped value w(0) (Fig. 12). This is in line with the view that
warping confers sensitivity to aspects of entrainment beyond the mere phase locking, as was
also indicated by the simulations of Rossler systems considered in Section IV. It should also
be noted that rank-, bin- and time-lag embedding based measures are inherently poorly
15
FIG. 12. Alternate synchronization measures. a) Linear correlation C, b) Generalized synchro-
nization L and c) Normalized mutual information N . See Section V for description.
FIG. 13. Effect of increasing the warping parameter c on the spatial pattern of warped phase
coherence across the electrodes located on different scalp regions. a) Average over all conditions
and participants hwjk(c)i. b) Rest-activation and c) left-right imagery contrasts expressed in units
of pooled standard deviation hzjk(c)i. Pericentral electrodes (FCx, Cx and CPx sites): j = 1 . . . 21.
Frontal electrodes (AFx, Fx and FPx sites): j = 22 . . . 38. Parietal-occipital electrodes (Px, POx
and Ox sites): j = 47 . . . 64. Electrodes overlaying the sensorimotor cortex: left (C5, C3, CP5 and
CP3 sites) j = 8, 9, 15, 16 and right (C4, C6, CP4 and CP6 sites) j = 13, 14, 20, 21.
suited for use on short time-series epochs, such as those considered next in Section VI (640
vs. > 10, 000 points).23 -- 25
VI. ELECTROENCEPHALOGRAM
Finally, we demonstrate the empirical value of the proposed measure for analyzing elec-
troencephalographic signals with the purpose of decoding imaginary actions, representing
a typical scenario in brain-computer interface systems.26 This is only presented here to
exemplify a hypothetical application. Explicit comparison with established approaches for
constructing connectivity matrices from these signals, such as nonlinear associations23,24,
16
FIG. 14. Representative example of the effect of warping on determining the level of synchronization
between segments of two electroencephalographic signals. a) Chosen signals and corresponding
time-windows for synchronization determination. b) Effect of increasing the warping parameter
c on warped phase coherence in the three time-windows (curves averaged over 100 evaluations of
Eq. (9)). For the intermediate time-window, the measured coherence is very close to that of the
left time-window before warping, but more appropriately intermediate between the left and right
time-windows after warping is applied. See Section VI for description.
17
FIG. 15. Warped phase coherence as a means of detecting brain states corresponding to imaginary
actions. a) Accuracy in classifying epochs of activity or idleness hAI(c). b) Accuracy in classifying
epochs of left or right hand motor imagery hLR(c).
Granger causality25 and wavelet bicoherence27, is mandatory but left for future work.
In electroencephalographic signals, determinism is considerably weaker compared to the
cases considered thus far, and there may be contamination by non-neural noise sources. We
re-analyzed a publicly-available dataset of 106 recordings from healthy volunteers, and we
focused on a task alternating idle wakefulness with actively imagining to repeatedly open
and close the right or left fist (respectively 45, 22 and 23 trials). This dataset yielded 4
s-long epochs sampled at 160 Hz (640 points) over n = 64 channels, from electrodes laid
out onto the scalp according to the international 10-10 system.28 -- 30
Given the dense electrode montage, spatial filtering was firstly applied by subtracting to
each signal sj(t) with j = 1 . . . n the average reference channel hs(t)i = 1
j=1 sj(t). Then,
temporal filtering was performed over the frequency range from 7 to 30 Hz using a second-
order band-pass Butterworth filter, which preserved activity in the α, µ- and β-bands. These
bands index different aspects of movement planning and execution, and allow attenuating
any slow drifts and electromyographic contamination.31 -- 33 To prevent selection bias, we
refrained from manually inspecting and rejecting any recording. Following calculation of
the analytic signal with Eq. (1), epoching was performed. The warped phase coherence
matrix was thereafter obtained with Eq. (8), separately for each epoch. Brain states were
binarily classified at individual-participant level, separately for each setting of the warping
parameter c. Sparse logistic regression with variational approximation was performed over
all (n2 − n)/2 synchronization values, and the binary contrasts active-idle and left-right
hand were considered. The corresponding decoding accuracy h(c) was quantified via the
pair-wise leave-one-out approach over 20 random combinations.34
Consideration of the non-warped value hw(0)i, averaged over all epochs and participants, re-
n Pn
18
vealed strong synchronization within the pericentral electrodes, especially between adjacent,
ipsilateral subsets (diagonal lines), as well as within and between the parietal-occipital and
the frontal electrodes. Particularly over the range c ∈ [0, 1.5], increasing the warping param-
eter c had a prominent effect, recalling the observations on the electronic oscillators: spatial
structure became considerably more evident and the value distribution became wider, ex-
panding from hw(0)i ∈ [0.1, 0.8] to hw(1.5)i ∈ [−0.4, 0.8]. The average coherence decreased
and preferential entrainment between ipsilaterial electrodes over the central-parietal re-
gion consequently became emphasized (Fig. 13a). This trend, which was similar for the
non-normalized values h w(c)i, reflected amplitude-related effects as well as average relative
phase.
It should be noted that, even though in principle volume conduction could pose
a problem since warping leads to higher coherence values for small phase lags, it remains
unsettled whether zero-lag connectivity contains useful information or not.35,36 On another
note, unlike for the electronic oscillators, here the spectra of sj(t) and rj(t) had limited
overlap as only the latter were 1/f -like, hence the warping direction had limited influence
(data not shown).
To aid understanding one of the possible effects of warping on determining the level of
synchronization between electroencephalographic signals, it is helpful to consider a repre-
sentative example, consisting of signals acquired over two pericentral sites. In particular,
let us consider a segment: towards the beginning the temporal activity appears relatively
disordered, low-amplitude and unsynchronized, and towards the end a volley (or "spin-
dle") of large-amplitude, more coherent µ-band oscillations occurs (Fig. 14a). It is helpful
to consider warped phase coherence within three overlapping 4 s-long time-windows: two
covering only the baseline activity and only the large-amplitude oscillations (respectively,
until and after t = 25 s), and one covering an intermediate interval wherein both types of
activity are found. For this latter time-window, in the absence of warping the measured
level of synchronization was very close to that of the baseline activity: it failed to represent
the fact that its second half contains the initial part of the large-amplitude volley (i.e.,
w(0) ≈ 0.56). As the warping parameter was elevated, the measured coherence increased
up to a level approximately half-way between the two non-overlapping time-windows (e.g.
w(2) = 0.69 vs. 0.61, 0.77): this more appropriately represented the presence of a combi-
nation of both types of activity within this intermediate window (Fig. 14b). Volleys of
coherent µ-band oscillations are knowingly generated in relation to motor tasks by event-
related synchronization phenomena, and, as introduced in Section II, their detection can be
enhanced by warping primarily as a reflection of their larger amplitude.32,37
The task-related contrasts were extracted from the non-normalized values h w(c)i, else classi-
fication accuracy would be reduced. Notably, task-related differences were also attenuated
when removing the amplitude differences between channels through individually setting
hψji = 1 (data not shown). The active-idle difference normalized in units of pooled
standard deviation zij (c) became visibly stronger with increasing c: during task perfor-
mance, coherence markedly emerged within the pericentral electrodes, and between them
and the rest of the network (Fig. 13b). As expected, the left-right imagery separation
was more subtle, but there was a clearly-identifiable pattern wherein warping attenuated
differences elsewhere on the scalp. Simultaneously, it enhanced a modulation of the level
of synchronization between all pericentral electrodes and specifically those directly over-
lying the sensorimotor cortex. This yielded a lateralized change in sign, which plausibly
reflected event-related synchronization/desynchronization ipsilaterally vs. contralaterally
to the imaginary movement (Fig. 13c).32,37
As the warping parameter was elevated, the recognition accuracy steadily improved, reach-
ing a plateau for c ≥ 2. For the active-idle contrast, the accuracy increased from hAI(0) =
0.65 ± 0.12 to hAI(2) = 0.73 ± 0.12 (µ ± σ; Fig. 15a). For the left-right hand contrast, it
increased from hLR(0) = 0.57± 0.13 to hLR(2) = 0.64± 0.15 (Fig. 15b). For both contrasts,
the effect of warping appeared more closely related to the amplitude fluctuations than to the
average phases between electrodes. In line with other works, considerable inter-individual
variability was present, plausibly reflecting issues related to task performance which, being
the task covert, could not be ascertained behaviorally.38 At group level, both comparisons
19
were strongly statistically significant (paired t-test p < 0.001), and there was no effect of
the warping direction (p > 0.1). The number of features selected by the sparse regression
decreased for the first contrast (52 ± 18 vs. 48 ± 22, p = 0.03) but not for the second one
(51 ± 13).
Confirming the validity of the results, no differences were found comparing the odd with
the even epochs of idleness, and all effects were obliterated by phase randomization pre-
serving both auto- and cross-correlations (multivariate surrogates, p > 0.8).39 Discarding
the amplitude information by assuming r = 1, the accuracy after warping was substantially
reduced (p < 0.001). Nevertheless, a marginal improvement with respect to the non-warped
value remained visible for both contrasts (p = 0.05), reflecting sensitivity to the average rel-
ative phases. Accordingly, the accuracy obtained with warped phase coherence significantly
exceeded that yielded by cross-correlation on sj(t) (p = 0.03, data not shown). Repeat-
ing the analyses limited to the α, µ-band activity (7 to 13 Hz), the accuracy of active-idle
classification dropped significantly (p < 0.001), but there was no effect on left-right hand
classification performance (p = 0.3). Repeating them again for the β-band activity (13 to 30
Hz), the accuracy of warped phase coherence in both contrasts was unaffected (p > 0.4).37
Lastly, we considered two other behavioral tasks from the same participants and experimen-
tal sessions, with an identical blocked design. Firstly, the same motor task but involving
actual hand movement. Also in this case, for the active-idle contrast the accuracy increased
from h(0) = 0.75 ± 0.13 to h(2) = 0.82 ± 0.10 and for the left-right hand contrast it in-
creased from h(0) = 0.60± 0.12 to h(2) = 0.68± 0.13 (p < 0.001). Secondly, a different task
comparing simultaneous imaginary movement of both fists or feet. Again, for the active-idle
contrast the accuracy increased from h(0) = 0.65 ± 0.11 to h(2) = 0.71 ± 0.11 and for the
fists-feet contrast it increased from h(0) = 0.62 ± 0.16 to h(2) = 0.67 ± 0.15 (p < 0.001).
VII. CONCLUSIONS
Warping yields sensitivity to amplitude fluctuations and differences in the average phase.
On the one hand, this complicates the physical interpretation of phase coherence, since
one loses its intuitive meaning associated with Kuramoto-like networks. As warping is in-
creased, the measure becomes more and more distant from representing phase locking, even
reminiscent of linear correlation. On the other hand, warping seems to have some em-
pirical value for studying collective dynamics under partial synchronization across diverse
systems having free amplitude. This stems from the fact that, in non-linear systems, a
stable phase relationship is only one of the hallmarks of entrainment. Knowingly, as the
coupling between two non-identical oscillators is strengthened, also amplitude fluctuations
become increasingly coherent and, in the absence of frustration or other obstacles, lags are
eventually overcome.1 -- 3
Given a certain level of phase entrainment, w(c) may increase, decrease or fluctuate with
the level of warping c; as illustrated by the given examples, these denote qualitatively dif-
ferent situations. As illustrated by the cases of coupled Rossler systems, due to dependence
on dynamics it is not straightforward to compare warped phase coherence measurements
between different systems. However, it was shown that, within a given network, this mea-
sure can aid the identification of which node pairs are most strongly entrained, surpassing
both the phase locking value and cross-correlation. In particular, the consistent advantage
observed when setting c ≥ 2 (assuming hri=1), plausibly related to the "twisting" effect of
warping and consequentially increased amplitude sensitivity, appears noteworthy.
Even though in this exploratory study it is impossible to be prescriptive, a warping with
c ≈ 2 would appear to be consistently appropriate, even across the rather diverse systems
that were considered. For smaller values, a larger dependence on the level of warping may
remain, whereas for larger values, wjk(c) in Eq. (8) and w′
jk(c) in Eq. (10) may become
undesirably close to unity, in turn affecting the accuracy of calculating wjk(c) in Eq. (9).
More generally, the impact of the phase and amplitude distributions and the assumptions
underlying and the implications of the normalization step require further elucidation (i.e.,
20
effects visible on wjk(c) vs.
wjk(c)). Furthermore, systematic comparison with differ-
ent synchronization measures, in particular other variants of the phase locking value, is
necessary.5,6,36 With reference to aiding the inference of structural couplings, confirmation
over diverse coupling strengths, more complex and denser topologies is required.17 -- 19 As
regards the electroencephalogram, the influences of volume conduction and reference choice
require additional analysis. In particular, one needs to consider the sensitivity to average
phase differences and its impact on the determination of synchronization between scalp sites
at which a given activity component is recorded with opposite polarity.33,35,36
In any case, the observed improvement in the classification accuracy for imaginary ac-
tions, on average ≈ 8% and considerably higher for certain participants, appears potentially
clinically-relevant. It is in line with that yielded by other algorithmic enhancements vastly
more demanding to deploy, such as transforming the signals from sensor to source space via
inverse modeling.38,40 Furthermore, the accuracy was maximized when including the entire
α, µ- and β-bands, thereby alleviating the need for individually-adjusted frequency filter
settings.31,32 It remains to be seen whether a similar improvement could be replicated in
other datasets, obtained when applying this measure in source space, and generalized in the
context of different sensorimotor and cognitive tasks. As indicated above, it also remains to
be determined how the present measure compares to more consolidated alternatives, par-
ticularly in the context of determining connectivity over short signal segments.23 -- 25,27
Nevertheless, consistent improvements were seen across three different motor tasks. Hence,
we posit that inclusion of warping in brain-computer and brain-machine interface systems
based on the phase locking value and similar indices could be an immediate means of en-
hancing their performance.
ACKNOWLEDGMENTS
This work was supported in part by JSPS under Grant KAKENHI 26112004 and Grant
In addition to employment by the Polish Academy of Sciences,
KAKENHI 17H05903.
Krak´ow, Poland, L. Minati gratefully acknowledges funding by the World Research Hub
Initiative, Institute of Innovative Research, Tokyo Institute of Technology, Tokyo, Japan.
The authors are grateful to G. Schalk, W.A. Sarnacki, A. Joshi, D.J. McFarland and J.R.
Wolpaw for collecting motor imagery EEG data and making them publicly available on
PhysioNet.org. Their work was supported by grants from the NIH/NIBIB (EB006356
GS, EB00856 JRW and GS) and PhysioNet is supported by the NIH/NIGMS/NIBIB
(2R01GM104987-09).
1M.G. Rosenblum, A.S. Pikovsky and J. Kurths, Phase Synchronization of Chaotic Oscillators, Phys. Rev.
Lett. 76, 1804 (1996).
2S. Boccaletti, J. Kurths, G. Osipov, D.L. Valladares and C.S. Zhou, The synchronization of chaotic
systems, Phys. Rep. 366, 1 (2002).
3S. Boccaletti, A.N. Pisarchik, C.I. del Genio and A. Amann, Synchronization: From Coupled Systems to
Complex Networks, Cambridge University Press, Cambridge UK (2018).
4G.V. Osipov, B. Hu, Ch. Zhou, M.V. Ivanchenko, J. Kurths, Three types of transition to phase synchro-
nization in coupled chaotic oscillators, Phys. Rev. Lett. 91, 024101 (2003).
5C.K. Kovach, A Biased Look at Phase Locking: Brief Critical Review and Proposed Remedy, IEEE Trans.
Signal Process. 65, 4468 (2017).
6K.Q. Lepage and S. Vijayan, The relationship between coherence and the phase-locking value, J Theor
Biol. 435, 106 (2017).
7E. Lowet, M.J. Roberts, P. Bonizzi, J. Karel and P. De Weerd, Quantifying Neural Oscillatory Synchro-
nization: A Comparison between Spectral Coherence and Phase-Locking Value Approaches, PLoS One.
11, e0146443 (2016).
8J. Garca-Prieto, R. Bajo and E. Pereda, Efficient Computation of Functional Brain Networks: toward
Real-Time Functional Connectivity, Front Neuroinform. 11, 8 (2017).
9C. Brunner, R. Scherer, B. Graimann, G. Supp and G. Pfurtscheller, Online Control of a Brain-Computer
Interface Using Phase Synchronization, IEEE Trans Biomed Eng 53, 2501 (2016).
10A.I. Sburlea, L. Montesano and J. Minguez, Advantages of EEG phase patterns for the detection of gait
intention in healthy and stroke subjects, J. Neural Eng. 14, 036004 (2017)
21
11B. Boashash, Estimating and interpreting the instantaneous frequency of a signal. II. Algorithms and
applications, Proc. IEEE 80, 540 (1992).
12L.V. Gambuzza, J. G´omez-Gardenes and M. Frasca, Amplitude dynamics favors synchronization in com-
plex networks, Sci. Rep. 6, 24915 (2016).
13A. Pikovsky, M. Rosenblum, and J. Kurths, Synchronization: A universal concept in nonlinear sciences,
Cambridge University Press, Cambridge UK (2001).
14R.T. Canolty, R.T. Knight, The functional role of cross-frequency coupling, Trends Cogn. Sci. 14, 506
(2010).
15O.E. Rossler, An Equation for Continuous Chaos, Phys. Lett. 57, 397 (1976).
16E. Ott, Chaos in Dynamical Systems, Cambridge University Press, Cambridge UK (2002).
17G. Tirabassi, R. Sevilla-Escoboza, J.M. Buld´u and C. Masoller, Inferring the connectivity of coupled
oscillators from time-series statistical similarity analysis, Sci. Rep 5, 10829 (2015).
18M. Timme, J. Casadiego, Revealing networks from dynamics: an introduction, J Phys A Math Theor.
47, 343001 (2014).
19N. Rubido, A.C. Mart´ı, E. Bianco-Martnez, C. Grebogi, M.S. Baptista and C. Masoller, Exact detection
of direct links in networks of interacting dynamical units, New J. Phys. 16, 093010 (2014).
20L. Minati, Experimental dynamical characterization of five autonomous chaotic oscillators with tunable
series resistance, Chaos 24, 033110 (2014).
21L. Minati, Synchronization, non-linear dynamics and low-frequency fluctuations: analogy between spon-
taneous brain activity and networked single-transistor chaotic oscillators, Chaos 25, 033107 (2015).
22See http://www.lminati.it/listing/2018/c/ for the experimental time-series.
23D. Chicharro and R. G. Andrzejak, Reliable detection of directional couplings using rank statistics, Phys.
Rev. E 80, 026217 (2009)
24L. Paninski, Estimation of entropy and mutual information, Neural Comput. 15, 1191 (2003)
25L. Faes, G. Nollo and K. H. Chon, Assessment of Granger causality by nonlinear model identification:
application to short-term cardiovascular variability, Ann Biomed Eng. 36, 381 (2008).
26L.M. Alonso-Valerdi, R.A. Salido-Ruiz and R.A. Ramirez-Mendoza, Motor imagery based braincomputer
interfaces: An emerging technology to rehabilitate motor deficits, Neuropsychologia 79, 354 (2015).
27X. Li, D. Li, L.J. Voss and J.W. Sleigh, The comodulation measure of neuronal oscillations with general
harmonic wavelet bicoherence and application to sleep analysis, Neuroimage 48, 501 (2009).
28A.L. Goldberger, L.A.N. Amaral, L. Glass, J.M. Hausdorff, P.Ch. Ivanov, R.G. Mark, J.E. Mietus, G.B.
Moody, C.-K. Peng and H.E. Stanley, PhysioBank, PhysioToolkit, and PhysioNet: Components of a New
Research Resource for Complex Physiologic Signals, Circulation 101, e215 (2000).
29G. Schalk, D.J. McFarland, T. Hinterberger, N. Birbaumer and J.R. Wolpaw, BCI2000: A General-
Purpose Brain-Computer Interface (BCI) System, IEEE Trans Biomed Eng. 51, 1034 (2004).
30See https://www.physionet.org/physiobank/database/eegmmidb/ and http://www.bci2000.org for the
experimental time-series.
31A.S. Aghaei, M.S. Mahanta and K.N. Plataniotis, Separable Common Spatio-Spectral Patterns for Motor
Imagery BCI Systems, IEEE Trans Biomed Eng. 63, 15 (2016).
32C. Neuper and G. Pfurtscheller, Evidence for distinct beta resonance frequencies in human EEG related
to specific sensorimotor cortical areas, Clin. Neurophysiol. 112, 2084 (2001).
33F. Chella, V. Pizzella, F. Zappasodi and L. Marzetti, Impact of the reference choice on scalp EEG
connectivity estimation, J Neural Eng. 13, 036016 (2016).
34O. Yamashita, M.A. Sato, T. Yoshioka, F. Tong and Y. Kamitani, Sparse estimation automatically selects
voxels relevant for the decoding of fMRI activity patterns, Neuroimage 42, 1414 (2008).
35W. Jian, M. Chen and D.J. McFarland, EEG based zero-phase phase-locking value (PLV) and effects of
spatial filtering during actual movement, Brain Res. Bull. 130, 156 (2017).
36R. Bruna, F. Maest´u and E. Pereda, Phase locking value revisited: teaching new tricks to an old dog, J.
Neural Eng. 15, 056011 (2018).
37D.J. McFarland, L.A. Miner, T.M. Vaughan and J.R. Wolpaw, Mu and Beta Rhythm Topographies
During Motor Imagery and Actual Movements, Brain Topogr. 12, 177 (2000).
38M. Ahn and S.C. Jun, Performance variation in motor imagery braincomputer interface: A brief review,
J. Neurosci. Methods 243, 103 (2015)
39T. Schreiber and A. Schmitz, Improved surrogate data for nonlinearity tests, Phys. Rev. Lett. 77, 635
(1996).
40N. Yoshimura, A. Nishimoto, A.N. Belkacem, D. Shin, H. Kambara, T.S. Hanakawa and Y. Koike,
Decoding of Covert Vowel Articulation Using Electroencephalography Cortical Currents, Front Neurosci.
10, 175 (2016).
|
1803.00643 | 1 | 1803 | 2018-03-01T21:51:39 | How strong are correlations in strongly recurrent neuronal networks? | [
"q-bio.NC",
"cond-mat.dis-nn",
"nlin.CD",
"physics.bio-ph"
] | Cross-correlations in the activity in neural networks are commonly used to characterize their dynamical states and their anatomical and functional organizations. Yet, how these latter network features affect the spatiotemporal structure of the correlations in recurrent networks is not fully understood. Here, we develop a general theory for the emergence of correlated neuronal activity from the dynamics in strongly recurrent networks consisting of several populations of binary neurons. We apply this theory to the case in which the connectivity depends on the anatomical or functional distance between the neurons. We establish the architectural conditions under which the system settles into a dynamical state where correlations are strong, highly robust and spatially modulated. We show that such strong correlations arise if the network exhibits an effective feedforward structure. We establish how this feedforward structure determines the way correlations scale with the network size and the degree of the connectivity. In networks lacking an effective feedforward structure correlations are extremely small and only weakly depend on the number of connections per neuron. Our work shows how strong correlations can be consistent with highly irregular activity in recurrent networks, two key features of neuronal dynamics in the central nervous system. | q-bio.NC | q-bio | How strong are correlations in strongly recurrent neuronal networks?
Ran Darshan1,2,3, Carl van Vreeswijk3 and David Hansel3∗
1. ELSC, The Hebrew University of Jerusalem, Israel;
2. Janelia Research Campus, Howard Hughes Medical Institute, Ashburn, VA, USA;
3. Center for Neurophysics, Physiology and Pathology; Cerebral Dynamics,
Plasticity and Learning Research Team; CNRS-UMR8119, Paris, France
8
1
0
2
r
a
M
1
]
.
C
N
o
i
b
-
q
[
1
v
3
4
6
0
0
.
3
0
8
1
:
v
i
X
r
a
ABSTRACT
Cross-correlations in the activity in neural networks
are commonly used to characterize their dynamical states
and their anatomical and functional organizations. Yet,
how these latter network features affect the spatio-
temporal structure of the correlations in recurrent net-
works is not fully understood. Here, we develop a gen-
eral theory for the emergence of correlated neuronal ac-
tivity from the dynamics in strongly recurrent networks
consisting of several populations of binary neurons. We
apply this theory to the case in which the connectivity de-
pends on the anatomical or functional distance between
the neurons. We establish the architectural conditions
under which the system settles into a dynamical state
where correlations are strong, highly robust and spatially
modulated. We show that such strong correlations arise
if the network exhibits an effective feedforward structure.
We establish how this feedforward structure determines
the way correlations scale with the network size and the
degree of the connectivity.
In networks lacking an ef-
fective feedforward structure correlations are extremely
small and only weakly depend on the number of connec-
tions per neuron. Our work shows how strong correla-
tions can be consistent with highly irregular activity in
recurrent networks, two key features of neuronal dynam-
ics in the central nervous system.
I.
INTRODUCTION
Two-point correlations are commonly used to charac-
terize collective dynamics in extended systems [1–7]. Re-
cent technical advances [8–10] make it possible to simul-
taneously record the activity of many neurons in net-
works in the brain. This allows for the measurement of
two-point correlations for large numbers of neuronal pairs
in spontaneous activity as well as upon sensory stimula-
tion or in controlled behavioral conditions.
Correlations in neuronal activity impact the ability of
networks to encode information [11–14]. Correlations are
also functionally important in performing sensory, mo-
tor, or cognitive tasks [15]. For instance, correlated os-
cillatory activity has been hypothesized to be involved
in visual perception [16].
In a recent study combining
modeling, electrophysiology and analysis of behavior, we
∗ Correspondance author:[email protected]
argued that non-oscillatory correlated neuronal activity
in the central nervous system is essential in the gen-
eration of exploratory behavior [17]. Correlations are
also important for the self-organization of neuronal net-
works through activity dependent plasticity [18]. Indeed,
changes in synaptic strength are thought to depend on
the temporal correlation of the activty of the pre and
postsynaptic neurons [19, 20].
Correlation strengths depend on the brain area, [21,
22], the layer in cortex [23], stimulus conditions, behav-
ioral states [24] and experience [25–27]. A wide range of
values for correlation coefficients, from negligible [21, 28]
to substantial [11, 29–33] have been reported in the last
two decades. Correlation coefficients are usually higher
for close-by neurons than for neurons that are far apart
[31, 34–37]. In cortex, they drop significantly over dis-
tances of 200 − 400 µm [35, 37]. Recent works have re-
ported correlations varying non-monotonically with dis-
tance [38] or correlations which are positive for close-by
neurons but negative for neurons farther apart [39]. Cor-
relation coefficients also depend on functional properties
of the neurons. Neurons which code for similar features
of sensory stimuli are more correlated [21, 31, 34, 37, 40].
Neurons in cortex receive recurrent inputs from several
hundreds [41] to a few thousands [42] of other neurons.
Individual connections can induce post-synaptic poten-
tials in a range of 0.1 mV to several mV [35]. Thus, 50
simultaneous inputs are sufficient to trigger or suppress
a spike in a neuron. These facts have been incorporated
√
in model networks with strongly recurrent connectivity
in which connection strengths are O(1/
K), where K is
the average number of inputs per neuron. This scaling is
in contrast to the one used in standard mean-field mod-
[43]) where connections are O(1/K) and thus
els (e.g.
√
are much weaker. Recent experiments in cortical cul-
tures are consistent with a 1/
K scaling of connection
strength [44].
van Vreeswijk and Sompolinsky [45, 46] showed that
strongly recurrent networks consisting of two populations
of neurons, one excitatory (E) and one inhibitory (I),
randomly connected on a directed Erdos-R´enyi graph,
operate in a state in which the strong excitation is bal-
anced by the strong inhibition. In this balanced regime,
neurons receive strong excitatory and inhibitory inputs,
each O(
K), but due to the recurrent dynamics these
inputs cancel each other at the leading order. This can-
cellation, which does not require fine-tuning, results in
O(1) net inputs into the neurons whereas spatial hetero-
geneities and temporal fluctuations in the inputs are also
O(1). As a result, the activity of the neurons remains
√
finite and exhibits strong temporal irregularity and het-
erogeneity [47, 48].
The latter results were established in two-population
sparsely connected networks. Renart et al.
[28] showed
that they also hold for densely connected networks. This
is because in unstructured and strongly recurrent net-
works the dynamics suppress correlations [49]. Despite
the fact that in these networks neurons share a finite
fraction of their inputs, they operate in an asynchronous
state with O(1/N ) correlations, where N is the number
of neurons in the network [50].
These previous studies focused on strongly recurrent
unstructured networks with two neuronal populations.
In the brain, however, neural networks comprise a diver-
sity of excitatory and inhibitory cell types, which differ in
their morphology, molecular signature and importantly
for our purpose, in their connectivity. These networks
are also structured at many levels.
In particular, the
probability of connection falls off with distance and de-
pends on functional properties of pre- and postsynaptic
neurons. For example, in mouse primary auditory cor-
tex the probability of excitatory neurons to be connected
decays to zero within ∼ 300µm [35]. In cat primary vi-
sual cortex neurons interact locally on range of ∼ 500µm,
whereas long range patchy connections are observed up
to several mms [51, 52].
In the present paper we investigate how structure in
network connectivity can give rise to strongly correlated
activity. Our goal is to explore the general architectural
features that control the strength of pair-wise correla-
tions in strongly recurrent neural circuits consisting of
several excitatory and inhibitory neuronal populations.
In Section II, we define the network architectures and
the neuronal model we use. In Section III, we establish
a set of constraints, the balanced correlation equations,
that need to be satisfied in any strongly recurrent net-
work if firing rates do not saturate. We derive in Sec-
tion IV explicit expressions for the Fourier components
of the correlations in two-population networks with spa-
tially modulated connectivity. We establish the condi-
tions under which correlations are strong and show that
these do not violate the balanced correlation equations.
Section V is devoted to networks with an arbitrary num-
ber of populations. We prove two theorems, which state
for which network architectures correlations are O(1/N )
and for which they increase with K, when K is large. In
Section VI, we apply these theorems to specific examples.
In Section VII, we assume that K = O(N γ) and derive
a bound on γ for which the scaling theorems still apply.
The paper closes with a discussion of our results.
II. THE NETWORK MODEL
II.1. Architecture
We consider a neuronal network with a ring architec-
ture, comprising D neuronal populations, some excita-
tory and others inhibitory (Fig. 1; [53, 54]). For sim-
2
plicity, we assume that all populations have the same
number of neurons, N . Neuron i, in population α (neu-
ron (i,α)) is located at angle θα
N with, i = 1, . . . , N
and α = 1, . . . , D. The probability, P αβ
ij , that a neuron
(j, β) projects to neuron (i, α) depends on their distance
j ),
N fαβ(θα
on the ring (Fig. 1b). We write: P αβ
j ) = N . In this paper we assume
where(cid:80)
j fαβ(θα
i − θβ
i − θβ
ij = K
i = 2πi
a finite number of non-zero Fourier modes in fαβ.
Thus, a neuron in population α receives, on average,
K inputs from neurons in each of the populations β. We
denote by Λ the adjacency matrix of the network con-
nectivity
(cid:40)
Λαβ
ij =
1 if (j, β) is presynaptic to (i, α)
0 otherwise
For simplicity we assume that all connections from pop-
ulation β to population α have the same strength, jαβ.
We thus define the connectivity matrix, J , as
ij = jαβΛαβ
J αβ
ij .
Note that jαβ is positive (negative) if population β is
excitatory (inhibitory).
In all this paper we focus on strongly recurrent net-
works, characterized by interactions which are O( 1√
)
[45, 46]. Thus, we scale the synaptic strengths with the
mean connectivity as
K
jαβ =
Jαβ√
K
.
(1)
where Jαβ is O(1).
II.2. Single neuron dynamics
i . When the neuron is quiescent, Sα
The state of neuron (i, α) is characterized by a binary
variable, Sα
i = 0,
while if it is active, Sα
i = 1. The total synaptic current
into neuron (i, α) at time t, hα
i (t), is the result of all
its recurrent interactions with the other neurons in the
network, as well as the feedforward inputs coming from
outside the network. It is given by
hα
i (t) =
j (t) + I α
i
(2)
D(cid:88)
N(cid:88)
β=1
j=1
ij Sβ
J αβ
√
where the external input I α
to be constant in time. Here, Iα is O(1) and thus I α
√
O(
KIα(θ) [46] is assumed
is
i =
K).
i
The neurons are updated with Glauber dynamics at
zero temperature [1, 28, 43]. Specifically, update times
for neuron (i, α) are Poisson distributed with rate 1/τ :
if neuron (i, α) is updated at time t, Sα
is set to 1 if
i (t) ≥ Tα and to 0 otherwise (for simplicity, we assume
i
hα
the same threshold, Tα, for all neurons in population α,
where hα(θα
gument. For large N it is given by:
i ) = [(cid:104)hα
3
i (cid:105)t]J is a smooth function of its ar-
√
K
hα(θ) =
Jαβ
fαβ(θ − θ(cid:48))mβ(θ(cid:48)) + Iα(θ)
D(cid:88)
β=1
(cid:90) dθ(cid:48)
2π
.
(5)
The second term in Eq. (4) is the quenched disorder in
the input. It satisfies
∆hα
i =
∆Λαβ
ij mβ(θβ
j ) + Λαβ
ij ∆Sβ
j
,
(6)
where ∆Λαβ
ij = Λαβ
ij ]J .
Finally, the temporal fluctuations in the inputs are
D(cid:88)
β=1
(cid:88)
(cid:104)
j
Jαβ√
K
ij − [Λαβ
D(cid:88)
δhα
i (t) =
Jαβ√
K
β=1
j
(cid:88)
(cid:105)
FIG. 1. Network architectures and dynamics. a. Net-
works consist of D populations of neurons, excitatory (E,red)
and inhibitory (I, blue), recurrently coupled and receiving
external drives. Triangles: Excitatory connections. Circles:
Inhibitory connections. b. The neurons in each population
are located on a ring. The probability that neuron (j, β) is
connected to neuron (i, α) depends on their distance. c. Left:
Neurons are binary units with zero temperature Glauber dy-
namics. Right: The network operates in the balanced regime.
Red: Excitatory input into a neuron. Blue: Inhibitory input
into the same neuron. Black: The net input is on the order of
the threshold (green, T = 1). Time is given in units of τ = 1.
and the same update rate for all neurons). Accordingly,
the transition probability, w, (Fig. 1c) can be written as
w(Sα
i (t) → 1 − Sα
i (t)) =
i (t) − Θ(hα
[Sα
i (t) − Tα)]2.
1
τ
ij δSβ
Λαβ
j (t).
(7)
√
Because we scale the synaptic strength as 1/
K, these
temporal fluctuations are O(1) when correlations are
√
weak. At first sight, hα(θ) is O(
K). This would imply
that, depending on whether hα(θ) is positive or nega-
tive, the neurons fire either at very high or at very low
rate. This happens unless the network settles into a state
in which excitation and inhibition are, on average, bal-
anced. In this case, the net average inputs to the neurons
are in fact O(1). On the other hand, large spatial and
temporal correlations could in principle lead to temporal
fluctuations and heterogeneities in the inputs of O(
K).
Thus in presence of strong correlations it is not sufficient
to require that the mean input, hα(θ), is O(1), for the
network to operate in a biologically relevant regime. For
that, we need both the mean net inputs and the input
fluctuations to be O(1) at any time and for all neurons.
When this happens we say that the network operates in
the balanced regime.
√
where Θ is the Heaviside function. We normalize time so
that τ = 1.
III.1. Balance equations for the quenched averaged
population activities
III. SPATIO-TEMPORAL PROFILE AND
CORRELATIONS OF THE ACTIVITY FOR
N, K → ∞
We write the state of neuron (i, α) as
Sα
i (t) = mα(θα
i ) + ∆Sα
i + δSα
i (t),
(3)
i (t)(cid:105)t]J is the average of Sα
where mα(θ) = [(cid:104)Sα
i (t) over
realizations of the network connectivity matrix and over
time. In Eq. (3), the term ∆Sα
is the quenched disorder,
i
∆Sα
i (t) represents
the temporal fluctuations in the activity of neuron (i, α),
δSα
i (t)(cid:105)t − mα(θα
i (t)(cid:105)t.
i (t) − (cid:104)Sα
i (t) = Sα
Similarly, we can write:
i ), whereas δSα
i = (cid:104)Sα
As in [55, 56], the requirement that the mean input is
√
fαβ(θ − θ(cid:48))mβ(θ(cid:48)) + Iα(θ) = O(1/
K),
O(1) yields
D(cid:88)
β=1
Jαβ
(cid:90) dθ(cid:48)
2π
(8)
for all α ∈ {1, . . . , D} and all θ ∈ [0, 2π). In the large K
limit this yields a set of linear equations which determines
the functions mα(θ) to leading order. These equations
can be written in Fourier space
J (n)
αβ m(n)
β + I (n)
α = O
.
(9)
(cid:19)
(cid:18) 1√
K
(cid:88)
β
where the superscript n denotes the nth Fourier mode
and we have used the short hand notation
hα
i (t) = hα(θα
i ) + ∆hα
i + δhα
i (t),
(4)
αβ = Jαβf (n)
J (n)
αβ .
(10)
distance between neurons (deg)0-909000.4Pαβw(01)w(10)1-w(01)1-w(10)100ms−20−1001020Exc inputInh inputnet input0100timeinput10Distance-dependent probability of connectionabcIIIEE which is the nth Fourier mode of the connectivity matrix.
In what follows, we consider J (n)
αβ as the elements of a
D × D matrix, J (n).
The spatial average of the activities, m(0)
α , must be
non-negative for the balanced state to exist. This im-
plies that the parameters Jαβ and the external inputs,
I (0)
α , must satisfy a set of inequalities. For example, for
networks with two populations, one excitatory (α = E)
and one inhibitory (α = I), these inequalities are [46]
I (0)
E
I (0)
I
>
JEI
JII
>
JEE
JIE
.
(11)
In general, Eq. (9), also implies additional
inequali-
ties that must be satisfied by f (n)
αβ to guarantee that
mα(θ) ≥ 0 for all α and θ [55, 56]. In the present work
we focus on the case where the external inputs are spa-
α = 0 for n ≥ 1, and the
tially homogeneous. Thus, m(n)
only condition which is required for the balanced state
to exist is: m(0)
α ≥ 0.
To study the stability of the homogeneous balanced
state it is useful to introduce the interaction matrix
¯J αβ
ij = gαJαβfαβ(θα
i − θβ
j )
(12)
where gα is the population averaged gain (see Appendixes
A,D; [46, 49]).
Small perturbations, δmα(θ, t), of the activity profile
around this homogeneous state evolve according to [46]:
dδm(n)
α
dt
= −δm(n)
α +
√
K
¯J (n)
αβ δm(n)
β ,
(13)
(cid:88)
β
α (t) and ¯J (n)
where δm(n)
δmα(θ, t) and ¯J αβ
ij . Since each row of ¯J αβ
non-zero elements which are O(1/K), ¯J (n)
αβ are the nth Fourier modes of
ij has O(K)
αβ is O(1).
The stability of the balanced state with respect to per-
turbations in mα(θ) requires that, for all n, all the eigen-
values of the matrices J (n) have real parts smaller than
1. For instance, for a two population (E,I) network one
must have in the large K limit
JEEJIIf (n)
EEf (n)
II − JIEJEIf (n)
EI f (n)
IE ≤ 0.
(14)
for all n. Note that the gain of the neurons, gα, dropped
from this equation. The balanced state undergoes a Tur-
ing bifurcation when for some n ≥ 1, JEEJIIf (n)
II −
JIEJEIf (n)
IE crosses 0 and becomes positive.
EEf (n)
EI f (n)
III.2. Balance equation for pair-wise correlations
The time-lagged auto- and cross-correlation functions
of the activities of a pair of neurons, (i, α), (j, β), for
(j, β) (cid:54)= (i, α), are
i (τ ) = (cid:104)δSα
aα
ij (τ ) = (cid:104)δSα
cαβ
i (t + τ )(cid:105)t
j (t + τ )(cid:105)t
i (t)δSα
i (t)δSβ
(16)
(15)
In what follows, it is convenient to define cαα
ii = 0.
4
Due to the randomness of the connectivity, the number
of excitatory and inhibitory inputs varies from neuron to
neuron, resulting in heterogeneous firing rates between
neurons even when they belong to the same population
[46, 48]. This structural randomness also results in het-
erogeneity of the auto-correlations of single neuron ac-
tivities. It also contributes to the heterogeneity in pair
cross-correlations. The latter is further enhanced by the
spatial variability in the number of inputs shared by pair
of neurons. A full characterization of the distributions
of the auto- and cross-correlations is beyond the scope of
this paper. Instead, here we will focus on their popula-
tion averages.
The population average auto-correlation, Aα, is given
by Aα(τ ) = 1
i (τ ), which in the thermodynamic
N
limit is also Aα(τ ) = [aα
i (τ )]J . With the architecture we
use, the probability that neurons (i, α) and (j, β) share
common inputs from a third neuron, (k, γ) is
i aα
(cid:80)
Pr(Λαγ
K 2
N 2 fαγ(θα
ik = 1 ∧ Λβγ
i − θγ
k )fβγ(θβ
jk = 1) =
j − θγ
k )
(17)
As is evident from this equation, the number of shared
inputs averaged over all pairs of neurons separated by
the same distance on the ring, ∆, depends only on
∆. Thus we define the average cross-correlations as
Cαβ(∆, τ ) = (cid:104)cαβ
ij (τ )(cid:105), where (cid:104).(cid:105) denotes the average over
i − θβ
pairs with θα
j = ∆. In the thermodynamic limit,
this quantity does not depend on the specific realization
of the network connectivity matrix. The Fourier expan-
sion of this function is
N−1(cid:88)
n=0
(cid:80)
Cαβ(∆, τ ) =
αβ (τ )e−in∆,
C (n)
(18)
where C (n)
αβ (τ ) = 1
N 2
kj cαβ
kj (τ )ein(θα
k −θβ
j ).
Equation (9), which determines the population aver-
aged firing rates, stems from the constraint that the net
input in every neuron must be O(1) when K is large.
The condition that for any pair of neurons the correla-
tion in their inputs is finite in that limit leads to another
constraint, which we now derive.
Let us consider the N D × N D matrix Q defined by
ij = [(cid:104)δhα
Qαβ
for (i, α) (cid:54)= (j, β) and Qαα
(cid:88)
(cid:88)
JαγJβγ(cid:48)
Qαβ
ij =
γγ(cid:48)
K
kk(cid:48)
i (t)δhβ
j (t)(cid:105)t]J ,
(19)
ii = 0. Using Eq.(7), one finds
k(cid:48) (t)(cid:105)t]J
jk(cid:48) ]J [(cid:104)δSγ
ik Λβγ(cid:48)
[Λαγ
k (t)δSγ(cid:48)
for (i, α) (cid:54)= (j, β). Similarly to Cαβ(∆), Qαβ
of ∆ = θα
i − θβ
ij
j only.
Eq.(20)
Expanding
in
Fourier
(Q(n)
αβ
≡
(20)
is a function
(cid:80)
1
N 2
kj Qαβ
kj ein(θα
k −θβ
j )), yields
Q(n)
αβ = K
αγ J (−n)
J (n)
βγ(cid:48) C (n)
γγ(cid:48) +
(cid:88)
γγ(cid:48)
(cid:88)
γ
αγ J (−n)
J (n)
βγ Aγ
K
N
which can be written in matrix form
Q(n) = KJ (n)C(n)[J (n)]† +
J (n)A[J (n)]†
K
N
(21)
Here X denotes the D × D matrix Xαβ, Aαβ = Aαδαβ
and the superscript † denotes Hermitian conjugation
([J (n)
αβ ]† = [J (−n)
βα ]).
with
Note that the diagonal of the matrix Q is not the vari-
ance of the inputs. The latter is (see Appendix E4)
(cid:88)
nγγ(cid:48)
(cid:88)
γ
σ2
α = K
αγ J (n)
J (n)
αγ(cid:48)C (n)
γγ(cid:48) +
J 2
αγAγ
(22)
The requirement that crosscorrelations of the inputs
into pair of neurons are at most O(1) implies that all the
quantities Q(n)
αβ are also at most O(1). This yields
J (n)C(n)[J (n)]† = O(1/K)
(23)
5
can be larger than O(1/K). In fact we will see that in
those cases correlations can even increase with K.
Finally, we note that one can also write a balanced cor-
relation equation for the quenched disorder of the neural
activity, using the fact that [∆hα
also must be at
most O(1). This requirement leads to the condition
i ∆hβ
j ](n)
J
J (n)Γ(n)[J (n)]† = O(1/K)
[Γ(n)]αβ ≡ [∆Sα
i ∆Sβ
j ](n)
J
(24)
(25)
IV. CORRELATIONS IN TWO-POPULATION
NETWORKS
In Appendix A we derive an equation for the spatial
Fourier modes of the equal-time quenched average corre-
lation functions, Cαβ(∆, 0). It yields (omitting the sec-
ond argument)
√
(cid:16) ¯J (n)C(n) + C(n)[ ¯J (n)]†(cid:17)
(cid:16) ¯J (n)A + A[ ¯J (n)]†(cid:17)
√
,
K
+
K
N
2C(n) =
(26)
We call Eq. (23), the balanced correlation equation. It
implies that for all n for which the matrix J (n) is in-
vertible and has entries O(1), C(n) is smaller or equal to
O(1/K) and is thus weak. For a broad class of network
architectures, we show below that correlations are in fact
O(1/N ) and barely depend on K, for large K. When,
however, J (n) is singular for some n ≥ 0, correlations
There we also show that the solution of this equation is
a fixed point of the dynamics of the correlations which
is always stable when the balanced state is stable with
respect to perturbation in the population rates.
This equation holds for a network with an arbitrary
number of neuronal populations.
In the case of two
populations, it can be solved explicitly, yielding after a
straightforward calculation (see Appendix E)
−2AE + AE
+
1
N
EE = − AE
(cid:16)
C (n)
N
EI = − 1
C (n)
N
II = − AI
C (n)
N
+
1
N
EI
IE + AI ¯J (n)
(cid:16)
AE ¯J (n)
−2 + 3T (n)
−2AI +
√
EI )2(cid:17)
II
√
−2 + 3T (n)
T (n) + 2 ¯J (n)
K −(cid:16)
K −(cid:16)
AE(∆(n) + ( ¯J (n)
(cid:16)
II )2) + AI ( ¯J (n)
(cid:17)
(cid:17)√
K −(cid:0)(T (n))2 + 2∆(n)(cid:1) K + T (n)∆(n)K 3/2
(cid:17)√
K −(cid:0)(T (n))2 + 2∆(n)(cid:1) K + T (n)∆(n)K 3/2
(cid:17)√
K −(cid:0)(T (n))2 + 2∆(n)(cid:1) K + T (n)∆(n)K 3/2
AI (T (n) + 2 ¯J (n)
−2 + 3T (n)
EE )2)) + AE( ¯J (n)
AI (∆(n) + ( ¯J (n)
¯J (n)
EI + AE ¯J (n)
K −(cid:16)
AI ¯J (n)
¯J (n)
√
K
EE
EE
IE
II
IE )2(cid:17)
K
K
(27)
where T (n) and ∆(n) are the trace and the determinant
of the matrix ¯J (n).
The expansion of these expressions for large K gives
AI ( ¯J (n)
N
K
EE = − AE
N − 1
√
C (n)
AE ¯J (n)
√
C (n)
EI = 1
N
II = − AI
N − 1
√
C (n)
K
II
N
K
T (n)∆(n)
EI )2+AE (∆(n)+( ¯J (n)
¯J (n)
IE +AI ¯J (n)
¯J (n)
T (n)∆(n)
IE )2+AI (∆(n)+( ¯J (n)
EE
EI
T (n)∆(n)
AE ( ¯J (n)
EE )2)
II )2)
N K )
+ O( 1
+ O( 1
+ O( 1
N K )
N K )
(28)
6
are not negligible, which happens when T (1)∆(1) is suf-
ficiently small, short range correlations can be positive
and longer range correlations negative (Eq. (30)). This
is the case in Fig. 2a, dashed line.
Figure 2b compares simulations (circles) and analyti-
cal results (solid lines) for the dependence on N of N C (1)
αβ
(parameters as in panel a, red solid line). It shows that
the spatial modulation of the correlations in the simula-
tion is close to the large K analytical results. Figure 2c-d
depict the dependence of N Cαβ on K. In the whole range
of K considered here simulations and analytical results
are close. For Cαα the nearby correlations and modu-
lation amplitudes increase with K and are much larger
than that of CEI.
K.
EE and C (n)
√
II by a factor 1/
Thus in general when K is large C (n)
II are very
small, namely O(1/N ), with negative prefactors which
EI , it is smaller than C (n)
do not depend on K. As for C (n)
and C (n)
It should be noted that to derive Eq. (28) we assumed
that T (n) (cid:54)= 0 and ∆(n) (cid:54)= 0. Equation (27) indicates,
however, that when T (n) = 0 and ∆(n) (cid:54)= 0, or T (n) (cid:54)= 0
and ∆(n) = 0, C (n)
EI are O(1/N ).
II and C (n)
EE, C (n)
The situation is different if T (n) = 0 and ∆(n) = 0.
Equation (27) shows that in this case it is possible to get
correlations which are O(K/N ).
EE
In the rest of this section we consider in detail two-
population networks in which for the probabilities of con-
nection only the first two Fourier modes are non-zero
(cid:104)
(cid:16)
(cid:17)(cid:105)
P αβ
ij =
K
N
1 + 2f (1)
αβ cos
i − θβ
θα
j
,
(29)
with α = E, I, β = E, I.
IV.1. T (1) (cid:54)= 0 and ∆(1) (cid:54)= 0
For f (1)
αβ such that T (1) (cid:54)= 0 and ∆(1) (cid:54)= 0 we have in
the large N, K limit
CEE(∆) = − AE
N
√
1
(1 + 2 cos ∆) + O(
√
1
)
K
N
EI cos(∆)) + O(
(30)
1
N K
)
N
( ¯C (0)
EI + 2 ¯C (1)
CEI (∆) =
K
CII (∆) = − AI
(1 + 2 cos ∆) + O(
N
√
√
1
)
K
N
KC (n)
EI ≡ N
EI are O(1). Thus, in that limit,
where ¯C (n)
the spatial average and modulation of the correlations
within the E and I populations do not depend on K,
at the leading order. Moreover, CEE(∆) and CII (∆)
depend on the synaptic strengths only because the auto-
correlations AE and AI depend on these parameters.
Figure 2 depicts simulation results for N = 40000 and
K = 2000. Figure 2a plots CEE(∆) for f (1)
αβ = 0.25
(α, β ∈ E, I) and two sets of values for the interaction
strengths (solid and dashed lines). For comparison we
also plots the results of a simulation when the connec-
αβ = 0; α, β ∈ E, I; gray line).
tivity is unstructured (f (1)
In all these cases CEE(∆) is very small (note the scale
on thr y-axis). When the connectivity is spatially mod-
ulated CEE(∆) varies with distance. However, the spa-
tial averages are comparable in the three cases consid-
ered ( 1
2π
ment with Eq. (30) since, to the leading order, the auto-
correlations, AE and AI , are not expected to depend on
whether the connectivity is spatially modulated or not.
Note that according to Eq. (30), CEE(∆) and CII (∆)
are negative for close-by neurons (∆ small) and positive
for neurons that are far apart. This is the case for the set
of parameters corresponding to the solid line in Fig. 2a
(see also Fig. 2b-d). However, when finite K corrections
(cid:82) CEE(∆)d∆ ∼ −0.2 × 10−5). This is in agree-
IE = f (1)
EI = f (1)
EI = f (1)
EE = f (1)
FIG. 2. Correlations in two-population E-I networks.
Connection probabilities are as in Eq. (29). Panels a, c,
d: N = 40000. a. Top panel: The network architecture.
αβ (cid:54)= 0). Bottom
All connections are spatially modulated (f (1)
panel: Simulation results for CEE(∆) with K = 2000. Red:
f (1)
EE = f (1)
II = 0.25. For comparison the correla-
tion is also plotted for f (1)
II = 0 (Gray).
For red-solid and gray lines other parameters are JEE =
0.3, JIE = 3, JEI = 2.5, JII = 5; IE = 0.3, II = 0.3, TE =
I (cid:39) 0.13,
1, TI = 0.7. With these parameters m(0)
gE (cid:39) 0.22, gI (cid:39) 0.1 and AE (cid:39) 0.1, AI (cid:39) 0.1. For red-dashed
line other vparameters are: JEE = 1; JIE = 2; JEI = 1; JII =
1.5; IE = 0.2; II = 0.08. b. N C (1)
αβ vs. N for K = 1000. c.
N C (1)
αβ vs. K. d. N Cαβ(0) vs. K (N = 40000). In panels
b, c, d: Solid line: Analytics, Eq. (27). Circles: Simulations.
Red: α = β = E. Blue: α = β = I. Green: α = E, β = I.
Parameters are as for red-solid line in panel a.
E (cid:39) 0.12, m(0)
IE = f (1)
6-600180-180x10-680004000200-0.30-0.2-0.1∆ (deg)CEE(∆)# of inputs (K)ab100# of neurons (N)N • C(1)αβc0.1d204080160x103-0.2-0.1N • Cαβ(0)80004000200-0.10# of inputs (K)N • C(1)αβEI IV.2. T (1) = ∆(1) = 0
II = f (1)
(cid:54)= 0 and
We now consider a network in which f (1)
EI
f (1)
EE = f (1)
IE = 0. The spatially modulated compo-
nent of the interaction has therefore an explicit feedfor-
ward structure (Fig. 3a, top panel) and T (1) = ∆(1) = 0.
Solving Eq. (27) shows that the correlations are on
average O(1/N ) and that their modulations are
K
AI
C (1)
EE =
√
N
2
EI = −
K
C (1)
N
C (1)
II = 0
¯J (1)
EI 2
¯J (1)
EI
AI
2
(31)
As a result, correlations in the E population are spatially
modulated and O(K/N ). They are positive for short
range and negative for long range. This is in contrast
to the correlations in the inhibitory population which
are not spatially modulated and O(1/N ) while the EI
√
correlations are spatially modulated and O(
K/N ).
Figure 3 compares analytical results with simulations.
Panel a plots simulation results for Cαβ(∆) for f (1)
EI =
0.25. Other parameters are as in Fig. 2a (grey solid line).
Thus, the locally averaged firing rates are the same as in
simulations in the latter figure. Correlations in the in-
hibitory neurons (blue) are extremely weak and are not
spatially modulated. In contrast, the correlations of ex-
citatory pairs (red) are larger by two orders of magni-
tude compared to those in Fig. 2. Correlations between
E and I neurons (green) are weaker than those between
E neurons. They are negative for nearby neurons while
for nearby excitatory pairs they are positive. All these
features are in agreement with our analytical results,
In this range, C (1)
The spatial modulation of the EE correlations in-
creases with K (Fig. 3b, circles). There is quantitative
agreement between simulations and theory (Eq. (31)) up
to K ≈ 4000 for N = 40000.
EE in
the simulations varies linearly with K. For larger values
of K, C (1)
EE is larger than predicted by Eq. (31). This
is because this equation was derived by linearizing the
dynamics, which is only valid when correlations are not
too large.
In fact, simulation results for fixed K devi-
ate less from Eq. (31) when N is increased (Fig.3b-c).
When K increases, C (1)
EI , becomes more negative (Fig. 3c;
green). Here too, for N = 40000 simulations agree well
with Eq. (31) up to K ≈ 4000 and deviations are smaller
when N is larger. The spatial modulation of the II cor-
relations in the simulations are extremely small (Fig. 3c;
blue) as the theory predicts.
EE, increases quadratically with f (1)
Finally, according to Eq. (31), the spatial modula-
tion, C (1)
EI whereas C (1)
EI
varies linearly with this parameter. Our simulations
are in very good agreement with these analytical results
(Fig. 3d).
7
IE = f (1)
EE = f (1)
FIG. 3. Correlations in a two population network with
an explicit feedfworward structure. Connection proba-
bilities as in Eq. (29). Parameters: f (1)
II = 0.
Panels a, b, c: f (1)
EI = 0.25. Interaction strengths as in Fig. 2.
a. Top panel: The network architecture. Green: the connec-
EI (cid:54)= 0). Connections
tions which are spatially modulated (f (1)
in gray are unstructured (f (1)
αβ = 0). Bottom panel: Simula-
tion results for Cαβ(∆); N = 40000, K = 2000 .b, c. N C (1)
αβ
vs. K. Solid lines: Analytical results, Eq. (31). Simulation
results are plotted for N = 20000 (plus), N = 40000 (circles)
and N = 80000 (crosses). b. N C (1)
II . Green:
N C (1)
EI . Solid lines: Analytical results,
Eq. (31); N = 40000, K = 400.
EE. c. Blue; N C (1)
EI . d. N C (1)
αβ vs. f (1)
IV.3. ∆(1) = 0, T (1) (cid:54)= 0
The network investigated in IV.2 has an explicit feed-
forward structure. Adding any spatial modulation to the
(cid:54)= 0, top panel of Fig. 4a) destroys
II connectivity (f (1)
II
this structure and now T (1) (cid:54)= 0 (while ∆(1) = 0). Solv-
ing Eq. (26) for that case, one finds:
¯J (1)
EI 2
√
K ¯J (1)
II )(1 +
¯J (1)
EI
√
KJ (1)
II )(1 +
¯J (1)
II
√
K ¯J (1)
II )
K
N
EI = − K
C (1)
N
II = − K
C (1)
N
K ¯J (1)
II )
√
KJ (1)
II )
(2 +
√
EE =
K(1 +
√
AI
C (1)
AI
(32)
√
AI
K(2 +
In the large N, K limit, CEE(∆) and CII (∆) are both
O(1/N ) and do not depend on K, whereas CEI (∆) is
2-200180-180x10-480004000400010-0.51.500.20.40.501∆ (deg)CII(∆),CEE(∆),CEI(∆)# of inputs (K)modulation (f (1)EI)abdN • C(1)αβN • C(1)αβ80004000400-10# of inputs (K)cN • C(1)αβEI √
O(1/(N
K)). Therefore the addition of a spatial mod-
ulation in the II interactions suppresses the correlations
that inhibitory projections induce in the E population.
Eq.(32). This yields
Q(1)
EE =
AI
N
K( ¯J (1)
EI )2
√
K ¯J (1)
II
1 +
8
(33)
Thus, when II interactions are spatially modulated, a
cancellation between terms which are O(K/N ), reduces
√
EE by a factor of O(1/
Q(1)
EE is much
smaller than when II interactions are not modulated. A
similar argument explains the suppression in C (1)
EI .
K). As a result, C (1)
Fig. 4a depicts simulation results for f (1)
EI = 0.25 and
three values of f (1)
II . It demonstrates the suppression of
correlations in the excitatory population when II interac-
tions are also modulated. The dependence of this effect
on f (1)
II , decreases
II
the modulation of all the correlations in very good agree-
ment with the analytical results (compare circles and
solid lines).
is depicted in Fig. 4b. Increasing f (1)
√
K (cid:39) 1/ ¯J (1)
According to Eq. (32), the correlation in the E pop-
ulation always increases linearly with K, for small K.
The cross-over between this regime, where CEE(∆) is
O(K/N ), and the large K regime, where CEE(∆) is
O(1/N ), occurs for
II . Figure 4c depicts this
crossover in numerical simulations. Thus, although in the
large N, K limit a transition from O(K/N ) to O(1/N )
(cid:54)= 0, this qualitative
correlations occurs as soon as f (1)
II
difference is significant only when K is sufficiently large.
In other words, for moderately large number of inputs
per neuron, correlations can exhibit a close to linear in-
crease even if the structure of spatial modulation of the
interaction matrix is not completely feedforward.
V. SCALING CORRELATION THEOREMS
In the previous section we studied networks with two
neuronal populations. In this case, it is straightforward
to analytically derive explicit expressions for the correla-
tions. These expressions are simple enough to fully clas-
sify how the network structure affects the scaling (with
K and N ) of the correlations. For networks consisting of
more than two populations, analytical expressions for the
correlations can in principle be derived. However, deal-
ing with these expressions becomes rapidly impractical
as the number of populations increases. In the following
we adopt an alternative approach. We prove two general
theorems, which, given the network architecture, allow
us to determine how correlations vary with K, without
computing these correlations explicitly.
Jordan basis of ¯J (n). We write
¯J (n) = U (n)J (n)
[ ¯J (n)]† = V (n)[J (n)
To prove these theorems, we rewrite Eq. (26) in the
jor [U (n)]−1
jor ]∗[V (n)]−1
(34)
Here x∗ denotes the complex conjugate of x, U (n)
(V (n)) are matrices whose rows are the generalized eigen-
vectors of ¯J (n), and J (n)
jor is the Jordan normal form of
FIG. 4. Spatial modulation in the II interactions sup-
presses the correlations in the E populations. Same
network as in Fig. 3 but f (1)
II = 0.25. N = 40000. a.
Top panel: The network architecture.
Interactions plotted
in green are spatially modulated. Bottom panel: Simula-
tion results for CEE(∆). N = 40000, K = 2000. Gray line:
f (1)
II = 0; Dashed- red: f (1)
II = 0.25.
b. N C (1)
II . Circles: Simulations. Solid lines: Solution
of Eq. (32). N = 40000, K = 400. Colors are as in Fig. 3.
c. Left panel: N C (1)
EE vs. K. Solid lines: Eq. (32). Circles:
Simulations. Gray line: f (1)
II = 0.025;
Dashed red: f (1)
II = 0.05; Solid red : f (1)
II = 0.25. Right panel:
Same as left panel but in a log-log scale.
II = 0.05; Solid-red : f (1)
II = 0; Dotted red: f (1)
αβ vs. f (1)
To understand further the origin of this, let us consider
the quenched average correlations of the inputs, Qαβ(∆)
(Eq. (19)). Using Eq. (21), one finds
EE = K( ¯J (1)
Q(1)
EI = K ¯J (1)
Q(1)
II = K( ¯J (1)
Q(1)
EI
II )2(C (1)
II +
EI )2(C (1)
II +
¯J (1)
II (C (1)
II +
AI
N
)
AI
N
)
AI
N
)
When ¯J (1)
On the other hand, when ¯J (1)
II = 0, C (1)
II = 0, therefore Q(1)
EE = K
(cid:54)= 0, C (1)
N ( ¯J (1)
EI )2AI .
is given by
II
II
0180-1802-20x10-4800040004000126# of inputs (K)N • C(1)EE∆ (deg)CEE(∆)ac00modulation (f (1)II)b0.5N • C(1)αβ0.30.6EI10-210010210-4100102# of inputs (K) ¯J (n). This implies that [J (n)
jor ]∗ is the Jordan form of
[ ¯J (n)]† is the [ ¯J (n)]†. We can write
9
[J (n)
jor ]µν = λ(n)
with µ, ν ∈ 1, ..., D and (n)
and zero otherwise.
Equation (26) then yields
µ δµν + (n)
µ δµ,ν−1
µ = 1 within a Jordan block
(cid:104)
2 −
√
(cid:105) C (n)
µν =
K(λ(n)
µ + [λ(n)
ν ]∗)
+
√
K((n)
√
µ
K
N
C (n)
µ+1,ν + (n)
ν−1
C (n)
µ,ν−1)
((n)
µ
A(n)
µ+1,ν + (n)
ν−1
A(n)
µ,ν−1) +
√
K
N
(λ(n)
µ + [λ(n)
ν ]∗) A(n)
µν
(35)
where we have defined
(n)
C
(n)
A
= [U (n)]−1C(n)V (n)
= [U (n)]−1A(n)V (n)
(36)
Note that while the matrix C(n) is symmetric and the
matrix A(n) is diagonal, this is not in general the case
for C
and A
(n)
(n)
.
Let us assume that the network is in a stable balanced
has no zero elements. In
(n)
state in which the matrix A
Appendix B we prove
Correlation Theorem 1: The nth Fourier mode of the
correlation matrix scales as O(1/N ) if and only J (n)
jor does
not have a Jordan block whose real part is a shift matrix
(A shift matrix, S, of dimension P is a P × P matrix of
the form Sµν = δµ+1,ν).
Correlation Theorem 2: If ¯J (n) has at least one Jordan
block whose real part is a shift matrix, the nth Fourier
mode of the correlation matrix is O(K P (n)−1/N ), where
P (n) is the dimension of the largest block in J (n)
jor , whose
real part is a shift matrix.
Corollary: The nth Fourier mode of the correlation ma-
trix is O(K D−1/N ) if and only if J (n) is nilpotent of
degree D.
In Appendix B we also show how to extend these re-
sukts to case where A
(n)
has zero elements.
The matrix U (n) can be viewed as a transformation
of the original network of D populations into a network
of D effective populations. The condition that ¯J (n) has
a Jordan block which is a shift matrix of size P , can
be interpreted as the existence of P effective populations
whose effective interaction matrix is feedforward in its
nth Fourier mode. In other words, the original network
exhibits a hidden feedforward structure, which is embed-
ded in the nth Fourier mode of its connectivity. Theorems
1 and 2 therefore implies that only when such a structure
exists, the nth mode of the correlation matrix increases
with K. To know which elements in this matrix increase
with K one has to compute the matrices U (n) and V (n)
(see Eq. (36)).
VI. APPLICATIONS OF THE CORRELATION
THEOREMS
In this section we consider networks comprising D pop-
ulations with connection probabilities
(cid:16)
(cid:17)
P αβ
ij =
K
N
1 + 2f (1)
αβ cos(θα
i − θβ
j )
(37)
with α = 1, ..., D and β = 1, ..., D.
VI.1. Two population networks
(cid:18) λ1 0
(cid:19)
For a network of two populations the Jordan form
0 λ
0 λ2
(cid:19)
jor =
jor =
(cid:18) λ 1
of the matrix ¯J (1) has the form: J (1)
if ¯J (1) is diagonalizable. Otherwise, it has the form:
J (1)
with λ real. Theorem 1 and 2 imply
that only in the second case with λ = 0, some of the
correlations Cαβ(∆) are O(K/N ). Otherwise, all corre-
lations are O(1/N ).
It is equivalent to say that some
correlations are O(K/N ) if and only if ¯J (1) (cid:54)= 0 and
T (1) = ∆(1) = 0. We therefore recover the result we de-
rived in Section IV without explicit computation of the
correlations.
As noted above, the correlation theorems do not tell
us which of the elements of the correlation matrix are
O(K/N ) when J (1)
. However, since a 2 × 2
matrix has such a Jordan form if and only if it is nilpo-
jor =
0 0
(cid:18) 0 1
(cid:19)
= 0) we have to consider two types of
tent (
connectivity:
Type 1: The network has an explicit feedforward
structure, i.e., the interaction matrix is either ¯J (1) =
(cid:16) ¯J (1)(cid:17)2
(cid:18)
(cid:18) 1
(cid:19)
or ¯J (1) =
(cid:18) 0
0
¯J (1)
IE 0
0 ¯J (1)
EI
0
0
(cid:19)
.
(cid:19)
(cid:18)
(cid:19)
0
, whereas V (1) =
In the former case,
0 1/ ¯J (1)
the matrix U (1)
1/ ¯J (1)
Eq. (36) gives (see Appendix B, Eq. (B8))
K/N ) O(K/N )
K/N )
0
(cid:18) O(
√
√
C
=
(1)
EI
(cid:19)
is U (1) =
0
1
EI 0
. Then
(38)
O(
EE = O(K/N ), C (1)
Using Eq. (36), one finds that C (1)
II = 0,
whereas C (1)
K/N ), in agreement with Eq. (31).
A similar calculation in the latter case (when the mod-
√
EI = O(
ulation is in the IE interactions) gives C (1)
O(K/N ), and C (1)
Eq. (27).
√
EI is O(
II =
K/N ), all in agreement with
EE = 0, C (1)
Type 2: ¯J (1) = c
where a, c > 0.
(cid:18) 1 −a
1/a −1
(cid:19)
In this case the network has no explicit feedforward
structure since all four interactions (EE, EI, IE, II) are
spatially modulated. It has, however, a hidden feedfor-
ward structure, as revealed by the Jordan form of the
interaction matrix.
For this network, the transformation matrices are
U (1) =
and V (1) =
. Using
(cid:18) −1 −1/c
(cid:19)
a
0
(cid:18) a a/c
(cid:19)
1
0
(1)
Eq. (36) and Eq. (38), it is clear that the transforma-
√
tion from C
to C(1) mixes elements which are 0 and
K/N ), with those which are O(K/N ). Thus, while
O(
is O(K/N ), all the ele-
in C
ments of the correlation matrix C(1) are O(K/N ). This
is also in line with Eq. (27).
only the element C (1)
12
(1)
We consider an example of such a network in Fig. 5.
The parameters Jαβ and the external
inputs are as
in Fig. 2-4. Therefore, to leading order, the pop-
ulation averaged activities, mα, the autocorrelations,
Aα, and the population gains, gα are the same as in
Figs. 2-4. The modulation of the connection probability,
f (1)
EE, f (1)
II , are all non-zero (Fig. 5a) and tuned
so that:
IE , f (1)
EI , f (1)
(cid:18) 1 −1/2
2 −1
(cid:19)
¯J (1) =
1
20
The Jordan form of this matrix is graphically repre-
sented in Fig. 5b. The top panel in Fig. 5c depicts
simulation results for the correlations in this network.
They are all positive for close-by neurons and their spa-
tial modulations increase approximately linearly with K
(Fig. 5c, bottom). The top panel of Fig. 5d shows that
most of the power in these correlations results from the
√
element C12 (red). The latter increases linearly with K
(Fig. 5d, bottom), while C11, C22 = O(
K/N ) and C21
is two order of magnitude smaller and does not exhibit
significant change with K (note the log-log scale in this
10
figure). These simulations are in quantitative agreement
with Eq. (27) up to K ≈ 4000 and the deviations for
larger K decrease with N (as in Fig 4).
The interaction matrix, ¯J (1), depends on the matrix
J (1), and on the population gains, gE and gI . If the con-
nectivity matrix has an explicit feedforward structure,
the interaction matrix has also such a structure, indepen-
dantly of the population gains. Thus, although changing
the external inputs, IE, II , modifies this gains, this does
not destroy the feedforward structure and thus does not
change the scaling of the correlations with K and N . In
contrast, in networks with a hidden feedforward struc-
ture, this scaling is sensitive to perturbations in the ex-
ternal inputs since the hidden feedforward structure is
destroyed unless the ratio gE/gI remains constant.
EE = 0.373, f (1)
EI = 0.0224, f (1)
FIG. 5. Correlations in a two population network with
a hidden feedforward structure. a. Network architec-
ture. All connections are modulated. Connection probabili-
ties are as in Eq. (29) with f (1)
II =
0.0487, f (1)
IE = 0.1623. Connection strengths are as in Figs. 2-
3. b. Hidden feedforward structure of the mode n = 1 as re-
vealed in the Jordan basis. c. Top panel: Simulation results
for Cαβ(∆). N = 40000, K = 2000. Bottom panel: N C (1)
αβ vs.
K. Solid lines: Analytical results (Eq. (27)). Simulations are
plotted for N = 20000 (plus), N = 40000 (circles), N = 80000
(cross) and N = 160000 (asterisks). d. Top panel: Correla-
tion matrix in the Jordan basis of ¯J (1). Purple: C11(∆);
Brown: C22(∆); Gray: C21(∆); Red: C12(∆). Bottom panel:
N C (1)
µν vs. K in log-log scale. Solid line: Linear fit. Dashed
line: Fit to a square root function.
80004000400# of inputs (K)N • C(1)αβ012N • C(1)µν1-100180-180x10-5∆ (deg)a10310410-410-2100~K~√K# of inputs (K)bcd1-100180-180∆ (deg)C11(∆),C22(∆),C21(∆),C12(∆)x10-4102CII(∆),CEE(∆),CEI(∆)EI VI.2. Examples with three populations or more
The two networks depicted in Fig. 6 consist of one exci-
tatory and two inhibitory populations. The positivity of
the spatial averaged activity of these three populations,
m(0)
α (α = 1, 2, 3), constrains the parameters (see Sec-
tion III), Jαβ, through a set of inequalities. We leave the
calculation of these conditions to the reader.
11
In both cases, the first Fourier mode of the population
average connectivity matrix has the form
0 J (1)
13
12 J (1)
22 J (1)
0 J (1)
23
0
0
0
J (1) =
(39)
with J (1)
12 , J (1)
13 , J (1)
23 < 0 (left panels of Fig 6a-c).
In the network of Fig. 6a, J (1)
22 = 0. Therefore J (1)
and J (1) are nilpotent of degree 3. According to the
Corollary of Section V, the first mode of the correlation
matrix is O(K 2/N ).
The Jordan form of ¯J (1) is
0 1 0
0 0 1
0 0 0
.
J (1)
jor =
This is graphically represented in Fig. 6a, middle panel.
The matrix C
satisfies (see Appendix B, Eq.(B8) )
(1)
O(K/N ) O(K 3/2/N ) O(K 2/N )
O(K 1/2/N ) O(K/N ) O(K 3/2/N )
O(K 1/2/N ) O(K/N )
0
(1)
C
=
Using the transformation matrices (Eq.(34)), one can
show that correlations are O(K 2/N ) only within the ex-
citatory population.
In the network in Fig. 6b, J (1)
interaction matrix, ¯J (1), is not nilpotent.
form is
22 < 0. Therefore, the
Its Jordan
0 1
0
0 0
0
0 0 J (1)
22
.
J (1)
jor =
The upper Jordan block is a shift matrix of degree 2.
The corresponding feedforward structure is graphically
represented in Fig. 6b (middle panel). According to The-
is O(K/N ). It satisfies (see Appendix B,
orem 2, C
Eq. (B8))
(1)
O(
√
(1)
C
=
K/N ) O(K/N ) O(1/N )
√
O(
K/N ) O(1/N )
0
O(1/N ) O(1/N )
O(1/N )
FIG. 6. Examples of networks with three populations
and their Jordan representations. Probability of connec-
tions are as in Eq. (37). a-b. Left: A network of three popula-
tions, two inhibitory and one excitatory. Gray: Unstructured
connections. Green: Spatially modulated connections. Mid-
dle: The Jordan representation of the n = 1 Fourier mode
of the network connectivity (left panel). Right: The scaling
23 (cid:54)= 0. Other
of the strongest correlations. b. J (1)
entries of the matrix J (1) are zero (see main text). b. Same
as (a), but with J (1)
13 , J (1)
12 , J (1)
22 (cid:54)= 0.
This approach can be generalized to arbitrary number
of populations to classify the scaling of the correlation
matrix for different architectures. Examples of networks
with four populations are depicted in Fig 7, together with
the graphic representations of their Jordan forms and the
maximum order of the correlations.
VII. CONSTRAINT ON SCALING OF THE
NUMBER OF INPUTS WITH THE NETWORK
SIZE
In this section we assume that K and N scale together
K ∝ N γ
(40)
with 0 < γ ≤ 1.
Equation(26), which determines the correlations of the
neuronal activities, is obtained under the Ansatz that
the correlations between the inputs hα
j (t) are
sufficiently small, namely o(1) (see Appendix A). This
condition is more stringent than the balance correlation
equation, Eq. (23). It can be written as
i (t) and hβ
J (n)C(n)[J (n)]† = o(1/N γ)
(41)
and using the transformation matrices one can show that
is O(K/N ). Other correlations are either
√
only C(1)
11
O(
K/N ) or O(1/N ).
This condition constrains γ as we now show.
According to Theorem 1, if the Jordan form, J (n)
jor ,
has no Jordan block whose real part is a shift matrix for
any n, correlations will be O(1/N ). This will also be
a'pop'1'pop'2'pop'3J(1)13b'pop'1'pop'2'pop'3Network architectureJordan formOrder of correlationsK/NK2/NJ(1)13IJ(1)12J(1)23EIIJ(1)22J(1)12J(1)23EI 12
FIG. 7. Examples of networks with four populations and their Jordan representations. Probability of connections
are as in Eq. (37). a-c. Left: The network consists of two coupled E-I networks. Gray: Unstructured connections. Green:
Spatially modulated connections. Middle: The Jordan representation of the n = 1 Fourier mode of the connectivity (network
of the left panel. Right: The scaling of the strongest correlation. d. Same as (a-c), with a population averaged connectivity
matrix as in Eq. (46). Middle: The Jordan form is complex (Eq. (47)). Black lines corresponds to complex eigenvalues.
the order of J (n)C(n)[J (n)]†. Therefore, Eq. (41) only
requires γ < γmax = 1.
If the Jordan form, J (n)
jor , contains a Jordan whose real
part in the shift matrix, we have to apply Theorem 2. In
this theorem, P (n) is the dimension of the largest block
in J (n)
jor , with a real part which is a shift matrix (see
Section V). Let us denote by Pmax the largest P (n) over
all Fourier modes, i.e.,
and by Cmax the corresponding block in C. Equation
(41) then yields
For instance, for Pmax = 2, we have (see Eq. (B8))
S CmaxS = o(N−γ)
(cid:18) O(N γ/2−1) O(N γ−1)
(cid:19)
0
O(N γ/2−1)
Cmax =
(44)
Pmax = max
n
P (n)
and thus
(42)
S CmaxS = 0
Equation (41) implies
¯J (n)C(n)[ ¯J (n)]† = o(N−γ)
(cid:88)
m,k
for all n. This yields in the Jordan basis
For Pmax = 3, we have
[J (n)
jor ]µm C (n)
mk [[J (n)
jor ]∗]kν = o(N−γ)
(43)
Cmax =
and
for all µ, ν.
By definition of Pmax, for at least one Fourier mode,
n, the matrix J (n)
jor has at least one block whose real part
is a shift matrix of degree Pmax. In general, there can be
several such Jordan blocks. For example, in the network
depicted in Fig. 7b, for which Pmax = 2, there are two
Jordan blocks with P = 2.
We first assume that all blocks which are a shift matrix
of size Pmax are real. Since for such blocks in J jor CJ ∗
jor
scale the same with K, it is sufficient to consider the case
where there is only one such block. We denote it by S
Therefore, for this block Eq. (41) is always satisfied. The
latter equation, however, also applies to other Jordan
blocks and Fourier modes. This implies that γ < γmax =
1
O(N γ/2−1) O(N γ−1) O(N 3γ/2−1)
O(N γ/2−1) O(N γ−1)
O(N γ−1) O(N 3γ/2−1) O(N 2γ−1)
0 O(N γ/2−1) O(N γ−1)
O(N γ/2−1)
0
S CmaxS =
0
0
0
0
0
In general, for a Pmax × Pmax shift matrix S CmaxS is
Thus, Eq. (41) is satisfied only if γ < γmax = 1/2.
O(N γ(Pmax−2)−1). This implies that γ < γmax with
γmax =
1
Pmax − 1
(45)
EIEIK/NEIEIK/NEIEIEIEINetwork architectureJordan formOrder ofcorrelationsNetwork architectureJordan formOrder ofcorrelationsK2/NK/Nabcd .
Let us now consider networks in which there is at least
one pair of complex conjugate Jordan blocks whose real
parts are a shift matrix of size Pmax. An example, of
such a network is depicted in Fig. 7d. The first Fourier
mode of the population averaged connectivity matrix in
this example is
0
0
0
0 −a
0
b
b
0
0
0
0
0
0 −a c
iω 1
0 iω
0
0
0
0
0
0
0 −iω
1
0 −iω
0
,
J (1) =
J jor =
√
with a, b, c real and positive. For this network
(46)
(47)
13
network operates in the balanced regime [45]. For sim-
plicity we considered networks with a one dimensional
ring architecture with connection probabilities solely de-
pendent on distance and on the nature (excitatory or
inhibitory) of the pre- and postsynaptic populations.
We present a balanced correlation equation, which to-
gether with the balanced rate equation, define the bal-
anced regime and insure that mean inputs to the neu-
rons and their fluctuations are both O(1). We derive a
set of equations that determine the equilibrium values
of the quenched averaged correlations and we show that
the solution of these equations is stable provided that the
solution of the balanced rate equations is stable.
Key results of our work are two scaling correlation the-
orems. The first shows that generically, all the Fourier
modes of the quenched average correlations are small
when N and K are large. They are of O(1/N ), and inde-
pendent of K to leading order. This is true in the large
N limit even if we take K = pN , provided that p is not
too large. However, the second theorem states that there
are recurrent network architectures in which some of the
Fourier modes of the quenched averaged correlations in-
crease with K. These architectures are characterized by
an explicit, or a hidden, feedforward structure in those
modes. This structure is revealed by the Jordan form of
the interaction matrix averaged over realizations. If this
Jordan form contains a block whose real part is a shift
matrix of size P > 1, the corresponding mode in the
correlation increases at least as O( KP −1
N ). Importantly,
in these cases the network still operates in the balanced
regime provided that K (cid:46) N
P , see Sec-
tion VII). Corrections to the theory become important
when K approaches N
P −1 (or K (cid:46) N 1
P −1 (or N 1
P ).
1
1
For simplicity, we assumed that synaptic weights de-
pend only on the identities of the populations to which
pre and postsynaptic neurons belong. The results, how-
ever, will not change if the weights are heterogeneous
with distributions that depend on the pre and postynap-
tic populations, as long as the mean and the variance of
all these distributions are finite.
For notational simplicity we assumed that all popu-
lations have the same number of neurons, N , and each
population receives, on average, inputs from K neurons
in every population. The theory can be easily extended to
networks in which population α has Nα = ναN neurons
and the average number of connections from population
β to population α is Kαβ = καβK. This will not affect
the scaling of the correlations with N and K. Prefac-
tors, however, will be different. For instance, assuming
four times fewer inhibitory than excitatory neurons in
the two-population networks of subsection IV.1, with-
out changing the number of connections per neuron, the
correlations of the inhibitory neurons will increase by a
factor of 4.
We focused on networks with a one-dimensional ring
(49)
VIII.2. Generality of the results
ab.
with ω =
These complex conjugate blocks can in general be writ-
ten as ±iωI + S, where I is the identity matrix of size
Pmax. Thus
J max
jor
+S CmaxS
Cmax[J max
jor ]∗ = ω2 Cmax ± iω[ CmaxS − S Cmax]
(48)
As shown above, Cmax = O(N γ(Pmax−1)−1) and
S CmaxS = O(N γ(Pmax−2)−1).
It is straightforward to
also show that CmaxS − S Cmax = O(N γ(Pmax−3/2)−1).
jor ]∗ = O(N γ(Pmax−1)−1). Equa-
Therefore J max
jor
tion (41) is then satisfied only if γ < γmax, with
Cmax[J max
1
γmax =
Pmax
According to Theorem 2, if J (n)
jor has a block whose
real part is a shift matrix for at least one mode n,
C = O(N−α) where α = 1 − γ(Pmax − 1). If γ < γmax,
correlations in the activity will decrease more slowly than
1/N , when N is increased. If γ > γmax correlations will
increase with N and the network will not operate in the
balanced regime. Finally, if γ = γmax, our theory will
give O(1) correlations in the input which is inconsistent
with the Ansatz in Eq. (41). In this case substantial cor-
rections to the Eq. (26) should be taken into account.
A different approach, similar to the one in [28] must be
adopted to self-consistently determine these correlations.
VIII. DISCUSSION
VIII.1. Main results
We developed a theory for the emergence of correla-
tions in strongly recurrent networks of binary neurons.
Each neuron receives on average K inputs from each of D
populations, with probabilities which are spatially mod-
ulated. The synaptic strengths scale as 1/
K and the
√
architecture and connection probabilities which are solely
distance dependent. This greatly simplifies the problem,
because when averaged over realizations, correlations de-
pend solely on distance. Furthermore, because of the
linearity of the self-consistent equation for the correla-
tions, the different Fourier modes decouple, allowing us
to analyze each mode separately. However, our analyti-
cal approach does not require rotation invariance. It can
be extended to any network architecture for which the
Jordan normal form of the interaction matrix, averaged
over realizations, can be established.
VIII.3. Robustness and self-consistency of the
results
The theory presented here makes the Anzatz that cor-
relations are sufficiently small so that the dynamics of the
crosscorrelations can be linearized (Appendix A). If this
Anzatz is correct, the theory is self-consistent. When the
theory predicts correlations which are O(1), non-linear
terms contribute and the correlations start to deviate
from the theoretical value (see for example simulation
results for N = 40000 in Fig. 3b). Nevertheless, the or-
der of the correlations is still correctly predicted.
In the large N, K limit, only networks with feedforward
structures -hidden or explicit- can exhibit correlations
that increase with the average number of inputs. How-
ever, when K and N are only moderately large, as is the
case in biological systems, a strict tuning of the architec-
ture is not necessary. This is because there is a crossover
between the regimes of strong and weak correlations as
K is increased (see Fig. 4c). As shown in Appendix B1,
the value of K for which this crossover occurs depends
on the eigenvalues of the interaction matrix.
VIII.4. Relation to previous works
Non-interacting neurons can exhibit correlations if
they share feedforward inputs [57, 58]. For instance, Lit-
vak et al. [59] investigated a chain of layers of integrate-
and-fire neurons lacking any recurrent interactions and
coupled only feedforwardly. In their model, each neuron
in a layer receives inputs from the same number of excita-
tory and inhibitory neurons in the previous layer in such
a way that their temporal averages exactly balance. They
found a build up of correlations along the chain. This is
because the correlations induced by shared feedforward
inputs are not suppressed during the activity propaga-
tion since the network lacks any recurrent interactions
and is thus purely feedforward.
Cortes and van Vreeswijk [60] studied a chain of
strongly recurrent unstructured E-I subnetworks coupled
through excitatory unstructured feedforward projections.
They found a gradual build up of correlations along the
chain. These correlations, however, decrease if the con-
nectivity, K, and the sub-networks size, N , increase to-
gether. This result is in agreement with our theory which
14
predicts that the correlations are O(1/N ) through the
whole chain.
In [43] Ginzburg and Sompolinsky considered networks
of binary neurons with finite temperature Glauber dy-
namics, unstructured, dense (K = pN ) connectivity and
weak interactions, i.e., of the order of O(1/K) and not
√
O(1/
K) as in our work. Mean Field theory shows that
in these networks correlations are O(1/N ) with a pref-
actor which diverges in the zero temperature limit. This
is in contrast to what happens in the strongly recurrent
unstructured networks we considered here where correla-
tions also scale as O(1/N ) but with a prefactor, which is
finite despite the fact that we assumed zero temperature
Glauber dynamics. This is because in strongly recurrent
networks, intrinsic noise emerges from the deterministic
dynamics of the network.
They also demonstrated that in their model the cor-
relations amplify up to O(1) at Hopf bifurcation onsets
[43]. At such onsets the dynamics exhibit critical slow-
ing down and thus this amplification is accompanied by a
divergence of the decorrelation times. Our work demon-
strates a different amplification mechanism: it occurs at
a point where the Jordan form of the interaction matrix,
¯J , contains a block which is a shift matrix. Since there
is no critical slowing down at such a point, the decorrela-
tion times are finite and on the order of the update time
constant (data not shown). Our theory may thus ac-
count for substantial correlations with short time scales,
as frequently observed in the brain [28, 31, 61].
Renart et al.
[28] and Helias et al.
[49] investigated
strongly recurrent unstructured networks with one exci-
tatory and one inhibitory population of binary neurons
which are densely connected, i.e with K ∝ N . They
found that in these networks mean pair correlations were
O(1/N ). As they only considered dense connectivity they
could not, however, disentangle the dependence on N and
K. Here, we considere different relations between K and
N and show that in unstructured networks the correla-
tions are O(1/N ), and in practice do not depend on K.
This last result is remarkable since one would expect cor-
relations to increase with the degree of connectivity. This
is not the case: the balance of excitation and inhibition
prevents that to occur in unstructured networks.
On the other hand, and somewhat surprisingly, we
found that even if the fraction of common inputs shared
by neurons is very small, a build up of correlations can
still occur for some network architectures. For exam-
ple, in a network of four populations with a feedforward
structure, correlations would be of O(K 3/N ), and thus,
to satisfy the balanced correlation equation, the scaling
of K with the network size can be at most K = O(N 1/3).
However, with this architecture and scaling, the proba-
bility of two neurons to share their inputs is O(N−4/3)
(Eq. (17)) and the number of inputs shared by two neu-
rons is therefore O(N−1/3). Thus, although the number
of shared inputs goes to zero in the large N limit, correla-
tions get amplified up to O(1) thanks to the feedforward
architecture.
Rosenbaum et al.
[39] have recently investigated
how feedforward excitation can drive correlations in spa-
tially structured E-I networks operating in the balanced
regime. The specific architecture they considered is rem-
iniscent of the particular example presented in Fig. 7a.
In their study the fluctuations which drove the corre-
lated activity were those in the feedforward inputs and
the contribution to the correlations of the fluctuations
generated by the recurrent dynamics was neglected. In
contrast, our work focuses on the role of the recurrent
dynamics in the emergence of correlations.
We recently studied the emergence of correlations in
a network consisting of two strongly recurrent E-I sub-
networks, the first projecting to the second with topo-
graphically organized feedforward connections [17]. The
architecture in that work is also reminiscent of the ex-
ample presented in Fig. 7a. We showed that while in the
first subnetwork correlations were weak, O(1/N ), in the
second subnetwork the activity was self-organized in such
a way that correlations in macroscopic sub-populations
of excitatory neurons were finite and did not depend on
N and K when the latter were sufficiently large. This
does not contradict our theory. In the architecture con-
sidered in [17] subsets of neurons in the first subnetwork
are projecting to a macroscopic fraction of neurons in the
second subnetwork. This organization generates correla-
tions between elements of the connectivity matrix, unlike
in the models considered here. Generalizing our theory
to such cases is possible but beyond the scope of this
paper.
VIII.5. Directions for future work
The theorems of Section V, tell us how the Fourier
modes of the averaged correlations scale for large N and
K. In the examples considered in Sections IV, VI and
VII, we focused on connectivities whose Fourier expan-
sion involve only two modes. In these cases, the scaling
of the spatial correlations can be immediately deduced
from that of the Fourier modes. This will also be the case
for interactions described by a finite number of Fourier
modes. However, if the interactions are described by an
infinite number of modes, inferring the scaling of the cor-
relations from those of their modes can be more compli-
cated. For instance, if one takes the large N limit with
K ∝ N γ, the convergence of the Fourier series may not
be uniform in N . In that case the scaling with N of the
spatial correlations may be highly non-trivial. We will
address this issue in an upcoming paper.
The present paper focuses on locally averaged two-
point correlation functions. Recent progress in experi-
mental techniques will create large data sets of neuronal
activities from which distributions of pairwise correla-
tions can be extracted. The power of theoretical ap-
proaches to interpret such data will be greatly enhanced
if they provide not only locally averaged correlations but
also higher order statistics of their distributions. Thus
15
it would be interesting to extend our approach to esti-
mate the scaling of higher order moments of correlations
in binary networks.
Several previous studies investigated EI networks (e.g.
[49, 62]) in which inhibition and excitation were unstruc-
tured and their strengths were only a function of the
presynaptic neurons, i.e., JEE = JIE and JII = JEI .
Other studies assumed unstructured connectivity with
JαE = −JαI , α ∈ {E, I} (e.g.
In both cases
the interaction parameters are on the edge of the region
where the network evolves towards the balanced state
(see Eq. (11)). The network dynamics may be qualita-
tively different on the edge of this region than inside it.
It is thus not clear that the scaling theorems presented
here apply to these tuned cases. A further investigation
of the correlation structure in such networks is a subject
for future research.
[63]).
Do the conclusions derived here for networks of bi-
nary neurons hold for networks with more realistic single
neuron dynamics? To approach this question we per-
formed extensive numerical simulations of strongly recur-
rent networks consisting of one inhibitory and one excita-
tory population of leaky integrate-and-fire (LIF) neurons.
The detailed analysis of these simulations will be pre-
sented elsewhere. In brief, we found in our simulations,
that in these networks the averaged pairwise correlations
scale with K and N in manner that is consistent with
the theory presented here for binary networks. It would
be very interesting to extend our analytical approach to
these type of networks.
To conclude, van Vreeswijk and Sompolinsky [45, 46]
investigated strongly recurrent networks of binary neu-
rons with unstructured sparse connectivity. They showed
how these network dynamics evolve into a balanced state.
Due to the sparseness of the connectivity, correlations are
negligible in these networks. Subsequent studies [28, 49]
extended these results to unstructured networks with
dense connectivity, and found that here too the network
dynamics evolve to a balance state in which reverber-
ations keep the correlations very small. In contrast, as
shown here, in networks with structured connectivity cor-
relations can be large. The theory presented here gives
the conditions on the network architecture to evolve into
a balanced state with strong correlations and show how
they depend on the network size and number of connec-
tions.
ACKNOWLEDGMENTS
We thank Gianluigi Mongillo and German Mato for
fruitful discussions. Work conducted in the framework
of the France Israel Laboratory of Neuroscience (FILN).
Grants: ANR/CRCNS-BASCO, ANR-BALWM, ANR-
BALAV1, France-Israel High Council for Science and
Technology, LIA-FILN (CNRS), IRN-FICNC (CNRS).
16
[1] Roy J Glauber, "Time-dependent statistics of the ising
model," Journal of mathematical physics 4, 294–307
(1963).
[2] Mark C Cross and Pierre C Hohenberg, "Pattern forma-
tion outside of equilibrium," Reviews of modern physics
65, 851 (1993).
[3] David Hansel and Haim Sompolinsky, "Chaos and syn-
chrony in a model of a hypercolumn in visual cortex,"
Journal of computational neuroscience 3, 7–34 (1996).
[4] Udo Seifert, "Stochastic thermodynamics, fluctuation
theorems and molecular machines," Reports on Progress
in Physics 75, 126001 (2012).
[5] Fotis K Diakonos, AK Karlis, and P Schmelcher, "A
universal mechanism for long-range cross-correlations,"
EPL (Europhysics Letters) 105, 26004 (2014).
[6] Brent Doiron, Ashok Litwin-Kumar, Robert Rosenbaum,
Gabriel K Ocker, and Kresimir Josi´c, "The mechanics
of state-dependent neural correlations," Nature neuro-
science 19, 383–393 (2016).
[7] Tatjana Tchumatchenko, Aleksey Malyshev, Theo
Geisel, Maxim Volgushev, and Fred Wolf, "Correlations
and synchrony in threshold neuron models," Physical Re-
view Letters 104, 058102 (2010).
[8] Edwin M Maynard, Craig T Nordhausen, and Richard A
Normann, "The utah intracortical electrode array: a
recording structure for potential brain-computer inter-
faces," Electroencephalography and clinical neurophysi-
ology 102, 228–239 (1997).
[9] Simon Peron, Tsai-Wen Chen,
and Karel Svoboda,
"Comprehensive imaging of cortical networks," Current
opinion in neurobiology 32, 115–123 (2015).
[10] Cyrille Rossant, Shabnam N Kadir, Dan FM Goodman,
John Schulman, Maximilian LD Hunter, Aman B Saleem,
Andres Grosmark, Mariano Belluscio, George H Denfield,
Alexander S Ecker, et al., Spike sorting for large, dense
electrode arrays, Tech. Rep. (Nature Publishing Group,
2016).
[11] Ehud Zohary, Michael N Shadlen, and William T New-
some, "Correlated neuronal discharge rate and its impli-
cations for psychophysical performance," (1994).
[12] LF Abbott and Peter Dayan, "The effect of correlated
variability on the accuracy of a population code," Neural
computation 11, 91–101 (1999).
[13] Haim Sompolinsky, Hyoungsoo Yoon, Kukjin Kang, and
Maoz Shamir, "Population coding in neuronal systems
with correlated noise," Physical Review E 64, 051904
(2001).
[14] Bruno B Averbeck, Peter E Latham,
and Alexan-
dre Pouget, "Neural correlations, population coding and
computation," Nature Reviews Neuroscience 7, 358–366
(2006).
[15] Gyorgy Buzsaki, Rhythms of the Brain (Oxford Univer-
sity Press, 2006).
[16] Charles M Gray, Peter Konig, Andreas K Engel, and
Wolf Singer, "Oscillatory responses in cat visual cor-
tex exhibit inter-columnar synchronization which reflects
global stimulus properties," Nature 338, 334–337 (1989).
and
D Hansel, "A canonical neural mechanism for behavioral
variability." Nature communications 8, 15415 (2017).
[17] R Darshan, WE Wood, S Peters, A Leblois,
computational biology 12, e1005056 (2016).
[19] Donald Olding Hebb, The organization of behavior: A
neuropsychological theory (Psychology Press, 2005).
[20] Guo-qiang Bi and Mu-ming Poo, "Synaptic modification
by correlated activity: Hebb's postulate revisited," An-
nual review of neuroscience 24, 139–166 (2001).
[21] Alexander S Ecker, Philipp Berens, Georgios A Keliris,
Matthias Bethge, Nikos K Logothetis, and Andreas S
Tolias, "Decorrelated neuronal firing in cortical micro-
circuits," science 327, 584–587 (2010).
[22] Marlene R Cohen and Adam Kohn, "Measuring and in-
terpreting neuronal correlations," Nature neuroscience
14, 811–819 (2011).
[23] Bryan J Hansen, Mircea I Chelaru, and Valentin Dragoi,
"Correlated variability in laminar cortical circuits," Neu-
ron 76, 590–602 (2012).
[24] Maria C Dadarlat and Michael P Stryker, "Locomotion
enhances neural encoding of visual stimuli in mouse v1,"
Journal of Neuroscience 37, 3764–3775 (2017).
[25] Gideon Rothschild, Lior Cohen, Adi Mizrahi, and Israel
Nelken, "Elevated correlations in neuronal ensembles of
mouse auditory cortex following parturition," Journal of
Neuroscience 33, 12851–12861 (2013).
[26] Takaki Komiyama, Takashi R Sato, Daniel H OConnor,
Ying-Xin Zhang, Daniel Huber, Bryan M Hooks, Mari-
ano Gabitto, and Karel Svoboda, "Learning-related fine-
scale specificity imaged in motor cortex circuits of behav-
ing mice," Nature 464, 1182–1186 (2010).
[27] James M Jeanne, Tatyana O Sharpee, and Timothy Q
Gentner, "Associative learning enhances population cod-
ing by inverting interneuronal correlation patterns," Neu-
ron 78, 352–363 (2013).
[28] Alfonso Renart, Jaime de la Rocha, Peter Bartho, Liad
Hollender, N´estor Parga, Alex Reyes, and Kenneth D
Harris, "The asynchronous state in cortical circuits," sci-
ence 327, 587–590 (2010).
[29] Timothy J Gawne and Barry J Richmond, "How inde-
pendent are the messages carried by adjacent inferior
temporal cortical neurons?" Journal of Neuroscience 13,
2758–2771 (1993).
[30] Timothy J Gawne, Troels W Kjaer, John A Hertz, and
Barry J Richmond, "Adjacent visual cortical complex
cells share about 20% of their stimulus-related informa-
tion," Cerebral Cortex 6, 482–489 (1996).
[31] Matthew A Smith and Adam Kohn, "Spatial and tem-
poral scales of neuronal correlation in primary visual
cortex," The Journal of Neuroscience 28, 12591–12603
(2008).
[32] Diego A Gutnisky and Valentin Dragoi, "Adaptive coding
of visual information in neural populations," Nature 452,
220–224 (2008).
[33] Wyeth Bair, Ehud Zohary, and William T Newsome,
"Correlated firing in macaque visual area mt: time scales
and relationship to behavior," Journal of Neuroscience
21, 1676–1697 (2001).
[34] Daeyeol Lee, Nicholas L Port, Wolfgang Kruse,
and
Apostolos P Georgopoulos, "Variability and correlated
noise in the discharge of neurons in motor and parietal
areas of the primate cortex," The Journal of neuroscience
18, 1161–1170 (1998).
[18] Neta Ravid Tannenbaum and Yoram Burak, "Shaping
neural circuits by high order synaptic interactions," PLoS
[35] Robert B Levy and Alex D Reyes, "Spatial profile of
excitatory and inhibitory synaptic connectivity in mouse
primary auditory cortex," The Journal of Neuroscience
32, 5609–5619 (2012).
[36] Elodie Fino and Rafael Yuste, "Dense inhibitory connec-
tivity in neocortex," Neuron 69, 1188–1203 (2011).
[37] Hiroki Tanaka, Hiroshi Tamura,
and Izumi Ohzawa,
"Spatial range and laminar structures of neuronal cor-
relations in the cat primary visual cortex," Journal of
neurophysiology 112, 705–718 (2014).
[38] Shervin Safavi, Abhilash Dwarakanath, Vishal Kapoor,
Joachim Werner, Nicholas G Hatsopoulos, Nikos K Lo-
gothetis,
and Theofanis I Panagiotaropoulos, "Non-
monotonic spatial structure of interneuronal correlations
in prefrontal microcircuits," bioRxiv , 128249 (2017).
[39] Robert Rosenbaum, Matthew A Smith, Adam Kohn,
Jonathan E Rubin,
and Brent Doiron, "The spatial
structure of correlated neuronal variability," Nature Neu-
roscience 20, 107–114 (2017).
[40] Daniel J Denman and Diego Contreras, "The structure
of pairwise correlation in mouse primary visual cortex
reveals functional organization in the absence of an ori-
entation map," Cerebral Cortex 24, 2707–2720 (2014).
[41] Tom Binzegger, Rodney J Douglas, and Kevan AC Mar-
tin, "A quantitative map of the circuit of cat primary
visual cortex," Journal of Neuroscience 24, 8441–8453
(2004).
[42] Moshe Abeles, Corticonics: Neural circuits of the cerebral
cortex (Cambridge University Press, 1991).
[43] Iris Ginzburg and Haim Sompolinsky, "Theory of corre-
lations in stochastic neural networks," Physical review E
50, 3171 (1994).
[44] J´er´emie Barral and Alex D Reyes, "Synaptic scaling rule
preserves excitatory-inhibitory balance and salient neu-
ronal network dynamics," Nature neuroscience 19, 1690–
1696 (2016).
[45] Carl van Vreeswijk and Haim Sompolinsky, "Chaos in
neuronal networks with balanced excitatory and in-
hibitory activity," Science 274, 1724–1726 (1996).
[46] Carl van Vreeswijk and Haim Sompolinsky, "Chaotic bal-
anced state in a model of cortical circuits," Neural com-
putation 10, 1321–1371 (1998).
[47] Michael Monteforte and Fred Wolf, "Dynamic flux tubes
form reservoirs of stability in neuronal circuits," Physical
Review X 2, 041007 (2012).
[48] Alex Roxin, Nicolas Brunel, David Hansel, Gianluigi
Mongillo, and Carl van Vreeswijk, "On the distribution
of firing rates in networks of cortical neurons," The Jour-
nal of neuroscience 31, 16217–16226 (2011).
[49] Moritz Helias, Tom Tetzlaff,
and Markus Diesmann,
"The correlation structure of local neuronal networks in-
trinsically results from recurrent dynamics," PLoS Com-
put Biol 10, e1003428 (2014).
[50] David Hansel and Haim Sompolinsky, "Synchronization
and computation in a chaotic neural network," Physical
Review Letters 68, 718 (1992).
[51] Zolt´an F Kisv´arday, E Toth, Martin Rausch, and Ulf T
Eysel, "Orientation-specific relationship between popula-
tions of excitatory and inhibitory lateral connections in
the visual cortex of the cat." Cerebral cortex (New York,
NY: 1991) 7, 605–618 (1997).
[52] Armen Stepanyants, Luis M Martinez, Alex S Ferecsk´o,
and Zolt´an F Kisv´arday, "The fractions of short-and
long-range connections in the visual cortex," Proceed-
ings of the National Academy of Sciences 106, 3555–3560
(2009).
17
[53] Rani Ben-Yishai, David Hansel, and Haim Sompolinsky,
"Traveling waves and the processing of weakly tuned in-
puts in a cortical network module," Journal of computa-
tional neuroscience 4, 57–77 (1997).
[54] David Hansel and Haim Sompolinsky, "13 modeling fea-
ture selectivity in local cortical circuits," (1998).
[55] C Van Vreeswijk and H Sompolinsky, "Irregular activity
in large networks of neurons," Methods and models in
neurophysics. Amsterdam: Elsevier (2005).
[56] Robert Rosenbaum and Brent Doiron, "Balanced net-
works of spiking neurons with spatially dependent recur-
rent connections," Physical Review X 4, 021039 (2014).
[57] Eric Shea-Brown, Kresimir Josi´c, Jaime de La Rocha,
and Brent Doiron, "Correlation and synchrony transfer
in integrate-and-fire neurons: basic properties and conse-
quences for coding," Physical review letters 100, 108102
(2008).
[58] Jaime De La Rocha, Brent Doiron, Eric Shea-Brown,
Kresimir Josi´c, and Alex Reyes, "Correlation between
neural spike trains increases with firing rate," Nature
448, 802–806 (2007).
[59] Vladimir Litvak, Haim Sompolinsky, Idan Segev, and
Moshe Abeles, "On the transmission of rate code in
long feedforward networks with excitatory–inhibitory
balance," The Journal of neuroscience 23, 3006–3015
(2003).
[60] Nelson Cortes and Carl van Vreeswijk, "Pulvinar thala-
mic nucleus allows for asynchronous spike propagation
through the cortex," Frontiers in computational neuro-
science 9 (2015).
[61] Matthew A Smith, Xiaoxuan Jia, Amin Zandvakili, and
Adam Kohn, "Laminar dependence of neuronal correla-
tions in visual cortex," Journal of neurophysiology 109,
940–947 (2013).
[62] Nicolas Brunel, "Dynamics of sparsely connected net-
works of excitatory and inhibitory spiking neurons,"
Journal of computational neuroscience 8, 183–208 (2000).
[63] Kanaka Rajan and LF Abbott, "Eigenvalue spectra of
random matrices for neural networks," Physical review
letters 97, 188104 (2006).
Appendix A: Correlations in binary networks
Here we calculate the equilibrium value and the stabil-
ity of the quenched average correlations.
We define, for (j, β) (cid:54)= (i, α), the out of equilibrium
auto- and crosscorrelations, cαβ
ij (t, τ ) as
ij (t, τ ) ≡ (cid:104)δSα
cαβ
i (t, τ ) ≡ (cid:104)δSα
aα
i (t)δSβ
i (t)δSα
j (t + τ )(cid:105)init
i (t + τ )(cid:105)init
where (cid:104)·(cid:105)init denotes averaging over many initial condi-
tions choosen with a probability measure that, for sim-
plicity, we choose such that (cid:104)Sα
i (0)(cid:105)init = (cid:104)Sα
i (t)(cid:105)t. It is
also convenient for the notation to define cαα
ii (t, τ ) = 0.
In this paper we focus on the equal time correlations,
ij (t) ≡ cαβ
cαβ
For networks of binary neurons the dynamics of the
ij (t, 0).
equal-time crosscorrelations is given by [1, 28, 43]
Thus, for large N , the quenched average of Eq. (A1)
18
yields
d
dt
Cαβ(∆, t) = − 2Cαβ(∆, t) +
(cid:88)
(cid:90) d∆(cid:48)
(cid:104)Jαγ(∆ − ∆(cid:48))Cγβ(∆(cid:48), t) +
2π
γ
(cid:105)
+ Jγβ(−∆(cid:48))Cαγ(∆ − ∆(cid:48), t)
+ Jαβ(∆)
+ Jβα(−∆)
Aα
N
Aβ
N
+
.
(A2)
Here we have not taken into account that cαα
ii (t) = 0. It
is easy to see, however, that in the cross-correlations this
neglects O(1/N 2) corrections. We show below that these
corrections are indeed negligeable in the large N limit.
αβ (t), satisfies
The nth Fourier mode of Cαβ(∆, t), C (n)
d
dt
(cid:88)
αβ (t) = − 2C (n)
C (n)
αβ (t)
(cid:104)J (n)
+
αγ C (n)
γβ (t) + C (n)
αγ (t)[J (n)
γβ ]†(cid:105)
γ
+ J (n)
αβ
Aβ
N
+
Aα
N
[J (n)
αβ ]†,
where J (n)
Jαβ(∆) and J (cid:62)
αβ and [J (n)
αβ(∆), respectively.
αβ ]† are the nth Fourier mode of
This can be written more compactly as
τ
d
dt
C(n)(t) = − 2C(n)(t) + J (n)C(n)(t) + C(n)(t)[J (n)]†
+ J (n) A
N
+
A
N
[J (n)]†.
(A3)
where X denotes the D × D matrix, Xαβ, and A is the
D × D matrix Aαβ ≡ δα,βAα. The equilibrium values of
the Fourier components of equal-time correlation func-
tions thus satisfy Eq. (26).
dcαβ
ij (t)
dt
= − 2cαβ
+ (cid:104)δSα
ij (t) + (cid:104)δΘ[hα
i (t)δΘ[hβ
i (t) − T ]δSβ
j (t)(cid:105)init
j (t) − T ](cid:105)init,
i (t) − T ] − (cid:104)Θ[hα
i (t) − T ] ≡ Θ[hα
i (t) − T ](cid:105)init.
If we make the Ansatz that correlations are weak, we
j (t)(cid:105)init =
i (t) − T ]δSβ
is the gain of neuron
i (t) − T ](cid:105)init, which is, to leading
where δΘ[hα
can, to leading order, take (cid:104)δΘ[hα
i (cid:104)δhα
gα
(i, α), gα
order, independent of the correlations [28, 43, 49].
i (t)δSβ
i = ∂h(cid:104)Θ[hα
j (t)(cid:105)init, where gα
i
Thus,
dcαβ
ij (t)
dt
j (t)(cid:105)init
= −2cαβ
+ gβ
= −2cαβ
i (t)δhβ
i (t)δSβ
ij (t) + gα
j (cid:104)δSα
(cid:88)
ij (t) + gα
i
i (cid:104)δhα
(cid:88)
j (t)(cid:105)init
ik (cid:104)δSγ
J αγ
jk (cid:104)δSγ
J βγ
k (t)δSα
γ,k
+ gβ
j
γ,k
k (t)δSβ
i (t)(cid:105)init.
j (t)(cid:105)init +
Using (cid:104)δSα
i (t)δSβ
dcαβ
ij (t)
dt
= −2cαβ
j (t)(cid:105)init = cαβ
(cid:88)
(cid:104) J αγ
ij (t) +
ij (t) + δα,βδi,jaα
i (t) yields
(cid:105)
ik cγβ
kj (t) + J βγ
jk cαγ
ik (t)
+
γ,k
+ J αβ
ij aβ
j (t) + J βα
ji aα
i (t),
(A1)
where J αβ
i J αβ
ij = gα
ij . Note that, for weak correlations
(cid:104)Sα
i (t)(cid:105)init does not depend on the correlations so that,
(cid:104)Sα
i (t)(cid:105)init = (cid:104)Sα
i (t)(cid:105)t, for our choice of initial condi-
i (0)(cid:105)init = (cid:104)Sα
tions ((cid:104)Sα
i (t)(cid:105)t). Hence, the equal time
autocorrelation aα
is independent of time and given by
i (t)(cid:105)2
i (t)(cid:105)t − (cid:104)Sα
i = (cid:104)Sα
i
aα
t .
We now average over the quenched disorder. Due to
the rotational symmetry of the connection probabilities,
P αβ
ij , [aα
i ]J is a constant
[aα
i ]J = Aα
Appendix B: Correlation theorems
whereas [cαβ
in the location of neurons (i, α) and (j, β)
ij ]J and [ J αβ
ij ]J are functions of the difference
[cαβ
ij (t)]J = Cαβ(θα
i − θβ
j , t)
Jαβ(θα
i − θβ
j )
1
N
[ J αβ
ij ]J =
√
where Jαβ(∆) =
KgαJαβfαβ(∆). Here we have as-
sumed that the correlations in the quenched disorder in
the inputs to the neurons are small, such that, to leading
order, the expected value of the gain does not depend on
the neuronal position. We comment on this Ansatz in
Appendix C.
In Appendix A we derived N 2D2 coupled linear differ-
ential equations that determine the evolution of equal-
time cross-correlations for all neuronal pairs in the net-
work. Averaging these equations over the quenched dis-
order and using the rotation invariance of the connec-
tion probabilities yields a set of N D2 coupled equations
for the quenched averaged correlations. In Fourier space
these equations lead to N independent sets of D2 coupled
equations for the correlations. Here we prove Theorems 1
and 2 (see Section V) which state how these correlations
scale with N and K.
To leading order, A is independent of K and N , but
J (n) is proportional to
K. Accordingly, we define
¯J (n) ≡ J (n)/
K. Thus,
we can rewrite the evolution equation of the nth mode of
K and [ ¯J (n)]† ≡ [J (n)]†/
√
√
√
19
Therefore CD,1(t), converges to its equilibrium value.
The evolution equations of CD,2 can be written as
CD,2(t) = −ΛD,2 CD,2(t) + MD,2 CD,2(t)
d
dt
where MD,2 depends on CD,1(t), AD,1 and AD,2. Since
CD,1(t) converges to its equilibrium value, MD,2 con-
verges to a constant. Because Re(ΛD,2) > 0, CD,2 also
∞
converges to its equilibrium value, C
D,1. A similar ar-
gument shows that likewise CD−1,1 converges to C∞
D−1,1.
One then sees by recursion that the whole matrix C con-
verges to its equilibrium value, C
. Hence, Eq. (26)
determines the stable equilibrium values of the correla-
tions.
∞
From here we only consider the correlations at equilib-
rium and, for notational simplicity, we drop the super-
script ∞.
In the Jordan basis the equilibrium values of the cor-
relations satisfy
Λµν Cµν =
√
(cid:104)
(cid:104)
K
√
+
K
N
µ Cµ+1,ν + ν−1 Cµ,ν−1
+
(λµ + λ∗
ν) Aµν+
(cid:105)
(cid:105)
+µ Aµ+1,ν + ν−1 Aµ,ν−1
(B3)
Let us now consider the case where ¯J is diagonizable,
so that µ = 0 for all µ. In this case Cµν = 0 for µ (cid:54)= ν
and
,
√
K(λµ + λ∗
µ)
K(λµ + λ∗
µ)]
Aµµ,
Cµµ =
(cid:17)
N [2 − √
(cid:16)
N→∞ N Cµµ
lim
lim
K→∞
√
(cid:16)
ν is O(1/
lim
K→∞
N→∞ N Cµν
lim
(cid:17)
= − Aµµ
(B4)
= 0
(B5)
the correlations as
C(n)(t) = − 2C(n)(t)
d
dt
(cid:104) ¯J (n)C(n)(t) + C(n)(t)[ ¯J (n)]†(cid:105)
(cid:104) ¯J (n)A(t) + A(t)[ ¯J (n)]†(cid:105)
.
(B1)
√
K
√
K
N
+
+
The D × D matrix, ¯J (n) can be written as
jor [U (n)]−1.
¯J (n) = U (n)J (n)
where J (n)
is the Jordan normal form of ¯J (n) and
[U (n)]−1 is the transformation matrix to the Jordan ba-
sis.
jor
jor can be written as
The matrix J (n)
[J (n)
are the eigenvalues of ¯J (n) and (n)
jor ]µν = λ(n)
µ δµ,ν + (n)
µ δµ,ν−1.
where λ(n)
µ = 1 in-
µ
side a Jordan block and is 0 otherwise (for clarity, in the
Jordan basis we use the subscripts µ and ν, rather than
α and β that we use in the original basis).
Importantly, the Jordan form of a matrix and of its
Hermitian conjugate are complex conjugate. We thus
can write
[ ¯J (n)]† = V (n)[J (n)
jor ]∗[V (n)]−1,
For notational convenience we will suppress the super-
script (n) in the rest of this Appendix. Defining C as
C = U−1CV
(cid:105)
Cµν(t) = − Λµν Cµν(t)+
d
dt
(cid:104)
√
K
√
+
K
N
(cid:104)
(λµ + λ∗
ν) Aµν +
+ µ Aµ+1,ν + ν−1 Aµ,ν−1
(cid:105)
,
(B2)
where we have defined Λµν = 2 − √
A = U−1AV .
K(λµ + λ∗
ν) and
√
We now assume that the connectivity is such that the
system is stable to perturbations of the locally averaged
rates. This implies that the real part of all eigenvalues of
¯J are less than 1/
K. The real part of Λµν is therefore
positive and thus C(t) converges to an equilibrium value,
C
∞
To see this, first consider CD,1(t) which satisfies
.
√
CD,1(t) = −ΛD,1 CD,1(t) +
d
dt
K
N
(λD + λ∗
1) AD,1.
and inserting into Eq. (B1) yields
Thus
µ Cµ+1,ν(t) + ν−1 Cµ,ν−1(t)
unless λµ + λ∗
K), in which case
In the first situation Cµµ = O(1/N ). In the second situ-
ation Cµµ = O(1/N ).
Let us now assume that ¯J is not diagonizable. Then
the D× D Jordan form of ¯J consists of B Jordan blocks
(1 ≤ B < D) that we denote by [ ¯J jor]i, i = 1, ..., B.
The size of the ith block will be denoted by s(i) × s(i).
Without loss of generality, we can assume that the blocks
are ordered in increasing size.
1 +(cid:80)i−1
j=1 s(j) and h(i) = (cid:80)i
The indices µ and ν of the elements of the Jordan
block i, take values between l(i) and h(i), with l(i) =
j=1 s(j). All the diagonal
elements of this Jordan block are equal one of the eigen-
values of J , which we denote by λi. The off-diagonal
elements, µ, are all equal to 1 except for µ = h(i)
(i = 1, .., B − 1) for which µ = 0.
The matrix C consists of B2 sectors that we denote
by Sij. In the sector Sij, µ ∈ {l(i), . . . , h(i)} and ν ∈
{l(j), . . . , h(j)}.
Let us consider Eq. (B3) for µ, ν in sector Sij. Since
h(i) and l(j)−1 are zero, the equation in this sector does
not depend on elements of C outside of it. Thus, we can
solve Eq. (B3) recursively to determine all the elements
of C in this sector. The recursion goes as follows. First,
one solves for Ch(i),l(j)
√
2 − √
(cid:16)λi + λ∗
(cid:17)
(cid:16)λi + λ∗
(cid:17) Ah(i)l(j)
Ch(i),l(j) =
(B6)
1
N
K
K
j
j
We can the solve Eq. (B3) to get Cµν for µ, ν = h(i) −
1, l(j) and µ, ν = h(i), l(j) + 1. This process can be
repeated until all the elements of C in the sector Sij are
determined.
1.
Ah(i),l(j) are all non-zero
A similar recursion can be performed to estimate the
order of magnitude of all the elements of C. First we
obtain
Ch(i),l(j) = O
(cid:18) 1
(cid:19)
.
N
= O
(cid:33)
(cid:32) Ah(i),l(j)
(cid:18) 1
N
(cid:19)
Cµν = O
N
For λi + λ∗
j = O(1) the recursion shows that we have
20
for all i, j ∈ {1, ..., B}. According to Eq. (B7), correla-
tions are at most O(1/N ) in all of the B2 sectors of the
matrix C. As a result, C(n) = O(1/N ). On the other
hand, if in each of the B2 sectors of the matrix C corre-
lations are at most O(1/N ), this is also the case for the
element of C in all the sectors. In this case, Eq. (B7)
implies that there is no Jordan block in ¯J for which
Re[λi] = 0.
Restoring the index n of the Fourier mode, one sees
that if ¯J (n) has at least one Jordan block whose real
part is a shift matrix and denoting by P (n) the size
of the largest shift, Eq. (B9) implies that Cl(i),h(i) =
O(K P (n)−1/N ), and thus also C
= O(K P (n)−1/N ).
This proves Correlation Theorem 2.
(n)
(n)
According to Eq.(B9), the scaling of the correlation
= O(K D−1/N ) if and only if P (n) = D. When
is C
¯J (n) is a real matrix (when the probability of connections
are symmetric in ∆), it means that the Jordan form of
¯J (n) is a shift matrix of size D. This is equivalent for
saying that ¯J (n) is nilpotent of degree D. On the other
hand, if ¯J (n) is a shift matrix of degree D, it has a Jordan
form which is a shift matrix of degree D. According to
Eq.(B9), the scaling of the correlation is then C
O(K D−1/N ). This proves the Corollary in section V .
=
j2) there is a crossover in sector
Sij of the matrix C between weak correlations (O(1/N ))
and strong correlations (O(K (s(i)+s(j)−2)/2/N )). To see
this, we note that if K = λi + λ∗
When K = O(1/λi+λ∗
Γ , for Γ ≤ 1/2
(n)
(B7)
whereas for 1/2 < Γ ≤ 1
for all µ, ν in sector Sij.
If λi + λ∗
(cid:17)
other elements in sector Sij are
j = 0, Ch(i),l(j) = 0 and by recursion all the
Cµν = O(cid:16)
Cl(i),h(j) = O(cid:16)
K (cid:96)/2/N
,
(B8)
where (cid:96) = ν − µ + h(i)− l(j). Thus, in this case Cl(i),h(j)
is the largest entry in the sector. It satisfies
K (s(i)+s(j)−2)/2/N
(cid:17)
(B9)
for (cid:96) (cid:54)= 0 and
for (cid:96) = 0
Cµν = O
Cµν = O
Cµν = O
N
(cid:18) K Γ(cid:96)
(cid:19)
j− 1
(cid:18) K (cid:96)/2
(cid:19)
(cid:19)
(cid:18) K 1/2−Γ
N
.
N
The scaling of the elements in sector Sij thus de-
pends λi + λ∗
j . The stability of the dynamics imposes
that Re(λi) < 0 for all blocks (or positive but at most
√
O(1/
K), see next subsection) and thus the condition
λi + λ∗
j = 0 implies that, for large K, the real part of
λi and λj are zero.
In other words, the real part of
[J jor]i and [J jor]j are shift matrices of size s(i)×s(i) and
s(j)× s(j), respectively. For the sector Sii it means that
the real part of [J jor]i is a shift matrix of size s(i)× s(i).
Proving Correlation Theorem 1 is now straightforward.
If the real part of all the B Jordan blocks of ¯J are differ-
j = O(1) (cid:54)= 0
ent from a shift matrix, one finds that λi + λ∗
2. Some elements of A are zero
So far, we have assumed that all the elements in A
are non zero. The derivation can, however, be extended
to include also situations where this is not the case as
follows.
j = O(1)
might result in some elements of C in that sector to be
Elements Aµν = 0 in a sector where λi + λ∗
equal to 0 rather than O(cid:0) 1
(cid:1). However, this will not
affect the overall scaling of the correlations.
Elements Aµν = 0 in a sector Sij where λi + λ∗
j = 0,
do not change the scaling in the matrix, provided that
N
√
Ah(i),l(j) (cid:54)= 0.
If Ah(i),l(j) = 0 the effect depends on
Ah(i)−1,l(j) and Ah(i),l(j)+1.
If at least one of them is
nonzero, the order of the largest entry in the sector is
decreased by
K. If both of them are zero, the order
of this entry decreases at least by a factor of K. Re-
flecting on further element being zero, one reaches the
following conclusion: Let {µij, νij} be the indices {µ, ν}
in Sij for which Aµν (cid:54)= 0, which maximize µ − ν. Then,
the maximal order of Cµν in the sector is O(K Pij−1/N ),
where
Pij = 1 + (µij − l(i) + h(i) − νij)/2.
(B10)
In conclusion: The highest order in C is O(K P−1/N ),
where P is the maximum value of the Pij (i, j ∈
{1, . . . , B}) defined by: 1) Pij = 1 for sector Sij for which
λi +λ∗
j = 0
and all Aµν (cid:54)= 0. 3) Pij is given by Eq. (B10) if in sector
Sij, λi + λ∗
j (cid:54)= 0. 2) Pij = P for sector Sij in which λi +λ∗
j = 0 but some Aµν are zero.
21
for (i, α) (cid:54)= (j, β) and ∆Sα
tion similar to that in Appendix A yields
i defined in Eq.(3). A deriva-
2Γαβ
ij =
jk Γαγ
ik
+
(cid:105)
(cid:104)J αγ
(cid:88)
γ,k
kj + J βγ
ik Γγβ
ij qβ/N + J βα
+ J αβ
ji qα/N,
(C2)
i )2]J . This equation can be solved us-
where qα = [(∆Sα
ing the same approach as in Appendices A-B. This anal-
ysis shows that correlations of the quenched disorder and
correlations of the temporal fluctuations are of the same
order. Thus, when the temporal fluctuations are small,
the Ansatz in Appendix A, were we neglected the spa-
tial fluctuations in the neuronal gain, is justified. If the
correlations are too strong, the Ansatz may no longer
be satisfied, but neither is the linearization assumed to
derive equation (A2).
Appendix D: Self-consistent equations for the
autocorrelation and the gain
We follow the notations from [46]:
mα =
1
2
erfc
Tα − hα
(cid:113)
(cid:32)
(cid:112)2σ2
2(σ2
α
α + σ2
qα)
(cid:33)
Tα − hα − σqαx
Dx erfc2
Aα = mα − qα
(D1)
(D2)
(D3)
i −
(D4)
(D5)
Appendix C: Correlations of the quenched disorder
Let us define the correlation of the quenched disorder
in the outputs of the neurons as
(cid:90)
qα =
1
2
Γαβ
ij = [∆Sα
i ∆Sβ
j ]J
(cid:88)
β
σ2
α =
αβ(mβ − qβ) + K
J 2
(C1)
(cid:88)
ββ(cid:48)
JαβJαβ(cid:48)
N 2
(cid:88)
j(cid:54)=j(cid:48)
where the variance of the input noise, σ2
(cid:104)hα
i (cid:105)t)2(cid:105)t]J , is
α = [(cid:104)(hα
fαβ(θα
i − θβ
j )fαβ(cid:48)(θα
i − θβ(cid:48)
j(cid:48) )Cββ(cid:48)(θβ
j − θβ(cid:48)
j(cid:48) )
and the quenched disorder in the inputs, σ2
qα =
[((cid:104)hα
i (cid:105)t − [(cid:104)hα
i (cid:105)t]J )2]J , is
(cid:88)
β
σ2
qα =
(cid:90) dθ
2π
αβ(qβ − p
J 2
(cid:88)
ββ(cid:48)
JαβJαβ(cid:48)
1
N 2
(cid:88)
j(cid:54)=j(cid:48)
f 2(θ)m2
β) + K
fαβ(θα
i − θβ
j )fαβ(cid:48)(θα
i − θβ(cid:48)
j(cid:48) )[∆Sβ
j ∆Sβ(cid:48)
j(cid:48) ]J
Finally, the gain of the neurons is
gα =
1√
2πσT α
e
−(Tα−hα )2
2(σ2
qα)
α+σ2
(D6)
Equations (D1)-(D3) need to be solved self consistently.
For simplicity, when we solved them we neglect O(p)
terms (but see [49]).
Appendix E: Cross-correlations in two-population
and
networks
For a two-population network Eq. (26) yields
=
B(n)
EE
C (n)
EI
C (n)
II
C (n)
2(1 − J (n)
−J (n)
EE )
IE
0
with
B(n) =
(cid:19)
(cid:18) AE
AI
1
N
D(n)
(E1)
D(n) =
2J (n)
J (n)
IE
0
EE
Solving this equation one gets
0
−J (n)
2(1 − J (n)
II )
EI
22
0
J (n)
2J (n)
EI
II
2 − (J (n)
EI
−2J (n)
EE + J (n)
II )
−2J (n)
√
IE
−2AE ¯J (n)
EE
C (n)
EE =
1
N
EI = − 1
C (n)
N
−2AI ¯J (n)
(AE ¯J (n)
−2 + 3T (n)
√
II
C (n)
II =
1
N
√
EI )2 + AE( ¯J (n)
K + (−AI ( ¯J (n)
−2 + 3T (n)
IE + AI ¯J (n)
¯J (n)
EI )
IE )K
K − ((T (n))2 + 2∆(n))K + T (n)∆(n)K 3/2
EE T (n) + ∆(n) + ¯J (n)
¯J (n)
EI + AE ¯J (n)
K − (AI ¯J (n)
√
√
K − ((T (n))2 + 2∆(n))K + T (n)∆(n)K 3/2
EE
EE
II
¯J (n)
II ))K − AET (n)∆(n)K 3/2
K + (−AE( ¯J (n)
√
−2 + 3T (n)
IE )2 + AI ( ¯J (n)
II T (n) + ∆(n) + ¯J (n)
EE
¯J (n)
II ))K − AI T (n)∆(n)K 3/2
K − ((T (n))2 + 2∆(n))K + T (n)∆(n)K 3/2
where T (n) = T r ¯J (n), ∆(n) = det ¯J (n). After some al-
gebra, this equation can be rewritten as Eq. (27).
|
1111.3065 | 1 | 1111 | 2011-11-13T21:35:50 | Towards a Probabilistic Definition of Seizures | [
"q-bio.NC",
"physics.med-ph"
] | This writing: a) Draws attention to the intricacies inherent to the pursuit of a universal seizure definition even when powerful, well understood signal analysis methods are utilized to this end; b) Identifies this aim as a multi-objective optimization problem and discusses the advantages and disadvantages of adopting or rejecting a unitary seizure definition; c) Introduces a Probabilistic Measure of Seizure Activity to manage this thorny issue. The challenges posed by the attempt to define seizures unitarily may be partly related to their fractal properties and understood through a simplistic analogy to the so-called "Richardson effect". A revision of the time-honored conceptualization of seizures may be warranted to further advance epileptology. | q-bio.NC | q-bio | 1
Towards a Probabilistic Definition of Seizures
Ivan Osorio1*, Alexey Lyubushin2, Didier Sornette3
1* Corresponding Author, Department of Neurology, University of Kansas Medical Center,
390 Rainbow Boulevard, Kansas City, Kansas 66160
913 5884529 office
913 5884585 fax
[email protected]
2 Institute of Physics of the Earth, Russian Academy of Sciences, 123995, Russia, Moscow,
B.Gruzinskaya, 10
3ETH Zurich, Chair of Entrepreneurial Risks, D-MTEC, D-PHYS and D-ERDW, Kreuzplatz
5, CH-8032 Zurich, Switzerland
Key Words: Seizure Definition; Seizure Detection; Average Indicator Function; Wavelet
Transform Maximum Modulus-Stepwise Approximation; Probabilistic Measure of Seizure
Activity; Temporally Fluctuating Correlations; Multi-objective Optimization
2
This writing: a) Draws attention to the intricacies inherent to the pursuit of a universal
seizure definition even when powerful, well understood signal analysis methods are
utilized to this end; b) Identifies this aim as a multi-objective optimization problem
and discusses the advantages and disadvantages of adopting or rejecting a unitary
seizure definition; c) Introduces a Probabilistic Measure of Seizure Activity to manage
this thorny issue.
The challenges posed by the attempt to define seizures unitarily may be partly related
to their fractal properties and understood through a simplistic analogy to the so-called
“Richardson effect”. A revision of the time-honored conceptualization of seizures
may be warranted to further advance epileptology.
Abstract
3
The task of automated detection of epileptic seizures is intimately related to and dependent on
the definition of what is a seizure, definition which to date is subjective and thus inconsistent
within and among experts [1,2,3]. The lack of an objective and universal definition not only
complicates the task of validation and comparison of detection algorithms, but possibly more
importantly, the characterization of the spatio-temporal behavior of seizures and of other
dynamical features required to formulate a comprehensive epilepsy theory.
The current state of automated seizure detection is, by extension, a faithful reflection of the
power and limitations of visual analysis, upon which it rests. The subjectivity intrinsic to
expert visual analysis of seizures and its incompleteness (it cannot quantify or estimate
certain signal features, such as power spectrum) confound the objectivity and reproducibility
of results of signal processing tools used for their automated detection. What is more, several
of the factors, that enter into the determination of whether or not certain grapho-elements
should be classified as a seizure, are non-explicit (“gestalt-based”) and thus difficult to
articulate, formalize and program into algorithms. Most, if not all, existing seizure detection
algorithms are structured to operate as expert electroencephalographers. Thus, seizure
detection algorithms that apply expert-based rules are at once useful and deficient; useful as
they are based on a certain fund of irreplaceable clinical knowledge and deficient as human
analysis biases propagate into their architecture. These cognitive biases which pervade human
decision processes and which have been the subject of formal inquiry [4-6] are rooted in
common practice behaviors such as: a) The tendency to rely too heavily on one feature when
making decisions (e.g., if onset is not sudden, it is unlikely to be a seizure because these are
paroxysmal events); b) To declare objects as equal if they have the same external properties
(e.g., this is a seizure because it is just as rhythmical as those we score as seizures) or c)
Classify phenomena by relying on the ease with which associations come to mind (e.g., this
pattern looks just like the seizures we reviewed yesterday).
The seizure detection algorithms’ discrepant results (Osorio et al, this issue) makes
attainment of a unitary or universal seizure definition ostensibly difficult; the notion that
expert cognitive biases are the main if not only obstacle on the path to “objectivity” is
rendered tenuous by these results. These divergences in objective and reproducible results
may be attributable in part, but not solely, to the distinctiveness in the architecture and
parameters of each algorithm. The fractal or multi-fractal structures of seizures [7,8] accounts
at least in part for the differences in results and draws attention to the so-called “Richardson
effect”. Richardson [9] demonstrated that the length of borders between countries (a natural
fractal) is a function of the size of the measurement tool, increasing without limit as the tool’s
size is reduced. Mandelbrot, in his seminal contribution “How long is the coast of Britain”
[10] stressed the complexities inherent to the Richardson’s effect, due to the dependency of
particular measurements on the scale of the tool used to perform them. Although defining
seizures as a function of a detection tool would be acceptable, this approach may be
impracticable when comparisons between, for example, clinical trials or algorithms are
warranted. Another strategy to bring unification of definitions is to universally adopt the use
of one method, but this would be to the detriment of knowledge mining from seizure-time
series and by extension to clinical epileptology.
A Probabilistic Measure of Seizure Activity (PMSA) is proposed as one possible strategy for
characterization of the multi-fractal, non-stationary structure of seizures, in an attempt to
eschew the more substantive limitations intrinsic to other alternatives.
The PMSA relies in this application on “indicator functions” (IFs) denoted χalgo for each
algorithm ‘algo’ and also on an Average Indicator Function (AIF):
4
The subscripts Val, r2, STA/LTA and WTMM refer to four different algorithms described
briefly below and more extensively in (Osorio et al, 2011, this issue) [Ivan Osorio, Alexey
Lyubushin and Didier Sornette, Automated Seizure Detection: Unrecognized
Challenges, Unexpected Insights, Epilepsy & Behavior (2011)]. An algorithm’s IF
equals 1 for time intervals (0.5 sec in this application) “populated” by ictal activity and 0 by
inter-ictal activity. The IF’s are used to generate four stepwise time functions, one for each
of: a) A 2ndorder auto-regressive model (r2); b) The Wavelet Transform Maximum Modulus
(WTMM); c) The ratio of short-to-long term averages (STA/LTA) and d) The Validated
algorithm (Val). With these IFs, the AIF is computed (its values may range between [0-1]
with intermediate1 values of 0.25, 0.5 and 0.75 in this application). These values [0-1] are
estimates of the probability of seizure occurrence at any given time.
The dependencies of AIF values on the detection algorithm applied to the ECoG are
illustrated in figure 1a-d and reflect the probability that grapho-elements are ictal in nature;
the higher the AIF value, the greater the probability that the detection is a seizure. AIF values
of 1 (the activity is detected by all algorithms as a seizure) or 0 (none of the algorithms
classifies the grapho-elements as a seizure) pose no ambiguity, but as shown in this study, are
likely to be less prevalent than intermediate values [0 < AIF < 1]. By way of example,
cortical activity may be classified as a seizure if the AIF value is 0.75, having been detected
by the majority (¾) of methods. In the study published in this issue (Osorio et al.), the four
different methods (r2, WTMM, STA/LTA, and Val) were investigated, but this number may
vary according to the task at hand; for warning for the purpose of allowing operation of a
motor vehicle, application of a larger number of detection algorithms to cortical signals and
an AIF value of 1 would be desirable while, for automated delivery of an innocuous, power
inexpensive therapy, less algorithms and much lower AIF values would be tolerable.
The cross-correlation between each pair of algorithm’s IF and their average function (AIF)
were calculated; since each of these is a step function (see figure 1), the Haar wavelet
transform was applied to them to facilitate visualization of their value (y-axis) as a function
of this wavelet’s logarithmic time scale (x-axis (Figure 2). The correlations (indicative of the
concordance level) between each IF pair and between each method’s IF and the AIF,
increases monotonically, reaching a maximum between 20-30s, after which they decrease
also monotonically (except for AIF vs. r2 ): The WTMM and
methods have the highest
correlations with AIF for time scales exceeding 100 sec. Since estimating the probability
measure of seizure activity based on the AIF requires the output of at least two detection
algorithms, a simpler approach is to apply only one, a Wavelet Transform Maximum
Modulus-Stepwise Approximation (WTMM-SAp).
be a logarithm of the standard deviation of differentiated ECoG computed within
Let
the time moments corresponding to right-
and
“small” adjacent time windows of length
hand ends of these windows. Thus,
values are given within the step
, where
is an
ECoG time interval.
1 Intermediate AIF values are functions of the number of algorithms applied to the signal. Since in this study 4
methods were used and the range of the indicator function is [0-1], the intermediated values are [0.25, 05, 0.75].
be a WTMM-SAp computed for the dyadic sequence of
Let
scale thresholds:
5
dimensionless
(26)
and
be their mean value:
(27)
The averaged WTMM-SAp
may reveal abrupt changes of
for different
scales (the use of a dyadic sequence (26) suppresses “outliers”). The background is estimated
by a simple average within a moving time window of the radius of
discrete values of
:
Seizures correspond to positive peaks of
values:
(28)
above background
. Thus, the
(29)
are regarded as a Measure of Seizure Activity (MSA). In order to make this measure
probabilistic (PMSA), consider an empirical probability distribution function:
(30)
be the
-quantile of the function (30), i.e. the root of the equation:
and let
The PMSA is defined by the formula:
(31)
(32)
It should be underlined that the PMSA (32) is defined within sequences of “small” time
and
are discrete time values, corresponding to right-hand
intervals of length
ends of these time windows.
The method of constructing a PMSA based on the WTMM-SAp utilizes the following
parameters whose values are shown in parentheses:
1)
of adjacent samples for computing the logarithm of the standard
for differentiated ECoG increments (
).
The number
deviations
2)
3)
6
for setting the dyadic sequence of WTMM scale thresholds
, e.g., the following scale thresholds were used:
The values of
in the formula (26) (
5, 10, 20, 40, 80 and 160).
The number
of
values for the radius of the moving averaging in formula
(28) (
Hz, the averaging length
and
, e.g., for
within formula (28) equals 401 sec).
The probability level
for calculating a quantile in formula (31) (
).
4)
The results of the estimations of PMSA using WTMM-SAp (Figure 3) differ in one aspect
(lower number of events with probability 1) from those obtained with the PMSA-AIF, given
the dissimilarities between these two approaches, but are alike in uncovering the
dependencies of PMSA on seizure duration: in general, the shorter the duration of a detection,
the larger the discordance between detection methods, a “trait” that interestingly, is also
shared by expert epileptologists (Osorio et al, 2002). Inter-algorithmic concordance as
evidenced by the cross-correlation values between PMSA-WTMM-SAp and PMSA-AIF
(Figure 4) grow quasi-linearly (albeit non-monotonically) with the temporal length of
seizures, reaching a maximum value (0.73) at 250 s. Worthy of comment is the decay in
cross-correlation values for seizure exceeding a certain length for both PMSA-AIF and
PMSA-WTMM-SAp
The crafting of, or “convergence” towards, a unitary seizure definition would be
epistemologically expensive and may thwart/delay deeper understanding of the dynamics of
ictiogenesis and of the spatio-temporal behavior of seizures at relevant time-scales. In the
absence of a universal definition, substantive gains are feasible through steps entailing, for
example, the application of advanced signals analyses tools to ECoG, to hasten the
identification of properties/features that would lead to the probabilistic discrimination of
seizures from non-seizures with worthwhile sensitivity and specificity for the task at hand.
Tools such as those available through cluster analysis of multidimensional vectors of relevant
features would aid in the pursuit of automated seizure detection and quantification. To even
have a modicum of success, this approach should not ignore the non-stationarity of seizures
and strike some sort of balance between supervised (human) and unsupervised machine-
learning) approaches. The resulting multidimensional parameter space, expected to be broad
and intricate, may also foster discovery of hypothesized (e.g. pre-ictal) brain sub-states.
Seizure detection belongs to a class of optimization problems known as “multi-objective”
[11] due to the competing nature between objectives; improvements in specificity of
detection invariably degrade sensitivity and vice-versa. Attempts to achieve a universal
seizure definition are likely to be fraught with similar competing objectives, but imaginative
application of tools from the field of multi-objective optimization, among others, are likely to
make this objective more tractable.
7
References
1. D’Ambrosio R, Hakimian S, Stewart T, Verley DR, Fender JS, Eastman CL, Sheerin
AH,Gupta P, Diaz-Arrastia R, Ojemann J, Miller JW. Functional definition of seizure
provides new insight into post-traumatic epileptogenesis. Brain 2009: 132; 2805–
2821
2. Dudek FE, Bertram EH. Counterpoint to “What Is An Epileptic Seizure”? by
D’AMbrosio and Miller. Epilepsy Curr. 2010; 10: 91–94
3. Frei MG, Zaveri HP, Arthurs S, Bergey GK, Jouny CD, Lehnertz K, Gotman J,
Osorio I, Netoff TI, Freeman WJ, Jefferys J, Worrell J, Le Van Quyen M, Schiff SJ,
Mormann F. Controversies in epilepsy: Debates held during the Fourth International
Workshop on Seizure Prediction. Epilepsy Behav. 2010 ;19:4-16.
4. Tversky A, Kahneman D. The framing of decisions and the Psychology of choice.
Science 1981; 211:453-58
5. Tversky A, Kahneman D. Judgement under uncertainty: Heuristics and Biases. In:
Conolly I, Arkes HR, Hammond KR, editors. Judgement and Decision Making. New
York: Cambridge University Press, 2000. p. 35-52.
6. Groopman J. How Doctors Think. Boston: Mifflin, 2007.
7. Osorio I, Frei MG, Sornette D, Milton J. Pharmaco-resistant seizures: self-triggering
capacity, scale-free properties and predictability? Eur J Neurosci. 2009;30:1554–
1558.
8. Osorio I, Frei MG, Sornette D, Milton J, Lai YC. Epileptic seizures: Quakes of the
brain? Phys Rev E. 2010;82:021919.
9. Richardson LF. The problem of contiguity: An appendix to Statistic of Deadly
Quarrels, General systems: Yearbook of the Society for the Advancement of General
Systems Theory 1961; 61:139–187
10. Mandelbrot, B. B. How long is the coast of Britain? Statistical self-similarity and
fractional dimension. Science 1967; 156: 636-638
11. Coello Coello CA, Lamont BG, van Velhuizen DA. Basic Concepts. In: Evolutionary
Algorithms for Solving Multi-objective Problems (Genetic and evolutionary
Computation ) New York: Springer, 2001. p. 1-60.
8
Figure 1a
9
Figure 1b
10
Figure 1c
11
Figure 1d
Figure 1a-d. Average Indicator Function value (AIF; grey step-wise functions) of the
probability that cortical activity (black oscillations) is a seizure over a certain time interval.
The AIF value (0-1) of this function is calculated based on the output of each of the four
detection algorithms used. Notice that the larger amplitude, longer oscillations are the only
ones to have an AIF value of 1, indicative of “consensus” among all detection algorithms (x-
axis: time; y-axis: AIF values)
12
Figure 2. Plots of time scale-dependent correlations between Haar wavelet coefficients of
the indicator functions (IFs) between pairs of detection methods and between each method
and the averaged indicator function (AIF). Notice that r2, STA/LTA and WTMM act as labels
for both columns (label on top) and rows (label to the right of each row), whereas Val
designates only the column below it and AIF the row to its left. This graph may be viewed as
the lower half of a square matrix; this triangle’s vertices are: the top left-most plot depicts the
correlation between Val and r2, the bottom left-most plot the correlation between Val and AIF
and the bottom right-most graph, that between WTMM and AIF; all other correlations lie
within these vertices (y-axes: Correlation values; x-axes: Logarithmic time scale).
13
Figure 3a
14
Figure3b
15
Figure 3c
16
Figure 3d
Figure 3 a-d. Probability Measure of Seizure Activity estimated using the Wavelet Transform
Maximum Modulus - Stepwise Approximations. Panels 3a-3d correspond to panels 1a-1d.
The oscillations in black are cortical activity and the grey stepwise function, the probability
value they correspond to seizures (x-axes: time; y-axes PMSA values.
17
Figure 4. Graphic of time scale-dependent correlations between PMSA-AIF and PMSA-SA
after smoothing of their step-wise functions with Haar wavelets. Correlation value increase
as a function of time before decaying steeply after approximately 250s.
|
1807.08686 | 1 | 1807 | 2018-07-23T15:54:33 | Storing and retrieving long-term memories: cooperation and competition in synaptic dynamics | [
"q-bio.NC",
"physics.bio-ph"
] | We first review traditional approaches to memory storage and formation, drawing on the literature of quantitative neuroscience as well as statistical physics. These have generally focused on the fast dynamics of neurons; however, there is now an increasing emphasis on the slow dynamics of synapses, whose weight changes are held to be responsible for memory storage. An important first step in this direction was taken in the context of Fusi's cascade model, where complex synaptic architectures were invoked, in particular, to store long-term memories. No explicit synaptic dynamics were, however, invoked in that work. These were recently incorporated theoretically using the techniques used in agent-based modelling, and subsequently, models of competing and cooperating synapses were formulated. It was found that the key to the storage of long-term memories lay in the competitive dynamics of synapses. In this review, we focus on models of synaptic competition and cooperation, and look at the outstanding challenges that remain. | q-bio.NC | q-bio | July 24, 2018
0:58
Advances in Physics
review
To appear in Advances in Physics
Vol. 00, No. 00, Month 20XX, 1 -- 28
8
1
0
2
l
u
J
3
2
]
.
C
N
o
i
b
-
q
[
1
v
6
8
6
8
0
.
7
0
8
1
:
v
i
X
r
a
REVIEW ARTICLE
Storing and retrieving long-term memories: cooperation and
competition in synaptic dynamics
Anita Mehta∗
Dipartimento di Fisica, Universit`a di Roma La Sapienza, P. A. Moro 2, 00185 Roma, Italy
and Institut fur Informatik, IZBI, Universitat Leipzig, Hartelstrasse 16 -- 18, 04107 Leipzig,
Germany
(Received 00 Month 20XX; final version received 00 Month 20XX)
We first review traditional approaches to memory storage and formation, drawing on the
literature of quantitative neuroscience as well as statistical physics. These have generally
focused on the fast dynamics of neurons; however, there is now an increasing emphasis on
the slow dynamics of synapses, whose weight changes are held to be responsible for memory
storage. An important first step in this direction was taken in the context of Fusi's cascade
model, where complex synaptic architectures were invoked, in particular, to store long-term
memories. No explicit synaptic dynamics were, however, invoked in that work. These were
recently incorporated theoretically using the techniques used in agent-based modelling, and
subsequently, models of competing and cooperating synapses were formulated. It was found
that the key to the storage of long-term memories lay in the competitive dynamics of synapses.
In this review, we focus on models of synaptic competition and cooperation, and look at the
outstanding challenges that remain.
PACS: 87.18.Sn Neural networks and synaptic communication, 87.19.lv Learning and
memory, 89.75.Da Systems obeying scaling laws, 05.40.-a Fluctuation phenomena, random
processes, noise, and Brownian motion
Keywords: Synaptic plasticity; competitive learning; power-law forgetting; competitive
synaptic dynamics; long-term memory
1.
Introduction
Memory [1, 2] and its mechanisms have always attracted a great deal of interest [3]. It is
well known that memory is not a monolithic construct, and that memory subsystems cor-
responding to episodic, semantic or working memory exist [4]. We focus here on explicit
memory, which is the memory for events and facts.
Models of memory have, themselves, long been studied in the field of mathematical
psychology: the article by Raaijmakers and Shiffrin [5] provides a valuable review of
models that existed well before the neural network models with which most physicists
are familiar, began to appear. Here, memory was assumed to be distributed over a large
set of nodes and an item was defined by the pattern of activation over a set of nodes.
This was propagated through a network of links whose geometry and weights determined
the output. Such models of storage and retrieval are discussed at length in [5], but in
the interests of a historical presentation, we briefly describe the earliest example known
∗Email: [email protected]
July 24, 2018
0:58
Advances in Physics
review
as the 'brain state in a box' model, or BSB [6]. In this model, items are vectors while
learning is represented by changes in synaptic strengths. For any such pair of items, the
synaptic strengths between the input and output layers are modified in such a way that
considerable storage and retrieval is possible, even in the presence of noise. There have
in parallel been a lot of suggestions regarding the way in which working memory actually
functions: from the point of view of the current review, the most important distinction
between these is that forgetting involves temporal decay in the research of Baddeley
and co-workers [7, 8] and that it does not, in the work of Nairne and co-workers [9, 10].
Although a detailed discussion of these psychological (and somewhat empirical) models
is beyond the scope of this review, they do indeed offer fertile ground for mathematical
modellers who would wish to construct quantitative models of working memory.
In general, memories are acquired by the process of learning. Simply put, patterns of
neural activity change the strength of synaptic connections within the brain, and the
reactivation of these constitutes memory [11]. In this context, we first review the differ-
ent kinds of learning to which a network can be subjected [12]. These are respectively:
supervised, reinforcement, and unsupervised learning. In supervised learning, the goal is
to learn a mapping between given input and output vectors, as, for instance, when we
classify the identity of items in a list. In reinforcement learning, the goal is to learn a
mapping between a set of inputs or actions in a particular environment and some mea-
sure of reward. In unsupervised learning, the network is provided with no feedback at
all. Rather, synaptic strength changes occur according to a learning rule based only on
pre- and post-synaptic activity, with no reference to any desired output. The pattern of
synaptic strengths that results in this case, depends on the nature of the learning rule
and the statistical structure of the inputs presented. It is this kind of learning with which
this review will be chiefly concerned.
The somewhat bland statement above, of memories being acquired by a process of
learning, actually pushes a lot of puzzles under the rug. Why is it that some memories
are quickly forgotten, while others last a lifetime? One hypothesis is that important
memories are transferred, via their synaptic strengths, to different parts of the brain
that are less exposed to ambient noise. In particular, during a process known as synaptic
consolidation [13]1, memories that are first stored in the hippocampus are transferred to
other areas of the cortex [14, 15]; this transfer can happen while the events are rerun
during sleep [16]. The case of the famous patient HM [17] whose hippocampus was
removed following epilepsy reinforces this hypothesis: HM retained old memories from
before his surgery, but he could barely acquire any new long-term memories.
There is yet another mechanism for memory consolidation which happens at the synap-
tic level, involving the mechanism of synaptic plasticity, whereby synapses change their
strength. Short-term plasticity (STP) occurs when the change lasts up to a few min-
utes, while long-lasting increases/decreases of synaptic strength are known respectively
as long-term potentiation/depression (LTP/LTD); LTP was first discovered experimen-
tally by Bliss and Lomo [18] in 1973. Long-term plasticity is further subdivided into
early-long-term plasticity (e-LTP) when synaptic changes last up to a few hours and
late-long-term plasticity (l-LTP), when they last from beyond typical experimental du-
rations of 10 hours to possibly a lifetime. Such late-long-term plasticity also falls within
the terminology of synaptic consolidation [19]; here, relevant memories are consolidated
within the synapses concerned, so that new memories can no longer alter previously
consolidated ones. The two most important theoretical models of this second kind of
1In the literature, this is sometimes referred to as systems consolidation, while synaptic consolidation is tradition-
ally used to describe the molecular mechanism that leads to the maintenance of synaptic plasticity.
2
July 24, 2018
0:58
Advances in Physics
review
synaptic consolidation involve a process called synaptic tagging [19 -- 21]. The hypothesis
is that a single, brief burst of high-frequency stimulation is enough to induce e-LTP,
and its expression does not require protein synthesis. On the other hand, l-LTP can be
induced by repeated bursts of high-frequency stimulation, which leads to an increase in
synaptic strength until saturation is reached. There is also a view [21] that more stim-
ulation does not increase the amount of synaptic weight change at individual synapses,
but rather increases the duration of weight enhancement. In this case, it has been shown
that protein synthesis is triggered at the time of induction. Also, it was found that e-LTP
at one synapse could be converted to l-LTP if repeated bursts of high-frequency stimu-
lation are given to other inputs of the same neuron during a short period before or after
the induction of e-LTP at the first synapse [22]. This discovery led to the hypothesis
that such stimulation initiates the creation of a 'synaptic tag' at the stimulated synapse,
which is thought to be able to capture plasticity-related proteins. The general framework
for these hetero-synaptic effects is called synaptic tagging and capture, for the details of
which the reader is referred to [19, 21].
It should be mentioned here that because of the interdisciplinary nature of the field,
much of the discussion in the literature [23, 24] involves terminology such as 'plasticity
induction and maintenance', to refer respectively to short-term and long-term plasticity
changes. Specifically, in [24], the author's findings reinforce the intuition that LTP induc-
tion and maintenance would lead respectively to short- and long-term memory. Thus in
the following, models manifesting short-term memory involve only plasticity induction,
while plasticity maintenance is responsible for the manifestation of long-term memory in
the models that form the core of this review.
Finally, some of the most recent developments in the modelling of memory acqui-
sition and maintenance involve the concept of engrams [25]; here, memories may be
reconstructed by single neuronal activation.The underlying idea is that a big network of
neurons is involved in memory acquisition, with several connections being modified; these
may be lost over time or in an activity-dependent manner such that memory is virtu-
ally supported by a single connection, and later reconstructed. This mechanism suggests
that memory reactivation may not rely on the same network involved in its acquisition,
but rather on the reconnection of neurons that may have similar responses. The authors
of [25] also suggest that memories at the time of acquisition are already stored in the
cortex, instead of being transferred from the hippocampus to the cortex as suggested
in [14, 15].
To sum up: memory formation is a complicated phenomenon related to neural activi-
ties, brain network structure, synaptic plasticity [26] and synaptic consolidation [19].
We will provide an overview of some of the more traditional approaches, involving
neural networks -- both those based on detailed biophysical principles, and those that were
explored by statistical physicists starting from the seminal work of Hopfield [27]. Much of
this has already been extensively reviewed, so the focus of the present review comprises
questions like: how can short-term and long-term memory coexist in our brains? While
it is known that short-term memory is ubiquitous, what are the synaptic mechanisms
needed for long-term memory storage?
It is well known that too much plasticity causes the erasure of old memories, while too
little plasticity does not allow for the quick storage of new memories. This palimpsest
paradox [28, 29] has been at the heart of the quandary faced by modellers of synaptic dy-
namics. While synaptic consolidation does indeed provide some insights into this, neuro-
scientists [30, 31] have typically focused on synaptic plasticity [32], for which increasingly
sophisticated models have emerged over the years [33 -- 35]. There are two broad classes:
biophysical models, which incorporate details at the molecular level, and phenomenolog-
3
July 24, 2018
0:58
Advances in Physics
review
ical models, which relate neuronal activity to synaptic plasticity. It is the latter class
of models that we will focus on in this review, both because they are more amenable
to statistical physical techniques and because they account for higher-level phenomena
like memory formation. Such modelling, while it may not include details of specificities
involving chemical and biological processes in the brain, can outline possible mechanisms
that take place in simplified structures. For example, the study of neural networks [33 --
35], while it greatly simplifies biological structures in order to make them tractable, has
still been able to make an impact on the parent field. In particular, neural networks such
as the Hopfield model [27, 36] have been extensively investigated via methods borrowed
from the statistical physics of disordered and complex systems [37 -- 39]. In these models,
memories are stored as patterns of neural activities, which correspond both to low-energy
states and to attractors of the stochastic dynamics of the model.
What this class of phenomenological models lacks in biological detail, it typically makes
up for in minimalism. Abbott, one of the pioneers in this field, summed up its virtues
thus [40]:
Identifying the minimum set of features needed to account for a particular phenomenon
and describing these accurately enough to do the job is a key component of model building.
Anything more than this minimum set makes the model harder to understand and more
difficult to evaluate. The term 'realistic' model is a sociological rather than a scientific term.
The truly realistic model is as impossible and useless a concept as Borges' map of the empire
that was of the same scale as the empire and that coincided with it point for point.
Within this class of models, there is yet another divide; there are models which focus on
the fast dynamics of neurons, and then those that focus on the slow dynamics of synapses.
We will review each one in turn. In particular, in the second case, we will focus on the
nature of synaptic dynamics, which involve competition and cooperation [41]. There is
abundant evidence that correlation-based rules of synaptic cooperation, which lead to the
outcome 'neurons that fire together, wire together', are followed in many organisms; the
latter is known as Hebb's rule, due to the pioneering work of Hebb in establishing it [42].
In synaptic cooperation therefore, synapses that work together are rewarded by being
strengthened. However, synapses also have a competitive side: while some synapses grow
stronger and prosper, others, which left to themselves would also have strengthened,
instead weaken. (An example of this can be seen in the process of ocular dominance
segregation [43], where competitive correlations ensure that inputs to the left and right
eye, though they fire together, do not wire together). Of these two processes, synaptic
cooperation is by far the more commonly used in mathematical modelling; however, its
unbridled prevalence leads to instabilities, for which synaptic competition provides a
cure. From a more biological standpoint, synaptic competition is a concept that has long
found favour with the neuroscience community [31]; however, its use is relatively recent
in the context of statistical physics models. The present review accordingly emphasises
those approaches where synaptic cooperation and competition are key.
We begin this article with a review of the Hopfield model (Section 2), where we de-
scribe the model as well as its use in storing and retrieving random patterns. We then
turn to phenomenological models of synaptic plasticity (Section 3), which are further
classified as rate-based models (Section 3.1) and spike-time-dependent plasticity models
(Section 3.2), where the synaptic strength is always treated as a continuous variable.
A change of key sets in in Section 4, when synapses are discretised, with the further
possibility (Section 4.1) of occupying a multiplicity of states. In the following section
(Section 5), we present an extensive review of the neuroscience literature to do with
the perceived need for synaptic competition. These ideas are implemented in Section 6
4
July 24, 2018
0:58
Advances in Physics
review
where, in particular, synaptic strengths are discretised and competitive dynamics em-
bedded, using tools from statistical physics. In the Discussion (Section 7), we summarise
the state of the literature, and discuss some future challenges.
2. The Hopfield model
Appropriately for the readership of this journal, we start by reviewing the Hopfield
model, both because this is one of the seminal contributions of physics to the field, and
also because it is the basis on which a large class of models (STDP, cf. Section 3.2) is
based.
In 1982, John Hopfield introduced an artificial neural network to store and retrieve
memory like the human brain [27, 36]. In such a fully connected network of N neurons,
there is a connectivity (synaptic) weight Jij between any two neurons i and j, which is
symmetric so that Jij = Jji, and Jii = 0. Such a network is initially trained to store a
number of patterns or memories. It is then able to recognise any of the learned patterns
by exposure to only partial or even some corrupted information about that pattern, i.e.,
it eventually settles down and returns the closest pattern or the best guess.
We present here a simple picture of memory storage and retrieval along the lines of [44].
Each neuron is characterised by a variable S which takes the value +1 if the neuron is
firing and −1 if the neuron is not firing. At time t + 1 the neuron labelled by the index i,
where i = 1, 2, 3, . . ., N for a system of N cells, fires or does not fire based on whether
the total signal it is receiving from other cells to which it is synaptically connected is
positive or negative. Thus, the basic dynamical rule is
Si(t + 1) = sgn N
Xj=1
JijSj(t)!,
(1)
where Jij is a continuous variable representing the strength of the synapse connecting
cell j to cell i. The basis of a network associative memory is that the above dynamics can
map an initial state of firing and non-firing neurons, Si(0), to a fixed pattern, ξi, which
remains invariant under it. Various memory patterns ξµ
for µ = 1, 2, 3, . . ., P which do
i
not change under the above transformation act as fixed-point attractors; initial inputs
Si(0) are mapped to an associated memory pattern ξµ
i Si(0)/N is
i
close enough to one. How close this overlap must be to one, or equivalently how well the
initial pattern must match the memory pattern in order to be mapped to it and thus
associated with it, is determined by the radius of the domain of attraction of the fixed
point. The issue of domains of attraction associated with a fixed point has never been
completely resolved. The sum of all synaptic inputs at site i,
if the overlap P ξµ
hµ
i =
N
Xj=1
Jijξµ
j ,
(2)
known as the local field, is the signal which tells cell i whether or not to fire when Sj = ξµ
j
for all j 6= i. In order for a memory pattern to be a stable fixed point of the dynamics,
the local field must have the same sign as ξµ
i or equivalently
i ξµ
hµ
i > 0 .
5
(3)
July 24, 2018
0:58
Advances in Physics
review
i ξµ
We will call the quantities hµ
i the aligned local fields. It seems reasonable to assume
that the larger the aligned local fields are for a given µ value the stronger the attraction
of the corresponding fixed point ξµ
i and so the larger its domain of attraction. This
reasoning is almost right, but it leaves out an important feature of the above dynamics.
Multiplying Jij by any constants has absolutely no effect since the dynamics depends
only on the sign and not on the magnitude of the quantity P JijSj. Since the quantities
hµ
i ξµ
i change under this multiplication they alone cannot determine the size of the basin
of attraction. Instead, it has been found that quantities known as stability parameters
and given by
where we define
γµ
i =
hµ
i ξµ
i
Ji
,
Ji = N
Xj=1
J 2
ij!1/2
,
(4)
(5)
i . Roughly speaking, the larger the values of the γµ
provide an important indicator of the size of the basin of attraction associated with the
fixed point ξµ
i the larger the domain of
attraction of the associated memory pattern. In order to construct an associative memory
one must find a matrix of synaptic strengths Jij which satisfies the condition of stability
of the memory fixed points and has a specified distribution of values for the γµ
i giving
the domain of attraction which is desired.
Notice that in the above, the synapses are used for storage and retrieval of memories,
as well as a way of updating the neuronal states; in other words, they are not explicitly
updated.
3. Phenomenological models of synaptic plasticity
We move on now to models where plasticity is invoked, i.e., where synapses are explicitly
updated. The assumption here is that neuronal firing rates are, in their turn, responsible
for synaptic strengthening or weakening. The basic principle at work is Hebb's rule [42],
which as mentioned above, says that 'cells that fire together, wire together'. Another
way of viewing this rule is to say that simultaneous events over a period of time suggest
a causal link, and many rate-based models of synaptic plasticity have been formulated
on this basis. However, and more recently, a great deal of attention has been paid to
a much stricter definition of causality via the field of spike-timing-dependent plasticity
(STDP) [45, 46]: here, synaptic strengthening only occurs if one of the neurons is sys-
tematically active just before another one. In addition to realising the Hebbian condition
that a synapse should be strengthened only if it constitutes a causal link between the
firing of pre- and post-synaptic neurons, STDP also leads to the weakening of synapses
which connect neurons whose firings are temporally correlated, but where the firing is
not causally ordered.
We briefly review these two classes of models below.
6
July 24, 2018
0:58
Advances in Physics
review
3.1. Rate-based models
Here, the rate of pre- and post-synaptic activities measured over some time period de-
termines the sign and magnitude of synaptic plasticity. The activities are modelled as
continuous variables, corresponding to a suitable average of neuronal firing rates. The
rate of change of synaptic strength or weight Ji at synapse i is modelled as a function
of the pre-synaptic input xi at that synapse, the post-synaptic output activity y, the
weight itself, and, in the most general case, the weights of other synapses:
dJi
dt
= f (xi, y, Ji, Jj ).
(6)
Without the competition from other synapses Jj, synaptic weights could grow un-
controllably. Before the explicit inclusion of synaptic competition, this instability was
combated in two ways; in Oja's [47] model, Hebbian plasticity was augmented with a
decay term, so that weights equilibrated to the first principal component of the input
correlation matrix. Another way forward was shown by the BCM model [48] which explic-
itly included both LTP and LTD regions, with a sliding threshold separating them; when
synaptic weights became too large, the threshold shifted so that any further activation
led to synaptic depression. Subsequently, indirect ways of including synaptic competi-
tion (Section 5), such as the normalisation of the total synaptic weights, were included
in the modelling; more recently, there have been a number of approaches where synaptic
weights are discretised (Section 4) and competition explicitly implemented (Section 6).
3.2. Models of spike-timing-dependent-plasticity
Spike-timing-dependent-plasticity (STDP) provides the answer to the following question:
For neurons embedded in a network which are bombarded with millions of inputs, which
ones are important? Which information should a given neuron 'listen' to, and pass along
to downstream neurons? These are the formidable questions that the vast majority of
neurons in the brain have to solve during brain development and learning. The crucial
link is causality -- if one of the cells is active systematically just slightly before another,
the firing of the first one might have a causal link to the firing of the second one and this
causal link could be remembered by increasing the wiring of connections. Theoreticians in
the mid-1990's realized just how important temporal order was for conveying and storing
information in neuronal circuits, and experimenters saw how the synaptic connections of
the brain should be acutely sensitive to timing. Thus the field of STDP was born, via
the key studies of Markram and Gerstner [45, 46]. With STDP, a neuron embedded in
a neuronal network can 'determine' which neighbouring neurons are worth connecting
with, by potentiating those inputs that predict its own spiking activity, and effectively
ignoring the rest [49]. The net result is that the sample neuron can integrate inputs with
predictive power and transform this into a meaningful predictive output, even though
the meaning itself is not strictly known by the neuron.
An early example of using such models in associative memory can be found in [50].
This introduces 'spiking' neurons in a Hopfield [27, 36] network: by the term spiking,
three main features are implied, which are: a) a neuron fires when a given threshold is
reached; b) it then undergoes a period of rest, which is referred to as 'refractoriness';
c) noise may be added to the firing rates. The synapses connecting the neurons follow
a Hebbian learning rule (with no explicit competition) whereby incoming patterns are
learnt, and their retrieval analysed along the lines of Section 2 as a function of various
7
July 24, 2018
0:58
Advances in Physics
review
parameters.
While models of neurons themselves are the subject of considerable discussion [51],
these early models have been greatly refined in recent times, and are usefully sum-
marised in [52]. However, as pointed out in [49], these theories are limited by the types
of plasticity invoked in the models concerned. Indeed, in [53], it is tacitly acknowledged
that without appropriate compensatory mechanisms (referred to there as being 'non-
Hebbian'), Hebbian learning alone is not able to account for the reliable storage and
recall of memories; the necessary mechanisms invoked in [53] involve, in addition to the
Hebbian LTP/LTD, the (implicitly competitive) mechanism of heterosynaptic up and
down regulation of synapses, as well as transmitter-induced plasticity and consolidation.
This indeed reinforces the perceived need for some form of competition, as well as a
somewhat more parsimonious form of modelling where possible.
Before concluding, we also mention that most STDP models can be averaged and re-
duced to rate-based models with certain assumptions: if all nodes interact with each
other, they can be reduced to correlation-based models [54] whereas if nearest-neighbour
interactions exist, the models that result are similar to the BCM model [55]. However,
the fast dynamics of neurons, on which the STDP models are based, continue to attract a
lot of research interest. Typically, models of integrate-and-fire neurons on networks have
been extensively studied, and their different dynamical regimes explored [56]. In [57], the
memory performance of a class of modular attractor neural networks has been exam-
ined, where modules are potentially fully-connected networks connected to each other
via diluted long-range connections. Interest in this fast dynamical regime has also been
fuelled by the discovery of neuronal avalanches in the brain [58], which was followed by
several dynamical models of neural networks [59, 60], where the statistics of avalanches
were investigated [61 -- 66] and reviewed in [67]. In fact, the field of spiking neurons is now
so well-established that it is the subject of textbooks -- of which an excellent example is
the one by two of the most important workers in the field, Gerstner and Kistler [35].
4. State-based models
An alternative to considering unbounded and continuous synaptic weights -- as is done
in Sections 2 and 3 -- is to consider discrete synapses, with a limited number of synaptic
states, whose weights are bounded. This has experimental support [26, 68], and also has
the advantage that binary synapses, say, may be more robust to noise than continuous
synapses [69]. An essential property of these models as well as real neural networks is
that their capacity is finite. Such bounded synapses have the palimpsest property, i.e.,
new memories are stored at the cost of old ones being overwritten [29]. This is in marked
contrast to the case of unbounded synapses where the overall quality of both old and new
memories degenerate as new information is processed. For bounded synapses, therefore,
forgetting is an important aspect of continued learning [26, 28, 29, 70 -- 73]. This situation
-- that of discrete, bounded synapses with an explicit forgetting mechanism -- is what we
will focus on in the rest of this review.
Van Rossum and coworkers [21, 74] have done a body of work on such state-based
models; they have shown in particular that there is not an overwhelming reduction
in the storage capacity of discrete synapses as compared to continuous ones. In their
work, each synapse is described with a state-diagram and each state has an associated
synaptic weight. The simplest case of binary synapses ('synaptic switches'), has been ex-
tensively used in earlier mathematical models [21, 75 -- 77]. Interactions between synapses
are incorporated in the state diagrams. Typically, Markov descriptions are used, and
8
July 24, 2018
0:58
Advances in Physics
review
the eigenvalues of the Markov transition matrices give the decay times of the synaptic
weights.
The above mechanism of synaptic plasticity has, however, been shown to be rather
inefficient when synapses change permanently [78]. Pure plasticity indeed does not pro-
vide a mechanism for protecting some memories while leaving room for other, newer,
memories to come in, hence leading to the need for the mechanism of metaplasticity [70 --
72]2. In order to improve performance, Fusi et al
[79] proposed a cascade model of
a synapse with many hidden states, which they claimed was able to store long-term
memories more efficiently, with a decay that was power-law rather than exponential in
time.The pathbreaking idea behind the work of [79] was that the introduction of 'hidden
states' for a synapse would enable the delinking of memory lifetimes from instantaneous
signal response: while maintaining quick learning, it would also enable slow forgetting. In
the original cascade model of [79], this was implemented by the storage of memories at
different 'levels': the relaxation times for the memories increased as a function of depth.
It was assumed that short-term memories, stored at the uppermost levels, would decay
as a consequence of their replacement by other short-term memories ('noise'). On the
other hand, longer-lasting memories remained largely immune to such noise as they were
stored at the deeper levels, which were accessible only rarely. This hierarchy of timescales
models the phenomenon of metaplasticity [80, 81], and will be discussed in detail below.
4.1. Fusi's cascade model: A quantitative formulation
Fusi's model [79] of a metaplastic binary synapse with infinitely many hidden states
was formulated quantitatively and investigated in [82]. Each state is here labelled by its
depth n = 0, 1, . . . , At every discrete time step t, the synapse is subjected either to an
LTP signal (encoded as ε(t) = +1) or to an LTD signal (encoded as ε(t) = −1), where
ε(t) = ±1 is the instantaneous value of the input signal at time t.
The model, portrayed in Figure 1, is defined as follows: The application of an LTP
signal can have three effects [82]:
• If the synapse is in its − state at depth n, it may climb one level (n → n − 1) with
probability αn. (This move was absent in the original model.)
• If it is in its − state at depth n, it may alternatively hop to the uppermost + state
with probability βn.
• If it is already in its + state at depth n, it may fall one level (n → n + 1) with
probability γn.
Long-term memories will be stored in the deepest levels of the synapse, because of
the persistent application of unimodal signals. The effect of noise on such a long-term
memory here is to replace a long-term memory by a short-term memory of the opposite
kind. If, for example, the signal is composed of all + + + + + + +, an isolated − event
could be seen to represent the effect of noise. In this case, the Fusi model [79] predicts
that the signal is thrown from a deep positive level of the synapse to the uppermost level
of the negative pole. Seen differently, this mechanism converts a long-term memory of
one kind to a short-term memory of the opposite kind.
Along the lines of [79, 82], the transition probabilities of this model are assumed to
2An older use of the term 'metaplasticity' relates to changes in synapses that are not expressed as changes in
synaptic efficacy, but rather alter their responses to subsequent stimuli, an example of this being the sliding
threshold of plasticity described in the BCM model [48].
9
July 24, 2018
0:58
Advances in Physics
review
β
0
β
1
β
2
α
1
α
2
β
0
β
1
β
2
α
1
α
2
γ
0
γ
1
γ
0
γ
1
ε = +1 (LTP)
ε = −1 (LTD)
Figure 1. Schematic representation of Model I. Arrows denote possible transitions in the presence of an LTP
signal (ε = +1, left panel) and of an LTD signal (ε = −1, right panel). Corresponding transition probabilities are
indicated. In each panel, the left (resp. right) column corresponds to the − (resp. +) state. The model studied in
this work is actually infinitely deep (after Ref. [82]).
decay exponentially with level depth n:
αn = αe−(n−1)µd ,
βn = βe−nµd,
γn = γe−nµd.
The corresponding characteristic length,
ξd =
1
µd
,
(7)
(8)
is one of the key ingredients of the model, which measures the number of fast levels at
the top of the synapse. It will be referred to as the dynamical length of the problem. The
choice made in [79] corresponds to e−µd = 1
2 , i.e., µd = ln 2. A different characteristic
length, the static length ξs, is given by
ξs =
1
µs
.
(9)
This is referred to as the static length of the problem, and gives a measure of the effective
number of occupied levels in the default state [82]. The regime of most interest is where ξs
is moderately large, so that the default state extends over several levels. The mean level
depth
hnist =
1
eµs − 1
= ξs − 1
2 + · · ·
is then essentially given by the static length.
The level-resolved output signal of level n at time t:
Dn(t) = Qn(t) − Pn(t)
10
(10)
(11)
July 24, 2018
0:58
Advances in Physics
review
and the total output signal at time t:
D(t) =Xn≥0
Dn(t)
(12)
can be expressed in terms of the probabilities Pn(t) (or Qn(t)) for the synapse to be in
the − state (or the + state) at level n = 0, 1, . . . at time t = 0, 1, . . .
We now describe the effect of an LTP signal, i.e., a sustained input of potentiating
pulses lasting for T consecutive time steps (ε(t) = +1 for 1 ≤ t ≤ T ) on the model
synapse. The synapse, assumed to be initially in its default state [82] will get almost
totally polarized in response to the persistent signal.
This saturation phenomenon is illustrated in Figure 2, which shows the output signal
D(t) for several durations T of the LTP signal. The synapse slowly builds up a long-term
memory in the presence of a long enough LTP signal, as the memorized signal moves
to deeper and deeper levels. At the end of the learning phase (t = T ), the polarisation
profile will have the form of a sharply peaked traveling wave, around a typical depth
which grows according to the logarithmic law [82]
n(T ) ≈ ξd ln γT.
(13)
After the signal is switched off, the total output signal decays. The late stages of the
forgetting process are characterized by a universal power-law decay of the output signal:
This is known as power-law forgetting [83 -- 85]. The forgetting exponent
D(t) ∼ t−θ.
θ = 1 +
ξd
ξs
(14)
(15)
is always larger than unity and depends on the ratio of the dynamical and static lengths
ξd and ξs. As Equation (14) shows, it has no dependence on the the duration of the
learning phase, in keeping with the requirements of universality.
1
0.8
0.6
0.4
0.2
)
t
(
D
1
3
10
30
0
0
20
40
60
80
100
t
Figure 2. Plot of the output signal D(t) against time t, for several durations T of the LTP signal for parameter
values β = 0.2, γ = 0.5, and ξs = ξd = 5 (after Ref. [82]).
11
July 24, 2018
0:58
Advances in Physics
review
4.2. Comparison of cascade model with experiment
The cascade model and its variants have frequently been criticised for being somewhat
abstract; one response has been to come up with ever-more sophisticated models for
synaptic consolidation which incorporate the multiple timescales inherent in the cascade
model. A three-layered model of synaptic consolidation has been proposed that accounts
for data across a large range of experimental conditions [86]; while it has a daunting
number of parameters -- 17 -- , it is able to incorporate the retention of long-term memo-
ries. Fusi's own recent extension of the cascade model is also rather intricate: memories
are stored and retained through complicated coupled processes operating on multiple
timescales. This is achieved by combining multiple dynamical processes that initially
store memories in fast variables, and then progressively transfer them to slower vari-
ables. It has the advantage of getting a larger memory capacity, while the corresponding
disadvantage is that it is even more abstract than his earlier model, so that involved
biological processes have to be explained via systems of communicating vessels [87].
We choose here instead to highlight a link with an experiment [88] whose findings are
explained by the complex synaptic architectures of Fusi's original model [79], to combat
the proposition that the cascade model is 'too abstract' to be useful. In particular the
experiment involves a single synapse connecting two cells, so that the Fusi model of a
single synapse is appropriate. Specifically, in a system comprising an excitatory synapse
between Lymnaea pre- and postsynaptic neurons (visceral dorsal 4 (VD4) and left pedal
dorsal 1 (LPeD1- Excitatory)), a novel form of short-term potentiation was found, which
was use-, but not time-dependent [88]. Following a tetanic stimulation (∼ 10 Hz) in
the presynaptic neuron with a minimum of seven action potentials, the synapse became
potentiated whereby a subsequent action potential triggered in the presynaptic neuron
resulted in an enhanced postsynaptic potential. Further, if an inducing tetanic stimu-
lation was activated, but a subsequent action potential was not triggered, the synapse
was shown to remain potentiated for as long as 5 hours. However, once this action po-
tential was triggered, the authors found that the synaptic strength rapidly returned to
baseline levels. It was also shown that this form of synaptic plasticity relied on the presy-
naptic neuron, and required pre- (but not post-) synaptic Ca2+/calmodulin dependent
kinase II (CaMKII) activity. Hence, this form of potentiation shares induction and de-
potentiation characteristics similar to other forms of short-term potentiation, but exhibits
a time-frame analogous to that of long-term potentiation.
In [89], this experiment was interpreted via a variant of the cascade model described
above, as follows: after a process of tetanic stimulation, the initial action potentials,
interpreted as a non-random signal, cumulatively built up a long-term memory of the
signal in the deepest synaptic levels. The synapse dynamics were then frozen so that fur-
ther discharge was prevented. When a further action potential was applied, the synaptic
dynamics restarted ('use'-dependence): the release of the accumulated memory from the
deepest levels of the synapse constituted the observed enhancement of the output sig-
nal described in [88]. While this enhancement is plausibly accounted for by the model
of metaplastic synapses [82], the explanation of the freezing of the synaptic dynamics
and its subsequent use-dependence needed the introduction of a stochastic and bistable
biological switch to model the role of kinase (CaMKII) in the actual experiment [88].
Specifically, the synapse (Figure 1), assumed to be initially in its default state, is
subjected to a sustained LTP signal of duration T1 (i.e., the application of T1 action
potentials), and to a single action potential at a much later time (T2 ≫ T1). It is subjected
to a random input at all the other instants of time (ε(t) = +1 for 1 ≤ t ≤ T1 and for
t = T2, else ε(t) = 0). In the regime where the number of action potentials T1 of the initial
12
July 24, 2018
0:58
Advances in Physics
review
signal is larger than some characteristic time T0 of the switch, the freezing probability
of the switch at the end of the LTP period is very high, i.e., very close to unity. During
this learning phase, the output signal D(t) grows progressively from D(0) = 0 to a
large value D(T1). The high value of the freezing probability at the end of this phase
typically freezes the synaptic dynamics, ensuring that this enhanced output signal is not
discharged. When the next action potential is applied at time T2, the switch is turned
off, and the synapse then relaxes via the full discharge of the stored, enhanced output
signal.
)
t
(
D
1
0.8
0.6
0.4
0.2
0
0
3 APs
11 APs
20
40
60
t
80
100
120
140
Figure 3. An integrative figure showing the predictive model (upper panel) and sharp-electrode electrophysiology
recordings of a VD4/LPeD1 synaptic pair (two lower panels). While three action potentials triggered during tetanic
stimulation are insufficient to result in potentiation of a subsequent excitatory postsynaptic potential (EPSP) in
the LPeD1 neuron, eleven action potentials elicited during tetanic stimulation result in a potentiated response, as
predicted by the model (after Ref. [89]).
Figure 3 shows a quantitative comparison between the theoretical predictions of [82]
(upper panel) with sharp-electrode electrophysiology recordings of a VD4/LPeD1 synap-
tic pair (two lower panels) [89]. The black theoretical curve corresponds to 3 APs triggered
during tetanic stimulation, which are insufficient to result in potentiation of a subsequent
excitatory postsynaptic potential (EPSP) in the LPeD1 neuron (T1 = 3 ≪ T0, so that
the switch remains off). The red theoretical curve corresponds to 11 APs, resulting in
a potentiated subsequent response (T1 = 11 ≫ T0, so the switch is turned on and the
synapse is frozen). The model biological switch used to model the action of kinase in [89]
displays an essential bistability so that the phenomenon described above is observed more
or less frequently depending on the difference between the duration T1 of the initial LTP
signal and the characteristic time T0 of the switch.
Thus, despite its seeming abstraction, the basic ideas of Fusi's cascade model can
indeed be related to real experimental data; in fact, such complex synaptic architectures
provide fertile ground for the inclusion of multiple timescales which are essential to the
modelling of long-term memory.
13
July 24, 2018
0:58
Advances in Physics
review
5. Synaptic dynamics: the need for competition
In the models of the preceding section, while synapses have been central to the acquisition
and recall of long-term memory, there has been no mention of their embedding networks,
in particular to do with the neurons that synapses connect. In this section we return to
the concepts of Section 2, and to the explicit mechanisms of synaptic strengthening and
weakening that result from neuronal firing within a network. We have already discussed
in Section 3 several phenomenological models of synaptic plasticity, where the need for
competitive dynamics has been made clear. In the following, we elaborate on several
ways in which these have been implemented in the neuroscience literature.
In the following, we follow the lines of argument of Van Ooyen's excellent review article
on synaptic competition [90], where a distinction is first made between independent and
interdependent competition. In interdependent competition, victors emerge as a result of
interactions between participants, such as in a sporting event. Interdependent competi-
tion is frequently considered, for example, in population biology; here, two species are
said to compete if they try to limit the growth of each others' population. In independent
competition, on the other hand, the participants do not interact, but are rather chosen on
the basis of some sort of contest. This kind of competition is reminiscent of competitive
learning which was introduced by Kohonen [91], and which will form the basis of the rest
of this article.
In neural network models based on competitive learning, only synapses connected to
the neurons most responsive to stimuli have their strengths changed. What is implicit here
is that these stimuli come from presynaptic neurons so that their correlated transmission
to postsynaptic neurons causes the corresponding synapses to be strengthened [92]. Such
synaptic competition [31] often arises through Hebbian learning so that when the synaptic
strength of one input grows, the strength of the others shrinks. Whereas many models
phenomenologically enforce competition by requiring the total strength of all synapses
onto a postsynaptic cell to remain constant [41], others implement biochemical processes
and modified Hebbian learning rules.
To see how competition between input connections can be enforced, consider n inputs,
with synaptic strengths Ji(t)(i = 1, . . ., n), impinging on a given postsynaptic cell at
time t. Simple Hebbian rules for the change ∆Ji(t) in synaptic strength in a time interval
∆t state that the synaptic strength should grow in proportion to the product of the
postsynaptic activity level y(t) and the activity level xi(t) of the ith input. Thus
∆Ji(t) ∝ y(t)xi(t)∆t .
(16)
If two inputs activate a common target, one needs competition to make one of the
synaptic strengths grow at the expense of the other. A common method to achieve
this is to constrain the total synaptic strength via synaptic normalisation -- this is the
constraint that
J p
i (t) = K,
n
Xi
(17)
with K constant and the integer p usually taken to be 1 or 2. Specifically, p = 1 conserves
the total synaptic strength, whereas p = 2 conserves the length of the weight vector. At
each time interval ∆t, following a phase of Hebbian learning, in which Ji(t + ∆) = Ji(t)+
∆Ji(t), the new synaptic strengths are forced to satisfy the normalization constraint of
Equation (17). Typically this can be enforced by one of two processes: multiplicative
14
July 24, 2018
0:58
Advances in Physics
review
or subtractive normalisation. These ensure that synaptic strengths do not grow without
bounds.
In subtractive normalization [43, 93], the same amount is subtracted from each weight
to enforce the constraint. In multiplicative normalization [94 -- 97] on the other hand, each
synaptic weight Ji(t + ∆t) is scaled in proportion to its size. A two-layer model is there
proposed, where the stimuli in neurons of the input layer are sent to an output layer of
neurons. If the neuronal inputs are above some specified threshold, then the responses in
the output layer are calculated, taking into account the pattern of synaptic connections;
weights are updated by a Hebbian rule after this neuronal activity stabilises. The final
outcome of development may of course differ depending on whether multiplicative or
subtractive normalization is used [12, 98].
Kohonen [91] proposed a drastic but effective simplification of the approach of [94].
In the latter, a few hotspots of activity typically emerged in the output layer following
the iterations of the input activity via the lateral synapses. To obviate the considerable
time taken to ensure the convergence of these iterations, Kohonen proposed the cen-
tering of the activity in the output layer on the so-called 'winning' neurons, followed
by standard Hebbian learning. This important simplification is vital to the statistical
physics approaches that will be presented in Section 6.1. Another way of viewing this
is to regard it as yet another nonlinear approach to competitive learning; if the layer of
output neurons is assumed to be connected by inhibitory synapses, the neuron with the
largest initial activity can be said to suppress the activity of all other output neurons.
The competitive approaches described in the above paragraphs are often described as
hard, in the sense of being 'winner-take-all'. In soft competitive learning, all neurons in
the output layer are updated by an amount that takes into account both their feed-
forward activation and the activity of other output neurons. This will also be seen to
have equivalences with agent-based learning models in the statistical physics approaches
of Section 6.1.
Another approach for achieving competition is to modify the simple Hebbian learning
rule of Equation (16) so that both increases in synaptic strength (LTP) and decreases
in synaptic strength (LTD) can take place. If we assume that the presynaptic activity
level xi(t) as well the postsynaptic activity level y(t) must be above some thresholds,
respectively θx, θy, to achieve LTP (and otherwise yield LTD), then a suitable synaptic
modification rule is [41]
∆Ji(t) ∝ [y(t) − θy][xi(t) − θx]∆t .
(18)
Thus, if both y(t) and xi(t) are above their respective thresholds, LTP occurs; if one
is below its threshold and the other is above, LTD occurs. For this to qualify as proper
competition, the synaptic strength lost through LTD must roughly equal the strength
gained through LTP. This can only be achieved with appropriate input correlations,
which makes simple LTD a fragile mechanism for achieving competition [41]. Another
mechanism which ensures that when some synaptic strengths increase, others must cor-
respondingly decrease - so that competition occurs - is to make one of the thresholds
variable. If the threshold θi
x increases sufficiently as the postsynaptic activity y(t) or
synaptic strength Ji(t) increases, conservation of synaptic strength is achievable [41].
Similarly, if the threshold θy increases faster than linearly with the average postsynap-
tic activity, then the synaptic strengths will adjust to keep the postsynaptic activity
near a limiting value [48]. This, however, results in temporal competition between input
patterns, rather than spatial competition between different sets of synapses.
So far, causal links between seemingly correlated firings of neurons have been as-
15
July 24, 2018
0:58
Advances in Physics
review
sumed. As before, spike-time dependent plasticity (cf. Section 3.2) makes this explicit
via its emphasis on the the relative timing of pre- and post-synaptic activity. In the
approach of [99], presynaptic activity that precedes postsynaptic spikes strengthens a
synapse, whereas presynaptic activity that follows postsynaptic spikes, weakens it. As a
consequence of the intrinsic nonlinearity of the spike generation mechanisms, and with
the imposition of hard limits on synaptic strengths, STDP has the effect of keeping the
total synaptic input to the neuron roughly constant, independent of the presynaptic fir-
ing rates. This approach, of rewarding truly correlated neuronal activity while penalising
its absence, has been taken into account in the models of synaptic dynamics presented
in Section 6.2.
6. Statistical physics models of competing synapses
The emergence of new areas in physics has strongly contributed to the development of
analytical tools; this is particularly true for the field of complex systems. A particular
area which is of relevance in the context of this review is that of agent-based modelling;
here, local interactions among agents may give rise to emergent phenomena on a macro-
scopic scale [100]. In these models, agents on the sites of appropriately defined lattices
interact with each other; their collective behaviour is then analysed in terms of global
outcomes. A typical example arises in, say, the context of financial markets; trading rules
between different agents at an individual level can result in specific sets of traders, or
their representative strategies, winning over their competitors. This makes for interesting
analogies with competitive learning; approaches based on this have therefore successfully
been used to investigate a wide variety of topics, ranging from the diffusion of innova-
tions [101, 102] through gap junction connectivity in the pancreas [103] to the dynamics
of competing synapses [104 -- 106]. It is the latter which will concern us here, but in the
interests of completeness, we first briefly review an agent-based model of competitive
learning in the following [101].
6.1. An agent-based model of competitive learning
The underlying idea [101] is that the strategy of a given agent is to a large extent
determined by what the other agents are doing, through considerations of the relative
payoffs obtainable in each case. Agents are located at the sites of a regular lattice, and
can be associated with one of two types of strategies. Every agent revises its choice of
type at regular intervals, and in this it is guided by two rules: a majority rule, reflecting
the tendency of agents to align with their local neighborhood, followed by an adaptive
performance-based rule, via which the agent chooses the type that is more successful
locally.
Assuming that the agents sit at the nodes of a d-dimensional regular lattice with
coordination number z = 2d, the efficiency of an agent at site i is represented by an Ising
spin variable:
ηi(t) =(+1
−1
if i is + at time t,
if i is − at time t.
(19)
The evolution dynamics of the lattice is governed by two rules. The first is a majority
rule, which consists of the alignment of an agent with the local field (created by its
16
July 24, 2018
0:58
Advances in Physics
review
nearest neighbours) acting upon it, according to:
ηi(t + τ1) =
Here, the local field
+1
±1
−1
w.p.
1
2
if
if
if
hi(t) > 0,
hi(t) = 0,
hi(t) < 0.
hi(t) =Xj(i)
ηj(t)
(20)
(21)
is the sum of the efficiencies of the z neighbouring agents j of site i and τ1 is the associated
time step. Next, a performance rule is applied. This starts with the assignment of an
outcome σi (another Ising-like variable, with values of ±1 corresponding to success and
failure respectively) to each site i, according to the following rules:
if ηi(t) = +1,
then
if ηi(t) = −1,
then
−1 w.p.
σi(t + τ2) =(cid:26) +1 w.p.
σi(t + τ2) =(cid:26) +1 w.p.
−1 w.p.
p+,
1 − p+,
p−,
1 − p−,
(22)
where τ2 is the associated time step and p± are the probabilities of having a successful
outcome for the corresponding strategy. With N +
i denoting the total number of
neighbours of a site i who have adopted strategies + and − respectively, and I +
(I −
i )
i
denoting the number of successful outcomes within the set N +
i ), the dynamical
i
rules for site i are:
i and N −
(N −
if ηi(t) = +1
and
I +
i (t)
N +
i (t)
< I −
N −
i (t)
i (t)
,
then
ηi(t + τ3) =(cid:26) −1 w.p. ε+
+1 w.p. 1 − ε+,
< I +
N +
i (t)
i (t)
if ηi(t) = −1
and
i (t)
I −
N −
i (t)
,
then
ηi(t + τ3) =(cid:26) +1 w.p. ε−
−1 w.p. 1 − ε−.
(23)
Here, the ratios Ii(t)
Ni(t) are nothing but the average payoff assigned by an agent to each of
the two strategies in its neighbourhood at time t (assuming that success yields a payoff
of unity and failure, zero). Also, τ3 is the associated time step and the parameters ε±
are indicators of the memory associated with each strategy. In their full generality, ε
and p are independent variables: the choice of a particular strategy can be associated
with either a short or a long memory.
Setting the timescales
the above steps of the performance rule are recast as effective dynamical rules involving
τ2 → 0,
τ1 = τ3 = 1,
(24)
17
July 24, 2018
0:58
Advances in Physics
review
the efficiencies ηi (t) and the associated local fields alone:
if ηi(t) = +1,
if ηi(t) = −1,
−1 w.p.
1-w+[hi(t)],
then ηi(t + 1) =(cid:26) +1 w.p. w+[hi(t)]
then ηi(t + 1) =(cid:26) +1 w.p. w−[hi(t)]
−1 w.p.
1-w−[hi(t)].
(25)
The effective transition probabilities w±(h) are evaluated by enumerating the 2z possible
realizations of the outcomes σj of the sites neighbouring site i, and weighting them ap-
propriately. The specific transition probabilities computed will depend on the embedding
lattice of network chosen [101].
The above rules are appropriate for cases where the majority rule is clearly definable,
i.e., where there is a mix of agent types. The situation is less clear when there are large
areas of a single species, since then, at least with a sequential update, there is a tendency
for any exceptions to revert to the majority type, whatever their performance. The way
around this in [101] was to formulate a so-called 'cooperative' model, where, say, a more
successful agent surrounded by neighbours who had failed, was able to convert all of
them to the more successful type, thus stabilising his own success. This hard rule is like
the 'winner-takes-all' model of synaptic competition alluded to earlier in this review;
analogously to that case, there is also a soft rule, where, while a significant majority of
agents were coerced into changing their type, not all were so obliged. In [101] all these
models were explored via ordered sequential updates of the agents, and phase diagrams
of their extremely different dynamical behaviour in various regimes were presented. The
agents were there also deemed to be memoryless, i.e., they did not take earlier results
into account when they made their choices. These restrictions were progressively removed
in [107, 108], so that the behaviour of the model with different updates, different levels
of memory, as well as different interactions was explored.
6.2. A minimal model of synaptic dynamics with emergent long-term
memory
The diligent reader will have noted the resemblance between Equations (19) and (23)
above, and some of the equations governing neuronal and synaptic dynamics earlier
presented in this paper. Indeed, the detailing of the agent-based model of competitive
learning [101] was to motivate just such a comparison. For example, neuronal firings are
subject to the kind of local field embodied by Equation (21); the performance in both
cases (successful neuronal firings and successful outcomes in the model of [101]) in turn
lead to other dynamical changes, and result in global outcomes. These were precisely the
lines of thought that led to the use of such agent-based models of competitive learning in
some of the early, and somewhat simple-minded, models of synaptic dynamics [104, 105].
Let us now recall what is needed for a minimal model of memory, via synaptic dynam-
ics. Both cooperation and competition are needed for a meaningful model of synaptic
plasticity [41], with competition acting as a check on the unstable growth of synaptic
weights when cooperation alone is invoked [30, 109]. Since synapses have finite stor-
age capacities, one should also include a representation of the spontaneous relaxation
of synapses when space is created via the spontaneous decay of old memories (cf. the
palimpsest effect [28, 29]). This is indeed what is done in the model network of synapses
18
July 24, 2018
0:58
Advances in Physics
review
and neurons [106] that we will describe in the following. Like the Fusi [79] model, it
is a model of discrete rather than continuous synapses; unlike it, however, here, there
are explicit mechanisms of synaptic weight change via mechanisms of competing and
cooperating synapses that depend intimately on neuronal firing rates.
The dynamical regime chosen in [106] is that of slow synaptic dynamics, where neuronal
firings are considered stochastic and instantaneous; the synapses 'see' only the mean firing
rates of individual neurons, characterising them as active or inactive, on that basis. As a
result of this temporal coarse-graining, the overall effect of the microscopic noise can be
represented by spontaneous relaxation rates from one type of synaptic strength to the
other, so far as the palimpsest mechanism is concerned. Cooperation between synapses
is incorporated via the usual Hebbian viewpoint, while the most crucial and original
part of the formalism involves synaptic competition where, along the lines of Kohonen's
arguments [91], synapses are converted to the type most responsible for neural activity
in their neighbourhood [104, 105].
The choice of basis is that of a fully connected network, as depicted in Figure 4, so
that mean-field theory applies in the thermodynamic limit of an infinitely large network.
ν
i
σ
ij
ν
j
Figure 4. The fully connected network for N = 4. Neurons with activities νi = 0, 1 live on the nodes. Synapses
with strength types σij = ±1 live on the bonds (after Ref. [106]).
Neurons live on the nodes (sites) of the network, labelled i = 1, . . . , N . The activity
state of neuron i at time t is described by a binary activity variable:
νi(t) =(1
0
if i is active
at time t,
if i is inactive at time t.
(26)
Active neurons are those whose instantaneous firing rate exceeds some threshold.
Synapses live on the undirected bonds of the network. The synapse (ij) lives on the
bond joining nodes i and j. The strength Jij of synapse (ij) at time t is also described
by a binary variable:
σij(t) =(+1 if (ij) is strong at time t,
−1 if (ij) is weak at time t.
(27)
Strong synapses are those whose strength Jij(t) exceeds some threshold.
19
July 24, 2018
0:58
Advances in Physics
review
Neuronal dynamics
Neurons have an instantaneous stochastic response to their environment. The activity of
neuron i at time t reads
νi(t) =(1 w.p. F (hi(t)),
0 w.p. 1 − F (hi(t)),
(28)
where F (h) is an increasing response function of the input field hi(t). The latter is a
weighted sum of the instantaneous activities of all other neurons:
hi(t) =
1
N − 1Xj6=i
(a + bσij(t))νj(t).
(29)
Strong synapses (σij = 1) enter the sum through a synaptic weight a + b, while weak
ones (σij = −1) have a synaptic weight a − b. We assume a and b are constant all over
the network. All synapses are therefore excitatory for b > 0, and inhibitory for b < 0.
In the following, we focus our attention onto the slow plasticity dynamics of the synap-
tic strength variables σij(t). It will therefore be sufficient to consider the mean activities
νi(t) and the mean input field hi(t), defined by averaging over a time window which is
large w.r.t. the characteristic time scale of neuron firings, but short w.r.t. that of synaptic
dynamics. These mean quantities obey
and
νi(t) = F (hi(t))
hi(t) =
1
N − 1Xj6=i
(a + bσij(t))ν j(t).
(30)
(31)
In most of this work we shall consider a spatially homogeneous situation in the ther-
modynamic limit of a large network. In this case the key quantity is the mean synaptic
strength
J(t) =
2
N (N − 1)X(ij)
σij(t),
(32)
which does not fluctuate anymore. The mean neuronal activity ν(t) and the mean input
field h(t) are related to J(t) by the coupled non-linear equations
and
ν(t) = F (h(t))
h(t) = (a + bJ(t))ν(t).
(33)
(34)
Consider first the case where there are as many strong and weak synapses, so that
the mean synaptic strength vanishes (J = 0). We have then h = aν, so that the mean
neuronal activity ν obeys ν = F (aν). We assume that the solution to that equation is
20
July 24, 2018
0:58
Advances in Physics
review
ν = 1
2 , meaning that there are also as many active and inactive neurons on average. We
further simplify the problem by linearising the coupled equations (33), (34) around this
symmetric fixed point. We thus obtain the following expression:
ν(t) = f (J(t)) = 1
2 (1 + εJ(t)).
The slope of the effective response function,
ε =
bF ′( a
2 )
1 − aF ′( a
2 )
,
(35)
(36)
is one of the key parameters of the model.3 It has to obey ε < 1. It is positive in the
excitatory case (b > 0), so that f (J) is an increasing function of J, and negative in the
inhibitory case (b < 0), so that f (J) is a decreasing function of J.
Synaptic plasticity dynamics
Synaptic strengths evolve very slowly in time, compared to the fast time scale of the
firing rates of neurons. It is therefore natural to model synaptic dynamics as a stochastic
process in continuous time [110], defined in terms of effective jump rates between the
two values (strong or weak) of the synaptic strength.
The model includes the following three plasticity mechanisms which drive synaptic
evolution:
1. Spontaneous relaxation mechanism. Synapses may spontaneously change their
strength type, either from weak to strong (potentiation) or from strong to weak
(depression) as a result of noise This spontaneous relaxation mechanism, illustrated
in Figure 5, translates into
(σij = −1 → +1 with rate Ω,
σij = +1 → −1 with rate ω.
(37)
−1
(weak)
Ω
ω
+1
(strong)
Figure 5. The spontaneous relaxation plasticity mechanism, with its potentiation rate Ω and depression rate ω
(after Ref. [106]).
2. Hebbian mechanism. When two neurons are in the same state of (in)activity, the
synapse which connects them strengthens; when one of the neurons is active and the
other is not, the interconnecting synapse weakens. This is the well-known Hebbian
mechanism [42], which we implement as follows:
(νi(t) = νj(t) : σij = −1 → +1 with rate α,
νi(t) 6= νj(t) : σij = +1 → −1 with rate α.
(38)
3Here and throughout the following, primes denote derivatives.
21
July 24, 2018
0:58
Advances in Physics
review
3. Polarity mechanism. This is a mechanism to introduce synaptic competition, in-
troduced for the first time in [104, 105], which converts a given synapse to the
type of its most 'successful' neighbours, i.e., those which augment the firing of an
intermediate neuron. Thus: if a synapse (ij) connects two neurons with different
activities at time t, e.g. νi(t) = +1 and νj(t) = −1, it will adapt its strength to
that of a randomly selected synapse (ik) connected to the active neuron i. If the
selected synapse is strong, the update σij = −1 → +1 takes place with rate β; if it
is weak, the update σij = +1 → −1 takes place with rate γ. Therefore:
(σij = −1 → +1 with rate 1
σij = +1 → −1 with rate 1
2 β(1 + J(t)),
2 γ(1 − J(t)).
(39)
Mean-field dynamics
For a spatially homogeneous situation in the thermodynamic limit, the mean synaptic
strength J(t) obeys a nonlinear dynamical mean-field equation of the form
dJ
dt
= P (J).
(40)
The explicit form of the rate function P (J) is obtained by summing the contributions of
the above three plasticity mechanisms. In the most general situation, the model has five
parameters: the slope ε of the effective response function (35) and the rates involved in the
three plasticity mechanisms. The resulting rate function is a polynomial of degree 4 [106]:
P (J) = p4J 4 + p2J 2 − (Ω + ω + α)J + Ω − ω − δ,
with
p4 = −δε2,
p2 = (α + δ)ε2 + δ,
δ = 1
4 (γ − β).
(41)
(42)
The spontaneous relaxation mechanism yields a linear rate function, while the Hebbian
mechanism is responsible for a quadratic non-linearity and the polarity-driven compet-
itive mechanism is responsible for a quartic non-linearity. This modelling of synaptic
competition satisfies the requirement on nonlinearity set out in Section 5 for meaningful
synaptic dynamics.
The parameter ε only enters (42) through its square ε2. The model therefore exhibits an
exact symmetry between the excitatory case (ε > 0) and the inhibitory one (ε < 0). (Since
none of the plasticity mechanisms distinguishes between these two cases, this symmetry
is to be expected). More generally, the model is invariant if the effective response function
f (J) is changed into 1 − f (J).
Generic dynamics
The rate function P (J) has an odd number of zeros in the interval −1 < J < +1 (counted
with multiplicities), i.e., either one or three. These zeros correspond to fixed points of
the dynamics. As a consequence, the model exhibits two generic dynamical regimes, as
shown in Figure 6.
In Regime I (see Figure 6, left), there is a single attractive (stable) fixed point at J0.
The mean synaptic strength J(t) therefore converges exponentially fast to this unique
22
July 24, 2018
0:58
Advances in Physics
review
P(J)
P(J)
Regime I
+1
J0
−1
Regime II
+1
J
−1
J1
J3
J2
J
Figure 6. The two possible generic dynamical regimes. Left: Regime I (one single attractive fixed point, J0).
Right: Regime II (two attractive fixed points, J1 and J2, and an intermediate repulsive one, J3) (after Ref. [106]).
fixed point, irrespective of its initial value, according to
J(t) − J0 ∼ e−t/τ0 .
The corresponding relaxation time τ0 reads
τ0 = −
1
P ′(J0)
.
(43)
(44)
where τ0 and J0 are obtainable in terms of the model parameters [106].
In Regime II (see Figure 6, right), there are two attractive (stable) fixed points at J1
and J2, and an intermediate repulsive (unstable) one at J3. The mean synaptic strength
J(t) converges exponentially fast to either of the attractive fixed points, depending on
its initial value, namely to J1 if −1 < J(0) < J3 and to J2 if J3 < J(0) < +1. The
corresponding relaxation times read
τ1 = −
1
P ′(J1)
,
τ2 = −
1
P ′(J2)
.
(45)
In other words, Regime II allows for the coexistence of two separate fixed points, leading
to network configurations which are composed of largely strong/weak synapses. In fact, it
is the polarity-driven competitive mechanism which gives rise to the quartic non-linearity,
essential for such coexistence.
Critical dynamics
When two of the three fixed points merge at some Jc, the dynamical system (40) exhibits
a saddle-node bifurcation. In physical terms, the dynamics become critical. We have then
P (Jc) = P ′(Jc) = 0,
(46)
so that the critical synaptic strength Jc is a double zero of the rate function P (J) (see
Figure 7). There is a left critical case, where J1 = J3 = J (L)
, while J2 remains non-
critical, and a right one, where J2 = J3 = J (R)
, while J1 remains non-critical. The
critical synaptic strength obeys Jc > 1
3 [106]. We thus conclude that the critical point
is always strengthening, as Jc is always larger then the 'natural' initial value J(0) = 0,
corresponding to a random mixture of strong and weak synapses in equal proportions.
c
c
23
July 24, 2018
0:58
Advances in Physics
review
P(J)
Left critical
P(J)
Right critical
−1
Jc
+1
J2
J
−1
J1
Jc
+1
J
Figure 7. The two possible kinds of critical dynamical behaviour: left critical case (J1 = J3 = J (L)
critical case (J2 = J3 = J (R)
) (after Ref. [106]).
c
c
) and right
The mean synaptic strength exhibits a universal power-law relaxation to its critical
value, of the form
J(t) − Jc ≈
Ac
t
.
(47)
The asymptotic 1/t relaxation law (47) holds irrespective of the initial value J(0), pro-
vided it is on the attractive side of the critical point, i.e., −1 < J(0) < Jc in the left
critical case (where Ac < 0), or Jc < J(0) < +1 in the right critical one (where Ac > 0).
To sum up, the non-critical fixed points of Regimes I or II are characterised by expo-
nential relaxation; the corresponding relaxation times, whether long or short, are always
finite. Anywhere along the critical manifold, on the other hand, one observes a universal
power-law relaxation in 1/t. Such behaviour corresponds to an infinite relaxation time
at least in terms of the mean synaptic strength J.
In conclusion, this minimal model is able to show the emergence of power-law relaxation
or long-term memory. It is clear that the most crucial one of these is the mechanism of
synaptic competition, which is in reassuring accord with the importance given to such
competition by neuroscientists [31] (Section 5). Purely analytical work is able, however,
just to give a flavour of the emergence of long-term behaviour in this model via the
critical behaviour of the mean synaptic strength J. If realistic learning and forgetting of
patterns are to be implemented with this model, considerable computational work needs
to be done. Only the identification of the parameter spaces where criticality is obtained
in response to random input patterns will clarify, at least phenomenologically, the routes
to long-term memory in this relatively minimal model.
7. Discussion
Even quantitative approaches to the subject of memory are truly interdisciplinary; con-
tributions range from mathematical psychology through quantitative neuroscience to
statistical physics. The narrowing of focus to physics still provides a huge range of con-
tributions: from the seminal contributions on Hopfield networks with their spin-glass
analogies, through the emphasis on causality with spiking neurons, both of which in-
volve fast neuronal dynamics, to the synaptic-dynamics-centred approaches that have
followed, with the boundedness of synaptic weights on discrete synapses, involving mul-
tiple 'hidden' synaptic states, as well as the attribution of competitive and cooperative
dynamics to synapses in model networks. In this review, we have sought to highlight
24
July 24, 2018
0:58
Advances in Physics
review
those approaches which generate long-term memory; while short-term memory, charac-
terised by exponential relaxation times, is ubiquitous, long-term memory is characterised
by power-law forgetting, a much slower process.
Another emphasis of this review is on synaptic competition, whose importance has long
been understood by the neuroscience community, but which has only very recently been
explicitly included in model networks.This review has gone into as much detail in the
need for this mechanism, as its inclusion in biophysical as well as physics-based modelling.
In the latter case, the recent advent of agent-based modelling techniques derived from
game theory [111] and extended to cover nonequilibrium situations, has been particularly
useful.
What is still a matter of debate is the extent to which phenomenological models, on
which this review has focused, are useful in unravelling the phenomenon of memory
storage and recall. While it is certainly true that detailed biophysical models are overall
better in matching experimental data point by point, there is a great deal to be said in
favour of the formulation of minimal models. These can, unlike the former, at least benefit
from a few analytical insights, which can help both experimentalists and theorists identify
the parameters that are truly important in what are typically huge parameter spaces,
most recently believed to be in eleven dimensions [112]. While these large parameter
spaces are indeed inclusive by definition, their inner workings can only be described by
computer simulations, which do not always give unambiguous answers to the relative
importance of parameters, or answers to physical questions like, what are the crucial
mechanisms for memory storage? This is of course not to minimise their importance;
we wish only to underscore the complementarity of the insights obtained by minimal
physical models to the enigma of memory.
Acknowledgements
AM is very grateful to Dr. Jean-Marc Luck for a careful reading of this manuscript.
She also thanks the Institut fur Informatik, Leipzig, and the University of Rome 'La
Sapienza' for their warm hospitality during the course of this work. This project has
received funding from the European Research Council (ERC) under the European Unions
Horizon 2020 Research and Innovation Programme (grant agreement N. 694925).
References
[1] E.R. Kandel, In Search of Memory: The Emergence of a New Science of Mind, W.W.
Norton and Company, New York, 2006.
[2] E.R. Kandel, J.H. Schwartz, T.M. Jessell, S.A. Siegelbaum, and A.J. Hudspeth, Principles
of Neural Science, McGraw-Hill, New York, 2012.
[3] H. Ebbinghaus, New York Teachers College 39 (1913), original work published 1885; reprint
of translation published by Dover, New York, 1964.
[4] L.R. Squire, Neurobiology of Learning and Memory 82 (2004), pp. 171 -- 177.
[5] J.G.W. Raaijmakers and R.M. Shiffrin, Models of memory, in Stevens' Handbook of Exper-
imental Psychology, Wiley Online Library, 2002.
[6] J.A. Anderson, J.W. Silverstein, S.A. Ritz, and R.S. Jones, Psychological Review 84 (1977),
pp. 413 -- 451.
[7] A. Baddeley, Nature Reviews Neuroscience 4 (2003), pp. 829 -- 839.
[8] A. Baddeley and G.J. Hitch, Working memory, in Recent advances in learning and moti-
vation, G.A. Bower, ed., Vol. 8, Academic, 1974, pp. 47 -- 89.
25
July 24, 2018
0:58
Advances in Physics
review
[9] J.S. Nairne, Annual Review of Psychology 53 (2002), pp. 53 -- 81.
[10] I. Neath and G.D. Brown, Frontiers in Psychology 3 (2012), pp. 1 -- 3.
[11] S.J. Martin and R.G.M. Morris, Hippocampus 12 (2002), pp. 609 -- 636.
[12] H.D. Simpson, D. Mortimer, and G.J. Goodhill, Current Topics in Developmental Biology
87 (2009), pp. 1 -- 51.
[13] W.C. Abraham and A. Robins, Trends in Neurosciences 28 (2005), pp. 73 -- 78.
[14] C.B. Kirwan, J.T. Wixted, and L.R. Squire, Journal of Neuroscience 28 (2008), pp. 10541 --
10548.
[15] C.N. Smith and L.R. Squire, Journal of Neuroscience 29 (2009), pp. 930 -- 938.
[16] K. Diba and G. Buzsaki, Nature Neuroscience 10 (2007), pp. 1241 -- 1242.
[17] W.B. Scoville and B. Milner, Journal of Neurology Neurosurgery and Psychiatry 20 (1957),
pp. 11 -- 21.
[18] T. Bliss and T. Lomo, Journal of Physiology 232 (1973), pp. 331 -- 356.
[19] C. Clopath, Cognitive Neurodynamics 6 (2012), pp. 251 -- 257.
[20] C. Clopath, L. Ziegler, E. Vasilaki, L. Busing, and W. Gerstner, PLoS Computational
Biology 4 (2008), p. e1000248.
[21] A.B. Barrett, G.O. Billings, R.G.M. Morris, and M.C.W. Van Rossum, PLoS Computa-
tional Biology 5 (2009), p. e1000259.
[22] U. Frey and R.G. Morris, Nature 385 (1997), pp. 533 -- 536.
[23] M. Graupner and N. Brunel, Front. Comput. Neurosci. 4 (2010), pp. 1 -- 19.
[24] E. Miyamoto, J. Pharmacol. Sci. 100 (2006), pp. 433 -- 42.
[25] T. Kitamura, S.K. Ogawa, D.S. Roy, T. Okuyama, M.D. Morrissey, L.M. Smith, R.L.
Redondo, and S. Tonegawa, Science 356 (2017), pp. 73 -- 78.
[26] M.C.W. Van Rossum and M. Shippi, Journal of Statistical Mechanics: Theory and Exper-
iment 2013 (2013), p. P03007.
[27] J.J. Hopfield, Proceedings of the National Academy of Sciences 79 (1982), pp. 2554 -- 2558.
[28] J.P. Nadal, G. Toulouse, J.P. Changeux, and S. Dehaene, Europhysics Letters 1 (1986), pp.
535 -- 542.
[29] G. Parisi, Journal of Physics A: Mathematical and General 19 (1986), pp. L617 -- L620.
[30] G.G. Turrigiano and S.B. Nelson, Nature Reviews Neuroscience 5 (2004), pp. 97 -- 107.
[31] A. Van Ooyen, Nature Reviews Neuroscience 12 (2011), pp. 311 -- 326.
[32] T. Takeuchi, A.J. Duszkiewicz, and R.G.M. Morris, Philosophical Transactions of the Royal
Society of London B: Biological Sciences 369 (2014), p. 20130288.
[33] D.J. Amit, Modeling Brain Function: The World of Attractor Neural Networks, Cambridge
University Press, Cambridge, 1989.
[34] P. Dayan and L.F. Abbott, Theoretical Neuroscience: Computational and Mathematical
Modeling of Neural Systems, MIT Press, Cambridge, 2001.
[35] W. Gerstner and W.M. Kistler, Spiking Neuron Models: Single Neurons, Populations, Plas-
ticity, Cambridge University Press, Cambridge, 2002.
[36] J.J. Hopfield, Proceedings of the National Academy of Sciences 81 (1984), pp. 3088 -- 3092.
[37] D.J. Amit, H. Gutfreund, and H. Sompolinsky, Physical Review Letters 55 (1985), pp.
1530 -- 1533.
[38] D.J. Amit, H. Gutfreund, and H. Sompolinsky, Physical Review A 35 (1987), pp. 2293 -- 2303.
[39] D.J. Amit, H. Gutfreund, and H. Sompolinsky, Annals of Physics 173 (1987), pp. 30 -- 67.
[40] L.F. Abbott, Neuron 60 (2008), pp. 489 -- 495.
[41] K.D. Miller, Neuron 17 (1996), pp. 371 -- 374.
[42] D. Hebb, The Organization of Behavior, Wiley, New York, 1949.
[43] K.D. Miller, J.B. Keller, and M.P. Stryker, Science 245 (1989), pp. 605 -- 615.
[44] L.F. Abbott, Network: Computation in Neural Systems 1 (1990), pp. 105 -- 122.
[45] H. Markram and B. Sakmann, Society for Neuroscience Abstracts 21 (1995), p. 2007.
[46] W. Gerstner, R. Kempter, J.L. van Hemmen, and H. Wagner, Nature 383 (1996), pp. 76 -- 78.
[47] E. Oja, Journal of Mathematical Biology 15 (1982), pp. 267 -- 273.
[48] E.L. Bienenstock, Journal of Neuroscience 2 (1982), pp. 32 -- 48.
[49] H. Markram, W. Gerstner, and P.J. Sjostrom, Frontiers in Synaptic Neuroscience 4 (2012),
26
July 24, 2018
0:58
Advances in Physics
review
p. 2.
[50] W. Gerstner and J.L. van Hemmen, Network: Computation in Neural Systems 3 (1992),
pp. 139 -- 164.
[51] W. Gerstner and R. Naud, Science 326 (2009), pp. 379 -- 380.
[52] J. Sjostrom and W. Gerstner, Scholarpedia 5 (2010), p. 1362, revision #151671.
[53] F. Zenke, E.J. Agnes, and W. Gerstner, Nature Communications 6 (2014), pp. 6922 -- 6922.
[54] R. Kempter, W. Gerstner, and J.L. Van Hemmen, Physical Review E 59 (1999), pp. 4498 --
4514.
[55] E.M. Izhikevich and N.S. Desai, Neural Computation 15 (2003), pp. 1511 -- 1523.
[56] N. Brunel and V. Hakim, Neural Computation 11 (1999), pp. 1621 -- 1671.
[57] A.M. Dubreuil and N. Brunel, Journal of Computational Neuroscience 40 (2016), pp. 157 --
175.
[58] J.M. Beggs and D. Plenz, Journal of Neuroscience 23 (2003), pp. 11167 -- 11177.
[59] S. Bornholdt and T. Rohl, Physical Review E 67 (2003), p. 066118.
[60] A. Roxin, H. Riecke, and S.A. Solla, Physical Review Letters 92 (2004), p. 198101.
[61] L. de Arcangelis, C. Perrone-Capano, and H.J. Herrmann, Physical Review Letters 96
(2006), p. 028107.
[62] L. de Arcangelis and H.J. Herrmann, Proceedings of the National Academy of Sciences 107
(2010), pp. 3977 -- 3981.
[63] A. Levina, J.M. Herrmann, and T. Geisel, Nature Physics 3 (2007), pp. 857 -- 860.
[64] A. Levina, J.M. Herrmann, and T. Geisel, Physical Review Letters 102 (2009), p. 118110.
[65] M. Uhlig, A. Levina, T. Geisel, and J. Herrmann, Frontiers in Computational Neuroscience
7 (2013), p. 87.
[66] T.J. Taylor, C. Hartley, P.L. Simon, I.Z. Kiss, and L. Berthouze, Journal of Mathematical
Neuroscience 3 (2013), p. 5.
[67] M.I. Rabinovich, P. Varona, A.I. Selverston, and H.D.I. Abarbanel, Reviews of Modern
Physics 78 (2006), pp. 1213 -- 1265.
[68] D.H. O'Connor, G.M. Wittenberg, and S.S.H. Wang, Proceedings of the National Academy
of Sciences 102 (2005), pp. 9679 -- 9684.
[69] F. Crick, Nature 312 (1984), p. 101.
[70] D.J. Amit and S. Fusi, Network: Computation in Neural Systems 3 (1992), pp. 443 -- 464.
[71] D.J. Amit and S. Fusi, Neural Computation 6 (1994), pp. 957 -- 982.
[72] D.J. Amit and N. Brunel, Network: Computation in Neural Systems 8 (1997), pp. 373 -- 404.
[73] S. Romani, D.J. Amit, and Y. Amit, Neural Computation 20 (2008), pp. 1928 -- 1950.
[74] A.B. Barrett and M.C.W. van Rossum, PLoS Computational Biology 4 (2008), p. e1000230.
[75] S. Fusi and W. Senn, Chaos 16 (2006), p. 026112.
[76] S. Fusi and L.F. Abbott, Nature Neuroscience 10 (2007), pp. 485 -- 493.
[77] C. Baldassi, A. Braunstein, N. Brunel, and R. Zecchina, Proceedings of the National
Academy of Sciences (2007), pp. 11079 -- 11084.
[78] S. Fusi, Biological Cybernetics 87 (2002), pp. 459 -- 470.
[79] S. Fusi, P.J. Drew, and L.F. Abbott, Neuron 45 (2005), pp. 599 -- 611.
[80] W.C. Abraham and M.F. Bear, Trends in Neurosciences 19 (1996), pp. 126 -- 130.
[81] T.M. Fischer, D.E.J. Blazis, N.A. Priver, and T.J. Carew, Nature 389 (1997), pp. 860 -- 865.
[82] A. Mehta and J.M. Luck, Journal of Statistical Mechanics: Theory and Experiment 2011
(2011), p. P09025.
[83] J.T. Wixted and E.B. Ebbesen, Psychological Science 2 (1991), pp. 409 -- 415.
[84] J.T. Wixted and E.B. Ebbesen, Memory and Cognition 25 (1997), pp. 731 -- 739.
[85] C.T. Kello, G.D.A. Brown, R. Ferrer-i-Cancho, J.G. Holden, K. Linkenkaer-Hansen,
T. Rhodes, and G.C. Van Orden, Trends in Cognitive Sciences 14 (2010), pp. 223 -- 232.
[86] L. Ziegler, F. Zenke, D.B. Kastner, and W. Gerstner, Journal of Neuroscience 35 (2015),
pp. 1319 -- 1334.
[87] M.K. Benna and S. Fusi, Nature Neuroscience 19 (2016), pp. 1698 -- 1706.
[88] C.C. Luk, H. Naruo, D. Prince, A. Hassan, S.A. Doran, J.I. Goldberg, and N.I. Syed,
European Journal of Neuroscience 34 (2011), pp. 569 -- 577.
27
July 24, 2018
0:58
Advances in Physics
review
[89] A. Mehta, J.M. Luck, C.C. Luk, and N.I. Syed, PLoS One 8 (2013), p. e78056.
[90] A.V. Ooyen, Network: Computation in Neural Systems 12 (2001), pp. 1 -- 47.
[91] T. Kohonen, Biological Cybernetics 43 (1982), pp. 59 -- 69.
[92] N. Swindale, Network: Computation in Neural Systems 7 (1996), pp. 161 -- 247.
[93] K.D. Miller and M.P. Stryker, The Development of Ocular Dominance Columns: Mecha-
nisms and Models, in Connectionist modeling and brain function: The developing interface,
chap. 9, MIT Press, Cambridge (1990), pp. 255 -- 350.
[94] C. von der Malsburg, Biological Cybernetics 14 (1973), pp. 85 -- 100.
[95] C. von der Malsburg and D.J. Willshaw, Experimental Brain Research 1 (1976), pp. 463 --
469.
[96] D.J. Willshaw and C. von der Malsburg, Proceedings of the Royal Society of London B 194
(1976), pp. 431 -- 445.
[97] D.J. Willshaw and C. von der Malsburg, Philosophical Transactions of the Royal Society
of London B: Biological Sciences 287 (1979), pp. 203 -- 243.
[98] K.D. Miller and D.J.C. MacKay, Neural Computation 6 (1994), pp. 100 -- 126.
[99] S. Song, K.D. Miller, and L.F. Abbott, Nature Neuroscience 3 (2000), pp. 919 -- 926.
[100] C. Castellano, S. Fortunato, and V. Loreto, Reviews of Modern Physics 81 (2009), pp.
591 -- 646.
[101] A. Mehta and J.M. Luck, Physical Review E 60 (1999), pp. 5218 -- 5230.
[102] K. Chatterjee and S.H. Xu, Advances in Applied Probability 36 (2004), pp. 355 -- 376.
[103] P. Goel and A. Mehta, PLoS One 8 (2013), p. e70366.
[104] G. Mahajan and A. Mehta, Europhysics Letters 95 (2011), p. 48008.
[105] A.A. Bhat, G. Mahajan, and A. Mehta, PloS One 6 (2011), p. e25048.
[106] J.M. Luck and A. Mehta, Physical Review E 90 (2014), p. 032709.
[107] A.A. Bhat and A. Mehta, Physical Review E 85 (2012), p. 011134.
[108] G. Mahajan and A. Mehta, Theory in Biosciences 129 (2010), pp. 271 -- 282.
[109] G.G. Turrigiano, K.R. Leslie, N.S. Desai, L.C. Rutherford, and S.B. Nelson, Nature 391
(1998), pp. 892 -- 896.
[110] N.G. Van Kampen, Stochastic Processes in Physics and Chemistry, North-Holland, 1992.
[111] J. von Neumann and O. Oskar Morgenstern, Theory of Games and Economic Behavior,
Princeton University Press, Princeton, 2007.
[112] M.W. Reimann, M. Nolte, M. Scolamiero, K. Turner, R. Perin, G. Chindemi, P. Dlotko,
R. Levi, K. Hess, and H. Markram, Frontiers in Computational Neuroscience 11 (2017),
pp. 1 -- 16.
28
|
1806.01597 | 1 | 1806 | 2018-06-05T10:22:05 | Brain synchronizability, a false friend | [
"q-bio.NC",
"nlin.AO",
"physics.bio-ph",
"physics.soc-ph"
] | Synchronization plays a fundamental role in healthy cognitive and motor function. However, how synchronization depends on the interplay between local dynamics, coupling and topology and how prone to synchronization a network with given topological organization is are still poorly understood issues. To investigate the synchronizability of both anatomical and functional brain networks various studies resorted to the Master Stability Function (MSF) formalism, an elegant tool which allows analysing the stability of synchronous states in a dynamical system consisting of many coupled oscillators. Here, we argue that brain dynamics does not fulfil the formal criteria under which synchronizability is usually quantified and, perhaps more importantly, what this measure itself quantifies refers to a global dynamical condition that never holds in the brain (not even in the most pathological conditions), and therefore no neurophysiological conclusions should be drawn based on it. We discuss the meaning of synchronizability and its applicability to neuroscience and propose alternative ways to quantify brain networks synchronization. | q-bio.NC | q-bio | Brain synchronizability, a false friend
D. Papo1* and J.M. Buldú2,3
1SCALab UMR CNRS 9193, Université de Lille, Villeneuve d'Ascq, France
2Laboratory of Biological Networks, Center for Biomedical Technology (UPM), 28223, Pozuelo de Alarcón, Madrid, Spain
3Complex Systems Group & G.I.S.C., Universidad Rey Juan Carlos, 28933 Móstoles, Madrid, Spain
* Email : [email protected]
Synchronization plays a fundamental role in healthy cognitive and
motor function. However, how synchronization depends on the
interplay between local dynamics, coupling and topology and how
prone to synchronization a network with given topological
organization is are still poorly understood issues. To investigate
the synchronizability of both anatomical and functional brain
networks various studies resorted to the Master Stability Function
(MSF) formalism, an elegant tool which allows analysing the
stability of synchronous states in a dynamical system consisting of
many coupled oscillators. Here, we argue that brain dynamics does
not fulfil the formal criteria under which synchronizability is
usually quantified and, perhaps more importantly, what this
measure itself quantifies refers to a global dynamical condition
that never holds in the brain (not even in the most pathological
conditions), and therefore no neurophysiological conclusions
should be drawn based on it. We discuss the meaning of
synchronizability and its applicability to neuroscience and propose
alternative ways to quantify brain networks synchronization.
the synchronizability parameter, and proposed that during the
course of development human brain anatomy evolves towards
an organization that limits synchronizability [9]. The authors
suggested that as the brain evolves towards its mature state, it
reduces its ability to synchronize, while, at the same time claim
that this reduction
is a consequence of promoting the
controllability of the brain network. Furthermore, a few studies
focused on the effects of different pathologies on brain
synchronizability, such as epilepsy [14,18-23], Alzheimer's
disease [8,10], or schizophrenia [17], showing statistically
significant changes
in the synchronizability parameter in
association with these diseases. Interestingly, epilepsy was
associated with an increased synchronizability during interictal
activity [21], while it decreased during ictal activity [14].
Functional networks synchronizability has been reported to
(EEG) activity of
decrease
using
schizophrenic
magnetoencephalography (MEG) showed that synchronizability
values depend on the frequency band considered when
constructing functional networks [11].
the electroencephalographic
patients
Studies
[17].
Introduction
Consider a network in which each node is a dynamical system,
e.g. an oscillator, and the links are couplings between these
nodes. Can these oscillators synchronize with each other
creating a coherent state and, if so, under what circumstances is
this state stable? Given a particular dynamical system and
coupling scheme, the Master Stability Function (MSF) formalism
[1-3] allows relating the stability of the fully synchronized state
to the spectral properties of the underlying matrix of
connections, and assessing which network structures can
maintain complete synchronization of the whole network.
At
scales of
the macroscopic
typical non-invasive
neuroimaging techniques, brain activity can be thought of as the
collective dynamics of a set of coupled dynamical units.
Synchronization among these units has been suggested to be a
basic mechanism of healthy brain functioning [4]. Thus, at first
glance, the problem above may seem to apply to brain activity,
justifying the use of the MSF formalism to quantify brain
network synchronizability. But appropriate though they may
sometimes seem, formalisms are created to address very
specific questions and come with their own set of formal and
theoretical assumptions, the compliance with which ultimately
decides whether they can be used in a given context.
In the remainder, we argue that some essential characteristics
of the brain render the MSF framework difficult to apply to
neuroscience, review some misunderstandings about the
synchronizability construct and propose alternative ways to
understand synchronization in brain networks.
Brain synchronizability
The use of synchronizability,
initially designed to study
theoretical models, rapidly extended to the analysis of real
datasets and, in the context of neuroscience, to quantify the
ability of anatomical [5-9] and
functional [10-23] brain
networks to synchronize. For example, Tang and co-workers [9]
investigated how the human brain's anatomical organization
evolves from childhood to adulthood by measuring changes in
1
While changes in synchronizability clearly exist, is this
particular metric measuring what it is supposed to measure?
The Master Stability Function formalism
The meaning and scope of the synchronizability construct
should be understood in the MSF theoretical context it is
predicated upon.
Given a group of N coupled dynamical systems whose
dynamics in isolation follows (cid:1)(cid:2) = F(cid:6)(cid:7)x(cid:6)(cid:9), the evolution of the
whole system is given by the equation:
(cid:20)
(cid:1)(cid:2) (cid:6)(cid:7)(cid:10)(cid:9) = (cid:11)(cid:12)(cid:1)(cid:6)(cid:7)(cid:10)(cid:9)(cid:13) − (cid:15) ∑
(cid:18)(cid:21)(cid:22)
(cid:12)x(cid:18)(cid:13), (cid:24) = 1, … , (cid:27) [1]
(cid:17)(cid:6)(cid:18)(cid:19)
where (cid:1)(cid:6) is the n-dimensional state vector of the ith oscillator, (cid:15)
the coupling strength, (cid:19)(cid:7)x(cid:9) a vectorial output function and (cid:17)(cid:6)(cid:18)
the elements of the Laplacian matrix (cid:28) [24] describing how the
oscillators are coupled together. For identical systems with the
same coupling function (cid:19)(cid:7)x(cid:9), the synchronized state is a
solution of (cid:1)(cid:2) (cid:29) = (cid:11)(cid:7)(cid:1)(cid:30)(cid:9), with (cid:1)(cid:22) = (cid:1)(cid:31) = ⋯ = (cid:1)(cid:20) = (cid:1)(cid:29). A linear
stability analysis around the synchronization manifold allows to
obtain the MSF, !(cid:7)"(cid:9), where the independent variable " is
related to the non-zero eigenvalues (cid:1)# of the Laplacian matrix as
"# = σ(cid:1)# [1-3]. The MSF tells how dynamics through (cid:11) and
network topology through the second term on the right side of
equation [1] concur
in determining the stability of the
synchronization manifold. The term synchronizability refers to
the stability of the global synchronization state.
The synchronization manifold is stable when all "# associated
with the non-zero eigenvalues of the Laplacian matrix lie in a
region in which the MSF is negative. However, different
dynamical systems with different coupling functions lead to
qualitatively different MSFs (see Fig. 1a for details), which can
be classified as [2]: class I (always positive), class II (always
negative above a threshold) and class III (negative only within a
specific region). Interestingly, in the context of brain networks,
synchronizability is commonly evaluated as if the brain were a
class III system, although no proof of it exists. Thus, the lower
the ratio & between the
largest and smallest (non-zero)
eigenvalues (i.e. & = (cid:1)(cid:20) (cid:1)(cid:31)⁄ ), the more packed the eigenvalues of
the Laplacian are and the highest the ability to fall within a
window where the MSFs is always negative. In that sense, brain
networks' synchronizability is sometimes [9] quantified by the
dispersion of the eigenvalues of the Laplacian matrix (cid:28) [25].
Synchronizability: some common misconceptions
The meaning of synchronizability and the questions it allows
addressing are a matter of frequent confusion and numerous
misconceptions.
An important issue is whether synchronizability can be
measured when ignoring the characteristics of the dynamics.
Stability under perturbations exists when all eigenvalues of the
combinatorial Laplacian matrix ((cid:1)(cid:6)) fall within the region of
stability due to the fact that the coupling is strong enough to
guarantee that the MSF enters the region but weak enough to
guarantee that it does not leave this region from the other side.
Synchronizability is ultimately determined by the sign of the
MSF evaluated at points that are indeed given by the spectrum
of the Laplacian matrix and an overall coupling strength. The
functional form of the MSF crucially depends on the dynamics of
the coupled oscillators and the function that couples its state
variables to those of other oscillators [26,27]. Depending on the
shape of the MSF, dynamical systems may never synchronize,
always synchronize above a certain coupling strength or
synchronize only for coupling strength values within a certain
range [2,27]. While the MSF for various families of dynamical
systems is typically convex for generic oscillator systems, its
exact shape depends not only on the dynamical systems but also
on the kind of coupling between them. Thus, quantifying the
synchronizability of anatomical brain networks using a
parameter based on the eigenvalues of the Laplacian matrix
alone, without information about the underlying dynamical
oscillators and their coupling function and strength cannot
ensure that the whole system falls within MSF's stability region.
In other words, it is not the network structure per se that is
synchronizable, but the particular combination of dynamical
systems, coupling strength and network structure formed by the
connections between these systems. While the eigenvalues of
the Laplacian matrix
likely contain potentially valuable
information of some sort [8,10], eigenvalue dispersion of the
anatomical network alone without at least some information on
node dynamics cannot determine the system class one is dealing
with, and conclusions on its MSF are no more than guesses (see
Fig. 1a).
Two related important questions are: what are high or low
synchronizability values telling us? When can synchronizability
values be compared? An early study using synchronizability [11]
reported that the synchronizability parameter for anatomical
brain networks was close to the region where a series of
theoretical models reached the synchronization manifold, based
on which the authors claimed that the brain is "located
dynamically on a critical point of the order/disorder transition"
[11]. However, the bare comparisons of synchronizability values
across dynamical systems and the characterization of a given
topology as being more or less synchronizable than another are
potentially problematic: insofar as different dynamical systems
haven't necessarily got similar MSFs, the synchronizability
parameter of a brain network cannot be compared with others
as long as its MSF is unknown.
to
the
frequent
relates
confusion
Perhaps at the root of most other ones, a major problematic
issue
between
synchronizability and synchronization. This is for instance
apparent in Tang's discussion of synchronizability [9]: "brain
networks […] do not fully limit synchronizability, perhaps because
some finite amount of synchronization is needed for dynamical
coordination and cognition" (p. 8). The synchronizability
parameter does not tell if the system is synchronized or not: a
system
synchronizable without being
synchronized, and synchronized with a low synchronizability
parameter (see Fig. 1b).
can be highly
Fig 1. Master Stability Function Ψ(ν) as a function of the parameter ν. " is
related to the (N–1) non-zero eigenvalues (cid:1)(cid:6) of the network Laplacian matrix as
"(cid:6) = (cid:15)(cid:1)(cid:6) where σ is the coupling strength. The synchronization manifold is stable
when all "(cid:6) lie in a region where the MSF is negative. MSFs can be classified as [2]:
class I (always positive), class II (always negative above a threshold "(cid:31)) and class
III (negative only within a specific region ["(cid:31), "(cid:20)*). The stability region may even
not be unique [27]. (a) Qualitative example of "(cid:6) of a network that would
synchronize class I and II dynamical systems, but not class III (assuming that
"(cid:20) "(cid:31)⁄ ). (b) Counter-example showing that defining synchronizability
(cid:1)(cid:20) (cid:1)(cid:31) +
parameter as the inverse of the dispersion of the eigenvalues can be misleading:
network A has lower dispersion (i.e., higher synchronizability) but lies in the
region of the MSF where the synchronization manifold is unstable, while network
B, has higher dispersion (and lower synchronizability) but can synchronize.
⁄
Finally, it is worth stressing that the synchronizability
construct only applies to anatomical networks. This is because
the MSF formalism relies on a structural property, i.e. the
connectivity pattern between dynamical units, which should
complemented by
the
construction of functional networks relies on the reported
coordination between brain regions, i.e. a dynamical property.
Therefore, functional networks are not the cause of a certain
level of synchronizability, but their consequence.
the coupling strength. However,
Why synchronizability should not be used (in neuroscience)
Even discounting
issues discussed above,
fundamental reasons make the MSF-based synchronizability
inapplicable to neuroscience.
technical
the
Crucially, in its original formulation [1], the MSF applies to
i.e. all
diffusively-coupled
interacting units of the network should have the same variables
and internal parameters. However, irrespective of the scale at
identical dynamical
systems,
2
it
the brain
is observed,
is dynamically highly
which
heterogeneous, ruling out an application of the MSF. While the
MSF formalism can be generalized to heterogeneous systems,
this comes at the price of rather restrictive conditions
hampering its application to brain data [26,28].
Perhaps the most fundamental obstacle to the use of the MSF
in brain sciences is represented by two issues related to the
definition of synchronization.
First, while various kinds of synchronization, including phase
[4], generalized [29], and relay synchronization [30] have been
reported for brain dynamics, and may even coexist [31],
synchronizability refers to a specific synchronization mode,
complete synchronization. Complete synchronization requires
that all dynamical units have exactly the same phase and
amplitude once the synchronization manifold is reached, a state
that has never been reported in the brain (not even in its most
pathological conditions).
Second, physics and neuroscience understand synchronization
in fundamentally different ways: in the former, synchronization
refers to a global and stable state, while in the latter to a local
and transient one. While local complete synchronization may be
a relevant mechanism or a reasonable modelling representation
of functionally segregated regions or circuits, its dynamics is
necessarily transient. Brain dynamics has in general a complex
phase space geometry, and possibly no stricto sensu attractor at
all [32,33], a scenario that cannot be dealt with using the MSF in
its current form.
Towards
synchronizability
neurophysiologically
plausible
alternatives
to
their
and
function.
to
[34-38].
technical
dynamics
represents,
Synchronizability
Several
systems, e.g. power grids, wireless
communication systems, require stable synchronization of their
units
good
approximation, a construct that can be used to model and
regulate
However,
synchronizability refers to a type of synchronization that the
brain does not, certainly should not, and possibly cannot achieve
in a stable way. In addition to being incompatible with the
dynamical and
functional heterogeneity of normal brain
functioning, a completely synchronized state represents a
considerable loss of complexity, and would likely be associated
with an unphysiological energetic cost [39,40].
Before figuring out possible alternatives to the MSF-based
synchronizability, one should perhaps address the following
question: why is the MSF framework used although it so
evidently at odds with neurophysiological stylized facts? What
makes the MSF a convenient tool? While a unique coupling
function for all network nodes and some hypotheses on the
coupling matrix are convenient mathematical conditions which
ensure the existence of an invariant set representing the
complete synchronization manifold and considerably simplify
the analysis of its stability, using steady state dynamics and
complete synchronization dispenses with defining spatial
topography and temporal scales of the target process.
To
figure out possible alternatives to the MSF-based
synchronizability construct one needs to understand both the
role played by synchronization within
this conceptual
framework and the objective pursued by the studies using it and
the problems that they aim to address. On the one hand, while in
neuroscience synchronization typically refers to bivariate
coupling between two neuronal ensembles, the synchronization
referred to by synchronizability is in fact better thought of as a
process on a network. On the other hand, from a teleological
view-point, resorting to the MSF formalisms can be understood
3
in terms of the need to address the relationship between
anatomy (or, more precisely, the topology defined on it) and
dynamics in complex systems [3,41,42]. Given an observed
dynamics and topological organization, a construct teleologically
equivalent to synchronizability may possibly be framed in terms
of a networked system's propensity to enter a functionally
desirable state or regime. But what dynamical states or regimes
may represent a valuable target, the distance from which may be
used as a neurophysiologically meaningful benchmark?
oscillators
of heterogeneous
The true difficulty in finding alternatives to the MSF-based
synchronizability is that one loses the uniqueness and task-
invariance of the complete synchronization state and needs to
cope with brain dynamics' spatial heterogeneity and temporal
multiscaleness, and brain function's translational invariance. To
define a valid equivalent of synchronizability will likely require
three key
ingredients: neurophysiologically plausible and
functionally meaningful order parameters describing collective
brain activity; mechanisms through which they may emerge;
and, no less importantly, those through which they may wane.
On the one hand, this should for instance involve considering
networks
and plausible
synchronization processes,
compatible with metastable
dynamics. On the other hand, the mechanisms through which
neural assemblies interact and their role in human brain
function at various scales of brain structure and dynamics
should be better understood at both functional/computational
and algorithmic/dynamical levels. These mechanisms are likely
task-specific, and various ones may even coexist [31]. As a
consequence, the definition of a dynamical target may vary as a
function of the putative role of synchrony lato sensu in the target
activity. Dynamical references could be associated with cluster
synchronization or chimera-like states [43-45] which would
prescribe spatial scales. However, rather than statistically
stationary states, what is needed is an analysis of their dynamics,
stability, bifurcations, and symmetries [44,46]. Importantly, a
reference regime should also replicate the temporal scales of
some (task-specific or independent) reference brain activity.
The construct may for instance contain predictive information
on the properties of and on the conditions under which these
clusters form. Defining meaningful dynamical target processes,
predicting these states and defining some sort of distance from
them to given observed ones, understanding whether and the
extent to which these may emerge from interactions between
local dynamics and network topology are all highly non-trivial
but fundamental questions, finding answers to which will likely
keep the neuroscience community busy for some time to come.
Concluding remarks
We have argued that not only is the synchronizability construct
an inadequate tool to quantify brain networks' ability to
synchronize, but the problem itself to which it is supposed to
provide an answer appears to be ill-posed when studying brain
dynamics. More generally, the brain differs in many essential
ways from the systems (e.g. the electrical power-grid or the
Internet) most network theory constructs were originally
designed to account for. Neuroscience, a field where network
theory has only relatively recently come to the foreground [47],
has so far borrowed its tools and concepts without inspiring
fresh theory, and this has exposed it to the risks inherent in such
an application: over-, under- and misuse of existing tools [48,49].
Rather than simply resorting to an existing bag of tricks,
neuroscience should instead use the brain's unique properties
to promote a fundamental reformulation of network science, for
the benefit of both.
References
1. Pecora, L.M., & Carroll, T.L. Master stability functions for synchronized
coupled systems. Phys. Rev. Lett. 80, 2109–2112 (1998).
2. Boccaletti, S., Latora, V., Moreno, Y., Chavez, M., & Hwang, D.-U. Complex
networks: structure and dynamics. Phys. Rep. 424, 175–308 (2006).
3. Arenas, A., Díaz-Guilera, A., Kurths, J., Moreno, Y., & Zhou C.J.
Synchronization in complex networks. Phys. Rep. 469, 93–153 (2008).
4. Varela, F., Lachaux, J.-P., Rodriguez, E., & Martinerie, J. The brainweb:
phase synchronization and large-scale integration. Nat. Rev. Neurosci. 2,
229–239 (2001).
5. Chavez, M., Besserve, M., & Le Van Quyen. M. Dynamics of excitable
neural networks with heterogeneous connectivity. Progr. Biophys Mol.
Bio. 105, 29–33 (2011).
6. Zhao, M., Zhou, C., Lü, J., & Lai, C.H. Competition between intra-
community and inter-community synchronization and relevance in brain
cortical networks. Phys. Rev. E 84, 016109 (2011)
7. Ton, R., Deco, G., & Daffertshofer, A. Structure-function discrepancy:
inhomogeneity and delays in synchronized neural networks. PLoS
Comput. Biol. 10, e1003736 (2014).
8. Phillips, D.J., McGlaughlin, A., Ruth, D., Jager, L.R., & Soldan, A. Graph
theoretic analysis of structural connectivity across the spectrum of
Alzheimer's disease: the
importance of graph creation methods.
NeuroImage: Clinical 7, 377–390 (2015).
9. Tang, E., Giusti, C., Baum, G.L., Gu, S., Pollock, E., Kahn, A.E., Roalf, D.R.,
Moore, T.M., Ruparel, K., Gur, R.C., Gur, R.E., Satterthwaite, T.D., & Bassett,
D.S. Developmental increases in white matter network controllability
support a growing diversity of brain dynamics. Nat. Commun. 8, 1252
(2017).
10. de Haan, W., van der Flier, W.M., Wang, H., Van Mieghem, P.F.A., Scheltens,
P., & Stam, C.J. Disruption of functional brain networks in Alzheimer's
disease: what can we learn from graph spectral analysis of resting-state
magnetoencephalography? Brain Connect. 2, 45–55 (2012).
11. Bassett, D.S., Meyer-Lindenberg, A., Achard, S., Duke, T., & Bullmore, E.
Adaptive reconfiguration of fractal small-world human brain functional
networks. Proc. Natl. Acad. Sci. U.S.A. 51, 19518–19523 (2006).
12. Reijneveld, J.C., Ponten, S.C., Berendse, H.W., & Stam, C.J. (2007). The
application of graph theoretical analysis to complex networks in the
brain. Clin. Neurophysiol. 118, 2317–2331.
13. Stam, C.J., & Reijneveld, J.C. Graph theoretical analysis of complex
networks in the brain. Nonlin. Biomed. Phys. 1, 3 (2007).
14. Schindler, K.A., Bialonski, S., Horstmann, M.T., Elger, C.E., & Lehnertz, K.
Evolving functional network properties and synchronizability during
human epileptic seizures. Chaos 18, 033119 (2008).
15. Deuker, L., Bullmore, E.T., Smith, M., Christensen, S., Nathan, P.J.,
Rockstroh, B., & Bassett, D.S. Reproducibility of graph metrics of human
brain functional networks. NeuroImage 47, 1460–1468 (2009).
16. van Wijk, B.C.M., Stam C.J., & Daffertshofer, A. Comparing brain networks
of different size and connectivity density using graph theory. PLoS ONE 5,
e13701 (2010). doi:10.1371/journal.pone.0013701
17. Jalili, M., & Knyazeva, M.G. EEG-based
functional networks
in
schizophrenia. Comput. Biol. Med. 41, 1178–1186 (2011).
18. van Dellen, E., Douw, L., Hillebrand, A., Ris-Hilgersom, I.H.M., &
Schoonheim, M.M., et al. MEG network differences between low- and
high-grade glioma related to epilepsy and cognition. PLoS ONE 7, e50122
(2012).
19. Tahaei, M.S., Jalili, M., & Knyazeva, M.G. Epilepsy synchronizability of
EEG-based functional networks in early Alzheimer's disease. IEEE Trans.
Neural Syst. Rehabil. Eng. 5, 636–641 (2012).
20. Bialonski, S., Lehnertz, K. Assortative mixing in functional brain
networks during epileptic seizures. Chaos 3, 033139 (2013) doi:
10.1063/1.4821915.
21. Lehnertz, K., Ansmann, G., Bialonski, S., Dickten, H., Geier, C., & Porz, S.
Evolving networks in the human epileptic brain. Physica D 267, 7–15
(2014).
22. Niso, G., Carrasco, S., Gudín, M., Maestú, F., del-Pozo, F., & Pereda, E. What
graph theory actually tells us about resting state interictal MEG epileptic
activity. NeuroImage: Clinical 8, 503–515 (2015).
23. Khambhati, A.N., Davis, K. A., Lucas, T.H., Litt, B., & Bassett, D.S. Virtual
cortical resection reveals push-pull network control preceding seizure
evolution. Neuron 91, 1170–1182 (2016).
24. The (combinatorial) Laplacian matrix is defined as , = - − ., where - is
a diagonal matrix containing the degree (i.e., number of connections) of
the nodes of the network and . is the adjacency matrix, with .(cid:6)(cid:18) = 1 if
nodes i and j are connected and zero otherwise. An adaptation to
weighted networks can easily be obtained by including the weights of
the connections in . and replacing the node degree by the node strength
(i.e., the sum of the weights of node's links).
25. The dispersion of the eigenvalues of the Laplacian matrix is given the
following
with
(cid:1)1 = (cid:7)1 (cid:27) − 1
, (cid:15) being the coupling strength of the (cid:27) nodes in
the anatomical network and (cid:1)(cid:6) the i-th eigenvalue of the Laplacian
matrix.
1 /(cid:31) = (cid:15)(cid:31)(cid:7)(cid:27) − 1(cid:9) ∑ 0(cid:1)(cid:6) − (cid:1)10(cid:31)
⁄
(cid:20)2(cid:22)
(cid:6)(cid:21)(cid:22)
⁄
expression:
(cid:9) ∑
(cid:20)2(cid:22)
(cid:6)(cid:21)(cid:22)
(cid:1)(cid:6)
⁄
26. Nishikawa, T., & Motter, A.E. Network synchronization landscape reveals
compensatory structures, quantization, and the positive effect of
negative interactions. Proc. Natl. Acad. Sci. USA 107, 10342–10347
(2010).
27. Huang, L., Chen, Q., Lai, Y.-C., & Pecora, L.M. Generic behavior of master-
stability functions in coupled nonlinear dynamical systems. Phys. Rev. E
80, 036204 (2009).
28. Sun, J., Bollt, E.M., & Nishikawa, T. Master stability functions for coupled
nearly identical dynamical systems. EPL 85, 60011 (2009).
29. Stam, C.J., & van Dijk, B.W. Synchronization likelihood: an unbiased
measure of generalized synchronization in multivariate data sets.
Physica D 163, 236–251 (2002).
30. Vicente, R., Gollo, L.L., Mirasso, C.R., Fischer, I., & Pipa, G. Dynamical
relaying can yield zero time lag neuronal synchrony despite long
conduction delays. Proc. Natl. Acad. Sci. USA 105, 17157–17163 (2008).
31. Malagarriga, D., Villa, A.E., García-Ojalvo, J., & Pons, A.J. Consistency of
heterogeneous synchronization patterns in complex weighted networks.
Chaos 27, 031102 (2017).
32. Rabinovich, M., Huerta, R., & Laurent, G. Transient dynamics for neural
processing. Science 321, 48–50 (2008).
33. Tognoli, E., & Kelso, J.A.S. The metastable brain. Neuron 81, 35–48 (2014).
34. Kinzel, W., Englert, A., & Kanter, I. On chaos synchronization and secure
communication. Phil. Trans. R. Soc. A 368, 379–389 (2010).
35. Tyrrell, A., Auer, G., & Bettstetter, C. Emergent slot synchronization in
wireless networks. IEEE Trans. Mobile Comput. 9, 719–732 (2010).
36. Rohden, M., Sorge, A., Timme, M., & Witthaut, D. Self-organized
synchronization in decentralized power grids. Phys. Rev. Lett. 109,
064101 (2012).
37. Rohden, M., Sorge, A., Witthaut, D., & Timme, M. Impact of network
topology on synchrony of oscillatory power grids. Chaos 24, 013123
(2014).
38. Motter, A.E., Myers, S.A., Anghel, M. & Nishikawa, T. Spontaneous
synchrony in power-grid networks. Nat. Phys. 9, 191–197 (2013).
39. Torrealdea, F.J., Sarasola, C., d'Anjou, A., Moujahid, A., & de Mendizábal,
N.V. Energy efficiency of information transmission by electrically
coupled neurons. BioSystems 97, 60–71 (2009).
40. Moujahid, A., d'Anjou, A., Torrealdea, F.J., & Torrealdea, F. Energy and
information in Hodgkin-Huxley neurons. Phys. Rev. E 83, 031912 (2011).
41. Skardal, P.S., Taylor, D., & Sun, J. Optimal synchronization of complex
networks. Phys. Rev. Lett. 113, 144101 (2014).
42. Menck, P.J., Heitzig, J., Marwan, N., & Kurths, J. How basin stability
complements the linear-stability paradigm. Nat. Phys. 9, 89–92 (2013).
43. Zhou, C., & Kurths, J. Hierarchical synchronization in complex
networks with heterogeneous degrees. Chaos 16, 015104 (2006).
44. Abrams, D.M., Mirollo, R., Strogatz, S.H., & Wiley, D.A. Solvable model for
chimera states of coupled oscillators. Phys. Rev. Lett. 101, 084103 (2008).
45. Bi, H., Hu, X., Boccaletti, S., Wang, X., Zou, Y., Liu, Z., & Guan, S. Coexistence
of quantized, time dependent, clusters in globally coupled oscillators.
Phys. Rev. Lett. 117, 204101 (2016).
46. Pecora, L.M., Sorrentino, F., Hagerstrom, A.M., Murphy, T.E., & Roy, R.
Cluster synchronization and isolated desynchronization in complex
networks with symmetries. Nat. Commun. 5, 4079 (2014).
47. Bullmore, E., & Sporns, O. Complex brain networks: Graph theoretical
analysis of structural and functional systems, Nat. Rev. Neurosci. 10, 186–
198 (2009).
48. Papo, D., Zanin, M., Pineda-Pardo, J.A., Boccaletti, S., & Buldú, J.M.
Functional brain networks: great expectations, hard times, and the big
leap forward. Phil. Trans. R. Soc. B 369, 20130525 (2014).
49. Papo, D., Zanin, M., Martínez, J.H., & Buldú, J.M. Beware of the Small-
World neuroscientist! Front. Hum. Neurosci. 10, 96 (2016).
4
|
1009.1828 | 3 | 1009 | 2011-05-23T13:30:08 | The Ising decoder: reading out the activity of large neural ensembles | [
"q-bio.NC",
"cond-mat.dis-nn"
] | The Ising Model has recently received much attention for the statistical description of neural spike train data. In this paper, we propose and demonstrate its use for building decoders capable of predicting, on a millisecond timescale, the stimulus represented by a pattern of neural activity. After fitting to a training dataset, the Ising decoder can be applied "online" for instantaneous decoding of test data. While such models can be fit exactly using Boltzmann learning, this approach rapidly becomes computationally intractable as neural ensemble size increases. We show that several approaches, including the Thouless-Anderson-Palmer (TAP) mean field approach from statistical physics, and the recently developed Minimum Probability Flow Learning (MPFL) algorithm, can be used for rapid inference of model parameters in large-scale neural ensembles. Use of the Ising model for decoding, unlike other problems such as functional connectivity estimation, requires estimation of the partition function. As this involves summation over all possible responses, this step can be limiting. Mean field approaches avoid this problem by providing an analytical expression for the partition function. We demonstrate these decoding techniques by applying them to simulated neural ensemble responses from a mouse visual cortex model, finding an improvement in decoder performance for a model with heterogeneous as opposed to homogeneous neural tuning and response properties. Our results demonstrate the practicality of using the Ising model to read out, or decode, spatial patterns of activity comprised of many hundreds of neurons. | q-bio.NC | q-bio |
The Ising decoder: reading out the activity of large neural
ensembles
Michael T Schaub and Simon R Schultz
Department of Bioengineering, Imperial College London,
South Kensington, London SW7 2AZ, UK.
[email protected]
May 29, 2018
Abstract
The Ising Model has recently received much attention for
the statistical description of neural spike train data. In this
paper, we propose and demonstrate its use for building de-
coders capable of predicting, on a millisecond timescale,
the stimulus represented by a pattern of neural activity.
After fitting to a training dataset, the Ising decoder can
be applied "online" for instantaneous decoding of test
data. While such models can be fit exactly using Boltz-
mann learning, this approach rapidly becomes compu-
tationally intractable as neural ensemble size increases.
We show that several approaches, including the Thouless-
Anderson-Palmer (TAP) mean field approach from statis-
tical physics, and the recently developed Minimum Prob-
ability Flow Learning (MPFL) algorithm, can be used for
rapid inference of model parameters in large-scale neural
ensembles. Use of the Ising model for decoding, unlike
other problems such as functional connectivity estima-
tion, requires estimation of the partition function. As this
involves summation over all possible responses, this step
can be limiting. Mean field approaches avoid this prob-
lem by providing an analytical expression for the parti-
tion function. We demonstrate these decoding techniques
by applying them to simulated neural ensemble responses
from a mouse visual cortex model, finding an improve-
ment in decoder performance for a model with heteroge-
neous as opposed to homogeneous neural tuning and re-
sponse properties. Our results demonstrate the practical-
ity of using the Ising model to read out, or decode, spatial
patterns of activity comprised of many hundreds of neu-
rons.
1
Introduction
Interpreting the patterns of activity fired by populations
of neurons is one of the central challenges of modern sys-
tems neuroscience. The design of decoding algorithms
capable of millisecond-by-millisecond readout of sensory
or behavioural correlates of neuronal activity patterns
would be a valuable step in this direction. Such decoding
algorithms, as well as helping us to understand the neural
code, may have further practical application, as the basis
of communication neural prostheses for severely disabled
patients such as those with "Locked In" syndrome.
At the heart of such a decoding algorithm must lie -
whether explicit or implicit - a description of the con-
ditional probability distribution of activity patterns given
stimuli or behaviours. Making this description is nontriv-
ial, as the brain, like other biological systems, exhibits
enormous complexity. This results in a very large number
of possible states or configurations exhibited by the sys-
tem, making the description of such systems by simply
measuring the probabilities of each state unfeasible. Ex-
cept for very small patterns, a model-based approach of
some kind is essential.
New technologies in neuroscience such as high-density
multi-electrode array recording and multi-photon calcium
imaging now make it possible to monitor the activity of
1
large numbers of neurons simultaneously. Analysis tools
for such high dimensional data have however lagged be-
hind the experimental technology, as most approaches are
limited to very small population sizes. While consider-
able advances have been made in the use of information-
theoretic approaches to characterise the statistical struc-
ture of small neural ensembles (Gawne et al. 1996, Panz-
eri, Schultz, Treves & Rolls 1999, Schultz & Panzeri
2001, Panzeri & Schultz 2001, Reich et al. 2001, Petersen
et al. 2001, Pola et al. 2003, Montani et al. 2007), finite
sampling limitations have made results for larger ensem-
bles much more difficult to obtain.
For the statistical description of multivariate neural
spike train data, parametric models able to capture most
of the interesting features of real data while still being
of empirically accessible dimensionality are highly desir-
able. One promising approach has emerged from statis-
tical mechanics: the use of Ising (or Ising-like) models,
exploiting an analogy between populations of spike trains
and ensembles of interacting magnetic spins (Shlens et al.
2006, Shlens et al. 2009, Schneidman et al. 2006).
Our aim here is to devise an algorithm for "millisecond-
by-millisecond" neural decoding, on the basis that in-
formation processing in the nervous system appears to
make use of such fine temporal scales (Carr 1993, Bair
& Koch 1996). The timescale of the "symbols" used in
information processing is thus likely to be somewhere be-
tween 1 and 20 ms for most purposes (Butts et al. 2007).
For time bins on this scale, neural spike trains are ef-
fectively binarized, and the simplest binary model (in
the maximum entropy sense) that captures pairwise cor-
relations is the Ising model. The Ising model is thus
a natural way to describe the statistics of neural spike
patterns at the timescale of interest. Fitting of such a
model to the observed neural data has the advantage that
it does not implicitly assume some non-measured struc-
ture in the data, i.e. maximum entropy models express the
most uncertainty about the modelled data given the (ex-
plicit) chosen constraints (e.g. that certain moments of the
measured distribution agree with the model distribution)
(Jaynes 1957). It can be shown that this is mathematically
equivalent to maximizing the likelihood of the model pa-
rameters to explain the observed data (Berger et al. 1996).
By using this approach to fit a model to the conditional ac-
tivity pattern distribution, in conjunction with maximum a
posteriori decoding (Foldi´ak 1993, Oram et al. 1998), it is
possible to train a decoder which takes as its input a pat-
tern of spiking activity, and gives as its output the stimulus
that it determines to have elicited that spike pattern.
A major obstacle to the use of Ising models for neu-
ral decoding, is that, in general, it is necessary to com-
pute a partition function (or normalization factor), involv-
ing a sum over all possible states. This can be numeri-
cally challenging, and for large numbers of neurons, un-
feasible.
In the present study, we adopted several ap-
proaches for circumventing this problem. Firstly, we
make use of mean field approximations, including both
the 'naive' mean field approximation and the Thouless
et al. (1977) (TAP) extension to it, following Roudi, Tyr-
cha & Hertz (2009). Secondly, we compare this with
the recently proposed Minimum Probability Flow Method
(Sohl-Dickstein et al. 2009) for learning model param-
eters. To assess the relative performance of these ap-
proaches in the context of a discrete decoding problem,
we simulated the activity of a population of neurons in
layer V of the mouse visual cortex during an experiment
in which a discrete set of orientation stimuli were pre-
sented. Using this simulation, we evaluated the relative
performance characteristics of the different decoding al-
gorithms in the face of limited data, exploring decoding
regimes with up to 1000 neurons. We demonstrate, for
the first time, the use of the Ising model to effectively de-
code discrete stimulus states from large-scale simulated
neural population activity.
Ising models of neural spike trains
2 Methods
2.1
Activity states in an Ising model are Boltzmann dis-
tributed, i.e. they are distributed according to the negative
exponential of the "energy" associated with each state.
This distribution,
µ λµfµ, is the maximum entropy distribution
subject to the set of constraints imposed by Lagrange mul-
tiplers λµ on variables fµ.
Imposing these constraints
upon firing rates and pairwise correlations gives
P ∝ e−(cid:80)
pIsing(rs) =
1
Z(s)
exp
2
(cid:88)
i
(cid:88)
i(cid:54)=j
hi(s)ri +
1
2
,
Jij(s)rirj
(1)
(cid:88)
r∈R
(cid:88)
i
.
(cid:88)
i(cid:54)=j
where r = (r1, r2, . . . , rC)T and each binary response
variable ri ∈ {0, 1} indicates the firing/not firing of neu-
ron i in the observed time interval. The parameters hi
(known in statistical physics as 'external fields') and Jij
('pairwise couplings') have to be fit to the data such that
the model displays the same means and pairwise correla-
tions as the data:
(cid:104)ri(cid:105)Ising =(cid:104)ri(cid:105)Data ,
(cid:104)rirj(cid:105)Ising =(cid:104)rirj(cid:105)Data ,
(2a)
(2b)
where (cid:104)·(cid:105)model denotes expectation with respect to the
specified distribution. Z(s) is the partition function,
which acts as a normalisation factor. i.e.:
Z(s) =
exp
hi(s)ri +
1
2
Jij(s)rirj
(3)
Note that the first sum is over all possible (as opposed to
observed) responses, given by the set R.
In statistical physics it is more common to use a sym-
metric representation σi ∈ {−1, 1} for the 'spins' that
describe the activation of neuron i (with −1 indicating
'no spike' and 1 indicating 'spike'), which simply corre-
sponds to a change of variables σi = 2ri − 1. Accord-
ingly the fields, couplings and partition functions change.
As it is occasionally more convenient to work in one or
the other representation we will denote the fields and cou-
plings in the spin representation with hi and Jij.
Standard Monte Carlo techniques for fitting these
model parameters, such as Boltzmann learning, which
can in principle provide an exact solution - given the
number of samples is high enough - become computa-
tionally very expensive if not intractable as the num-
ber of cells increases. We have found in previous work
(Seiler et al. 2009), that the Boltzmann learning approach
becomes computationally too expensive in our case for
ensemble sizes larger than 30 cells. This poor scaling
behaviour is mainly due to the exponentially increasing
number of states with the number of cells. Speeding up
the model fitting process is hence an essential requirement
to utilize Ising models for studies with large ensembles
of neurons. Solutions to speed up the "classical" Boltz-
mann learning approach have been suggested (Broderick
et al. 2007). However these are still associated with a high
computational cost.
3
2.2 Neural Decoding
In this paper we consider the problem of decoding which
of a number of different stimuli has elicited a neural spike
pattern. This can be seen as a discrete classification task:
we have a set of S stimuli s ∈ S = {s1, s2, . . . , sS}. De-
coding in this scenario means that we have to provide a
decision rule that estimates which stimulus has been the
input to the system, given an observed spike pattern robs.
The particular example to which we apply this is a simula-
tion of the spike pattern responses elicited by visual stim-
uli across the receptive field of a visual cortical neuron:
in this case each stimulus si represents a different orien-
tation of a sinusoidal grating. Our main aim with this sim-
ulation was to validate our methodology in a neurophys-
iologically realistic coding regime, relevant to datasets to
which our methodology might be applied. As a supple-
mentary goal, we hoped to gain some insight into whether
some aspects of the model affect decoding performance
- such as heterogeneity of tuning, observed in real neural
recordings but often ignored in population coding models.
For decoding, we use the maximum a posteriori (MAP)
rule:
s = arg max
s
= arg max
s
= arg max
s
p(srobs)
p(robss)p(s)
p(robss)p(s),
p(robs)
(4)
(5)
(6)
where the second step is the application of Bayes' theo-
rem and the third equality holds because p(robs) is inde-
pendent of s and is hence irrelevant for maximising the
given expression, i.e. just a constant factor with respect
to s that scales the maximum accordingly. In the case we
examine here we assume we are in control of the stimulus
distribution p(s), and thus we can choose it to be uniform,
i.e. to exhibit the same constant probability for each stim-
ulus and therefore be independent of s, as well. Hence
our decoding rule simplifies further to the maximum like-
lihood (ML) rule:
s = arg max
s
p(robss).
(7)
Within this setting, the task of creating a neural decoder
reduces to the modelling of the stimulus dependent dis-
tributions p(rs). Once these are obtained, we can apply
our ML decoding rule (Equation 7) to estimate the given
input stimulus s.
We have used two different statistical models to fit the
observed spike patterns for each stimulus. Firstly, we have
used an Ising model for p(rs), i.e. we assume that for
each stimulus, the spike pattern distribution can be de-
scribed by a (different) Ising model. Secondly, we have
used an independent model distribution pind, assuming
that given a stimulus, each cell is independent of the oth-
ers:
C(cid:89)
pind(rs) =
p(ris).
(8)
i=1
The independent model is the binary maximum entropy
model of first order, i.e.
it takes into account only the
first order moments (the constraints on the means given
by Equation 2a) and is therefore a natural comparison for
the Ising model. As it is very easy to fit the independent
model, we used this as a control method, to test whether
the more complex Ising model could enhance decoding
performance. Note that the numerical values for the prob-
abilities can get very small for large cell ensembles, and
therefore to evade finite precision problems we use in this
case an equivalent log-likelihood decoding rule instead of
the ML rule, i.e. maximise the logarithm of the likelihood
instead of maximising the likelihood directly.
2.3 Training and Testing
To train and test the decoders, we proceed as follows:
1. For each stimulus we simulate a set of possible re-
sponse vectors. The details of the simulation are de-
scribed in the following subsection.
2. We separate the simulated response patterns into
training data, which is used to fit the model and test
data which we use to evaluate the decoding perfor-
mance of the obtained models.
3. The whole testing procedure is performed with 10
fold cross-validation, i.e. we divide the whole data
for each stimulus into 10 equally sized parts. We
then use 9 parts of the data to train our model and
the remaining one for testing. We repeat this pro-
cess again with all 10 possible test/training data set
combinations of this kind to reveal if our results gen-
eralize to the whole dataset.
4
2.4 Simulation of Evoked Spike Patterns
from Mouse Primary Visual Cortex
We simulated the transient response patterns of activity
evoked by visual (orientation) stimuli in layer V pyrami-
dal neurons of the anaesthetized mouse visual cortex. The
orientation direction were chosen to be n · 180/S, where
S is the number of stimuli and n ∈ {0, 1, . . . , S − 1}. The
properties of our simulation are motivated by the results
reported by Niell & Stryker (2008). We simulated dif-
ferent models by augmenting a basic model with mostly
homogeneous response characteristics, with some come
controlled heterogeneous characteristics.
Our model is defined as follows, with parameters spec-
ified in Table 1. The spontaneous activity of each neu-
ron was set to 1.7 spikes per second, corresponds to the
reported median value for layer V neurons in (Niell &
Stryker 2008). We assumed that neuronal direction pref-
erences were uniformly spaced around the circle. Each
neuron's tuning curve was defined by a von Mises func-
tion (circular Gaussian) with half width at half maximum
(HWHM) fit to experimental data (Niell & Stryker 2008).
The direction selectivity index (DSI) was set to 0.1 for all
layer 5 neurons in our model. Sustained firing rates were
fit to the distributions reported in (Niell & Stryker 2008).
To reflect that we are considering a situation in which a
stimulus is decoded from a short time window (20 ms) of
data, we multiplied these evoked rates by a fixed transient-
to-sustained ratio of 1.5, taken to reflect the onset re-
sponse of the neuron's response to a flashed stimulus. As
our model is fit to data from directional (drifting grat-
ing) stimuli, we took the arithmetic mean value of the two
corresponding diametrically opposite directions for each
neuron to compute the model response to a flashed orien-
tation.
The characteristics of the basic model were modulated
via two inhomogeneity parameters γ, ξ ∈ [0, 1], to in-
troduce a heterogeneous distribution of firing rates and
the tuning widths respectively, as described in Table 1.
We neglected inhomogeneity in other parameters. Thus
the parameter γ regulates firing rate heterogeneity, where
γ = 0 corresponded to the basic homogeneous mode,l
and γ = 1 to the most heterogeneous firing rate setting.
The parameter ξ was used analogously to regulate hetero-
geneity in the tuning widths of the neurons. The effects
of the two heterogeneity parameters γ, ξ are illustrated in
Parameter
Preferred tuning direction
Tuning width (HWHM)
Value
uniformly spaced 0◦−360◦
38 + ξΦi
Direction Selectivity Index
Spontaneous firing rate
0.1
1.7 + γXi (Hz)
Sustained evoked firing rate
7 + γYi (Hz)
Transient-to-sustained ratio
1.5
Comments
ξ ∈ [0, 1]; Φi normally distributed, fitted to Fig. 4f of Niell &
Stryker (2008), truncated minimum of 10.
fixed for all neurons; Niell & Stryker (2008) Fig. 5c.
γ ∈ [0, 1]; Xi normally distributed about zero, distribution fitted to
Fig. 8d of Niell & Stryker (2008); truncated to minimum of 0.3
γ ∈ [0, 1]; Yi normally distributed, fitted to data in Fig. 8 and supp.
fig. S3 of Niell & Stryker (2008); truncated to minimum of zero
fixed for all neurons
Table 1: Parameters of the mouse visual cortical model simulated. γ and determine the extent of heterogeneity in
the model, with γ, = 0 setting the model properties to homogeneous.
Figure 1.
Patterns of spikes fired by the neural population
were simulated using a dichotomized Gaussian approach
(Macke et al. 2009). Since we cannot estimate covariance
matrices from experimental data directly, and not every
positive definite symmetric matrix can be used as the co-
variance matrix of a multivariate binary distribution, we
adapted the following approach. First we compute upper
and lower covariance bounds for each pair of neurons, ac-
cording to (Macke et al. 2009)
max{−pq,−(1 − p)(1 − q)} ≤ Cov(ri, rj)
≤ min{(1 − q)p, (1 − p)q} ,
(9)
where p and q are the means (mean spiking probabilities)
of neuron i and j, respectively. We then choose a ran-
dom symmetric matrix A that lies between these bounds.
As in general this choice does not result in a permissible
correlation matrix for the underlying Gaussian, a Higham
(2002) correction is applied to find the closest correlation
matrix possible for the latent Gaussian (cf. Macke et al.
(2009)), to which we finally arithmetically add a random
correlation matrix with uniformly distributed eigenvalues
to adjust the mean correlation strength. Having estab-
lished a dichotomized Gaussian model we can thus draw
samples with high efficiency.
Where not otherwise stated in the text, 10000 trials per
stimulus were simulated, allowing 9000 training samples
and 1000 test samples with 10-fold crossvalidation. In the
absence of a detailed characterization of the correlation
structure of neural responses in the mouse visual cortex,
we assumed that the correlation in firing between each
pair of neurons was weak and positive. Our simulation
results in a mean correlation of 0.11 and a standard devi-
ation of 0.040 (measured with 100000 samples for differ-
ent ensemble sizes) for our basic model and similar levels
of correlation for nonzero γ, ξ. Due to limitations of the
dichotomized Gaussian simulation, we were not able to
specify the correlations of the spike trains/between the in-
dividual neurons exactly, thus all reported correlations are
measured and may be prone to small variations. However,
such limitations would be inherent to any simulation ap-
proach, as i) the covariance structure of a multivariate bi-
nary distribution is always constrained by the firing prob-
abilities of the cells and can not be chosen independently
of these firing rates (Macke et al. 2009), and ii) finite sam-
pling effects will always affect the simulated data, result-
ing in fluctuations in the correlation structure.
We were able to vary the (measured) mean absolute
correlation level in some simulations, allowing an assess-
ment of the relative effects of correlation strength for de-
coding. To do this we proceeded as follows: having es-
tablished the correlation matrix of the latent Gaussian as
described before allows us to sample in a regime with cor-
relations around 0.1. Likewise a latent Gaussian with a
correlation matrix given by the identity matrix would cor-
respond to the case where there are no correlations in the
latent Gaussian and thus in the simulated spike trains it-
self, which means that the simulated spike patterns for
each neuron are independent (note that due to the model
used in Macke et al. (2009) the correlation matrix and the
5
covariance matrix of the latent Gaussian are the same). Fi-
nite sampling effects might still introduce some nonzero
measured correlations in the spike patterns, but the un-
derlying distribution would still be with independent neu-
rons. Thus, by interpolating between these two cases we
can reduce the effective correlations in the data in a con-
trolled way, where at one end we yield our original model
and at the other end we yield a set of independent neurons.
2.5 Fitting the Ising Model Parameters
For fitting the model parameters in the Ising model case
we use two different strategies: mean field approxima-
tions and minimum probability flow learning. In earlier
work we used Boltzmann learning (Seiler et al. 2009),
however as this becomes rapidly computationally in-
tractable with an increasing number of neurons, we have
not reported it here.
2.5.1 Mean Field Methods
The suitability of different Mean Field approaches for fit-
ting the parameters of an Ising model have been recently
assessed by Roudi, Tyrcha & Hertz (2009). In their work,
Roudi et al. compared the learned model parameter as
inferred by Boltzmann learning, with the parameters in-
ferred by a number of different approximative algorithms.
Here we examine the utility for decoding of using succes-
sively higher order mean field approximations. Tanaka
(1998) demonstrated how to systematically obtain mean
field approximations of increasing order based on a Ple-
fka series expansion (Plefka 2006) of the Gibbs free en-
ergy. By truncating these series to terms up to n-th or-
der, and using the linear response correction, it is possible
to derive an n-th order mean field approximation, yield-
ing C equations for the external fields, and C(C − 1)/2
equations for the pairwise couplings (cf. Tanaka (1998)).
These equations can then be solved with respect to Jij
and hi. For higher order approximations, these equations
can have more than one solution. This problem can be re-
solved by considering that the Plefka series expansion is
effectively a Taylor series expansion and continuity of the
solutions is expected when higher order terms are gradu-
ally increased (Tanaka 1998).
Tanaka further provided an explanation for the "diago-
nal weight trick" as used by Kappen & Rodrguez (1998).
With this trick one introduces C extra equations for the
pairwise "self-couplings" Jii, which can be used to refine
the respective approximations. The success of this trick
can be explained (Tanaka 1998) by considering that us-
ing this diagonal weight trick in an n-th order method, is
effectively incorporating the dominant terms of the next
higher (n + 1) order expansion.
For fitting the parameters, we have compared different
mean field approximations from zeroth to second order
with and without the diagonal weight trick, thus incor-
porating with our highest approximations all second or-
der terms and the dominant third order terms of the free
energy expansion. The zeroth order method is thereby
equivalent to the independent model, thus providing an-
other way of thinking about the independent decoder.
First order methods (naive mean field approximation).
For the first order or naive mean field approximation, the
equations for the external fields and pairwise couplings
become:
J = −C−1
hi = tanh
−1 mi −(cid:88)
(i (cid:54)= j),
Jijmj,
j(cid:54)=i
(10a)
(10b)
where J is the estimated coupling matrix with elements
Jij, the 'magnetization' mi = (cid:104)σi(cid:105), and the covariance
matrix C is defined by Cij = (cid:104)σiσj(cid:105) − mimj. In the
following we will denote this naive mean field method by
nMF.
By incorporating the diagonal weight trick the above
equations change slightly into the following:
J = P−1 − C−1,
hi = tanh
−1 mi −(cid:88)
Jijmj,
(11a)
(11b)
j
where in addition to the matrices defined above, we have
defined the diagonal matrix Pij = (1 − m2
i )δij, with δij
being the Kronecker delta. Note that in the nomencla-
ture of Roudi, Tyrcha & Hertz (2009), this has been called
"naive mean field method". However we will in the fol-
lowing refer to it as naive mean field method with diago-
nal weight trick (nMFwd).
6
Second order methods (TAP approximation). The
equations of the second order method are also known as
the TAP equations (Thouless et al. 1977) and can be seen
as a correction to the naive mean field methods. Using
the TAP approach the equations for the model parameters
read:
(cid:88)
j(cid:54)=i
(i (cid:54)= j),
(12a)
ij (1 − m2
J 2
j ),
(12b)
ij mimj + Jij + (C−1)ij = 0
2 J 2
hi = tanh
−1 mi −(cid:88)
j(cid:54)=i
Jijmj + mi
where the first equation can be solved for the pairwise
couplings Jij. As mentioned previously the correct so-
lution has to be chosen according to continuity condi-
tions outlined in Tanaka (1998), from which then the
external fields hi can be computed. More precisely, if
mimj(C−1)ij < 0 we choose the solution, which is
closer to the original first order mean field solution.
If
mimj(C−1)ij > 0 we use the first order mean field solu-
tion directly. We use this procedure as it avoids pairwise
couplings becoming complex, and respects the continu-
ity of the inverse Ising problem for Jij as a function of
(C−1)ij. In the following this method is denoted by TAP.
Incorporating third order terms. The second order
method can also be augmented by the diagonal weight
trick, hence incorporating the leading third order terms
of the free energy expansion. The equations for the TAP
method with diagonal weight trick (TAPwd) are given by:
2 J 2
ij mimj + Jij + (C−1)ij = 0
(i (cid:54)= j),
Jii =
1 − m2
i
1
− (C−1)ii
−1 mi −(cid:88)
Jijmj,
hi = tanh
(i = j),
(13a)
(13b)
(13c)
j
where we have solved the equations in an analogous fash-
ion to that described for the normal TAP approach.
2.5.2 Minimum Probability Flow Learning
Sohl-Dickstein et al. (2009) recently proposed the Min-
imum Probability Flow Learning (MPFL) technique,
which provides a general framework for learning model
parameters. As this technique is also applicable to the
7
Ising model, we have used it to learn external fields and
pairwise couplings for our model. However, as the sam-
pling regime usually feasible in neurophysiological exper-
iments dictates a small number of samples compared to
the number of parameters in the model (which is O(C 2)
with C cells), the learning problem for the parameters be-
comes under-constrained already at intermediate neural
ensemble sizes, i.e. we are likely to have more param-
eters to fit than there are samples.
We therefore introduced a regularization term to their
original objective function to penalize model parameters
growing to large numbers, i.e. to avoid overfitting. Given
the original objective function K(θ) with θ being the pa-
rameters of our model, our regularized objective function
reads:
Kreg(θ) = K(θ) +
(cid:107)θ(cid:107)2
2,
λ
2
(14)
where (cid:107)·(cid:107)2 is the L2 norm, which is a common choice of
regularization term (Bishop 2007). So for the Ising model
case we have:
(cid:88)
i
(cid:88)
i(cid:54)=j
(cid:107)θ(cid:107)2
2 =
h2
i +
J 2
ij.
For the present work, the regularization parameter was
set to λ = 0.0127, after systematically assessing differ-
ent settings for an ensemble size of 100 cells via cross-
validation. We refer to this learning algorithm as rMPFL
(regularized MPFL) in the rest of the manuscript.
Other choices for the regularization term are possible
and might even result in better performance for decoding
purposes, e.g. two independent penalty terms for the ex-
ternal fields and the pairwise couplings. However an ex-
tensive assessment of different parameter settings would
be very time consuming due to the cost of calculating
the invoked partition function. We therefore have not
performed an exhaustive analysis of regularization. In a
real experiment the regularization term should however be
adapted to yield the best possible performance for the spe-
cific number of cells. We not that while it is not necessary
to compute the partition function for some applications of
MPFL (e.g.
if learning the Jij parameters is the end in
itself), it is required for decoding.
2.6 Partition Function Estimation
Estimating the partition function is a computationally ex-
pensive task, since the set of possible responses R grows
exponentially with the number of cells C, rendering an
analytical computation (Equation 3) intractable for large
neural ensembles.
As MPFL learning does not provide an estimate of the
partition function, we used the Ogata-Tanemura partition
function estimator (Ogata & Tanemura 1984, Huang &
Ogata 2001), which is based on Markov Chain Monte
Carlo (MCMC) techniques. However MCMC is still a
very time consuming technique. To speed up the model
fitting process, we estimated the partition function for
each stimulus only once in a 10 fold cross-validation run
when using MPFL, as ideally all samples for a specific
stimulus should come from the same distribution, thus ap-
proximately sharing the same partition function. We have
examined the effect of this approximation (see Results).
When fitting the model parameters with mean field the-
oretic approaches, we computed the (true) partition func-
tion Z(s) in the mean field approximation, as reported in
Kappen & Rodrguez (1998), Thouless et al. (1977), and
Tanaka (1998).
For the first order mean field approach this yields:
log Z =
(cid:17)(cid:17) −(cid:88)
(cid:16)hi + Wi
2 cosh
(cid:16)
log
Wimi
i
(15)
Jijmimj,
(cid:88)
j(cid:54)=i
Wi =
Jijmj.
(cid:88)
(cid:88)
i
+
i<j
with
Here each of the parameters is actually a function of the
stimulus s, which we omit for clarity.
For the second order methods, the corresponding equa-
(cid:16)
tion becomes:
log Z =
(cid:88)
(cid:88)
i
i<j
+
Jijmimj +
log
2 cosh
Limi
(cid:17)(cid:17) −(cid:88)
i
(cid:16)hi + Li
(cid:88)
1
2
ij (1 − m2
J 2
i )(1 − m2
j ),
i<j
(16)
8
with
(cid:88)
j(cid:54)=i
Li =
Jijmj − mi
(cid:88)
j(cid:54)=i
ij (1 − m2
J 2
j ).
2.7 Performance evaluation
The fraction of correctly decoded trials was the principal
method used to assess decoding performance. However,
as the fraction correct or accuracy does not by itself pro-
vide a complete description of decoder performance, we
sought used additional performance measures. The per-
formance of a decoder is fully described by its confusion
matrix, and we show how directly examining this matrix
can yield insight into its behaviour. However, it is advan-
tageous to be able to reduce this to a single number in
many cirumstances. We therefore additionally computed
the mutual information between the encoded and decoded
stimulus (Panzeri, Treves, Schultz & Rolls 1999) to char-
acterise the performance further. This provides a compact
summary of the information content of the decoding con-
fusion matrix.
We can write the mutual information (measured in bits)
as:
I(s, s) = H(s) − H(ss),
where H(s) is the entropy of the encoded stimulus
H(s) = −(cid:88)
s∈S
p(s) log2 p(s),
(17)
(18)
and H(ss) is the conditional entropy describing the dis-
tribution of stimuli s that have been observed to elicit each
decoded state s,
p(s, s) log2 p(ss).
(19)
H(ss) = − (cid:88)
s,s∈S
Since we have in the current study opted for a uniform
stimulus distribution, the entropy H(s) is simply given
by
H(s) = log2 S.
(20)
In general the conditional entropy H(ss) has to be com-
puted from the confusion matrix. We note that if we were
to assume that the correctly decoded stimuli and errors
are uniformly distributed for all stimuli, i.e. that the con-
ditional distribution p(ss) is of the form
fc
1 − fc
S − 1
p(ss) =
for s = s
for s (cid:54)= s,
then the conditional entropy simplifies to
H(ss) = −fc log2 fc − (1 − fc) log2
(cid:19)
(cid:18) 1 − fc
S − 1
.
This simplified expression has been used to characterize
decoder performance in the Brain Computer Interface lit-
erature (Wolpaw et al. 2002). Here, we present this equa-
tion only to make apparent the scaling behaviour, and
compute the decoded information using the more general
expression (Eqn. 17).
3 Results
We performed computer simulations as described in
Methods, to generate datasets for training and testing de-
coding algorithms. For all settings, 20 simulations were
performed with different random number seeds, in order
to characterize decoding performance. A number of met-
rics, including the fraction of correct decodings (accu-
racy) and mutual information between decoded and pre-
sented stimulus distributions, were used in order to char-
acterize and compare decoding performance.
Basic Model
The Ising model based decoders show better performance
than the independent decoder in nearly all cases (Fig. 3a),
in terms of the average fraction of correctly decoded stim-
uli. The performance of the standard (non-regularized)
MPFL technique, however, in our hands falls away rel-
atively quickly as the number of cells increases, fail-
ing to better the independent decoder after approximately
100 cells. This behaviour can be explained by consid-
ering that the problem of parameter estimation becomes
more and more underconstrained as the number of cells
increases while holding the number of training samples
fixed. Falsely learned model parameters moreover affect
the decoding performance by influencing the estimated
9
partition function and thus worsen the decoder perfor-
mance. As we have only estimated the partition function
once per stimulus when using MPFL, large fluctuations
in the training dataset can potentially have a big effect.
To compensate for this behaviour, a regularization term
can be included, which can stabilize the performance up
to a significantly larger number of neurons (as described
in Methods). Our regularized version of MPFL still de-
creases in performance after about 110 cells. However,
we have used a fixed value for the regularization param-
eter for all simulations here, whereas ideally the regular-
ization term should be adapted to the number of cells and
number of samples in the specific setting. Using such an
approach would most likely result in a better performance
for larger numbers of cells. As an example we have tested
for an optimized λ parameter for 150 cells and found that
we could increase performance to 0.874 in this case (av-
erage over 10 trials, not shown in figure).
One of the assumptions made for much of this pa-
per is that the partition function can be estimated only
once in each 10-fold crossvalidation run, thus speeding
up training. We studied the effect of this approximation
on performance, finding, as shown in Figure 4, that the
effect is marginal, at least under our operating conditions.
No statistically significant difference between the two ap-
proaches was observed. In a decoding regime where the
rMPFL method starts to infer the wrong model parameters
(e.g. due to overfitting) one would expect these effects to
become more pronounced. However, in such a scenario
where the rMPFL method becomes unreliable, a different
regularization term or a different inference method should
be considered.
The decoded information analysis reveals that the dif-
ference in the decoding performance of the independent
and Ising (as exemplified by TAP, TAPwd and rMPFL)
models becomes more pronounced as the number of stim-
uli is increased, as shown in Fig. 3b,c. As the number
of stimuli increase, the independent and Ising decoder
curves separate, indicating not only a difference in the ac-
curacy of both decoders, but also a difference in the con-
fusion matrices, i.e. in the distribution of errors between
the two approaches. This is considered in more detail in
section 3.2.
A further interesting behavior of the Ising decoder is
apparent in Fig. 3c: as the number of stimuli increases,
the relative decoding gain η, (which we define as the ra-
may be particularly advantageous as the decoding prob-
lem becomes more difficult and not easily discriminable
in terms of the neural firing rates. Real world performance
will of course be dependent upon the level of systematic
difference in correlation structure induced by stimuli in
any given dataset.
Performance of the Ising decoder is strongly dependent
on the number of training trials available (Fig.
3d i).
Here we found, for 70 neurons, that around 400 train-
ing samples were required to allow the Ising decoder to
outperform independent decoding. The independent de-
coder will necessarily have better sampling performance,
as it relies only upon lower order response statistics. (It is
worth recalling that a "full" decoder, which made use of
all aspects of spike pattern structure, would have far worse
scaling behaviour than either). Another way of looking at
this is to examine the estimated covariance matrix from
the simulated data (Fig. 3d ii). We define the mean rela-
tive error of the covariance matrix as
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) Cdata
ij,s − Cas
Cas
ij,s
ij,s
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12) ,
(21)
(cid:88)
s∈S
(cid:88)
i,j
E =
1
S
1
C 2
where Cdata
ij,s is the estimated covariance between unit i and
j under stimulus s from the (finite) simulated data (nor-
mally 10000 samples), and Cas
ij,s is the asymptotic covari-
ance, defined in the same way but computed with 100,000
samples. This error provides a measure of the finite sam-
pling bias we encounter for fitting the model.
Heterogeneous Models
The performance characteristics in a heterogeneous sce-
nario, i.e. with nonzero ξ, γ, are for the most part broadly
similar to the homogeneous case, so we report here only
on the observed differences. The overall classification
performance both in terms of fraction correct and mutual
information, is slightly improved for both independent
and Ising decoders with heterogeneous neural ensembles.
As an example, the accuracy as a function of ensemble
size is compared for the basic model and the "fully hetero-
geneous" (γ = ξ = 1) models in Fig.5, for both TAPwd
and Independent decoders. The greater performance for
nonzero γ, ξ can be explained by the greater variability
of cell properties, allowing more specific response pat-
terns than in the homogeneous scenario. Such a scenario
10
Figure 4: Decoding performance effects of reduced par-
tition function estimation for rMPFL method as Box-
Whisker plot. Interquartile range is indicated by blue box,
the median value by a red line. Whiskers are used to visu-
alize the spread of the remaining data for values at most
within 1.5 times the interquartile range from the ends of
the box. Left: results for computing the partitioning func-
tion for each stimulus every round for 10-fold crossval-
idation. Right: results for computing the partition func-
tion only in the first round for each stimulus. Data from
20 simulation runs, 70 cells, basic model, 10000 samples
simulated per stimulus
tio between the "actual" and "chance" fraction correct)
keeps increasing with the number of stimuli for the Ising
model case, whereas it saturates for the independent de-
coder. This effect can be explained as follows: the inde-
pendent decoder relies on the fact that two different stim-
uli will result in different firing rates for each cell. With an
increasing number of stimuli however, the difference be-
tween two adjacent stimuli becomes smaller and smaller,
and thus the difference in the firing rates discriminating
two adjacent stimuli becomes smaller and smaller. There-
fore the decoder performance of the independent decoder
rapidly decreases as the number of stimuli increases. As
two adjacent stimuli can result in neural responses that
have quite different correlation structure despite having
very close firing rates, the Ising decoder can additionally
make use of this information in the data and thus yield
better performance. This suggests that the Ising decoder
0.80.8050.810.8150.820.8250.83every roundonly onceFraction Correct Figure 5: Decoder performance is enhanced in the hetero-
geneous scenario (ξ = γ = 1). Comparison of Fraction
correct vs. number of cells for basic model and fully het-
erogeneous model. The TAPwd algorithm was used to
train the decoder.
is presumably relevant to many real-world decoding prob-
lems, suggesting that decoder performance analysed with
homogeneous test data may slightly under-represent real-
world performance.
We also assessed the relative influence of the individ-
ual parameters γ, ξ on the decoding behavior as shown in
Figure 6. Our analysis shows that, while the heterogene-
ity in the firing rates as specified by γ has a positive effect
on the decoding performance, an increased tuning width
heterogeneity slightly decreases the performance of the
decoder. However, the relative influence of the parameter
ξ is small.
3.1 Dependence on level of correlation
As the advantage of the Ising Decoder over the Inde-
pendent Decoder stems from its ability to take advantage
of information contained in pairwise correlations, we ex-
amined the dependence of this advantage on the average
strength of correlation. Although we have set the average
level of correlation to what has traditionally been thought
to be a reasonable level for cells in the same neighbour-
hood (Zohary & Shadlen 1994), there is an ongoing de-
bate about the level and stimulus dependence of correla-
Figure 6: Decoder performance dependence on parame-
ters ξ, γ. With increasing γ the decoding performance is
enhanced. The tuning width heterogeneity as specified by
ξ reduces the performance of the decoder (values from up-
per left to lower right curve ξ = {0, 0.2, 0.4, 0.6, 0.8, 1}).
Data for 70 cells, TAPwd algorithm used to train the de-
coder.
tion relevant for cortical function (Renart et al. 2010, Bair
et al. 2001, Nase et al. 2003, Ecker et al. 2010). This
is of course critical for the performance of the Ising de-
coder. If there were no (noise) correlations present in the
data, i.e. an independent decoder were the correct model
for the data, there would be no benefit to using any de-
coder including correlation such as the Ising decoder. In
fact, any decoder including correlations would most likely
perform worse in practice than an independent decoder,
as due to finite sampling effects one would most likely
falsely estimate some small correlation in the data. Over-
fitting to these falsely learned correlations would then re-
sult in a performance decrease compared to an indepen-
dent decoder, which by construction does not include any
correlations, i.e. would implement the correct model.
This effect be seen in Figure 7, where we have varied
the mean absolute correlation strength as outlined in the
methods section. It can be seen that as the level of corre-
lation increases for specified spike count, the independent
decoder loses some discriminative capacity. The Ising de-
coders, however, take advantage of the higher level and
spread of correlation values for discrimination between
11
Figure 7:
Ising decoder benefits from a correlation
regime. Upper panel: Fraction correct as a function of
the measured average absolute Pearson correlation coeffi-
cient for four decoding algorithms. Lower panel: Average
Standard Deviation of the Pearson Correlation coefficient
vs. average Pearson correlation coefficient. All averages
taken over 20 simulations; 70 cells; basic model; correla-
tions measured with 10000 samples.
stimuli. As described in Methods, at the lower end of this
curve the underlying latent Gaussian has an identity corre-
lation matrix, and thus the individual neurons are actually
independent, although the average measured absolute cor-
relation does not completely decrease to zero. This pro-
vides an explanation of why the Ising decoder fails to beat
the independent decoder: by assuming the measured cor-
relations are due to the true distribution, it overfits to these
correlations, and thus performs worse compared to the
(by construction correct) independent model. It should be
noted, however, that the performance drop of the Ising de-
coder compared to the independent decoder is small even
for a regime where cells are effectively independent.
To mimic the richer correlation structures potentially
found in real neural recordings, we performed further test-
ing. We simulated larger ensembles of neurons, while
keeping the number of observed cells fixed, thus effec-
tively creating "hidden" neurons. The visible neurons
were simulated as described in methods, while for every
unobserved cell, the preferred tuning direction was cho-
Figure 8: Decoder robustness to correlations due to hid-
den units. Fraction Correct f c for 100 cells, normalized to
the fraction correct f c0 where 100 out of 100 cells are vis-
ible, is displayed as a function of the ratio between Chidden
hidden cells and Cvisible observed cells. For simulations
the basic model was used.
sen randomly according to a uniform distribution. By us-
ing this scheme we could assure that the observed Cvisible
cells always had the same characteristics except the cor-
relation structure, which was effectively changed by the
introduction of Chidden unobserved neurons. It should be
noted that introducing hidden cells alters not only the sec-
ond order statistics but also higher order correlations in
the data, thus providing a much richer statistical structure
in the data.
The different decoding (and training) methods vary in
their robustness towards the introduction of such higher
order correlation structure (Fig. 8). While the indepen-
dent decoder is relatively unaffected by the introduction
of additional unobserved cells, the Ising decoder model
is more sensitive to such changes. However, significant
differences between the different training strategies can
be observed: in our case the rMPFL method showed the
least performance drop (about 10%) in a scenario where
approximately 100 out of 500 cells were observed, while
the fraction correct for the standard TAP approach drops
by more than 15% in this scenario. TAPwd is less af-
fected than TAP, consistent with its inclusion of the lead-
ing terms for the next order expansion.
12
rMPFLAverage Standard DeviationPearson CoefficientAverage Absolute Pearson Coefficient10−210−11001010.60.70.80.91Chidden/Cvisiblefc / fc0 IndTAPTAPwdrMPFL 3.2 Confusion Matrix Analysis
The confusion matrix provides complete information
about the decoding error distribution. Confusion matri-
ces for the Ising decoder and the Independent decoder
respectively are shown in Fig. 9. The Ising decoder has
higher diagonal terms, corresponding to better decoding
accuracy. The overall appearance of the Ising decoder
confusion matrix is fairly similar to the independent de-
coder. However, by comparing the two confusion ma-
trices it can be seen that the Ising decoder mainly gains
its performance benefit over the independent decoder by
avoiding confusion of adjacent stimuli. This shows that
as the difference in adjacent stimulus directions becomes
less with an increasing number of stimuli, the Ising model
decoder can utilize the correlation patterns to enhance the
decoding accuracy, i.e.
to distinguish between adjacent
stimulus directions more precisely. This effect of course
depends on the correlation model used - here the correla-
tion between each pair of neurons was resampled for each
stimulus in the simulation.
If (noise) correlations were
not at all stimulus-dependent, or if the model was quite
different, then the Ising decoder may not be able to take
advantage of this potential performance advantage.
3.3 Comparison with linear decoding
While most work in the Brain-Machine Interface litera-
ture has focused on continuous decoders, there has been
some work on the discrete decoding problem, although
to date with a focus on the analysis of either contin-
uous data, such as electroencephalographic (EEG) data
(Wolpaw et al. 2002), or on longer time windows of multi-
electrode array data, in which spike counts are far from
binary (Santhanam et al. 2009). As the discrete decoding
problem can be viewed as a classification problem (in the
same sense as the continuous decoding problem can be
seen as a regression problem), it is of interest to compare
the performance of our approach with traditional classi-
fication approaches such as the Optimal Linear Classifer
(OLC).
Following Bishop (2007), each stimulus class can be
described by its own linear model, so that
yk = wT
s (cid:126)σ + wk0 ,
(22)
where s = 1, . . . , S. Using a 1-of-S binary coding
scheme (i.e. we denote stimulus class si by a "target" col-
umn vector t with all zeros except the i-th entry, which is
one) the weights ws can be trained such as to minimize a
sum of squares error function for the target stimulus vec-
tor.
This is done in Fig. 10. It can be seen that, under the
conditions we test here (10000 trials, simulated data as
described previously for the basic model), the OLC un-
derperforms both the Ising (as exemplified by the TAPwd
approach) and Independent classifiers. The former is not
unexpected, but it may seem initially counter-intuitive that
the OLC does not yield identical performance to the In-
dependent decoder, as the latter is in effect performing a
linear classification.
However, the following differences must be noted for
these two algorithms. As their implementation details
differ, they may have markedly different sensitivity to
limited sampling - the independent decoder, as we have
constructed it here (product of marginals) has remark-
able sampling efficiency. Most importantly however, the
OLC decoder assumes by construction a Gaussian error
in the stimulus class target vectors, i.e. it corresponds to
a maximum likelihood decoding when assuming that the
target vectors follow a Gaussian conditional distribution
(Bishop 2007). This assumption is clearly not valid in the
case of binary target vectors. Therefore the failure of the
OLC decoder should not be surprising.
4 Discussion
We have demonstrated, for the first time, the use of the
Ising model to decode discrete stimulus states from simu-
lated large-scale neural population activity. To do this, we
have had to overcome several technical obstacles, namely
the poor scaling properties of previously used algorithms
for learning model parameters, and similarly the poor
scaling behavior of methods for estimating the partition
function, which although not necessary for some appli-
cations of the Ising model in neuroscience, is required
for decoding. The Ising model has one particular advan-
tage over a simpler independent decoding algorithm:
it
can take advantage of stimulus dependence in the cor-
relation structure of neuronal responses, where it exists.
With the aid of a statistical simulation of neuronal ensem-
ble spiking responses in the mouse visual cortex, we have
13
models might not necessarily provide an exact descrip-
tion of neural spike train statistics, for decoding purposes
we only require that we can approximate their distribution
well enough to achieve good decoding performance.
The development of discrete neural population decod-
ing algorithms has two motivations. The first motivation
is the desire to develop brain-computer communication
devices for cognitively intact patients with severe motor
disabilities (Mak & Wolpaw 2009).
In this type of ap-
plication, an algorithm such as those we describe could
be used together with multi-electrode brain recordings to
allow the user to select one of a number of options (for
instance a letter from a virtual keyboard), or even in the
longer term to communicate sequences of symbols from
an optimized code directly into a computer system or
communications protocol. Given the short timescale to
which the Ising decoder can be applied (we have fixed
this at 20 ms here), and sufficiently large recorded ensem-
bles to saturate decoder performance, very high bit rates
could potentially be achieved.
The second motivation is more scientific: to use such
decoding algorithms to probe the organization and mech-
anisms of information processing in neural systems.
It
should be immediately be apparent that the Ising decoder
and related models can be used to ask questions about
the neural representation of sensory stimuli, motor states
or other behavioral correlates, by comparing decoding
performance under different sets of assumptions (for in-
stance, by changing the constraints in an Ising model to
exclude correlations, include correlations within 50 µm,
etc). This (commonly referred to as the "encoding prob-
lem") is essentially the same use to which Shannon in-
formation theory has been applied in neuroscience (see
e.g. Schultz et al. (2009) for a recent example), with
simply a different summary measure. Use of decoding
performance may be an intuitively convenient way to ask
such questions, but it is still asking exactly the same ques-
tion. However, there are other uses to which such algo-
rithms can be applied. For instance, combining sensory
and learning/memory experimental paradigms, once a de-
coder has been trained, it could be used subsequently to
read out activity patterns in different brain states such as
sleep, or following some period of time - for instance, to
"read out memories" by decoding the patterns of activ-
ity that represent them. The decoding approach may thus
have much to offer the study of information processing in
Figure 10: Comparison of the TAPwd Ising Decoder and
Independent Decoder with the Optimal Linear Classifier
(OLC).
demonstrated that correlational information can be taken
advantage of for decoding the activity of neuronal ensem-
bles of size in the hundreds by several algorithms, includ-
ing mean field methods from statistical physics and the
rMPFL algorithm.
Ising models have gained much attraction recently in
neuroscience to describe the spike train statistics of neu-
ral ensembles (Schneidman et al. 2006, Shlens et al. 2006,
Shlens et al. 2009). However, these findings have largely
been made only in relatively small neural ensembles (typ-
ically a few tens of cells), from which an extrapolation to
larger ensemble sizes might not be wise (Roudi, Niren-
berg & Latham 2009, Roudi, Aurell & Hertz 2009). The
principal reason for this limit has been the poor scaling of
the computational load of fitting the Ising model param-
eters, when algorithms such as Boltzmann learning are
used. Moreover, new findings suggests that pairwise cor-
relations (and thus Ising models) might not be sufficient to
predict spike patterns of small scale local clusters of neu-
rons (< 300µm apart), which have been observed to pro-
vide evidence of higher order interactions (Ohiorhenuan
et al. 2010). While the formalism for higher order mod-
els may be similar, scaling properties are guaranteed to be
even worse. There is thus the pressing need to develop
better algorithms for learning the parameters of Ising and
Ising-like models.
It should be noted that while Ising
14
1011020.50.550.60.650.70.750.80.850.9Number of CellsFraction Correct TAPwdIndependentOLC neural circuits.
Our results show that decoding performance is criti-
cally dependent on the sample size used for training the
decoder, as relatively precise characterization of pairwise
correlations is needed to fit a model that matches the sta-
tistical structure of as-of-yet unobserved data well. In the
"encoding problem", such finite sampling constraints re-
sult in a biased estimate of the entropy of the system. For
the decoding problem considered here, finite sampling
leads to overfitting of the model to the observed training
data, so that it does not generalize well to the unobserved
data, and accordingly fails to predict stimulus classes cor-
rectly during test trials. Such finite sampling constraints
mean that below a particular sampling size - which we
found to be 400 trials for 70 neurons in one particular
example we studied - there is no point in using a model
which attempts to fit pairwise (or above) correlations, one
may as well just use an independent model. This has im-
plications for experimental design. However, it should be
noted that the real brain has no such limitation - in effect,
many thousands or millions of trials are available over de-
velopment, and so a biological system should certainly be
capable of learning the correct correlations from the data
(Bi & Poo 2001) and thus may well be able to operate in a
regime where decoding benefits substantially from known
correlation structure.
We have shown that incorporating correlations in the
decoding process might be especially relevant for 'hard'
decoding problems, i.e. multi-class discrimination prob-
lems in which stimuli are not easily distinguishable by
just observing individual neuronal firing rates. In this sce-
nario including correlations could be a means to enhance
the precision of the decoding process by increasing the
discriminability between adjacent or similar stimuli. In-
cluding correlations could make the pattern distribution
more flat, or uniform, with low firing rates, leading to
greater energy efficiency of information coding (Baddeley
et al. 1997).
We must insert a note of caution concerning the pres-
ence of higher order correlations in data to be fitted. Such
higher order correlations might arise simply from the
presence of a large number of "hidden" neurons whose
activity has not been recorded, but which have a substan-
tial influence on the activity of those neurons recorded.
This is likely to occur frequently in real neurophysiolog-
ical situations. The performance of pairwise correlation
decoder models, such as the Ising model, is necessarily
affected detrimentally by such effects, as shown in Fig. 8.
Interestingly, an independent decoder is less affected (al-
though it may also be capturing less information anyway).
Obviously, it is possible to make use of higher order mod-
els to alleviate this problem, but a penalty then has to be
paid in sampling terms.
We note that the primary use of the Ising model in
neuroscience so far has been to model the empirical
statistics of neural spike train ensembles (Schneidman
et al. 2006, Shlens et al. 2006). There is of course no
requirement that a good decoder is also a good model of
neural spike train statistics - what matters is its perfor-
mance on the test dataset. Nevertheless, knowing how
well the model captures spike train pattern structure may
help to build better decoders, and of course may be of par-
ticular value when those decoders are used to study neu-
ral information processing (as opposed to being used for
a practical purpose such as brain-machine interface devel-
opment). Unfortunately, a direct comparison of empirical
to model spike pattern probabilities is not experimentally
feasible for large ensemble sizes - while this is relatively
simple for 15 cells, it is far from viable for 500 cells.
Further work is needed to determine how best to evaluate
the performance of decoders at predicting empirical spike
patterns, when decoding very large neural ensembles.
One caveat to the advantage provided by correlations
of using Ising over Independent decoders is that it de-
pends entirely upon the extent to which correlations are
found to depend upon the stimulus variables of interest.
While a previous study at longer timescales has found cor-
relations to improve neural decoding (Chen et al. 2006),
the jury is still out on the prevalence of stimulus de-
pendence of pairwise and higher order correlations in
the mammalian cortex. Stimulus-dependent correlations
have been found in mouse (Nase et al. 2003), cat (Das
& Gilbert 1999, Samonds & Bonds 2005) and monkey
(Kohn & Smith 2005, Kohn et al. 2009), where they
have been shown to contribute to information encoding
(Montani et al. 2007), but most recordings to date have
sampled relatively sparsely from the local cortical circuit,
due to limitations in multi-electrode array hardware. It is
possible that if one were to be able to record from a greater
proportion of neurons in the local circuit, then stronger
stimulus-dependent correlations might be observable.
In an experimental setting the performance could also
15
be enhanced by choosing an optimal timebin-width, a
question that we have not addressed in this paper. How-
ever care is needed for choosing the right bin-width. As
shown by Roudi et al.
(Roudi, Aurell & Hertz 2009,
Roudi, Nirenberg & Latham 2009), a small bin-width is
likely to yield a good fit, however choosing a too small
bin-width invalidates the underlying assumption of uncor-
related time-bins. Choosing a too large time-bin makes it
however harder to find a good fit for the model and addi-
tionally may violate the assumptions of binary responses.
A number of avenues present themselves for future de-
velopment of decoding algorithms. Firstly, algorithms
for reducing model dimensionality without losing dis-
criminatory power, may prove advantageous.
These
may include graph and hypergraph theoretic techniques
(Aghagolzadeh et al. 2010) for pruning out uninformative
dimensions (edges and nodes), and factor analysis meth-
ods for modeling conditional dependencies (Santhanam
et al. 2009). Such an approach may be particularly ad-
vantageous when experimental trials are limited, as the
dimensionality of the parameter set is the main reason for
Ising model performance not exceeding the Independent
model for limited trials. One difficulty with the use of
graph pruning approaches is that the usual pairwise cor-
relation matrix of neural recordings, unlike the graph in
many network analysis problems, tends not to be sparse.
It is of course a functional, as opposed to synaptic, con-
nectivity matrix, and one reason for this lack of sparseness
is its symmetric nature. It has recently been proposed that
the symmetry property of the Jij matrix can be relaxed in
the context of the (non-equilibrium) Kinetic Ising model,
which also provides a convenient way to take into account
space-time dependencies, or causal relationships (Hertz
et al. 2010, Roudi & Hertz 2011). Use of the Kinetic Ising
model framework for decoding would appear to be an in-
teresting future direction to pursue.
New experimental technologies are yielding increas-
ingly high dimensional multivariate neurophysiological
datasets, usually without concomitant increases in the du-
ration of data that can be collected. However, there is
some reason for optimism that we will be able to develop
new data analysis methods capable of taking advantage of
this data. Maximum entropy approaches to the fitting of
structured parametric models such as the Ising model and
its extensions would appear to be one approach likely to
yield progress.
Acknowledgements
We thank Phil Bream, H´el`ene Seiler and Yang Zhang
for their contributions to earlier work leading up to that
reported here, and Aman Saleem for useful discussions
and comments on this manuscript. We also thank Yasser
Roudi for useful comments on the TAP equations, and
Jascha Sohl-Dickstein, Peter Battaglino, and Michael R
DeWeese for useful MATLAB code and discussion of the
MPFL technique. This work was supported by EPSRC
grant EP/E002331/1 to SRS.
References
Aghagolzadeh, M., Eldawlatly, S. & Oweiss, K. (2010),
'Synergistic Coding by Cortical Neural Ensembles.',
IEEE Trans Inf Theory 56(2), 875–899.
Baddeley, R., Abbott, L. F., Booth, M. C., Sengpiel, F.,
Freeman, T., Wakeman, E. A. & Rolls, E. T. (1997),
'Responses of neurons in primary and inferior tem-
poral visual cortices to natural scenes.', Proc Biol
Sci 264(1389), 1775–83.
Bair, W. & Koch, C. (1996), 'Temporal precision of spike
trains in extrastriate cortex of the behaving macaque
monkey.', Neural Comput 8(6), 1185–202.
Bair, W., Zohary, E. & Newsome, W. T. (2001), 'Cor-
related firing in macaque visual area MT:
time
scales and relationship to behavior.', J Neurosci
21(5), 1676–97.
Berger, A. L., Pietra, V. J. D. & Pietra, S. A. D. (1996),
'A maximum entropy approach to natural language
processing', Comput. Linguist. 22(1), 39–71.
Bi, G.-q. & Poo, M.-m. (2001), 'Synaptic modification by
correlated activity: hebb's postulate revisited', An-
nual Review of Neuroscience 24(1), 139–166.
Bishop, C. M. (2007), Pattern Recognition and Machine
Learning (Information Science and Statistics), 1st
ed. 2006. corr. 2nd printing edn, Springer.
Broderick, T., Dud´ık, M., Tkacik, G., Schapire, R. E. &
Bialek, W. (2007), 'Faster solutions of the inverse
pairwise Ising problem', http://arxiv.org, Identifier:
0712.2437v2.
16
Butts, D. A., Weng, C., Jin, J., Yeh, C.-I. I., Lesica, N. A.,
Alonso, J.-M. M. & Stanley, G. B. (2007), 'Tempo-
ral precision in the neural code and the timescales of
natural vision.', Nature 449(7158), 92–5.
Carr, C. E. (1993), 'Processing of temporal information in
the brain', Annu Rev Neurosci 16, 223–43.
Kappen, H. J. & Rodrguez, F. B. (1998), 'Efficient Learn-
ing in Boltzmann Machines Using Linear Response
Theory', Neural Computation 10(5), 1137–1156.
Kohn, A. & Smith, M. A. (2005), 'Stimulus dependence
of neuronal correlation in primary visual cortex of
the macaque.', J Neurosci 25(14), 3661–73.
Chen, Y., Geisler, W. S. & Seidemann, E. (2006), 'Op-
timal decoding of correlated neural population re-
sponses in the primate visual cortex.', Nat Neurosci
9(11), 1412–20.
Kohn, A., Zandvakili, A. & Smith, M. A. (2009), 'Cor-
relations and brain states:
from electrophysiol-
ogy to functional imaging.', Curr Opin Neurobiol
19(4), 434–8.
Das, A. & Gilbert, C. D. (1999), 'Topography of con-
textual modulations mediated by short-range inter-
actions in primary visual cortex', Nature 399, 643–
4.
Ecker, A. S., Berens, P., Keliris, G. A., Bethge, M., Lo-
gothetis, N. K. & Tolias, A. S. (2010), 'Decorrelated
Neuronal Firing in Cortical Microcircuits', Science
327(5965), 584–587.
Foldi´ak, P. (1993), The 'ideal homunculus': statistical in-
ference from neural population responses, in 'Com-
putation and Neural Systems', Kluwer Academic
Publishers, Norwell, MA, pp. 55–60.
Gawne, T. J., Kjaer, T. W., Hertz, J. A. & Richmond,
B. J. (1996), 'Adjacent visual cortical complex cells
share about 20% of their stimulus-related informa-
tion.', Cereb Cortex 6(3), 482–9.
Hertz, J., Roudi, Y., Thorning, A., Tyrcha, J., Aurell, E.
& Zeng, H. L. (2010), 'Inferring network connectiv-
ity using kinetic Ising models', BMC Neuroscience
11(Suppl 1), P51.
Higham, N. J. (2002), 'Computing the nearest correla-
tion matrix–a problem from finance', IMA Journal
of Numerical Analysis 22(3), 329.
Huang, F. & Ogata, Y. (2001), 'Comparison of two meth-
ods for calculating the partition functions of var-
ious spatial statistical models', Australian & New
Zealand Journal of Statistics 43(1), 47–65.
Jaynes, E. T. (1957), 'Information Theory and Statistical
Mechanics', Phys. Rev. 106(4), 620–630.
Macke, J. H., Berens, P., Ecker, A. S., Tolias, A. S. &
Bethge, M. (2009), 'Generating Spike Trains with
Specified Correlation Coefficients', Neural Compu-
tation 21(2), 397–423.
Mak, J. N. & Wolpaw, J. R. (2009), 'Clinical Applications
of Brain-Computer Interfaces: Current State and Fu-
ture Prospects.', IEEE Rev Biomed Eng 2, 187–199.
Montani, F., Kohn, A., Smith, M. A. & Schultz, S. R.
(2007), 'The Role of Correlations in Direction and
Contrast Coding in the Primary Visual Cortex', J.
Neurosci. 27(9), 2338–2348.
Nase, G., Singer, W., Monyer, H. & Engel, A. K. (2003),
'Features of Neuronal Synchrony in Mouse Visual
Cortex', J Neurophysiol 90(2), 1115–1123.
Niell, C. M. & Stryker, M. P. (2008), 'Highly Selective
Receptive Fields in Mouse Visual Cortex', J. Neu-
rosci. 28(30), 7520–7536.
Ogata, Y. & Tanemura, M. (1984), 'Likelihood analysis
of spatial point patterns', Journal of the royal statis-
tical Society. Series B. 46(3), 496–518.
Ohiorhenuan, I. E., Mechler, F., Purpura, K. P., Schmid,
A. M., Hu, Q. & Victor, J. D. (2010), 'Sparse cod-
ing and high-order correlations in fine-scale cortical
networks', Nature 466(7306), 617–621.
Oram, M. W., Fldik, P., Perrett, D. I., Oram, M. W. & Sen-
gpiel, F. (1998), 'The 'Ideal Homunculus': decoding
neural population signals', Trends in Neurosciences
21(6), 259 – 265.
17
Panzeri, S. & Schultz, S. R. (2001), 'A unified approach
to the study of temporal, correlational, and rate cod-
ing.', Neural Comput 13(6), 1311–49.
Panzeri, S., Schultz, S. R., Treves, A. & Rolls, E. T.
(1999), 'Correlations and the encoding of infor-
mation in the nervous system', Proceedings of the
Royal Society of London. Series B: Biological Sci-
ences 266(1423), 1001–1012.
Panzeri, S., Treves, A., Schultz, S. & Rolls, E. T. (1999),
'On decoding the responses of a population of neu-
rons from short time windows.', Neural Comput
11(7), 1553–77.
Petersen, R. S., Panzeri, S. & Diamond, M. E. (2001),
'Population Coding of Stimulus Location in Rat So-
matosensory Cortex', Neuron 32(3), 503 – 514.
Plefka, T. (2006), 'Expansion of the Gibbs potential for
quantum many-body systems: General formalism
with applications to the spin glass and the weakly
nonideal Bose gas', Phys. Rev. E 73(1), 016129.
Pola, G., Thiele, A., Hoffmann, K. & Panzeri, S. (2003),
'An exact method to quantify the information trans-
mitted by different mechanisms of correlational
coding', Network-Computation in Neural Systems
14(1), 35–60.
Reich, D. S., Mechler, F. & Victor, J. D. (2001), 'Indepen-
dent and Redundant Information in Nearby Cortical
Neurons', Science 294(5551), 2566–2568.
Renart, A., de la Rocha, J., Bartho, P., Hollender, L.,
Parga, N., Reyes, A. & Harris, K. D. (2010), 'The
asynchronous state in cortical circuits.', Science
327(5965), 587–90.
Roudi, Y., Aurell, E. & Hertz, J. A. (2009), 'Statistical
physics of pairwise probability models', Frontiers in
Computational Neuroscience 3:22, 1–15.
Roudi, Y. & Hertz, J. (2011), 'Mean field theory for non-
equilibrium network reconstruction', Physical Re-
view Letters 106, 048702.
Roudi, Y., Nirenberg, S. & Latham, P. E. (2009),
'Pairwise Maximum Entropy Models for Study-
ing Large Biological Systems: When They Can
Work and When They Can't', PLoS Comput Biol
5(5), e1000380.
Roudi, Y., Tyrcha, J. & Hertz, J. (2009), 'Ising model for
neural data: Model quality and approximate meth-
ods for extracting functional connectivity', Phys.
Rev. E 79(5), 051915 (12 pages).
Samonds, J. M. & Bonds, A. B. (2005), 'Gamma oscil-
lation maintains stimulus structure-dependent syn-
chronization in cat visual cortex.', J Neurophysiol
93(1), 223–36.
Santhanam, G., Yu, B. M., Gilja, V., Ryu, S. I., Afshar, A.,
Sahani, M. & Shenoy, K. V. (2009), 'Factor-analysis
methods for higher-performance neural prostheses.',
J Neurophysiol 102(2), 1315–30.
Schneidman, E., Berry II, M. J., Segev, R. & Bialek, W.
(2006), 'Weak pairwise correlations imply strongly
correlated network states in a neural population',
Nature 440(7087), 1007–1012.
Schultz, S. R., Kitamura, K., Post-Uiterweer, A., Krupic,
J. & Hausser, M. (2009), 'Spatial Pattern Coding
of Sensory Information by Climbing Fiber-Evoked
Calcium Signals in Networks of Neighboring Cere-
bellar Purkinje Cells', J. Neurosci. 29(25), 8005–
8015.
Schultz, S. R. & Panzeri, S. (2001), 'Temporal correla-
tions and neural spike train entropy.', Phys Rev Lett
86(25), 5823–6.
Seiler, H., Zhang, Y., Saleem, A., Bream, P., Apergis-
Schoute, J. & Schultz, S. R. (2009), 'Maximum en-
tropy decoding of multivariate neural spike trains',
BMC Neuroscience 10(Suppl 1), P107.
Shlens, J., Field, G. D., Gauthier, J. L., Greschner, M.,
Sher, A., Litke, A. M. & Chichilnisky, E. J. (2009),
'The Structure of Large-Scale Synchronized Firing
in Primate Retina', J. Neurosci. 29(15), 5022–5031.
Shlens, J., Field, G. D., Gauthier, J. L., Grivich,
M. I., Petrusca, D., Sher, A., Litke, A. M. &
Chichilnisky, E. J. (2006), 'The Structure of Multi-
Neuron Firing Patterns in Primate Retina', J. Neu-
rosci. 26(32), 8254–8266.
18
Sohl-Dickstein, J., Battaglino, P. & DeWeese, M. R.
(2009),
'Minimum Probability Flow Learning',
http://arxiv.org/abs/0906.4779 arXiv:0906.4779v2.
Tanaka, T. (1998), 'Mean-field theory of Boltzmann ma-
chine learning', Phys. Rev. E 58(2), 2302–2310.
Thouless, D. J., Anderson, P. W. & Palmer, R. G. (1977),
'Solution of solvable model of a spin glass', Philos
Mag 35(3), 593–601.
Wolpaw,
J. R., Birbaumer, N., McFarland, D. J.,
Pfurtscheller, G. & Vaughan, T. M. (2002), 'Brain-
computer interfaces for communication and con-
trol.', Clin Neurophysiol 113(6), 767–91.
Zohary, E. & Shadlen, M. (1994), 'Correlated neuronal
discharge rate and its implications for psychophysi-
cal performance', Nature 370(6485), 140–143.
19
Figure 1: Tuning curves for a neural ensemble of 50 cells for different heterogeneity parameters γ, ξ. The shown
spiking rates correspond to the transient rates. a Tuning curves for basic model, γ = ξ = 0. b Tuning curves with
heterogeneous firing rates (γ = 1), but fixed tuning widths (ξ = 0) (NB the tuning width is defined relative to the
spontaneous activity). c Tuning curves with heterogeneous tuning widths (ξ = 1), but fixed firing rates (γ = 0). d
Tuning curves for fully heterogeneous scenario (γ = ξ = 1)
20
Figure 2: Neural ensemble responses simulated for basic model (γ = ξ = 0) for 50 cells. a Correlation matrices for
4 stimuli (computed with 100000 samples). Diagonal terms set to zero for visualization purposes only. b Simulated
population neural responses over 1000 trials for two different stimuli, with black indicated a spiking response from a
neuron on a given trial.
21
CellCellPearson CoefficientStimulus 1Stimulus 2Stimulus 3Stimulus 4CellTrialCellTrialab Figure 3: Performance of decoding algorithms for the basic model. a Fraction of correct decodings versus neural
ensemble size, for a training dataset size of 9000 samples per stimulus. b The relationship between fraction correct
and mutual information I(s, s) for varying stimulus set and ensemble size (C = {50, 80, 110, 140, 170, 200} varying
from bottom left to top right for each symbol type). Triangles denote the performance of the Independent decoder,
squares the TAP Ising Decoder with diagonal weight trick. c The dependence of decoding performance on stimulus
set size for 70 cells. i TAP, TAPwd, Independent and rMPFL decoders compared to random selection of stimuli. This
is replotted in ii as the gain in fraction correct over chance performance, η = pdec/pguess, making the performance
saturation for the independent decoder as problem difficulty increases more apparent. d Dependence of decoder
performance on training set sample size, for 4 stimuli and a neural ensemble size of 70. i Fraction correct as a function
of number of training samples. Below 450 training samples the Ising decoder fails to better the independent decoder.
ii Relation between mean relative covariance error E as a measure of finite sampling effects and fraction correct.
22
aciiiiiidIndChancebMean Relative Covariance Error Figure 9: Decoder confusion matrices. The case illustrated is for 16 Stimuli, a 200-neuron ensemble, under the basic
model (γ = ξ = 0). Relative frequency as an approximation to the conditional probability with which each stimulus
is decoded, for 10000 stimulus presentations. a Ising model decoder (TAPwd). b Independent decoder. c A difference
plot between a and b.
23
relative frequency |
1809.06441 | 3 | 1809 | 2019-01-07T16:30:50 | The physics of brain network structure, function, and control | [
"q-bio.NC",
"physics.bio-ph"
] | The brain is a complex organ characterized by heterogeneous patterns of structural connections supporting unparalleled feats of cognition and a wide range of behaviors. New noninvasive imaging techniques now allow these patterns to be carefully and comprehensively mapped in individual humans and animals. Yet, it remains a fundamental challenge to understand how the brain's structural wiring supports cognitive processes, with major implications for the personalized treatment of mental health disorders. Here, we review recent efforts to meet this challenge that draw on intuitions, models, and theories from physics, spanning the domains of statistical mechanics, information theory, and dynamical systems and control. We begin by considering the organizing principles of brain network architecture instantiated in structural wiring under constraints of symmetry, spatial embedding, and energy minimization. We next consider models of brain network function that stipulate how neural activity propagates along these structural connections, producing the long-range interactions and collective dynamics that support a rich repertoire of system functions. Finally, we consider perturbative experiments and models for brain network control, which leverage the physics of signal transmission along structural wires to infer intrinsic control processes that support goal-directed behavior and to inform stimulation-based therapies for neurological disease and psychiatric disorders. Throughout, we highlight several open questions in the physics of brain network structure, function, and control that will require creative efforts from physicists willing to brave the complexities of living matter. | q-bio.NC | q-bio | The physics of brain network structure, function, and control
Christopher W. Lynn1 & Danielle S. Bassett1,2,3,4,5,∗
1Department of Physics & Astronomy, College of Arts & Sciences, University of Pennsylvania,
Philadelphia, PA 19104, USA
2Department of Bioengineering, School of Engineering & Applied Science, University of Penn-
sylvania, Philadelphia, PA 19104, USA
3Department of Electrical & Systems Engineering, School of Engineering & Applied Science,
University of Pennsylvania, Philadelphia, PA 19104, USA
4Department of Neurology, Perelman School of Medicine, University of Pennsylvania, Philadel-
phia, PA 19104, USA
5Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadel-
phia, PA 19104, USA
9
1
0
2
n
a
J
7
]
.
C
N
o
i
b
-
q
[
3
v
1
4
4
6
0
.
9
0
8
1
:
v
i
X
r
a
1
The brain is a complex organ characterized by heterogeneous patterns of structural con-
nections supporting unparalleled feats of cognition and a wide range of behaviors. New
noninvasive imaging techniques now allow these patterns to be carefully and comprehen-
sively mapped in individual humans and animals. Yet, it remains a fundamental challenge
to understand how the brain's structural wiring supports cognitive processes, with major
implications for the personalized treatment of mental health disorders. Here, we review re-
cent efforts to meet this challenge that draw on intuitions, models, and theories from physics,
spanning the domains of statistical mechanics, information theory, and dynamical systems
and control. We begin by considering the organizing principles of brain network architec-
ture instantiated in structural wiring under constraints of symmetry, spatial embedding, and
energy minimization. We next consider models of brain network function that stipulate how
neural activity propagates along these structural connections, producing the long-range in-
teractions and collective dynamics that support a rich repertoire of system functions. Finally,
we consider perturbative experiments and models for brain network control, which leverage
the physics of signal transmission along structural wires to infer intrinsic control processes
that support goal-directed behavior and to inform stimulation-based therapies for neurolog-
ical disease and psychiatric disorders. Throughout, we highlight several open questions in
the physics of brain network structure, function, and control that will require creative efforts
from physicists willing to brave the complexities of living matter.
2
It is our good fortune as physicists to seek to understand the nature of the observable world
around us. In this inquiry, we need not reach to contemporary science to appreciate the fact that
our perception of the world around us is inextricably linked to the world within us: the mind.
Indeed, even Aristotle c. 350 B.C. noted that it is by mapping the structure of the world that the
human comes to understand their own mind 1. "Mind thinks itself because it shares the nature of
the object of thought; for it becomes an object of thought in coming into contact with and thinking
its objects, so that mind and object of thought are the same" 2. Over the ensuing 2000-plus years,
it has not completely escaped notice that the mappers of the world have unique contributions to
offer the mapping of the mind (from Thales of Miletus, c. 624 -- 546 B.C., to Leonardo Da Vinci,
1452 -- 1519). More recently, it is notable that nearly all famous physicists of the early 20th century
-- Albert Einstein, Niels Bohr, Erwin Schroedinger, Werner Heisenberg, Max Born -- considered
the philosophical implications of their observations and theories 3.
sophical musings turned to particularly conspicuous empirical contributions at the intersection of
In the post-war era, philo-
neuroscience and artificial intelligence, spanning polymath John von Neumann's work enhancing
our understanding of computational architectures 4 and physicist John Hopfield's invention of the
associative neural network, which revolutionized our understanding of collective computation 5.
In the contemporary study of the mind and its fundamental organ -- the brain -- nearly all
of the domains of physics, perhaps with the exception of relativity, are not only relevant but truly
essential, motivating the early coinage of the term neurophysics some four decades ago 6. The
fundamentals of electricity and magnetism prove critical for building theoretical models of neu-
rons and the transmission of action potentials 7. These theories are being increasingly informed by
mechanics to understand how force-generating and load-bearing proteins bend, curl, kink, buckle,
constrict, and stretch to mediate neuronal signaling and plasticity 8. Principles from thermodynam-
ics come into play when predicting how the brain samples the environment (action) or shifts the
distribution of information that it encodes (perception) 9. Collectively, theories of brain function
are either buttressed or dismantled by imaging, with common tools including magnetic resonance
imaging 10 and magnetoencephalography 11, the latter being built on superconducting quantum in-
terference devices and next-generation quantum sensors that can be embedded into a system that
3
can be worn like a helmet, revolutionizing our ability to measure brain function while allowing
free and natural movement 12. Moreover, recent developments in nanoscale analysis tools and in
the design and synthesis of nanomaterials have generated optical, electrical, and chemical methods
to explore brain function by enabling simultaneous measurement and manipulation of the activity
of thousands or even millions of neurons 13. Beyond its relevance for continued imaging advance-
ments 14, optics has come to the fore of neuroscience over the last decade with the development
of optogenetics, an approach that uses light to alter neural processing at the level of single spikes
and synaptic events, offering reliable, millisecond-timescale control of excitatory and inhibitory
synaptic transmission 15.
Such astounding advances, enabled by the intersection of physics and neuroscience, have
motivated the construction of a National Brain Observatory at the Argonne National Laboratory
(Director: Peter Littlewood, previously of Cavendish Laboratories) funded by the National Sci-
ence Foundation, as well as frequent media coverage including titles in the APS News such as
"Physicists, the Brain is Calling You."16 And as physicists answer the call, our understanding of
the brain deepens and our ability to mark and measure its component parts expands. Yet alongside
this growing systematization and archivation, we have begun to face an increasing realization that
it is the interactions between hundreds or thousands of neurons that generate the mind's functional
states 13. Indeed, from interactions among neural components emerge computation 17, commu-
nication 18, and information propagation 19. We can confidently state of neuroscience what Henri
Poincare, the French mathematician, theoretical physicist, and philosopher of science, states of sci-
ence generally: "The aim of science is not things themselves, as the dogmatists in their simplicity
imagine, but the relations among things; outside these relations there is no reality knowable."20 The
overarching goal of mapping these interactions in neural systems has motivated multibillion-dollar
investments across the United States (the Brain Initiative generally, and the Human Connectome
Project specifically 21), the European Union (the Blue Brain Project 22), China (the China Brain
Project 23), and Japan (Japan's Brain/MINDS project 24).
While it is clear that interactions are paramount, exactly how the functions of the mind arise
4
from these interactions remains one of the fundamental open questions of brain science 25. To the
physicist, such a question appears to exist naturally within the purview of statistical mechanics 26,
with one major caveat: the interaction patterns observed in the brain are far from regular, such as
those observed in crystalline structures, and are also far from random, such as those observed in
fully disordered systems 27. Indeed, the observed heterogeneity of interaction patterns in neural
systems -- across a range of spatial and temporal scales -- generally limits the utility of basic contin-
uum models or mean-field theories, which would otherwise comprise our natural first approaches.
Fortunately, similar observations of interaction heterogeneity have been made in other technologi-
cal, social, and biological systems, leading to concerted efforts to develop a statistical mechanics of
complex networks 28. The resultant area of inquiry includes criteria for building a network model
of a complex system 29, statistics to quantify the architecture of that network 30, models to stipulate
the dynamics that can occur both in and on a network 31 -- 33, and theories of network function and
control 34, 35.
Here, we provide a brief review for the curious physicist, spanning the network-based ap-
proaches, statistics, models, and theories that have recently been used to understand the brain.
Importantly, the interpretations that can be rationally drawn from all such efforts depend upon the
nature of the network representation 29, including its descriptive, explanatory, and predictive valid-
ity -- topics that are treated with some philosophical rigor elsewhere 36. Following a simple primer
on the nature of network models, we discuss the physics of brain network structure, beginning
with an exposition regarding measurement before turning to an exposition regarding modeling. In
a parallel line of discourse, we then discuss the physics of brain network function, followed by a
description of perturbation experiments and brain network control. In each section we separate our
remarks into the known and the unknown, the past and the future, the fact and the speculation. Our
goal is to provide an accessible introduction to the field, and to inspire the younger generation of
physicists to courageously tackle some of the most pressing open questions surrounding the inner
workings of the mind.
5
The physics of brain network structure
We begin with a discussion of the architecture, or structural wiring, of networks in the brain, fo-
cusing on the measurement and modeling of their key organizational features (see Box 1 for a
simple primer on networks). Each edge in a structural brain network represents a physical con-
nection between two elements. For example, synapses support the propagation of information
between neurons 37 and white matter tracts define physical pathways of communication between
brain regions 38. In physics, it has long been recognized that the organization of such structural
connections can determine the qualitative large-scale features of a system 28. In the Ising model, for
instance, a one-dimensional lattice remains paramagnetic across all temperatures 39, while in two
dimensions or more, the system spontaneously breaks symmetry, yielding the type of bulk magne-
tization exhibited by magnets on a refrigerator 40, 41. Similarly, the organization of structural wiring
in the brain largely determines the types of mental processes and cognitive functions that can be
supported 42 -- 46, from memory 47 -- 49 to learning 50, 51, and from vision 52 to motion 53. However,
unlike many physics applications, which assume simple lattice or random network architectures,
the wiring of the brain is highly heterogeneous, often making symmetry arguments and mean-field
descriptions far from applicable 27. While this heterogeneity presents a unique set of challenges,
in what follows we review some powerful experimental and theoretical tools that allow us to distill
the brain's structural complexity to a number of fundamental organizing principles.
[Box 1 here]
Measuring brain network structure. Some of the earliest empirical measurements of the brain's
structural connectivity can be traced to Camillo Golgi, who in 1873 soaked blocks of brain tissue
in silver-nitrate solution to provide among the first glimpses of the intricate branching of nerve
cells 54. Soon after, Santiago Ram´on y Cajal combined Golgi's method with light microscopy to
achieve stunning pictures establishing that neurons do not exist in solitude; they instead combine
to form intricate networks of physical connections 55. This notion that the brain comprises a com-
6
plex web of distinct components, known as the neuron doctrine 56, established the foundation upon
which modern network neuroscience has flourished. The introduction of the electron microscope
in the 1930s provided even more detailed measurements of the physical connections between neu-
rons. Perhaps the most impressive application remains the complete mapping of interconnections
between the 302 neurons in the nematode C. elegans 57. Since this achievement, reconstructions
of the synaptic connectivity in other animals have evolved rapidly, from a mapping of the optic
medulla in the visual system of the fruit fly Drosophila to the enumeration of connections between
950 distinct neurons in the mouse retina 52, 58. Efforts continue to press forward toward the ultimate
goal of reconstructing the neuronal wiring diagram of an entire human brain 59.
Concurrently with these achievements using electron microscopy, complimentary efforts in
tract tracing have revealed the mesoscale structure of the macaque 60, 61, cat 62, mouse 63, and
fly 64. Particularly important for our understanding of human cognition are recent advances in
noninvasive imaging that have allowed unprecedented views of the mesoscale structure of the
brain in vivo. Introduced in the 1970s, computerized axial tomography (CAT) provided among
the most detailed anatomic images of the human brain to date 65. Soon after, the development
of magnetic resonance imagining (MRI) sparked an explosion of refinements, a notable example
being diffusion tensor imaging (DTI) 66. While standard CAT and MRI techniques capture cross-
sectional images of the brain, DTI traces the diffusion of water molecules through white matter
tracts to reconstruct the large-scale neural pathways connecting distinct brain regions 67, 68. Given
measurements of the anatomical wiring connecting a set of neural elements, such as synapses
linking neurons or white matter tracts connecting brain regions, researchers can build a structural
brain network by forming edges between elements that share a physical connection (Fig. 1a).
Ongoing experimental efforts to acquire these measurements continue to provide rich network
datasets detailing the brain's structural organization.
Modeling brain network structure. A first glance at the brain's wiring reveals that it is far
from homogeneous -- a fact that is not surprising considering the array of physical, energetic, and
cognitive constraints that it is required to balance 69. To handle this heterogeneity, researchers have
7
Figure 1 Measuring and modeling brain network structure. a The measurement of brain
network structure begins with experimental data specifying the physical interconnections be-
tween neurons or brain regions. As an example, we consider a dataset of white matter tracts
measured via DTI. First, the data is discretized into non-overlapping gray matter volumes rep-
resenting distinct nodes. Then, one constructs an adjacency matrix A, where Aij represents
the connection strength between nodes i and j. This adjacency matrix, in turn, defines a
structural brain network constructed from our original measurements of physical connectivity.
b To capture an architectural feature of structural brain networks, we utilize generative net-
work models. The simplest generative network model is the Erdos -- R´enyi model, which has
no discernible non-random structure. Networks with modular structure, divided into commu-
nities with dense connectivity, are constructed using the stochastic block model. Small-world
networks, which balance efficient communication and high clustering, are generated using
the Watts -- Strogatz model. Networks with hub structure, characterized by a heavy-tailed de-
gree distribution, are typically constructed using a preferential attachment model such as the
Barab´asi -- Albert model. Spatially embedded networks, whose connectivity is constrained to
exist within a physical volume, are generated through the use of spatial network models.
8
Community structureStochastic blockSmall-world
(efficient communication)Watts -- StrogatzHub structure
(heavy-tailed deg. dist.)Barabási -- AlbertaMeasurementExample: White matter tracts (via DTI)Structural brain networkAdjacency matrixNetwork
typeModelingbGenerative
modelRandom
(no structure)Erdös -- RényiSpatially embeddedSpatial model increasingly turned to the field of network science for mathematical tools and intuitions 70, 71. The
primary goal of this interdisciplinary effort has been to distill the explosion of experimental data,
spanning structural brain networks in C. elegans 72, the mouse 73, cat 74, macaque 75, 76, and human
77, down to a number of cogent organizing principles. Here we review some important properties
that are thought to characterize structural brain networks and introduce several generative network
models that help to explain how these properties arise from underlying biological mechanisms
(Fig. 1b).
Random structure. While healthy members of a species exhibit anatomical similarities in brain
structure, the specific instantiation of physical connections in each individual is far from determin-
istic. Indeed, in vivo imaging techniques in humans, such as DTI described above, have revealed
not only stark differences in brain structure between individuals 78, but also within the same indi-
vidual over time 79, 80. Importantly, these structural differences have been linked to variability in
a wide range of behaviors 81, including empathy 82, introspection 83, fear acquisition 84, and even
political orientation 85. To study the mathematical properties of random networks, and to under-
stand the types of biological mechanisms that can give rise to qualitative structural properties, it
is useful to consider generative network models 70. The simplest and most common model for
generating random networks is the Erdos -- R´enyi (ER) model 86, wherein each pair of nodes is con-
nected independently with a fixed probability P . While the ER model has a number of interesting
mathematical properties, such as a binomial degree distribution, it has no discernible structure and
does not reflect the mechanisms by which most networks grow in the brain. Accordingly, if we
wish to understand some of the principles underlying naturally occurring brain networks, we must
consider generative models that yield networks with realistic properties.
Community structure. Perhaps the brain's most well-studied structural property is its division into
distinct anatomical regions, which are widely thought to be responsible for specialized cognitive
functions 87. Interestingly, by studying the large-scale structure of brain networks in several mam-
malian species, researchers have shown that the organization of connections tends to partition the
networks into densely-connected communities separated by sparse inter-community connectivity
9
88 -- 91. Moreover, these clusters of high connectivity closely resemble postulated anatomical subdi-
visions 89. It has therefore been argued that the so-called community structure of brain networks
segregates the brain into subnetworks with specific cognitive functions 92 -- 96. Practically speaking,
in order to extract the community structure of a real-world network, one must employ algorithms
for community detection -- a vibrant branch of research that is now applied throughout network
neuroscience 97, 98. From a complimentary perspective, to generate networks with a defined com-
munity structure, researchers predominantly use the stochastic block (SB) model, wherein nodes
are assigned to distinct communities and an edge is placed between each pair of nodes with a prob-
ability that depends on the nodes' community assignments 99, 100. Such SB networks are often used
as null models to distinguish between properties of brain networks that are implied simply by their
community structure and those that require additional biological mechanisms 70, 100.
Small-world structure. Seemingly in contradiction to their striking community structure, large-
scale brain networks also exhibit average path lengths between all nodes that are much shorter than
a typical random network 69, 101, 102. This competition between high clustering and short average
paths is thought to facilitate the simultaneous segregation and integration of information in the
brain 103, possibly minimizing the total number of computational steps needed to process external
stimuli 104, 105. Seeking an explanation for similar "small-world" topologies exhibited by other real-
world systems (most notably social networks 106), Duncan Watts and Steven Strogatz developed a
model for generating random networks with both high clustering and short average path lengths 107.
Generally, the Watts-Strogatz (WS) model supposes that small-world networks are an interpolation
between two extreme configurations: a ring lattice, wherein nodes are arranged along a circle and
connected to their k nearest neighbors on either side, and an ER random network. Notably, the
presence of small-world structure in the brain suggests that efficient communication emerges from
a finely-tuned balance of lattice-like organization and structural disorder.
Hub structure. In addition to their modular and small-world structure, many large-scale brain net-
works also feature high-degree "hubs", which form a densely interconnected structural core 108.
Acting as bridges between structurally distinct communities, these specialized hub regions are
10
thought to help minimize overall path lengths across the network 90 and facilitate the integration
of information 103. Supporting the notion of a centralized core, many studies have identified hubs
within the parietal and prefrontal regions, areas that are often active during a wide range of cog-
nitive functions 108, 109. Such core-periphery architecture is characterized by a heavy-tailed degree
distribution, such as that observed in scale-free networks, in some cases arising through preferen-
tial attachment mechanisms 110. In the Barab`asi -- Albert (BA) model 111, for instance, nodes are
added to a network in sequential order, and each new node i forms an edge with each existing node
j with a probability proportional to the degree of node j. In this way, new nodes preferentially at-
tach to existing nodes of high degree, creating a "rich club" of centralized hubs that link otherwise
distant regions of the network.
Spatial structure. Thus far, we have focused exclusively on the topological properties of brain
networks, which are thought to be driven primarily by the simultaneous functional pressures of in-
formation segregation and integration 103. However, brain networks are also physically constrained
to exist within a tight three-dimensional volume and their structural connections are metabolically
driven to minimize total wiring distance 69, 92, 105. Such physical and metabolic constraints are cap-
tured by spatial (or geometric) network models, which embed networks into three-dimensional
Euclidean space and penalize the formation of long-distance connections 70. The simplest such
model assumes that the probability of two nodes i and j forming an edge is proportional to d−α
ij ,
where dij is the physical distance between i and j, and α ≥ 0 tunes the metabolic cost associated
with constructing connections of a given length 112. If we keep the number of nodes and edges
fixed, one can see that, much like the WS model, this spatial model interpolates between a lattice-
like structure, in which nodes only connect to their nearest neighbors (α → ∞), and an ER random
network (α = 0).
Competition between structural properties. As the brain grows and adapts to changing cognitive
demands, it is widely thought that the underlying network evolves to balance the trade-off between
topological value and metabolic wiring cost 69. Thus, while the modular, small-world, heavy-
tailed, and inherently physical properties of brain networks provide simple organizing principles,
11
in reality the brain is constantly and dynamically weighing these pressures against one another.
Accordingly, an accurate generative model should aim to explain multiple real-world properties
at once 70. With this goal in mind, recent work has shown that an impressive range of topolog-
ical properties can be understood as arising from a competition between two competing factors:
a metabolic penalty for the formation of long-distance connections and a topological incentive
to connect regions with similar inputs 113. Notably, investigations of the human, C. elegans, and
mouse connectomes have revealed that the total wiring distance is consistently greater than mini-
mal, supporting the notion that brain networks weigh the costs of long-distance connections against
the functional benefits of an integrated network topology 92, 114. Together, these efforts toward a
comprehensive generative model are vital for our understanding of healthy brain network struc-
ture, with important clinical implications for the diagnosis, prognosis, prevention, and treatment
of disorders of mental health 115, 116.
The future of brain network structure. Current advances in neuroimaging techniques and net-
work science continue to expand our ability to measure and model the architecture of structural
connections in the brain. As experimental measurements become increasingly detailed, an impor-
tant direction is the bridging of brain network structure at different spatiotemporal scales 117 -- 119.
Such cross-scale approaches could link protein interaction networks within neurons to the wiring
of synaptic connectivity between neurons to mesoscale networks connecting brain regions and all
the way to social networks linking distinct organisms (Box 2). The goal of such cross-scale inte-
gration is to understand how the architecture of connectivity at each of these scales emerges from
the scale below. Practically, researchers have begun to address this goal by employing hierarchi-
cal network models 120, which treat each node at the macroscale as an entire subnetwork at the
microscale 121.
[Box 2 here]
Perhaps the most ambitious future goal is the reconstruction of the entire human connectome
at the scale of individual neurons, pressing the current boundaries of 3D electron microscopy
12
and statistical image reconstruction 59. Extensive mapping efforts in other species have revealed
notable and quantifiable neuronal diversity 122, 123, suggesting the importance of extending network
models to include non-identical units. At the mesoscale, advances in noninvasive imaging have
allowed researchers to begin tracking changes in structural connectivity over time 72, 124 -- 126. To
analyze these temporally ordered measurements, network scientists have extended standard static
graph theoretic tools to study networks with dynamically evolving connections 98. Notably, these
so-called temporal networks 127 were recently shown to be easier to control, requiring less energy
to attain a desired pattern of neural activity, than their static counterparts 128.
Properly modeling the dynamics of brain networks requires also understanding the functional
dynamics occurring on brain networks. For instance, dating to Donald Hebb's 1949 book The
Organization of Behavior, it has been posited that the strength of a synaptic connection increases
with the persistent synchronized firing of its pre- and postsynaptic neurons 129. Such Hebbian
plasticity has been observed in vitro 130 and is thought to explain many aspects of brain network
structure 131, 132. More generally, Hebb's postulate highlights the fact that a complete understanding
of the brain cannot simply include a description of its structural wiring; it must also stipulate the
types of dynamics supported by this physical circuitry.
The physics of brain network function
While structural brain networks represent the physical wiring between neural elements (e.g., be-
tween individual neurons or brain regions), knowledge of this circuitry alone is not sufficient to
understand how the brain works. For this reason, we turn our attention to models of brain network
function that stipulate how neural activity propagates along structural connections. Just as the neu-
ron doctrine postulates that the brain's structure is divided into a network of distinct nerve cells,
it is also widely expected that the brain's array of cognitive functions emerges from the collective
activity of individual neurons 13, 18, 25, 133, 134. To understand how the firing of simple nerve cells
can give rise to the brain's rich repertoire of cognitive functions 135, analogies are often drawn
with notions of emergence in statistical mechanics 25, 133, 136. Developed concurrently with the neu-
13
ron doctrine in the late 19th century, statistical mechanics established (among other achievements)
that the thermodynamic laws governing the macroscopic behavior of gas molecules can be derived
from the microscopic dynamics of the molecules themselves 137. Similarly, growing evidence sug-
gests that the dynamics of individual neurons and brain regions, when embedded in networks of
structural connections, can produce the types of long-range correlations and collective patterns of
activity that we observe in the brain 133, 138 -- 143. Here we traverse what is known about brain net-
work function in relatively broad strokes, from the dynamics of distinct neurons to the networked
activity of the entire brain.
Measuring brain network function. The first measurements of the brain's functional organiza-
tion date to 1815, when Marie-Jean-Pierre Flourens pioneered the use of localized lesions in the
brains of living animals to observe their effects on behavior. Through his experiments, Flourens
discovered that the cerebellum regulates motor control, the cerebral cortex supports higher cog-
nition, and the brain stem controls vital functions 144. The remainder of the 19th century brought
increasingly detailed measurements of the brain's functional organization, from the demonstration
that the occipital lobe regulates vision 145 to the discovery that the left frontal lobe is essential
for speech 146. These discoveries, combined with the early images of neural circuits captured by
Ram´on y Cajal 55, culminated in Thomas Scott Sherrington's book The Integrative Action of the
Nervous System, which proposed the idea that neurons behave in functional groups 87.
Meanwhile, in 1849 the physicist Hermann von Helmholtz achieved the first electrical mea-
surements of a nerve impulse 147, sparking a wave of experiments investigating the electrical prop-
erties of the nervous system. Through invasive measurements in animals using newly-developed
electroencephalography (EEG) techniques 148, it quickly became clear that individual neurons
communicate with one another via electrical signals 149 -- 151, thus providing a clear mechanism
explaining how information is propagated and manipulated in the brain. Today, scientists possess
a rich menu of experimental techniques for measuring brain dynamics across a range of scales. At
the neuronal level, the development of invasive methods in animals, such as electrophysiological
recordings of brain slice preparations in vitro 152, 153 and calcium imaging of neuronal activity in
14
vivo 154, 155, have vastly expanded our understanding of synaptic communication. At the regional
level, complimentary minimally-invasive imaging techniques have identified fundamental proper-
ties of information processing in humans 156. Interestingly, these advances in mesoscale functional
imaging can largely be traced to the efforts of physicists. MEG methods, for instance, use super-
conducting quantum interference devices (SQUIDS) to directly measure the magnetic fields gen-
erated by electrical currents in the brain 12, 157; and PET techniques measure the positron emission
of radioisotopes produced in cyclotrons to reconstruct the metabolic activity of neural tissue 158.
Over the last twenty years, measurements of brain dynamics have been increasingly dominated
by functional MRI (fMRI) 159, which estimates neural activity by calculating contrasts in blood
oxygen levels, without relying on the invasive injections and radiation that limit the applicability
of other imaging techniques 160. This modern progress in functional brain imaging has galvanized
the field of network neuroscience by making detailed datasets of large-scale neural activity widely
accessible.
One particularly important application of functional brain imaging has been the study of
so-called functional brain networks 161, which have allowed researchers to investigate the orga-
nization of neural activity using tools from network science. In functional brain networks, as in
their structural counterparts, nodes represent physical neural elements, ranging in size from indi-
vidual neurons to distinct brain regions 162. However, whereas structural brain networks define
the connectivity between elements based on physical measures of neural wiring (e.g., synapses
between neurons or white matter tracts between brain regions), functional brain networks define
connectivity based on the similarity between two elements' dynamics 162. To see how this works,
we briefly consider the common example of a large-scale functional brain network calculated from
fMRI measurements of regional activity 161 (Fig. 2a). First, blood oxygen levels indirectly re-
flecting neural activity are measured within three-dimensional non-overlapping voxels, spatially
contiguous collections of which each represent a distinct brain region. After preprocessing the
signal to correct for sources of systematic noise such as fluctuations in heart rate, the activity of
each brain region is discretized in time, yielding a vector (or time series) of neural activity. Fi-
nally, to quantify functional connectivity, one computes the similarity between each pair of brain
15
regions, for example using the quite simple Pearson correlation between the two regions' activity
time series 138, 163. The end result, even for different types of functional data and different choices
for the preprocessing steps and similarity metric, is a functional brain network representing the
organization of neural activity.
After constructing a functional brain network, researchers can utilize techniques from net-
work science to study its key organizing features. Such efforts have demonstrated that large-scale
functional brain networks, much like structural networks, exhibit signs of modular, small-world,
heavy-tailed, and metabolically constrained organization 161, 164 -- 167. The existence of strong func-
tional community structure, for instance, further supports the hypothesis that brain networks segre-
gate into subnetworks with specialized cognitive functions 168, 169. Moreover, the presence of high
clustering and short average path lengths, combined with the existence of high-degree hub regions,
highlights the competing functional pressures of information segregation and integration in the
brain 166, 170. Metabolic constraints on the brain's structural wiring are also evident in its functional
connectivity 171, with spatially localized brain regions generally supporting more strongly corre-
lated activity than distant regions 69. In light of the similarities between the brain's functional and
structural organization, it is tempting to suspect that functional brain networks closely resemble
the physical wiring upon which they exist 172, 173. However, the relationship between brain func-
tion and structure is highly nonlinear 174, and understanding how a functional brain network arises
from its underlying structural connectivity remains a subject of intense academic focus 119, 175.
Modeling brain network function. To understand how the web of physical connections in the
brain gives rise to its functional properties, statistical mechanical intuition dictates that we should
begin by studying the dynamics of individual elements. Once we have settled on accurate mod-
els of the interactions between individual neurons and brain regions, we can link these elements
together in a network to predict macroscopic features of the brain's function from its underlying
structure 36, 71. Interestingly, the history of modeling in neuroscience has followed precisely this
path, beginning with models of neuronal dynamics 17, 176, 177, then increasing in scale to mean-field
neural mass models of distinct brain regions 178, 179, and eventually achieving models of entire
16
Figure 2 Measuring and modeling brain network function. a The measurement of brain
network function begins with experimental data specifying the activity of neurons or brain re-
gions. As an example, we consider variations in blood oxygen level in different parts of the
brain measured via fMRI. Calculating the similarity (e.g., correlation or synchronization) be-
tween pairs of activity time series, one arrives at a similarity matrix. This matrix, in turn, de-
fines a functional brain network constructed from our original measurements of neural activity.
b We divide models of neural activity into two classes: abstract models with artificial dynam-
ics (left) and biophysical models with realistic dynamics (right). Models of artificial neurons,
such as the MP neuron, typically take in a weighted combination of inputs and pass the inputs
through a nonlinear threshold function to generate an output. Networks of artificial neurons,
from deep neural networks to Hopfield networks, have been shown to reproduce key aspects
of human information processing, such as learning from examples and storing memories. By
contrast, biophysical models of individual neurons, such as the Hodgkin -- Huxley or FitHugh --
Nagumo models, capture realistic functional features such as the propagation of the nerve
impulse. When interconnected with artificial synapses, researchers are able to simulate en-
tire neuronal networks. Complimentary mesoscale approaches, including neural mass models
such as the Wilson -- Cowan model, average over all neurons in a population to derive a mean
firing rate. To simulate the large-scale activity of an entire brain, researchers use neural mass
models to represent brain regions and embed them into a network with connectivity derived
from measurements of neural tracts (e.g., as measured via DTI).
17
ModelingbFunctional brain networkSimilarity matrixaMeasurementExample: Blood oxygen level (via fMRI)ActivityBiophysical dynamicse.g. Wilson -- CowanFine-scalee.g. Hodgkin -- Huxley,
FitzHugh -- NagumoMesoscaleinputse.g. McCulloch -- Pitts"dog"e.g. image classifierArtificial neural netoutputinputsthresholdArtificial neuronArtificial dynamicsoutputLarge-scaleneuronneuronal
networkneural massregional network networks of neurons and brain regions 5, 141, 180. Here we review important developments in the
modeling of neural dynamics, dividing the modeling techniques into two complimentary classes:
those with artificial dynamics and those with biophysically realistic dynamics (Fig. 2b). As we
will see, models from each of these two classes are able to reproduce important aspects of neural
activity and system function that have been observed in a range of physiological and behavioral
experiments.
Artificial models. One of the earliest mathematical models of neural activity whatsoever was pro-
posed in the mid-1940s by Warren McCulloch and Walter Pitts to describe the logical functioning
of an individual neuron 17. Known as the MP neuron, their model accepted binary inputs, com-
bined these inputs using linear weights, and produced a binary output reflecting whether or not
the weighted sum of inputs exceeded a given threshold (Fig. 2b). Albeit a simple caricature of
neuronal dynamics, this model has been shown to reproduce some important qualitative features
of neuronal activity, including the linear summation of excitatory inputs 181 and the "all-or-none"
response to the resulting integrated signal 182. Moreover, by connecting the inputs and outputs of
multiple MP neurons, researchers have achieved deep insights about how brain networks perform
basic cognitive functions. For example, soon after the introduction of the MP model, researchers
demonstrated that networks of artificial neurons could be used to represent any Boolean function
(i.e., any function mapping a list of binary variables to a binary output), thereby establishing the
basic capability of neural networks to perform logical computations 6.
While their ability to perform basic computations was quickly realized, it was not clear at
the outset whether artificial neural networks could reproduce other cognitive functions, such as
the ability to learn or store memories. The former was established by Frank Rosenblatt in 1957,
when he showed that the weights on the inputs to an MP neuron could be tuned such that the
output defines a binary classifier. Known as the perceptron, this algorithm enabled a single MP
neuron to segregate incoming data into one of two classes by learning from past examples. This
remarkable result directly inspired more advanced learning algorithms, including support vector
machines 183 and artificial neural networks 184, effectively setting in motion the study of machine
18
learning. Today, deep neural networks, consisting of multiple layers of artificial neurons feeding
in one direction from the input layer to the output layer (Fig. 2b), are able to learn a wide range
of impressive cognitive functions that we have come to expect from the brain 185. While the list
of applications is ever-expanding, deep neural networks have been used to process and identify
images of objects, scenes, and people 186; recognize, interpret, and respond to spoken language 187;
and formulate strategies and make decisions in adversarial settings 188.
In addition to performing computations and learning from examples, the physicist John Hop-
field showed in 1982 that neural networks can also store and recall memories. Specifically, Hop-
field demonstrated that the synaptic weights connecting a set of MP neurons could be adjusted
in a Hebbian fashion such that the network is able to "memorize" a number of desired activity
states 5 (i.e., configurations of the network in which each neuron is either active or inactive). No-
tably, the number of memorized states grows linearly with the number of neurons in the network
189, and errors in recall often yield states that are semantically similar to the target state, a phe-
nomenon commonly observed in humans 190. Interestingly, the memorized activity states can be
interpreted as local minima of an associated energy function, making each Hopfield network equiv-
alent to an Ising model at zero temperature 41. More recently, Ising-like models have also been
used to explain the critical or avalanche-like behavior of activity in neural ensembles 191, which
is thought to support adaptation to environmental changes 192, information storage 193, optimal
information transmission 194, maximal dynamic range 195, 196, and computational power 197. Fur-
ther building upon this connection to statistical mechanics, scientists have recently used maximum
entropy techniques to construct data-based models of neuronal dynamics. These maximum en-
tropy models, which are equivalent to networks of Ising spins with specially-chosen external fields
and interaction strengths, have been shown to predict the observed long-range correlations within
naturally occurring networks of neurons and brain regions 141, 198. Together, artificial models of
neural dynamics, from simple MP neurons to artificial neural networks and data-driven maximum
entropy models, continue to inform our understanding of brain networks as information processing
systems.
19
Biophysical models. While artificial models continue to generate insights about the nature of neural
computation, they only vaguely resemble the complex biophysical mechanisms that guide observ-
able neural activity. Among the first biophysically realistic models of the electrical behavior of
an individual neuron was achieved nearly a decade after the introduction of the MP neuron by
physiologists Alan Lloyd Hodgkin and Andrew Fielding Huxley 176. Beginning from a principled
description of the initiation and propagation of action potentials in living neurons, the Hodgkin --
Huxley (HH) model explains important qualitative aspects of neuronal behavior 6, including the
spontaneous emergence of limit cycles or oscillations in activity 199 and the presence of a Hopf bi-
furcation in the neuronal firing rate, which is thought to underlie the all-or-none principle 176 (Fig.
2c). Subsequent extensions of the HH model expand biophysical realism by incorporating multiple
ion channel populations 200, the complex geometries of dendrites and axons 201, and more realistic
stochastic dynamics yielding thermodynamic and hybrid HH models 202, 203. Concurrent with these
descriptive improvements, several simplified neuronal models were also developed, including the
notable FitzHugh -- Nagumo model 177, 204, facilitating efficient large-scale simulations of groups of
neurons.
Simplifications in neuronal modeling, paired with fine-scale measurements of the synaptic
wiring in several animals, have spurred large-scale simulations of real neuronal circuits (Fig. 2b).
For example, on the heels of mapping the entire C. elegans connectome 57, researchers began sim-
ulating the 302-neuron network at the cellular level 205, eventually even including the nematode's
entire muscular system and representations of its physical environment 206. Despite these and other
efforts simulating the Drosophila brain 207 and the rat's neocortical column 208, it remains unclear
how networks of neurons combine to generate the complex range of behaviors observed even in
these relatively simple organisms. This contrast between the simplicity of neuronal dynamics and
the apparent complexity of large-scale neural behavior hints at the crucial role of emergence. To
understand how macroscopic behaviors emerge within groups of neurons, researchers began devel-
oping mean-field descriptions of large neuronal populations. Known as neural mass models, these
efforts culminated in the foundational Wilson -- Cowan (WC) model of population dynamics 179.
Whereas previous neural mass models only considered excitatory interactions between neurons,
20
Wilson and Cowan also included inhibitory interactions, thereby enabling the WS model to predict
the collective neural oscillations observed in experiments as well as the emergence of other key
properties of neural behavior, including the existence of multiple stable states and hysteresis in the
neural response to stimuli 179. This progress was further extended to include spatial fluctuations in
activity, yielding neural field models that exhibit other behaviors typically observed in the brain,
including regions of localized activity 209 and traveling waves 210.
In much the same way that neuronal circuits have been modeled using observable synaptic
wiring in animals, one could imagine simulating a network of neural mass models whose connec-
tions are drawn based on non-invasive measures of regional connectivity in humans. By doing so,
researchers are now able to simulate whole sections of the human brain (Fig. 2c), opening the door
for comparisons with experimental measurements of regional activity. Precisely this approach has
driven a deeper understanding of the structure-function relationship, including the demonstration
that the broad spectrum of MEG/EEG recordings of electrical activity can be reproduced by net-
worked models of neural masses 211 and that the functional connectivity within such recordings
depends critically on the coupling strength between neural masses 212. To facilitate large-scale
simulations of the entire human brain, researchers have frequently turned to the Kuramoto model
of oscillatory dynamics as a simplified neural mass model 180, 213. These efforts have provided
insights about the spontaneous synchronization of neural oscillations 214, a phenomenon which is
thought to play a critical role in neural communication 215, information processing 216, and motor
coordination 217. Moreover, by embedding Kuramoto oscillators into a realistic map of the human
connectome, researchers have shown that even this simple model is able to reproduce the patterned
fluctuations in activity and long-range correlations observed in fMRI data 218. Detailed biophysi-
cal models of neural dynamics, from descriptions of the electrical activity of individual neurons to
networked neural mass models simulating the entire brain, continue to inform our understanding of
how collective neural behavior and high-level cognitive functions arise from the brain's underlying
physical circuitry.
The future of brain network function. Over the last two centuries, our understanding of the
21
brain's functional organization and information processing capabilities has progressed immensely.
Despite this progress, the modern neuroscientist remains fundamentally limited by the experimen-
tal and theoretical tools at their disposal 219, 220. Invasive techniques such as intracranial electrocor-
ticography, and even minimally invasive techniques such as stereotactic electroencephalography
(sEEG) 221 -- 223, provide immense precision in mapping human brain dynamics, but remain con-
strained to patients with medically refractory epilepsy. Other noninvasive imaging techniques all
suffer from trade-offs between spatial and temporal resolution 224; methods that directly measure
electromagnetic signals (e.g., EEG and MEG) have high temporal resolution but low spatial res-
olution, while measurements of blood flow and metabolic activity (e.g., via fMRI or PET) have
relatively high spatial accuracy but poor resolution in time. Even fMRI -- widely considered the
standard for high spatial resolution in humans -- integrates signals over hundreds of thousands of
neurons and several seconds 225. Consequently, any changes in neural activity that occur over tens
of thousands of neurons or even over the span of a second are imperceivable on a standard fMRI
scan.
To improve the precision of functional neuroimaging (fMRI in particular), recent efforts have
leveraged modern advances in image processing to strengthen the signal and reduce background
noise. For example, to minimize the inevitable effects of head movements and fluctuations in
blood flow during scanning, fMRI signals are increasingly corrected using techniques similar to
image stabilization in video cameras 226. Additionally, in order to draw general conclusions from
neuroimaging results across a group of subjects, impressive strides have been made to correct for
inter-subject heterogeneities in brain structure 227. Together, advances in image processing have
begun to push neuroimaging from a tool exclusively used for academic research to one that can
aid in the diagnosis and treatment of psychiatric disorders such as schizophrenia and Alzheimer's
disease.
Beyond data collection, data analysis and models in network neuroscience have historically
been limited to dyadic relationships between neural elements, such as synapses connecting pairs
of neurons or Pearson correlations between pairs of brain regions 36, 71. While these dyadic notions
22
of connectivity have provided important insights about the brain's circuitry, mounting evidence
suggests that higher-order interactions between three or more elements are also crucial for under-
standing the large-scale behavior of entire brain networks 198, 228, 229. In order to study these higher-
order connections, recent efforts have focused on generalizing traditional definitions and intuitions
from network science, primarily by adopting methods from algebraic topology 230. One notable
approach, known as persistent homology, has allowed researchers to extrapolate conclusions about
neural activity across scales, escape the problem of selecting appropriate thresholds for functional
edge strengths 231, and extract principled mesoscale features of network organization 229, 232.
Efforts have also been made to expand traditional metrics of functional connectivity, which
are typically based on correlation, to include more sophisticated notions of causality 167. Since
causality reflects the flow of information in a network from one element to another, efforts which
aim to uncover causal relationships between neurons and brain regions have naturally drawn inspi-
ration from concepts in information theory (see Box 3) 233. From mutual information to transfer
entropy, information theoretic notions of functional connectivity are increasingly being used to
quantify the flow of information in the brain 216, 234, 235. These measures of causality, in turn, have
real-world implications for controlling brain networks and intervening to treat neurological disease
and psychiatric disorders.
[Box 3 here]
Perturbation experiments and the physics of brain network control
Thus far, we have examined what is known about the structural circuitry connecting neural com-
ponents in the brain as well as the dynamical laws governing the interactions between these com-
ponents. An ultimate test of our understanding, however, lies in our ability to intervene and shift
the brain's dynamics to facilitate desirable behaviors. An important implication of the brain's net-
worked structure is that localized perturbations (e.g., targeted lesions or stimulation) do not just
23
In this way, the task of controlling brain dynamics requires knowledge of how signals
yield localized effects -- they also induce indirect effects that propagate along neural pathways
236, 237.
transmit along the brain's structural wires, making the problem inherently one of network control
238. Building upon targeted lesioning experiments in animals and clinical interventions in humans,
efforts toward a theory of network control in the brain have recently taken shape, inspiring several
fundamental questions 239. Are brain networks designed to facilitate control 240? What are the
principles that allow brain networks to control themselves toward desired activity states 241, 242?
Can we leverage these principles to inform stimulation-based therapies for neurological diseases
and psychiatric disorders 243 -- 246? To address these questions, here we review the current frontiers
in the physics of brain network control.
Targeted perturbations and clinical interventions. The first attempts to systematically control
brain dynamics date to the early 19th century, when Marie-Jean-Pierre Flourens noticed that tar-
geted lesions to the brain in living rabbits and pigeons yielded specific changes in the animals'
perception, motor coordination, and behavior 144. These efforts, in conjunction with other tar-
geted lesioning experiments in animals 145, 146, supported the notion of functional localization -- the
theory that specific cognitive functions are supported by specific parts of the brain. In humans,
evidence for functional localization has typically relied on patients with localized brain damage
(e.g., due to a stroke or head trauma). Historical studies of this kind have revealed, for instance,
that damage to one half of the occipital lobe often induces blindness in the opposite field of vision
247 and that lesions in the frontal lobe can result in memory loss and an increase in impulsivity and
risk taking 248. More recently, advances in non-invasive stimulation techniques such as transcranial
magnetic stimulation (TMS) 249, which induces "transient" lesions by disrupting the brain's nor-
mal electrical activity, have opened the door for the control of localized brain functions, including
perception 250, learning 251, language processing 252, and attention 253. These non-invasive transcra-
nial techniques have been supplemented by more invasive deep brain stimulation (DBS) methods
to provide targeted therapies for a number of psychiatric and neurological disorders 249, 254. By
focusing electromagnetic stimulation on the brain regions associated with specific disorders, both
TMS and DBS have been used to treat Parkinson's disease, epilepsy, depression, and schizophre-
24
nia, among other disorders that are resistant to traditional therapies 255, 256 (Fig. 3a). Despite these
therapeutic benefits, it remains unclear exactly how and why TMS and DBS are so effective 236, 254;
however, recent evidence suggests that the answers may rely on a deeper understanding of the
indirect effects of stimulation that are mediated by the brain's physical circuitry 257, 258.
With the recent development of whole-brain neuroimaging methods such as fMRI, evidence
continues to mount that brain regions are heavily interdependent on one another, often working in
unison to process information and formulate responses 103, 161. In a particularly clear demonstra-
tion of the brain's functional integration, Anthony Randall McIntosh and colleagues trained human
subjects to associate an auditory stimulus with a visual event. Later, when the auditory stimulus
was presented alone, the investigators observed increased activity in the occipital lobe, more tra-
ditionally thought of as being reserved for visual processing 259. Experiments such as these reveal
how activity or stimulation in one part of the brain can propagate along neural pathways to induce
activity in other distant parts. To understand the system-wide impacts of targeted stimulation, re-
searchers have increasingly drawn upon network models of brain dynamics 257, 258. These efforts
have resulted in the identification of neural circuits, rather than isolated regions, that are critical
for reducing the symptoms of Parkinson's disease 258, 260. Similar network-based approaches are
also being used to suppress epileptic seizures using DBS 261, non-invasively treat depression using
TMS 262, and modulate consciousness during surgery using anesthesia 263. Moreover, by stimulat-
ing and recording neural activity in several brain regions simultaneously, researchers have achieved
closed-loop strategies for dynamically updating targeted treatments 264, 265 (Fig. 3a). Meanwhile,
clinical applications are increasingly being informed by detailed computational simulations of per-
turbations to specific brain regions, typically employing networked biophysical models such as
those discussed in the previous section 266, 267. Together, these real-world and computational stud-
ies of targeted stimulation have opened the door for sophisticated strategies that aim to shift neural
activity with the ultimate goal of guiding healthy cognitive function.
Network control in the brain. To inform strategies for targeted stimulation and brain network
control, it helps to draw upon existing tools from control theory in mathematics and intuitions from
25
Figure 3 Targeted perturbations and brain network control. a Methods for targeted
control are used in the study, design, and optimization of external control processes, such as
transcranial magnetic stimulation and deep brain stimulation. These targeted perturbations
of neural activity are being utilized in clinical settings to treat major depression, epilepsy, and
Parkinson's disease. By simultaneously stimulating and measuring neural activity, researchers
can now perform closed-loop control, continuously updating stimulation strategies in real time.
b Controllability metrics provide summary statistics regarding the ease with which a given
node can enact influence on the network. Two common metrics are the average controllability,
which assesses the ease of moving the system to all nearby states, and the modal controlla-
bility, which assesses the ability to move the system to distant states (see Box 4). Notions of
controllability have proven useful in the study of the brain's internal control processes, such as
homeostatic regulation and cognitive control. For example, the human brain displays marked
levels of both average and modal controllability, and the proportion of average and modal con-
trollers differs across cognitive systems, suggesting the capacity for a diverse repertoire of
dynamics 241.
26
Targeted stimulationaCognitive controlbe.g. deep brain stimulationelectrodee.g. targeted treatment for
Parkinson's diseaseDefault modeFrontoparietalCingulo-opercularVentral attentionDorsal attentionAuditoryVisualSomatosensoryOther1%3%2%11%30%18%4%19%12%16%3%17%15%1%18%12%7%11%Average
controllabilityModal
controllabilityclosed-loop
control cognitive control in psychology. Given a mathematical model of a system, control theory seeks to
understand how the system can be influenced such that it moves toward a desired state 238, 268 (see
Box 4). Cognitive control, on the other hand, encompasses a broad class of processes by which the
brain enacts control over itself, typically to achieve an abstract goal or desired response 269. For
example, dating to the early 1970s neurophysiological studies revealed that the act of holding an
object in working memory induces a sustained neural response in the prefrontal cortex 270, 271. In
fact, the prefrontal cortex is now believed to play a key role in many cognitive control processes,
from the representation of complex goal-directed behaviors 272 to the support of flexible responses
to changes in the environment 273. But how do these notions of cognitive control (as defined by
psychologists and cognitive neuroscientists) compare to theories of network control (as defined
by physicists and engineers)? Furthermore, how can knowledge of the brain's intrinsic control
processes inform targeted therapies for mental illness?
[Box 4 here]
To address these questions, we begin by comparing cognitive notions of intrinsic control with
theoretical measures of control and controllability in brain networks (see Box 4). It is interesting,
for example, to ask which brain regions are most capable of inducing desired neural responses in
other brain regions that are responsible for common functions such as vision, audition, and motor
coordination. Toward this end, Gu et al. used methods from control theory to demonstrate that
the strongest driver nodes corresponded to brain regions with high communicability -- or many
topological paths through the brain network -- to the target brain regions 274. In a related study,
Betzel et al. used the structural wiring of the brain to simulate transitions between commonly
observed activity states 275. They found that optimal control nodes tended to have high degree in
the network, and that when this rich-club of hub regions was destroyed by simulated lesioning, the
ability of the brain to make common transitions was significantly reduced.
In addition to studying the roles of specific control trajectories, complementary approaches
have considered trajectory-independent metrics such as the average and modal controllabilities
27
discussed in Box 4 276. By comparing control theoretic measures of node controllability with
the cognitive functions associated with each brain region, researchers have observed that different
types of controllers are located in distinct areas of the brain (Fig. 3b) 241. For example, brain
regions with strong average controllability are disproportionately located in the default mode sys-
tem, which is associated with baseline neural activity; meanwhile, strong modal controllers are
primarily located in cognitive control systems. These observations are particularly interesting be-
cause they suggest that regions associated with the default mode are optimally positioned to push
the system into many easily reachable states, while regions associated with cognitive control are
optimally positioned to steer the system toward distant states.
As a final layer of abstraction, rather than studying the controllabilities of specific brain re-
gions, one could envision averaging over all regions to quantify the mean controllability of an
entire brain network. Interestingly, by taking precisely this approach, Tang et al. established that
brain networks as a whole are finely tuned to maximize both average and modal controllability,
thereby supporting a diverse range of possible control strategies 277. Furthermore, by comparing
subjects in different stages of adolescence, the researchers found that brain network controllability
increases with age, suggesting that neural circuitry evolves over time to support increasingly com-
plex dynamics. In related studies, metrics of network controllability were found to differ by sex
278 and to be altered in individuals with high genetic risk for bipolar disorder 242. Taken together,
these results demonstrate that network measures of optimal control and controllability correspond
closely to existing notions of intrinsic and cognitive control in neuroscience. This close correspon-
dence, in turn, suggests that network control theory, by taking into account the complex wiring of
the brain, has the promise to enrich our understanding of the brain's control principles 279.
The future of brain network control. Throughout this section, we have focused primarily on
targeted therapies that rely on the coarse-grained stimulation of entire brain regions and simple
control strategies that assume idealized linear dynamics. Emerging efforts in neuroscience and
control theory, however, are opening the door for a number of significant improvements, includ-
ing: (i) techniques for fine-scale control of neural activity 280 -- 283, even down to the level of in-
28
dividual neurons 284, 285, (ii) systems identification approaches that allow for the incorporation of
effective connectivity measurements to inform control, superseding solely structural explanations
286, and (iii) generalizations of linear control theory that include more realistic nonlinear dynamics
287, 288. Among recent advances in the manipulation of fine-scale neural activity, arguably the most
promising tool is optogenetics, which offers millisecond-scale optical control of specific cell types
within the brains of conscious animals 280, 281. Its striking precision 282, in some cases even down
to single-cell resolution 284, 285, has enabled researchers to investigate the nature of causal signals
between neurons and to study how these signals give rise to qualitative changes in animal behavior
283.
While linear control theory continues to provide critical insights about how signals propagate
along the brain's structural wiring 240, 241, 274, 275, interactions between neural components, from in-
dividual neurons to entire brain regions, are highly nonlinear (Fig. 2b) 119. Initial efforts to develop
a theory of nonlinear control, dating as early as the 1970s 289 -- 291, quickly converged on the conclu-
sion that results as strong and general as those derived for linear dynamics could not be obtained
for a general nonlinear system 238. Fortunately, concerted theoretic efforts have led to weaker no-
tions of nonlinear controllability 292, notable among which are techniques for linearizing nonlinear
systems around stable equilibrium states 287, 288 and methods for leveraging the symmetries of a
system 293 such as repeated network motifs to simplify control strategies 294. Additional efforts
have utilized advances in computing power to simulate the effects of external perturbations across
a range of model systems, including networks of FitzHugh -- Nagumo neurons 293, Wilson -- Cowan
neural masses 243, and Kuramoto oscillators 295 as well as artificial neural networks such as the
Ising model 296, 297. Together, recent advances in high-precision neural stimulation like optogenet-
ics and our emerging understanding of the principles governing nonlinear control are pushing the
boundaries of what is considered possible in the investigation of neural activity. Targeted control
of the brain's complex behavior -- once considered a topic of science fiction -- now has the promise
to shape targeted therapies for a range of psychiatric and neurological disorders.
29
Conclusions and future directions in the neurophysics of brain networks
The intricate inner workings of the brain remains one of the greatest mysteries defying resolution
by contemporary scientific inquiry. On the heels of decades of effort investigating the functions of
the brain's individual components 298, from neurons to neuronal ensembles and large-scale brain
regions, conclusive evidence points to the need for maps and models of the interactions between
these components in order to fundamentally understand the brain's ensemble dynamics, circuit
function, and emergent behavior 36, 299. Here we reviewed recent advances toward meeting this
challenge with an eclectic array of curios from the physicist's cabinet: statistical mechanics of
complex networks, thermodynamics, information theory, dynamical systems theory, and control
theory. In the course of our exposition, we considered the principles of small-worldness 27, in-
terconnected high-degree hubs 300, modularity 91, and spatial embedding 301 that provide useful
explanations for the architecture of structural brain networks. We then saw these same principles
reflected in the organization of long-range functional connectivity supporting information dissemi-
nation, and the computations that can result therefrom 38, 216. As with any physical system, a natural
next step is to probe the validity of our descriptive and explanatory models using perturbative ap-
proaches both in theory and experiment. Thus, we next summarized the utility of network control
theory in offering insights into internal control processes such as homeostatic regulation and cog-
nitive control, as well as external control processes such as neurostimulation, which are currently
being used to treat multiple disorders of mental health 279.
Throughout the exposition, we described current frontiers in the investigation of brain net-
work structure, function, and control. Although we will not reiterate those points here, we do wish
to offer the sentiment that, while the empirical advances laying the foundation of the field have
spanned several decades, the network physics of the brain is an incredibly young area, rich with
opportunities for discovery. And perhaps -- with a bit of courage -- we may even begin to provide
an empirical constitution to the deeper philosophical questions that humans have wrestled with
for millennia: What makes us unique and different from non-human animals 240, 302? How do we
represent abstract concepts such as value to ourselves 303 and others 304? How are representations
30
transmitted throughout the brain or reconfigured based on new knowledge 305? What makes a mind
from a brain? Physicists, the brain is calling you.
31
Box 1 A simple primer on networks. Here, we define what we mean by a network and describe
tools for summarizing its architecture. Importantly, a network is agnostic to the system that it rep-
resents 36, whether it be a brain, a granular material 306, or a quantum lattice 307. By far the simplest
network model is represented by a binary undirected graph in which identical nodes represent sys-
tem components and identical edges indicate relations or connections between pairs of nodes (see
the figure). Such a network can be encoded in an adjacency matrix A, where each element Aij
indicates the strength of connectivity between nodes i and j. When all edge strengths are unity, the
network is said to be binary. When edges have a range of weights, the network represented by the
(cid:124), the network is undirected; otherwise, the
adjacency matrix is said to be weighted. When A = A
network is directed.
One can extend this simple encoding to study multilayer, multislice, and multiplex networks
308; dynamic or temporal networks 127, 309; annotated networks 310; hypergraphs 311; and simplicial
complexes 230. One can also calculate various statistics to quantify the architecture of a network
and to infer the function thereof (see figure). Intuitively, these statistics range from measures of the
local structure in the network, which depend solely on the links directly emanating from a given
node (e.g., degree and clustering), to measures of the network's global structure, which depend
Intermediate statistics exist to study network organization at the mesoscale, such as cavity
on the complex pattern of interconnections between all nodes (e.g., path lengths and centrality)
30.
structure and community structure, the latter of which describes the presence of communities of
densely connected nodes 312 -- 314. As we will see, the encoding of a system as a network and the
quantitative assessment of its architecture can provide important insights into its function 34, 107.
32
33
DegreeClusteringPath lengthCentralityCommunitiesCavitiesNodeEdge Box 2 Bridging spatiotemporal scales. In the context of complex systems generally and neural
systems specifically, the cutting edge work relates to extending our tools, theories, and intuitions
from a single network to so-called multiscale, multilayer, and multiplex networks 308, 315. Perhaps
the most obvious context in which to make this extension is from regional networks to cellular-scale
neuronal networks 121. Large-scale brain activity provides a coarse-grained encoding of neural pro-
cesses, and the map from cellular dynamics to regional dynamics reflects rules of system function.
By combining these two layers we can address questions like, "How do cellular processes shape
circuit behavior?" The next logical extension is to move even further down the natural hierarchy
of scales to understand how molecular networks -- including gene coexpression networks 123, 316 -- 318
-- shape the behavior of cells 319. Understanding how molecular mechanisms affect large-scale
brain network function is critical for the development of effective pharmacological interventions
116, 320, 321. By extending the network model from regions to cells to molecular drivers, we can ask
questions like, "How do genetic codes and epigenetic drivers shape circuit behavior across spatial
scales?" And in a final extension, it is time to move up in the natural hierarchy of scales to combine
information from the connectivity within a single human brain to the connectivity between human
brains in large-scale social networks 304, 322 -- 324. While brain activity and structure offer biological
mechanisms for human behaviors, social networks offer external inducers or modulators of those
behaviors 325. By extending the network model to this larger scale, we can start to ask -- and po-
tentially answer -- questions like, "How do brains shape social networks? And how do social ties
shape the brain?" This extension will be important in understanding human behavior within the
broader contexts of culture and society.
34
Molecular networkNeuronal networkSocial networkRegional network Box 3 Information theory and network neuroscience. At its core the brain is an information
processing system, having evolved over millions of years to encode and manipulate a continuous
stream of sensory signals 326. As such, information theory -- the science of how signals are encoded
and processed -- provides a compelling lens through which to study the brain's function 327. Infor-
mation theory began with the 1948 paper "A Mathematical Theory of Communication," wherein
Claude Shannon proposed the entropy of a signal as the natural measure of its information content
and derived fundamental limits on the information capacity of a communication channel 328. Soon
after, MacKay and McCulloch adapted the concept of channel capacity to obtain limits on the rate
at which one neuron can transmit information to another 329, sparking the study of information
flow in the brain. Subsequent work by Attneave and Barlow proposed the idea that neural activity
is optimized for the transmission of sensory information 330, 331, providing the foundation for future
investigations of neural coding 135, 326.
Despite these initial efforts bridging information theory and neuroscience, progress slowed
primarily due to difficulties obtaining unbiased information estimates from neural systems. Im-
provements in experimental techniques, however, eventually sparked renewed interest 332, spurring
the introduction of robust methods for estimating information theoretic quantities 333 -- 335. On the
basis of these advancements, information theory has once again become a powerful tool for the net-
work neuroscientist. Recent attempts, for instance, to uncover causal relationships between neural
elements have successfully adapted notions of information flow, such as mutual information and
transfer entropy 336, 337. At the same time, efforts to understand large-scale correlations within neu-
ronal populations have utilized the principle of maximum entropy 338, resulting in Ising-like models
of collective neural behavior 141, 198. As information theory becomes increasingly integrated into
the fabric of neuroscience, physicists are uniquely positioned to pioneer exciting new techniques
for investigating the nature of information processing in the brain.
35
Box 4 Linear control and network controllability. To investigate the principles of control in
the brain, it is useful to understand the theory of network control generally. In network control,
the system in question typically comprises a complex web of interacting components, and the goal
is to drive this networked system toward a desired state by influencing a select number of input
nodes 238. The starting point for most control theoretic problems is the linear time-invariant control
system x(t + 1) = Ax(t) + u(t), where x(t) defines the state of the system (e.g., the BOLD signal
measured by fMRI), A is the interaction matrix (e.g., white matter tracts estimated using DTI),
and u(t) defines the input signal (e.g., electromagnetic stimulation using TMS or DBS) 339. Such
a system is said to be controllable if it can be driven to any desired state. Often, however, many
naturally occurring networks that are theoretically controllable cannot be steered to certain states
due to limitations on control resources 340, 341, motivating the introduction of control strategies
u∗(t) that minimize the so-called control energy E(u) =(cid:80)∞
t=0 u(t)2
2.
By limiting the control input to a single node, we can quantify the ability of that node to steer
the dynamics of the entire system. For example, the average controllability of a node represents
its capacity to drive the network to many nearby states 241, while a node's modal controllability
quantifies its ability to push the network toward distant hard-to-reach states 276 (see figure). Aver-
aging these metrics over all nodes in a system, one can estimate the inherent controllability of an
entire network itself. Control theoretic efforts such as these have only recently been applied to un-
derstand the locomotion of the nematode 342 and the networked behavior of the brain more broadly
237, 239, 279, promising new strategies for stimulation-based therapies and fresh insights about the
brain's capacity for intrinsic control.
36
Controllability metricsLinear controlLinear dynamicsuisignalAijnetworkxistateInitial
statex1x3x2Controlled
trajectoryUncontrolled
trajectoryAverage controllability:
Nearby statesModal controllability:
Distant statesControl
energyx1Initial
statex(t+1) = Ax(t) + u(t) References
1. Lear, J. Aristotle: The Desire to Understand (Cambridge University Press, 1988).
2. Aristotle. Metaphysics, vol. VII.7.
3. Stenger, V. J., Lindsay, J. A. & Boghossian, P. Physicists are philosophers, too. Scientific
American (2015).
4. von Neumann, J. The Computer and the Brain (1958).
5. Hopfield, J. J. Neural networks and physical systems with emergent collective computational
abilities. Proc. Nat. Acad. Sci. (USA) 79, 2554 -- 2558 (1982).
6. Scott, A. Neurophysics (Wiley, 1977).
7. Koch, C. & Poggio, T. A theoretical analysis of electrical properties of spines. Proc R Soc
Lond B Biol Sci 218, 455 -- 477 (1983).
8. Tyler, W. J. The mechanobiology of brain function. Nature Reviews Neuroscience 13, 867 --
878 (2012).
9. Friston, K., Kilner, J. & Harrison, L. A free energy principle for the brain. J Physiol Paris
100, 70 -- 87 (2006).
10. Plewes, D. B. & Kucharczyk, W. Physics of MRI: a primer. J Magn Reson Imaging 35,
1038 -- 1054 (2012).
11. Hari, R. & Salmelin, R. Magnetoencephalography: From SQUIDs to neuroscience. Neu-
roimage 20th anniversary special edition. Neuroimage 61, 386 -- 396 (2012).
12. Boto, E. et al. Moving magnetoencephalography towards real-world applications with a
wearable system. Nature 555, 657 -- 661 (2018).
13. Alivisatos, A. P. et al. Nanotools for neuroscience and brain activity mapping. ACS Nano 7,
1850 -- 1866 (2013).
37
14. Piazza, S., Bianchini, P., Sheppard, C., Diaspro, A. & Duocastella, M. Enhanced volumetric
imaging in 2-photon microscopy via acoustic lens beam shaping. J Biophotonics 11 (2018).
15. Boyden, E. S., Zhang, F., Bamberg, E., Nagel, G. & Deisseroth, K. Millisecond-timescale,
genetically targeted optical control of neural activity. Nat Neurosci 8, 1263 -- 1268 (2005).
16. Popkin, G. Physicists, the brain is calling you (2016).
17. McCulloch, W. S. & Pitts, W. A logical calculus of the ideas immanent in nervous activity.
Bull Math Biol 5, 115 -- 133 (1943).
18. Fries, P. Rhythms for cognition: Communication through coherence. Neuron 88, 220 -- 235
(2015).
19. Betzel, R. F. & Bassett, D. S. Specificity and robustness of long-distance connections in
weighted, interareal connectomes. Proc Natl Acad Sci U S A 115, E4880 -- E4889 (2018).
20. Poincare, H. Science and Hypothesis (London: Walter Scott Publishing, 1905).
21. Van Essen, D. C. et al. The WU-Minn Human Connectome Project: an overview. Neuroim-
age 80, 62 -- 79 (2013).
22. Markram, H. et al. Reconstruction and simulation of neocortical microcircuitry. Cell 163,
456 -- 492 (2015).
23. Poo, M. M. et al. China Brain Project: Basic Neuroscience, Brain Diseases, and Brain-
Inspired Computing. Neuron 92, 591 -- 596 (2016).
24. Okano, H., Miyawaki, A. & Kasai, K. Brain/MINDS: brain-mapping project in Japan. Philos
Trans R Soc Lond B Biol Sci 370, 20140310 (2015).
25. Bassett, D. S. & Gazzaniga, M. S. Understanding complexity in the human brain. Trends
Cogn Sci 15, 200 -- 209 (2011).
26. Sethna, J. P. Statistical Mechanics: Entropy, Order Parameters and Complexity (Oxford
University Press, 2006).
38
27. Bassett, D. S. & Bullmore, E. T. Small-world brain networks revisited. Neuroscientist Sep
21, 1073858416667720 (2016).
28. Albert, E. & Barabasi, A.-L. Statistical mechanics of complex networks. Rev. Mod. Phys. 74
(2002).
29. Butts, C. T. Revisiting the foundations of network analysis. Science 325, 414 -- 416 (2009).
30. Costa, L. d. F., Rodrigues, F. A., Travieso, G. & Villas Boas, P. R. Characterization of
complex networks: A survey of measurements. Advances In Physics 56, 167 -- 242 (2006).
31. Gross, T. & Blasius, B. Adaptive coevolutionary networks: a review. J R Soc Interface 5,
259 -- 271 (2008).
32. Zhang, X., Moore, C. & Newman, M. E. J. Random graph models for dynamic networks.
Eur. Phys. J. B 90, 200 (2017).
33. Hackett, A., Melnik, s. & Gleeson, J. P. Cascades on a class of clustered random networks.
Phys. Rev. E 83, 056107 (2011).
34. The structure and function of complex networks. SIAM REVIEW 45, 167 -- 256 (2003).
35. Motter, A. E. Networkcontrology. Chaos 25, 097621 (2015).
36. Bassett, D. S., Zurn, P. & Gold, J. I. On the nature and use of models in network neuroscience.
Nat Rev Neurosci Epub ahead of print (2018).
37. Pereda, A. E. Electrical synapses and their functional interactions with chemical synapses.
Nat Rev Neurosci 15, 250 -- 263 (2014).
38. Avena-Koenigsberger, A., Misic, B. & Sporns, O. Communication dynamics in complex
brain networks. Nat Rev Neurosci 19, 17 -- 33 (2017).
39. Ising, E. Beitrag zur theorie des ferromagnetismus. Zeitschrift fur Physik 31, 253 -- 258
(1925).
39
40. Onsager, L. Crystal statistics. i. a two-dimensional model with an order-disorder transition.
Phys. Rev. 65, 117 (1944).
41. Brush, S. G. History of the lenz-ising model. Rev. Mod. Phys. 39, 883 (1967).
42. Sporns, O., Chialvo, D. R., Kaiser, M. & Hilgetag, C. C. Organization, development and
function of complex brain networks. Trends Cogn Sci 8, 418 -- 425 (2004).
43. Medaglia, J. D., Lynall, M. E. & Bassett, D. S. Cognitive network neuroscience. J Cogn
Neurosci 27, 1471 -- 1491 (2015).
44. Sporns, O. Contributions and challenges for network models in cognitive neuroscience. Nat
Neurosci 17, 652 -- 660 (2014).
45. Petersen, S. E. & Sporns, O. Brain networks and cognitive architectures. Neuron 88, 207 -- 219
(2015).
46. Misic, B. & Sporns, O. From regions to connections and networks: new bridges between
brain and behavior. Curr Opin Neurobiol 40, 1 -- 7 (2016).
47. Wallace, E., Maei, H. R. & Latham, P. E. Randomly connected networks have short temporal
memory. Neural Comput 25, 1408 -- 1439 (2013).
48. Rajan, K., Harvey, C. D. & Tank, D. W. Recurrent network models of sequence generation
and memory. Neuron 90, 128 -- 142 (2016).
49. Chaudhuri, R. & Fiete, I. Computational principles of memory. Nat Neurosci 19, 394 -- 403
(2016).
50. Hermundstad, A. M., Brown, K. S., Bassett, D. S. & Carlson, J. M. Learning, memory, and
the role of neural network architecture. PLoS Comput Biol 7, e1002063 (2011).
51. Teileanu, T., Olveczky, B. & Balasubramanian, V. Rules and mechanisms for efficient two-
stage learning in neural circuits. Elife 6, e20944 (2017).
40
52. Takemura, S. Y. et al. A visual motion detection circuit suggested by drosophila connec-
tomics. Nature 500, 175 -- 181 (2013).
53. Zhen, M. & Samuel, A. D. C. elegans locomotion: small circuits, complex functions. Curr
Opin Neurobiol 33, 117 -- 126 (2015).
54. Golgi, C. Sulla fina anatomia degli organi centrali del sistema nervoso (S. Calderini, 1885).
55. y Cajal, S. R. Estructura de los centros nerviosos de las aves (1888).
56. Shepherd, G. M. Foundations of the neuron doctrine (Oxford University Press, 2015).
57. White, J. G., Southgate, E., Thomson, J. N. & Brenner, S. The structure of the nervous
system of the nematode Caenorhabditis elegans. Phil. Trans. R. Soc. Lond. B 314, 1 -- 340
(1986).
58. Helmstaedter, M. et al. Connectomic reconstruction of the inner plexiform layer in the mouse
retina. Nature 500, 168 -- 174 (2013).
59. Sporns, O., Tononi, G. & Kotter, R. The human connectome: a structural description of the
human brain. PLoS Comput Biol 1, e42 (2005).
60. Stephan, K. E. et al. Advanced database methodology for the collation of connectivity data
on the Macaque brain (CoCoMac). Philos Trans R Soc Lond B Biol Sci 356, 1159 -- 1186
(2001).
61. Markov, N. T. et al. A weighted and directed interareal connectivity matrix for macaque
cerebral cortex. Cereb Cortex 24, 17 -- 36 (2014).
62. Young, M. P., Scannell, J. W., Burns, G. A. & Blakemore, C. Analysis of connectivity: neural
systems in the cerebral cortex. Rev Neurosci 5, 227 -- 250 (1994).
63. Oh, S. W. et al. A mesoscale connectome of the mouse brain. Nature 508, 207 -- 214 (2014).
64. Shih, C. T. et al. Connectomics-based analysis of information flow in the Drosophila brain.
Curr Biol 25, 1249 -- 1258 (2015).
41
65. Hsieh, J. et al. Computed tomography: principles, design, artifacts, and recent advances
(SPIE Bellingham, WA, 2009).
66. Pierpaoli, C., Jezzard, P., Basser, P. J., Barnett, A. & Di Chiro, G. Diffusion tensor MR
imaging of the human brain. Radiology 201, 637 -- 648 (1996).
67. Basser, P. J., Pajevic, S., Pierpaoli, C., Duda, J. & Aldroubi, A. In vivo fiber tractography
using DT-MRI data. Magn Reson Med 44, 625 -- 632 (2000).
68. Behrens, T. E. & Johansen-Berg, H. Relating connectional architecture to grey matter func-
tion using diffusion imaging. Philos Trans R Soc Lond B Biol Sci 360, 903 -- 911 (2005).
69. Bullmore, E. & Sporns, O. The economy of brain network organization. Nat. Rev. Neurosci.
13, 336 -- 349 (2012).
70. Betzel, R. F. & Bassett, D. S. Generative models for network neuroscience: prospects and
promise. J R Soc Interface 14, 20170623 (2017).
71. Bassett, D. S. & Sporns, O. Network neuroscience. Nat Neurosci 20, 353 -- 364 (2017).
72. Nicosia, V., V´ertes, P. E., Schafer, W. R., Latora, V. & Bullmore, E. T. Phase transition in
the economically modeled growth of a cellular nervous system. Proceedings of the National
Academy of Sciences USA 110, 7880 -- 7885 (2013).
73. Henriksen, S., Pang, R. & Wronkiewicz, M. A simple generative model of the mouse
mesoscale connectome. Elife 5, e12366 (2016).
74. Beul, S. F., Grant, S. & Hilgetag, C. C. A predictive model of the cat cortical connectome
based on cytoarchitecture and distance. Brain Struct Funct 220, 3167 -- 3184 (2015).
75. Ercsey-Ravasz, M. et al. A predictive network model of cerebral cortical connectivity based
on a distance rule. Neuron 80, 184 -- 197 (2013).
76. Beul, S. F., Barbas, H. & Hilgetag, C. C. A predictive structural model of the primate con-
nectome. Sci Rep 7, 43176 (2017).
42
77. Betzel, R. F. et al. Generative models of the human connectome. Neuroimage 124, 1054 --
1064 (2016).
78. Thompson, P. M. et al. Genetic influences on brain structure. Nat. Neurosci. 4, 1253 (2001).
79. Raz, N. et al. Regional brain changes in aging healthy adults: general trends, individual
differences and modifiers. Cereb. Cortex 15, 1676 -- 1689 (2005).
80. Gong, G. et al. Age-and gender-related differences in the cortical anatomical network. J.
Neurosci. 29, 15684 -- 15693 (2009).
81. Kanai, R. & Rees, G. The structural basis of inter-individual differences in human behaviour
and cognition. Nat. Rev. Neurosci. 12, 231 (2011).
82. Banissy, M. J., Kanai, R., Walsh, V. & Rees, G. Inter-individual differences in empathy are
reflected in human brain structure. Neuroimage 62, 2034 -- 2039 (2012).
83. Fleming, S. M., Weil, R. S., Nagy, Z., Dolan, R. J. & Rees, G. Relating introspective accuracy
to individual differences in brain structure. Science 329, 1541 -- 1543 (2010).
84. Hartley, C. A., Fischl, B. & Phelps, E. A. Brain structure correlates of individual differences
in the acquisition and inhibition of conditioned fear. Cereb. Cortex 21, 1954 -- 1962 (2011).
85. Kanai, R., Feilden, T., Firth, C. & Rees, G. Political orientations are correlated with brain
structure in young adults. Curr. Biol. 21, 677 -- 680 (2011).
86. Erdos, P. & R´enyi, A. On the evolution of random graphs. Publ. Math. Inst. Hung. Acad. Sci
5, 17 -- 60 (1960).
87. Sherrington, C. S. The Integrative Action of the Nervous System (Yale University Press,
1906).
88. Sporns, O., Tononi, G. & Edelman, G. M. Theoretical neuroanatomy: relating anatomical
and functional connectivity in graphs and cortical connection matrices. Cereb. cortex 10,
127 -- 141 (2000).
43
89. Hilgetag, C.-C., Burns, G. A., O'Neill, M. A., Scannell, J. W. & Young, M. P. Anatomical
connectivity defines the organization of clusters of cortical areas in the macaque and the cat.
Phil. Trans. R. Soc. Lon. B 355, 91 -- 110 (2000).
90. Sporns, O. & Zwi, J. D. The small world of the cerebral cortex. Neuroinformatics 2, 145 -- 162
(2004).
91. Sporns, O. & Betzel, R. F. Modular brain networks. Annu Rev Psychol 67, 613 -- 640 (2016).
92. Bassett, D. S. et al. Efficient physical embedding of topologically complex information pro-
cessing networks in brains and computer circuits. PLoS Comput. Biol. 6, e1000748 (2010).
93. Taylor, P. N., Wang, Y. & Kaiser, M. Within brain area tractography suggests local modularity
using high resolution connectomics. Sci Rep 7, 39859 (2017).
94. Lesicko, A. M., Hristova, T. S., Maigler, K. C. & Llano, D. A. Connectional modularity
of top-down and bottom-up multimodal inputs to the lateral cortex of the mouse inferior
colliculus. J Neurosci 36, 11037 -- 11050 (2016).
95. Sohn, Y., Choi, M. K., Ahn, Y. Y., Lee, J. & Jeong, J. Topological cluster analysis reveals
the systemic organization of the Caenorhabditis elegans connectome. PLoS Comput Biol 7,
e1001139 (2011).
96. Azulay, A., Itskovits, E. & Zaslaver, A. The C. elegans connectome consists of homogenous
circuits with defined functional roles. PLoS Comput Biol 12, e1005021 (2016).
97. Betzel, R. F. & Bassett, D. S. Multi-scale brain networks. Neuroimage 160, 73 -- 83 (2017).
98. Khambhati, A. N., Sizemore, A. E., Betzel, R. F. & Bassett, D. S. Modeling and interpreting
mesoscale network dynamics. Neuroimage S1053-8119, 30500 -- 1 (2017).
99. Aicher, C., Jacobs, A. Z. & Clauset, A. Learning latent block structure in weighted networks.
Journal of Complex Networks 3, 221 -- 248 (2015).
44
100. Betzel, R. F., Medaglia, J. D. & Bassett, D. S. Diversity of meso-scale architecture in human
and non-human connectomes. Nature Communications 9, 346 (2018).
101. van den Heuvel, M. P. & Sporns, O. Network hubs in the human brain. Trends Cogn. Sci. 17,
683 -- 696 (2013).
102. Liao, X., Vasilakos, A. V. & He, Y. Small-world human brain networks: Perspectives and
challenges. Neurosci Biobehav Rev 77, 286 -- 300 (2017).
103. Deco, G., Tononi, G., Boly, M. & Kringelbach, M. L. Rethinking segregation and integration:
contributions of whole-brain modelling. Nat. Rev. Neurosci. 16, 430 (2015).
104. Latora, V. & Marchiori, M. Efficient behavior of small-world networks. Phys. Rev. Lett. 87,
198701 (2001).
105. Kaiser, M. & Hilgetag, C. C. Nonoptimal component placement, but short processing paths,
due to long-distance projections in neural systems. PLOS Comput. Biol. 2, e95 (2006).
106. Travers, J. & Milgram, S. The small world problem. Phychology Today 1, 61 -- 67 (1967).
107. Watts, D. J. & Strogatz, S. H. Collective dynamics of 'small-world' networks. Nature 393,
440 -- 442 (1998).
108. Gong, G. et al. Mapping anatomical connectivity patterns of human cerebral cortex using in
vivo diffusion tensor imaging tractography. Cereb. cortex 19, 524 -- 536 (2008).
109. Wedeen, V. J., Hagmann, P., Tseng, W.-Y. I., Reese, T. G. & Weisskoff, R. M. Mapping
complex tissue architecture with diffusion spectrum magnetic resonance imaging. Magn.
Reson. Med. 54, 1377 -- 1386 (2005).
110. de Solla Price, D. J. Networks of scientific papers. Science 149, 510 -- 515 (1965).
111. Barabasi, A. L. & Albert, R. Emergence of scaling in random networks. Science 286, 509 --
512 (1999).
45
112. Dall, J. & Christensen, M. Random geometric graphs. Physical Review E 66, 016121 (2002).
113. Vertes, P. E. et al. Simple models of human brain functional networks. Proc Natl Acad Sci
U S A 109, 5868 -- 5873 (2012).
114. Rubinov, M., Ypma, R., Watson, C. & Bullmore, E. Wiring cost and topological participation
of the mouse brain connectome. Proceedings of the National Academy of Sciences of the USA
doi/10.1073/pnas.1420315112 (2015).
115. Kaiser, M. Mechanisms of connectome development. Trends Cogn Sci 21, 703 -- 717 (2017).
116. Stam, C. J. Modern network science of neurological disorders. Nat Rev Neurosci 15, 683 -- 695
(2014).
117. Scholtens, L. H., Schmidt, R., de Reus, M. A. & van den Heuvel, M. P. Linking macroscale
graph analytical organization to microscale neuroarchitectonics in the macaque connectome.
J Neurosci 34, 12192 -- 12205 (2014).
118. Chaudhuri, R., Knoblauch, K., Gariel, M. A., Kennedy, H. & Wang, X. J. A large-scale
circuit mechanism for hierarchical dynamical processing in the primate cortex. Neuron 88,
419 -- 431 (2015).
119. Breakspear, M. Dynamic models of large-scale brain activity. Nat Neurosci 20, 340 -- 352
(2017).
120. Bentley, B. et al. The multilayer connectome of Caenorhabditis elegans. PLoS Comput Biol
12, e1005283 (2016).
121. Mejias, J. F., Murray, J. D., Kennedy, H. & Wang, X. J. Feedforward and feedback frequency-
dependent interactions in a large-scale laminar network of the primate cortex. Sci Adv. 2,
e1601335 (2016).
122. Seung, H. S. & Sumbul, U. Neuronal cell types and connectivity: lessons from the retina.
Neuron 83, 1262 -- 1272 (2014).
46
123. Arnatkeviciute, A., Fulcher, B. D., Pocock, R. & Fornito, A. Hub connectivity, neuronal
diversity, and gene expression in the Caenorhabditis elegans connectome. PLoS Comput Biol
14, e1005989 (2018).
124. Scholz, J., Klein, M. C., Behrens, T. E. & Johansen-Berg, H. Training induces changes in
white-matter architecture. Nat Neurosci 12, 1370 -- 1371 (2009).
125. Baum, G. L. et al. Modular segregation of structural brain networks supports the development
of executive function in youth. Curr Biol 27, 1561 -- 1572.e8 (2017).
126. Zuo, X. N. et al. Human connectomics across the life span. Trends Cogn Sci 21, 32 -- 45
(2017).
127. Holme, P. & Saramaki, J. Temporal networks. Phys. Rep. 519, 97 -- 125 (2012).
128. Li, A., Cornelius, S. P., Liu, Y.-Y., Wang, L. & Barab´asi, A.-L. The fundamental advantages
of temporal networks. Science 358, 1042 -- 1046 (2017).
129. Hebb, D. The Organization of Behavior (Wiley, 1949).
130. Magee, J. C. & Johnston, D. A synaptically controlled, associative signal for hebbian plas-
ticity in hippocampal neurons. Science 275, 209 -- 213 (1997).
131. Montague, P. R., Dayan, P. & Sejnowski, T. J. A framework for mesencephalic dopamine
systems based on predictive hebbian learning. J. Neurosci. 16, 1936 -- 1947 (1996).
132. Song, S., Miller, K. D. & Abbott, L. F. Competitive hebbian learning through spike-timing-
dependent synaptic plasticity. Nat. Neurosci. 3, 919 (2000).
133. Chialvo, D. R. Emergent complex neural dynamics. Nat. Phys. 6, 744 (2010).
134. Tononi, G., Boly, M., Massimini, M. & Koch, C. Integrated information theory: from con-
sciousness to its physical substrate. Nat Rev Neurosci 17, 450 -- 461 (2016).
135. Abbott, L. F. & Dayan, P. Theoretical Neuroscience (MIT Press, 2001).
47
136. Dechery, J. B. & MacLean, J. N. Emergent cortical circuit dynamics contain dense, interwo-
ven ensembles of spike sequences. J Neurophysiol 118, 1914 -- 1925 (2017).
137. Reif, F. Fundamentals of statistical and thermal physics (Waveland Press, 2009).
138. Brody, C. D. Correlations without synchrony. Neural Comput 11, 1537 -- 1551 (1999).
139. Brody, C. D. Disambiguating different covariation types. Neural Comput 11, 1527 -- 1535
(1999).
140. Sporns, O., Tononi, G. & Edelman, G. M. Connectivity and complexity: the relationship
between neuroanatomy and brain dynamics. Neural Netw 13, 909 -- 922 (2000).
141. Schneidman, E., Berry II, M. J., Segev, R. & Bialek, W. Weak pairwise correlations imply
strongly correlated network states in a neural population. Nature 440, 1007 (2006).
142. Levina, A., Herrmann, J. M. & Geisel, T. Dynamical synapses causing self-organized criti-
cality in neural networks. Nat. Phys. 3, 857 (2007).
143. Vuksanovic, V. & Hovel, P. Functional connectivity of distant cortical regions: role of remote
synchronization and symmetry in interactions. Neuroimage 97, 1 -- 8 (2014).
144. Flourens, P. Recherches exp´erimentales sur les propri´et´es et les fonctions du syst`eme nerveux
dans les animaux vert´ebr´es (Balli`ere, 1842).
145. Panizza, B. Osservazioni sul nervo ottico (Bernardoni, 1855).
146. Broca, P. Remarques sur le si`ege de la facult´e du langage articul´e, suivies dune observation
daph´emie (perte de la parole). Bulletin et Memoires de la Societe anatomique de Paris 6,
330 -- 357 (1861).
147. von Helmholtz, H. Vorlaufiger bericht uber die fortpflanzungs-geschwindigkeit der nerven-
reizung. Archiv fur Anatomie, Physiologie und wissenschaftliche Medicin 71 -- 73 (1850).
148. Haas, L. F. Hans berger (1873 -- 1941), richard caton (1842 -- 1926), and electroencephalogra-
phy. J. Neurol. Neurosurg. Psychiatry 74, 9 -- 9 (2003).
48
149. Caton, R. Electrical currents of the brain. J. Nerv. Ment. Dis. 2, 610 (1875).
150. Beck, A. Die strome der nervencentren. Centralbl Physiol 4, 572 -- 573 (1890).
151. Lorente de N´o, R. Studies on the structure of the cerebral cortex. ii. continuation of the study
of the ammonic system. Journal fur Psychologie und Neurologie (1934).
152. Green, D. J. & Gillette, R. Circadian rhythm of firing rate recorded from single cells in the
rat suprachiasmatic brain slice. Brain Res. 245, 198 -- 200 (1982).
153. Edwards, F. A., Konnerth, A., Sakmann, B. & Takahashi, T. A thin slice preparation for patch
clamp recordings from neurones of the mammalian central nervous system. Pflugers Archiv
414, 600 -- 612 (1989).
154. Stosiek, C., Garaschuk, O., Holthoff, K. & Konnerth, A. In vivo two-photon calcium imaging
of neuronal networks. Proc Natl Acad Sci U S A 100, 7319 -- 7324 (2003).
155. Grewe, B. F., Langer, D., Kasper, H., Kampa, B. M. & Helmchen, F. High-speed in vivo
calcium imaging reveals neuronal network activity with near-millisecond precision. Nat.
Methods 7, 399 (2010).
156. Penny, W. D., Friston, K. J., Ashburner, J. T., Kiebel, S. J. & Nichols, T. E. Statistical
parametric mapping: the analysis of functional brain images (Elsevier, 2011).
157. Hamalainen, M., Hari, R., Ilmoniemi, R. J., Knuutila, J. & Lounasmaa, O. V. Magnetoen-
cephalographytheory, instrumentation, and applications to noninvasive studies of the working
human brain. Rev. Mod. Phys. 65, 413 (1993).
158. Bailey, D. L., Maisey, M. N., Townsend, D. W. & Valk, P. E. Positron emission tomography
(Springer, 2005).
159. Raichle, M. E. Behind the scenes of functional brain imaging: a historical and physiological
perspective. Proc Natl Acad Sci U S A 95, 765 -- 772 (1998).
49
160. Zarahn, E., Aguirre, G. K. & D'Esposito, M. Empirical analyses of bold fmri statistics.
Neuroimage 5, 179 -- 197 (1997).
161. Van Den Heuvel, M. P. & Pol, H. E. H. Exploring the brain network: a review on resting-state
fmri functional connectivity. Eur Neuropsychopharmacol 20, 519 -- 534 (2010).
162. Bullmore, E. & Sporns, O. Complex brain networks: graph theoretical analysis of structural
and functional systems. Nat Rev Neurosci 10, 186 -- 198 (2009).
163. Zalesky, A., Fornito, A. & Bullmore, E. On the use of correlation as a measure of network
connectivity. Neuroimage 60, 2096 -- 2106 (2012).
164. He, Y. et al. Uncovering intrinsic modular organization of spontaneous brain activity in
humans. PloS one 4, e5226 (2009).
165. Salvador, R. et al. Neurophysiological architecture of functional magnetic resonance images
of human brain. Cerebral cortex 15, 1332 -- 1342 (2005).
166. Achard, S., Salvador, R., Whitcher, B., Suckling, J. & Bullmore, E. A resilient, low-
frequency, small-world human brain functional network with highly connected association
cortical hubs. J Neurosci 26, 63 -- 72 (2006).
167. Bettencourt, L. M., Stephens, G. J., Ham, M. I. & Gross, G. W. Functional structure of
cortical neuronal networks grown in vitro. Phys Rev E 75, 021915 (2007).
168. Sadovsky, A. J. & MacLean, J. N. Scaling of topologically similar functional modules defines
mouse primary auditory and somatosensory microcircuitry. J Neurosci 33, 14048 -- 14060
(2013).
169. Yue, Q. et al. Brain modularity mediates the relation between task complexity and perfor-
mance. J Cogn Neurosci 29, 1532 -- 1546 (2017).
170. Bassett, D. S. & Bullmore, E. Small-world brain networks. Neuroscientist 12, 512 -- 523
(2006).
50
171. Rosenbaum, R., Smith, M. A., Kohn, A., Rubin, J. E. & Doiron, B. The spatial structure of
correlated neuronal variability. Nat Neurosci 20, 107 -- 114 (2017).
172. Goni, J. et al. Resting-brain functional connectivity predicted by analytic measures of net-
work communication. Proceedings of the National Academy of Sciences 111, 833 -- 838
(2014).
173. Honey, C. et al. Predicting human resting-state functional connectivity from structural con-
nectivity. Proceedings of the National Academy of Sciences 106, 2035 -- 2040 (2009).
174. Medaglia, J. D. et al. Functional alignment with anatomical networks is associated with
cognitive flexibility. Nature Human Behaviour 2, 156 -- 164 (2018).
175. Park, H.-J. & Friston, K. Structural and functional brain networks: from connections to
cognition. Science 342, 1238411 (2013).
176. Hodgkin, A. L. & Huxley, A. F. A quantitative description of membrane current and its
application to conduction and excitation in nerve. J. Physiol. 117, 500 -- 544 (1952).
177. FitzHugh, R. Impulses and physiological states in theoretical models of nerve membrane.
Biophys. J. 1, 445 -- 466 (1961).
178. Beurle, R. L. Properties of a mass of cells capable of regenerating pulses. Phil. Trans. R.
Soc. Lond. B 240, 55 -- 94 (1956).
179. Wilson, H. R. & Cowan, J. D. Excitatory and inhibitory interactions in localized populations
of model neurons. Biophys. J. 12, 1 -- 24 (1972).
180. Kuramoto, Y. Chemical oscillations, waves, and turbulence, vol. 19 (Springer Science &
Business Media, 2012).
181. Cash, S. & Yuste, R. Linear summation of excitatory inputs by ca1 pyramidal neurons.
Neuron 22, 383 -- 394 (1999).
51
182. Ferrell, J. E. & Machleder, E. M. The biochemical basis of an all-or-none cell fate switch in
xenopus oocytes. Science 280, 895 -- 898 (1998).
183. Hearst, M. A., Dumais, S. T., Osuna, E., Platt, J. & Scholkopf, B. Support vector machines.
IEEE Intell. Syst. 13, 18 -- 28 (1998).
184. Kleene, S. C. Representation of events in nerve nets and finite automata. Tech. Rep., RAND
PROJECT AIR FORCE SANTA MONICA CA (1951).
185. Schmidhuber, J. Deep learning in neural networks: An overview. Neural Netw. 61, 85 -- 117
(2015).
186. Egmont-Petersen, M., de Ridder, D. & Handels, H. Image processing with neural networksa
review. Pattern Recognit. 35, 2279 -- 2301 (2002).
187. Hinton, G. et al. Deep neural networks for acoustic modeling in speech recognition: The
shared views of four research groups. IEEE Signal Process. Mag. 29, 82 -- 97 (2012).
188. Silver, D. et al. Mastering the game of go with deep neural networks and tree search. Nature
529, 484 (2016).
189. Newman, C. M. Memory capacity in neural network models: Rigorous lower bounds. Neural
Netw. 1, 223 -- 238 (1988).
190. Hertz, J., Krogh, A. & Palmer, R. G.
Introduction to the theory of neural computation.
(Addison-Wesley/Addison Wesley Longman, 1991).
191. Moosavi, S. A. & Montakhab, A. Structural versus dynamical origins of mean-field behavior
in a self-organized critical model of neuronal avalanches. Phys Rev E Stat Nonlin Soft Matter
Phys 92, 052804 (2015).
192. Woodrow, W. L. et al. Adaptation to sensory input tunes visual cortex to criticality. Nature
Physics 11, 659 -- 663 (2015).
52
193. Haldeman, C. & Beggs, J. M. Critical branching captures activity in living neural networks
and maximizes the number of metastable states. Phys. Rev. Lett. 94, 058101 (2005). URL
https://link.aps.org/doi/10.1103/PhysRevLett.94.058101.
194. Beggs, J. M. & Plenz, D.
Neuronal avalanches in neocortical circuits.
Jour-
nal of Neuroscience 23, 11167 -- 11177 (2003).
org/content/23/35/11167.
URL http://www.jneurosci.
http://www.jneurosci.org/content/23/
35/11167.full.pdf.
195. Kinouchi, O. & Copelli, M. Optimal dynamical range of excitable networks at criticality.
Nature Physics 2, 348 EP -- (2006). URL http://dx.doi.org/10.1038/nphys289.
196. Shew, W. L., Yang, H., Petermann, T., Roy, R. & Plenz, D. Neuronal avalanches im-
ply maximum dynamic range in cortical networks at criticality. Journal of Neuroscience
29, 15595 -- 15600 (2009). URL http://www.jneurosci.org/content/29/49/
15595. http://www.jneurosci.org/content/29/49/15595.full.pdf.
197. Bertschinger, N. & Natschlager, T. Real-time computation at the edge of chaos in re-
current neural networks. Neural Computation 16, 1413 -- 1436 (2004). URL https:
//doi.org/10.1162/089976604323057443. https://doi.org/10.1162/
089976604323057443.
198. Ganmor, E., Segev, R. & Schneidman, E. Sparse low-order interaction network underlies
a highly correlated and learnable neural population code. Proc Natl Acad Sci U S A 108,
9679 -- 9684 (2011).
199. Lee, S.-G., Neiman, A. & Kim, S. Coherence resonance in a hodgkin-huxley neuron. Phys.
Rev. E 57, 3292 (1998).
200. Hille, B. et al. Ion channels of excitable membranes, vol. 507 (Sinauer Sunderland, MA,
2001).
201. Plant, R. & Kim, M. Mathematical description of a bursting pacemaker neuron by a modifi-
cation of the hodgkin-huxley equations. Biophys. J. 16, 227 -- 244 (1976).
53
202. Andersen, S. S., Jackson, A. D. & Heimburg, T. Towards a thermodynamic theory of nerve
pulse propagation. Prog. Neurobiol. 88, 104 -- 113 (2009).
203. Pakdaman, K., Thieullen, M. & Wainrib, G. Fluid limit theorems for stochastic hybrid sys-
tems with application to neuron models. Adv. Appl. Probab. 42, 761 -- 794 (2010).
204. Nagumo, J., Arimoto, S. & Yoshizawa, S. An active pulse transmission line simulating nerve
axon. Proc. IRE 50, 2061 -- 2070 (1962).
205. Niebur, E. & Erdos, P. Theory of the locomotion of nematodes: control of the somatic motor
neurons by interneurons. Math. Biosci. 118, 51 -- 82 (1993).
206. Bryden, J. & Cohen, N. A simulation model of the locomotion controllers for the nematode
caenorhabditis elegans. In From Animals to Animats 8: Proceedings of the Eighth Interna-
tional Conference on the Simulation of Adaptive Behavior, 183 -- 192 (MIT Press, 2004).
207. Arena, P., Patan´e, L. & Termini, P. S. An insect brain computational model inspired by
drosophila melanogaster: simulation results. In Neural Networks (IJCNN), The 2010 Inter-
national Joint Conference on, 1 -- 8 (IEEE, 2010).
208. Markram, H. The blue brain project. Nat. Rev. Neurosci. 7, 153 (2006).
209. Kishimoto, K. & Amari, S.-i. Existence and stability of local excitations in homogeneous
neural fields. J. Math. Biol. 7, 303 -- 318 (1979).
210. Pinto, D. J. & Ermentrout, G. B. Spatially structured activity in synaptically coupled neuronal
networks: I. traveling fronts and pulses. SIAM J Appl Math 62, 206 -- 225 (2001).
211. David, O. & Friston, K. J. A neural mass model for meg/eeg:: coupling and neuronal dynam-
ics. NeuroImage 20, 1743 -- 1755 (2003).
212. David, O., Cosmelli, D. & Friston, K. J. Evaluation of different measures of functional
connectivity using a neural mass model. Neuroimage 21, 659 -- 673 (2004).
54
213. Kuramoto, Y. & Araki, H. Lecture notes in physics, international symposium on mathemati-
cal problems in theoretical physics (1975).
214. Ward, L. M. Synchronous neural oscillations and cognitive processes. Trends Cogn. Sci. 7,
553 -- 559 (2003).
215. Fries, P. A mechanism for cognitive dynamics: neuronal communication through neuronal
coherence. Trends Cogn. Sci. 9, 474 -- 480 (2005).
216. Palmigiano, A., Geisel, T., Wolf, F. & Battaglia, D. Flexible information routing by transient
synchrony. Nat Neurosci 20, 1014 -- 1022 (2017).
217. Schnitzler, A. & Gross, J. Normal and pathological oscillatory communication in the brain.
Nat. Rev. Neurosci. 6, 285 (2005).
218. Cabral, J., Hugues, E., Sporns, O. & Deco, G. Role of local network oscillations in resting-
state functional connectivity. Neuroimage 57, 130 -- 139 (2011).
219. Petersson, K. M., Nichols, T. E., Poline, J.-B. & Holmes, A. P. Statistical limitations in
functional neuroimaging. i. non-inferential methods and statistical models. Philos. Trans. R.
Soc. Lond., B, Biol. Sci. 354, 1239 -- 1260 (1999).
220. Petersson, K. M., Nichols, T. E., Poline, J.-B. & Holmes, A. P. Statistical limitations in
functional neuroimaging ii. signal detection and statistical inference. Philos. Trans. R. Soc.
Lond., B, Biol. Sci. 354, 1261 -- 1281 (1999).
221. Bancaud, J. & Talairach, J. Methodology of stereo eeg exploration and surgical intervention
in epilepsy. Rev. Otoneuroophtalmol. 45, 315 -- 328 (1973).
222. Chauvel, P., Vignal, J., Biraben, A., Badier, J. & Scarabin, J. Stereoelectroencephalography,
80 -- 108 (Springer Verlag, 1996).
223. Todaro, C., Marzetti, L., Valdes Sosa, P. A., Valdes-Hernandez, P. A. & Pizzella, V. Map-
ping brain activity with electrocorticography: Resolution properties and robustness of inverse
solutions. Brain Topogr Epub Ahead of Print (2018).
55
224. Menon, R. S. & Kim, S.-G. Spatial and temporal limits in cognitive neuroimaging with fmri.
Trends Cogn. Sci. 3, 207 -- 216 (1999).
225. Aguirre, G. K. Functional neuroimaging: technical, logical, and social perspectives. Hastings
Cent. Rep. 44, S8 -- S18 (2014).
226. Ciric, R. et al. Benchmarking of participant-level confound regression strategies for the
control of motion artifact in studies of functional connectivity. Neuroimage 154, 174 -- 187
(2017).
227. Avants, B. B. et al. A reproducible evaluation of ANTs similarity metric performance in
brain image registration. Neuroimage 54, 2033 -- 2044 (2011).
228. Amari, S.-i., Nakahara, H., Wu, S. & Sakai, Y. Synchronous firing and higher-order interac-
tions in neuron pool. Neural Comput. 15, 127 -- 142 (2003).
229. Sizemore, A. E. et al. Cliques and cavities in the human connectome. J Comput Neurosci
Epub Ahead of Print (2017).
230. Giusti, C., Ghrist, R. & Bassett, D. S. Two's company, three (or more) is a simplex:
Algebraic-topological tools for understanding higher-order structure in neural data. J Comput
Neurosci 41, 1 -- 14 (2016).
231. Giusti, C., Pastalkova, E., Curto, C. & Itskov, V. Clique topology reveals intrinsic geometric
structure in neural correlations. Proc Natl Acad Sci U S A 112, 13455 -- 13460 (2015).
232. Reimann, M. W. et al. Cliques of neurons bound into cavities provide a missing link between
structure and function. Front Comput Neurosci 11, 48 (2017).
233. Battaglia, D., Witt, A., Wolf, F. & Geisel, T. Dynamic effective connectivity of inter-areal
brain circuits. PLoS Comput Biol 8, e1002438 (2012).
234. Zylberberg, J., Pouget, A., Latham, P. E. & Shea-Brown, E. Robust information propagation
through noisy neural circuits. PLoS Comput Biol 13, e1005497 (2017).
56
235. Kirst, C., Timme, M. & Battaglia, D. Dynamic information routing in complex networks.
Nat Commun 7, 11061 (2016).
236. McIntyre, C. C., Savasta, M., Kerkerian-Le Goff, L. & Vitek, J. L. Uncovering the mecha-
nism (s) of action of deep brain stimulation: activation, inhibition, or both. Clin. Neurophys-
iol. 115, 1239 -- 1248 (2004).
237. Lozano, A. M. & Lipsman, N. Probing and regulating dysfunctional circuits using deep brain
stimulation. Neuron 77, 406 -- 424 (2013).
238. Liu, Y.-Y. & Barab´asi, A.-L. Control principles of complex systems. Rev. Mod. Phys. 88,
035006 (2016).
239. Schiff, S. J. Neural control engineering: the emerging intersection between control theory
and neuroscience (MIT Press, 2012).
240. Kim, J. Z. et al. Role of graph architecture in controlling dynamical networks with applica-
tions to neural systems. Nature Physics Epub Ahead of Print (2018).
241. Gu, S. et al. Controllability of structural brain networks. Nat Commun 6, 8414 (2015).
242. Jeganathan, J. et al. Fronto-limbic dysconnectivity leads to impaired brain network control-
lability in young people with bipolar disorder and those at high genetic risk. Neuroimage
Clin 19, 71 -- 81 (2018).
243. Muldoon, S. F. et al. Stimulation-based control of dynamic brain networks. PLoS Comput
Biol 12, e1005076 (2016).
244. Taylor, P. N. et al. Optimal control based seizure abatement using patient derived connectiv-
ity. Front Neurosci 9, 202 (2015).
245. Medaglia, J. D. et al. Network controllability in the inferior frontal gyrus relates to controlled
language variability and susceptibility to TMS. J Neurosci 38, 6399 -- 6410 (2018).
57
246. Holt, A. B., Wilson, D., Shinn, M., Moehlis, J. & Netoff, T. I. Phasic burst stimulation: A
closed-loop approach to tuning deep brain stimulation parameters for Parkinson's disease.
PLoS Comput Biol 12, e1005011 (2016).
247. Holmes, G. Disturbances of vision by cerebral lesions. The British journal of ophthalmology
2, 353 (1918).
248. Owen, A. M., Downes, J. J., Sahakian, B. J., Polkey, C. E. & Robbins, T. W. Planning
and spatial working memory following frontal lobe lesions in man. Neuropsychologia 28,
1021 -- 1034 (1990).
249. Walsh, V. & Cowey, A. Transcranial magnetic stimulation and cognitive neuroscience. Nat.
Rev. Neurosci. 1, 73 (2000).
250. Amassian, V. E. et al. Measurement of information processing delays in human visual cortex
with repetitive magnetic coil stimulation. Brain Res. 605, 317 -- 321 (1993).
251. Pascual-Leone, A., Grafman, J. & Hallett, M. Modulation of cortical motor output maps
during development of implicit and explicit knowledge. Science 263, 1287 -- 1289 (1994).
252. Pascual-Leone, A., Gates, J. R. & Dhuna, A. Induction of speech arrest and counting errors
with rapid-rate transcranial magnetic stimulation. Neurology 41, 697 -- 702 (1991).
253. Walsh, V., Ellison, A., Battelli, L. & Cowey, A. Task -- specific impairments and enhancements
induced by magnetic stimulation of human visual area v5. Proc. R. Soc. Lond., B, Biol. Sci.
265, 537 -- 543 (1998).
254. Kringelbach, M. L., Jenkinson, N., Owen, S. L. & Aziz, T. Z. Translational principles of
deep brain stimulation. Nat. Rev. Neurosci. 8, 623 (2007).
255. George, M. S., Lisanby, S. H. & Sackeim, H. A. Transcranial magnetic stimulation: applica-
tions in neuropsychiatry. Archives of General Psychiatry 56, 300 -- 311 (1999).
256. Perlmutter, J. S. & Mink, J. W. Deep brain stimulation. Annu. Rev. Neurosci. 29, 229 -- 257
(2006).
58
257. Tass, P. et al. Detection of n: m phase locking from noisy data: Application to magnetoen-
cephalography. Phys. Rev. Lett. 81, 3291 (1998).
258. Santaniello, S. et al. Therapeutic mechanisms of high-frequency stimulation in parkinsons
disease and neural restoration via loop-based reinforcement. Proceedings of the National
Academy of Sciences 112, E586 -- E595 (2015).
259. Zeki, S. A vision of the brain (Blackwell Scientific Publ., 1993).
260. Chiken, S. & Nambu, A. Disrupting neuronal transmission: mechanism of dbs? Front. Syst.
Neurosci. 8, 33 (2014).
261. Ber´enyi, A., Belluscio, M., Mao, D. & Buzs´aki, G. Closed-loop control of epilepsy by
transcranial electrical stimulation. Science 337, 735 -- 737 (2012).
262. Kedzior, K. K., Gierke, L., Gellersen, H. M. & Berlim, M. T. Cognitive functioning and deep
transcranial magnetic stimulation (dtms) in major psychiatric disorders: a systematic review.
J. Psychiatr. Res. 75, 107 -- 115 (2016).
263. Ching, S. et al. Real-time closed-loop control in a rodent model of medically induced coma
using burst suppression. Anesthesiology 119, 848 -- 860 (2013).
264. Holt, A. B. & Netoff, T. I. Origins and suppression of oscillations in a computational model
of parkinsons disease. J. Comput. Neurosci. 37, 505 -- 521 (2014).
265. Heck, C. N. et al. Two-year seizure reduction in adults with medically intractable partial
onset epilepsy treated with responsive neurostimulation: final results of the RNS System
Pivotal trial. Epilepsia 55, 432 -- 441 (2014).
266. Crinion, J. et al. Spatial normalization of lesioned brains: performance evaluation and impact
on fmri analyses. Neuroimage 37, 866 -- 875 (2007).
267. Santaniello, S., Fiengo, G., Glielmo, L. & Grill, W. M. Closed-loop control of deep brain
stimulation: a simulation study. IEEE Trans. Neural. Syst. Rehabil. Eng. 19, 15 -- 24 (2011).
59
268. Iudice, F. L., Garofalo, F. & Sorrentino, F. Structural permeability of complex networks to
control signals. Nat. Commun. 6, 8349 (2015).
269. Posner, M. I., Snyder, C. R. & Solso, R. Attention and cognitive control. Cogn. Psychol. 205
(2004).
270. Fuster, J. M. & Alexander, G. E. Neuron activity related to short-term memory. Science 173,
652 -- 654 (1971).
271. Goldman, P. S. & Rosvold, H. E. Localization of function within the dorsolateral prefrontal
cortex of the rhesus monkey. Exp. Neurol. 27, 291 -- 304 (1970).
272. Bechara, A., Damasio, A. R., Damasio, H. & Anderson, S. W. Insensitivity to future conse-
quences following damage to human prefrontal cortex. Cognition 50, 7 -- 15 (1994).
273. Dias, R., Robbins, T. & Roberts, A. Dissociation in prefrontal cortex of affective and atten-
tional shifts. Nature 380, 69 (1996).
274. Gu, S. et al. Optimal trajectories of brain state transitions. Neuroimage 148, 305 -- 317 (2017).
275. Betzel, R. F., Gu, S., Medaglia, J. D., Pasqualetti, F. & Bassett, D. S. Optimally controlling
the human connectome: the role of network topology. Sci. Rep. 6, 30770 (2016).
276. Pasqualetti, F., Zampieri, S. & Bullo, F. Controllability metrics, limitations and algorithms
for complex networks. IEEE Trans. Control Network Syst. 1, 40 -- 52 (2014).
277. Tang, E. et al. Developmental increases in white matter network controllability support a
growing diversity of brain dynamics. Nat. Commun. 8, 1252 (2017).
278. Cornblath, E. J. et al. Sex differences in network controllability as a predictor of executive
function in youth. arXiv 1801, 04623.
279. Tang, E. & Bassett, D. S. Control of dynamics in brain networks. Reviews of Modern Physics
In Press (2018).
60
280. Adamantidis, A. R., Zhang, F., Aravanis, A. M., Deisseroth, K. & De Lecea, L. Neural
substrates of awakening probed with optogenetic control of hypocretin neurons. Nature 450,
420 (2007).
281. Deisseroth, K. Optogenetics. Nat. Methods 8, 26 (2011).
282. Gunaydin, L. A. et al. Ultrafast optogenetic control. Nature neuroscience 13, 387 (2010).
283. Grosenick, L., Marshel, J. H. & Deisseroth, K. Closed-loop and activity-guided optogenetic
control. Neuron 86, 106 -- 139 (2015).
284. Prakash, R. et al. Two-photon optogenetic toolbox for fast inhibition, excitation and bistable
modulation. Nat. Methods 9, 1171 (2012).
285. Rickgauer, J. P., Deisseroth, K. & Tank, D. W. Simultaneous cellular-resolution optical
perturbation and imaging of place cell firing fields. Nat. Neurosci. 17, 1816 (2014).
286. Becker, C. O., Bassett, D. & Preciado, V. M. Large-scale dynamic modeling of task-fMRI
signals via subspace system identification. J Neural Eng Epub Ahead of print (2018).
287. Coron, J.-M. Control and nonlinearity. 136 (American Mathematical Soc., 2007).
288. Klickstein, I., Shirin, A. & Sorrentino, F. Locally optimal control of complex networks. Phys.
Rev. Let. 119, 268301 (2017).
289. Haynes, G. & Hermes, H. Nonlinear controllability via lie theory. SIAM J. Control 8, 450 --
460 (1970).
290. Sussmann, H. J. & Jurdjevic, V. Controllability of nonlinear systems. Differ. Equ. 12, 95 -- 116
(1972).
291. Hermann, R. & Krener, A. Nonlinear controllability and observability. IEEE Trans. Automat.
Contr. 22, 728 -- 740 (1977).
292. Cornelius, S. P., Kath, W. L. & Motter, A. E. Realistic control of network dynamics. Nat.
Commun. 4, 1942 (2013).
61
293. Whalen, A. J., Brennan, S. N., Sauer, T. D. & Schiff, S. J. Observability and controllability
of nonlinear networks: The role of symmetry. Phys. Rev. X 5, 011005 (2015).
294. Isidori, A. Nonlinear control systems (Springer Science & Business Media, 2013).
295. Chopra, N. & Spong, M. W. On exponential synchronization of kuramoto oscillators. IEEE
Trans. Automat. Contr. 54, 353 -- 357 (2009).
296. Lynn, C. W. & Lee, D. D. Statistical mechanics of influence maximization with thermal
noise. EPL 117, 66001 (2017).
297. Lynn, C. W. & Lee, D. D. Maximizing activity in ising networks via the tap approximation.
In Association for the Advancement of Artificial Intelligence, 679 -- 686 (2018).
298. Amunts, K. & Zilles, K. Architectonic mapping of the human brain beyond Brodmann.
Neuron 88, 1086 -- 1107 (2015).
299. Cohen, M. R. & Kohn, A. Measuring and interpreting neuronal correlations. Nat Neurosci
14, 811 -- 819 (2011).
300. van den Heuvel, M. P. & Sporns, O. Rich-club organization of the human connectome. J
Neurosci 31, 15775 -- 15786 (2011).
301. Stiso, J. & Bassett, D. S. Spatial embedding imposes constraints on the network architectures
of neural systems. arXiv 1807, 04691 (2018).
302. van den Heuvel, M. P., Bullmore, E. T. & Sporns, O. Comparative connectomics. Trends
Cogn Sci 20, 345 -- 361 (2016).
303. Persichetti, A. S., Aguirre, G. K. & Thompson-Schill, S. L. Value is in the eye of the be-
holder: early visual cortex codes monetary value of objects during a diverted attention task.
J Cogn Neurosci 27, 893 -- 901 (2015).
304. Dore, B. P. et al. Brain activity tracks population information sharing by capturing consensus
judgments of value. Cereb Cortex Aug 28 (2018).
62
305. Constantinescu, A. O., O'Reilly, J. X. & Behrens, T. E. J. Organizing conceptual knowledge
in humans with a gridlike code. Science 352, 1464 -- 1468 (2016).
306. Papadopoulos, L., Porter, M. A., Daniels, K. E. & Bassett, D. S. Network analysis of particles
and grains. Journal of Complex Networks 6, 485 -- 565 (2018).
307. Bianconi, G., Rahmede, C. & Wu, Z. Complex quantum network geometries: Evolution and
phase transitions. Phys Rev E Stat Nonlin Soft Matter Phys 92, 022815 (2015).
308. Kivel, M. et al. Multilayer networks. J. Complex Netw. 2, 203 -- 271 (2014).
309. De Domenico, M., Granell, C., Porter, M. A. & Arenas, A. The physics of spreading pro-
cesses in multilayer networks 12, 901 -- 906 (2016).
310. Newman, M. E. J. & Clauset, A. Structure and inference in annotated networks. Nature
Communications 7, 11863 (2016).
311. Bassett, D. S., Wymbs, N. F., Porter, M. A., Mucha, P. J. & Grafton, S. T. Cross-linked
structure of network evolution. Chaos 24, 013112 (2014).
312. Porter, M. A., Onnela, J.-P. & Mucha, P. J. Communities in networks. Notices of the AMS
56, 1082 -- 1097 (2009).
313. Fortunato, S. Community detection in graphs. Physics reports 486, 75 -- 174 (2010).
314. Fortunato, S. & Hric, D. Community detection in networks: A user guide. Physics Reports
659, 1 -- 44 (2016).
315. Gosak, M. et al. Network science of biological systems at different scales: A review. Phys
Life Rev 24, 118 -- 135 (2018).
316. Richiardi, J. et al. Correlated gene expression supports synchronous activity in brain net-
works. Science 348, 1241 -- 1244 (2015).
317. Romero-Garcia, R. et al. Structural covariance networks are coupled to expression of genes
enriched in supragranular layers of the human cortex. Neuroimage 171, 256 -- 267 (2018).
63
318. Whitaker, K. J. et al. Adolescence is associated with genomically patterned consolidation of
the hubs of the human brain connectome. Proc Natl Acad Sci U S A 113, 9105 -- 9110 (2016).
319. Hardingham, G. E., Pruunsild, P., Greenberg, M. E. & Bading, H. Lineage divergence of
activity-driven transcription and evolution of cognitive ability. Nat Rev Neurosci 19, 9 -- 15
(2018).
320. Luke, D. A. & Harris, J. K. Network analysis in public health: history, methods, and appli-
cations. Annu Rev Public Health 28, 69 -- 93 (2007).
321. Braun, U. et al. From maps to multi-dimensional network mechanisms of mental disorders.
Neuron 97, 14 -- 31 (2018).
322. Schmalzle, R. et al. Brain connectivity dynamics during social interaction reflect social
network structure. Proc Natl Acad Sci U S A 114, 5153 -- 5158 (2017).
323. Parkinson, C., Kleinbaum, A. M. & Wheatley, T. Similar neural responses predict friendship.
Nat Commun 9, 332 (2018).
324. Parkinson, C., Liu, S. & Wheatley, T. A common cortical metric for spatial, temporal, and
social distance. J Neurosci 34, 1979 -- 1987 (2014).
325. Falk, E. B. & Bassett, D. S. Brain and social networks: Fundamental building blocks of
human experience. Trends Cogn Sci 21, 674 -- 690 (2017).
326. Rieke, F., Warland, D., de Ruyter van Steveninck, R. & Bialek, W. Spikes: exploring the
neural code (MIT Press, 1997).
327. Cover, T. M. & Thomas, J. A. Elements of information theory (John Wiley & Sons, 2012).
328. Shannon, C. E. A mathematical theory of communication. Bell Syst. Tech. J. 27 (1948).
329. MacKay, D. M. & McCulloch, W. S. The limiting information capacity of a neuronal link.
Bull. Math. Biophys. 14, 127 -- 135 (1952).
64
330. Attneave, F. Some informational aspects of visual perception. Psychol. Rev. 61, 183 (1954).
331. Barlow, H. B. Possible principles underlying the transformations of sensory messages (1961).
332. van Steveninck, R. d. R. & Bialek, W. Real-time performance of a movement-sensitive neu-
ron in the blowfly visual system: coding and information transfer in short spike sequences.
Proc. R. Soc. Lond. B 234, 379 -- 414 (1988).
333. Strong, S. P., Koberle, R., van Steveninck, R. R. d. R. & Bialek, W. Entropy and information
in neural spike trains. Phys. Rev. Lett. 80, 197 (1998).
334. Paninski, L. Estimation of entropy and mutual information. Neural Comput. 15, 1191 -- 1253
(2003).
335. Nemenman, I., Bialek, W. & van Steveninck, R. d. R. Entropy and information in neural
spike trains: Progress on the sampling problem. Phys. Rev. E 69, 056111 (2004).
336. Schreiber, T. Measuring information transfer. Phys. Rev. Lett. 85, 461 (2000).
337. Vicente, R., Wibral, M., Lindner, M. & Pipa, G. Transfer entropy -- a model-free measure of
effective connectivity for the neurosciences. J. Comput. Neurosci. 30, 45 -- 67 (2011).
338. Jaynes, E. T. Information theory and statistical mechanics. Phys. Rev. 106, 620 (1957).
339. Kailath, T. Linear Systems (Prentice-Hall, Inc., 1980).
340. Liu, Y.-Y., Slotine, J.-J. & Barab´asi, A.-L. Controllability of complex networks. Nature 473,
167 -- 173 (2011).
341. Klickstein, I., Shirin, A. & Sorrentino, F. Energy scaling of targeted optimal control of
complex networks. Nat. Commun. 8, 15145 (2017).
342. Yan, G. et al. Network control principles predict neuron function in the Caenorhabditis
elegans connectome. Nature 550, 519 -- 523 (2017).
65
Acknowledgements We are grateful to Lia Papadopoulos, Jason Z. Kim, and Vivek Buch for helpful
comments on an earlier version of this manuscript. We also thank Ann E. Sizemore for artistic inspiration.
D.S.B. and C.W.L. acknowledge support from the John D. and Catherine T. MacArthur Foundation, the
Alfred P. Sloan Foundation, the ISI Foundation, the Paul Allen Foundation, the Army Research Laboratory
(W911NF-10-2-0022), the Army Research Office (Bassett-W911NF-14-1-0679, Grafton-W911NF-16-1-
0474, DCIST- W911NF-17-2-0181), the Office of Naval Research, the National Institute of Mental Health
(2-R01-DC-009209-11, R01-MH112847, R01-MH107235, R21-M MH-106799), the National Institute of
Child Health and Human Development (1R01HD086888-01), National Institute of Neurological Disorders
and Stroke (R01 NS099348), and the National Science Foundation (BCS-1441502, BCS-1430087, NSF
PHY-1554488 and BCS-1631550). We also thank Vecteezy for supplying vector art.
Competing interests The authors declare no competing interests.
Corresponding author Correspondence and requests for materials should be addressed to D.S.B.
([email protected]).
66
|
1309.2609 | 4 | 1309 | 2019-07-09T19:49:02 | Contextual Focus: A Cognitive Explanation for the Cultural Revolution of the Middle/Upper Paleolithic | [
"q-bio.NC"
] | Many elements of culture made their first appearance in the Upper Paleolithic. Previous hypotheses put forth to explain this unprecedented burst of creativity are found wanting. Examination of the psychological basis of creativity leads to the suggestion that it resulted from the onset of contextual focus: the capacity to focus or defocus attention in response to the situation, thereby shifting between analytic and associative modes of thought. New ideas germinate in a defocused state in which one is receptive to the possible relevance of many dimensions of a situation. They are refined in a focused state, conducive to filtering out irrelevant dimensions and condensing relevant ones. | q-bio.NC | q-bio | Abstract
Many elements of culture made their first appearance in
the Upper Paleolithic. Previous hypotheses put forth to
explain this unprecedented burst of creativity are found
wanting. Examination of the psychological basis of
creativity leads to the suggestion that it resulted from the
onset of contextual focus: the capacity to focus or
defocus attention in response to the situation, thereby
shifting between analytic and associative modes of
thought. New ideas germinate in a defocused state in
which one is receptive to the possible relevance of many
dimensions of a situation. They are refined in a focused
state, conducive to filtering out irrelevant dimensions
and condensing relevant ones.
Introduction: A Cultural Revolution
Human culture is widely believed to have begun
between 2 and 1.5 mya, at which time a rapid increase
in brain size coincides with onset of the use of fire and
sophisticated stone tools. The archaeological record
suggests that a perhaps even more profound cultural
transition occurred between 60,000 and 30,000 ka
during the Middle / Upper Paleolithic (Bar-Yosef,
1994; Klein, 1989; Mellars, 1973, 1989a, b; Mithen,
1996, 1998; Soffer, 1994; Stringer & Gamble, 1993;
White, 1993). Leakey (1984) writes:
Unlike previous eras, when stasis dominated, innovation is
now the essence of culture, with change being measured in
millennia rather than hundreds of millennia. Known as the
Upper Paleolithic Revolution, this collective archaeological
signal is unmistakable evidence of the modern human mind at
work (p. 93-94).
Mithen (1996) refers to this period as the 'big bang' of
human culture, claiming that it shows more innovation
than the previous six million years of human evolution.
It marks the beginning of a more strategic style of
hunting involving specific animals at specific sites. We
also see the colonization of Australia, the replacement
of Levallois tool technology by blade cores in the Near
East, and the first appearance of many forms of art in
Europe, including naturalistic cave paintings of
animals, bone and antler tools with engraved designs,
ivory statues of animals and sea shells, personal
decoration such as beads, pendants, and perforated
animal teeth, and elaborate burial sites. Some of these
items are associated with social change and the
beginnings of ritualized religion; White (1982) writes of
a "total restructuring of social relations" (p. 176).
Moreover, we see the kind of cumulative change that
Tomasello (1999) refers to as a Ratchet Effect.
What could have caused this unprecedented explosion
of creativity? Some have noted that it would make
things easier if this second cultural transition also
coincided with an increase in brain size (Mithen, 1998;
Richerson & Boyd, 2000). And in fact, human brain
enlargement does seem to have occurred in two spurts.
However, the second takes place between 500,000 and
200,000 (Aiello, 1996) or 600,000 and 150,000 ka
(Ruff, Trinkaus, & Holliday, 1997); at any rate, well
before the Upper Paleolithic. Thus the cultural
revolution cannot be directly attributed to a change in
the size or shape of the cranium. Leakey (1984) writes
of anatomically modern human populations in the
Middle East with little in the way of culture, and
concludes "The link between anatomy and behavior
therefore seems to break" (p. 95).
Existing Hypotheses
Let us review some of the explanations for the Upper
Paleolithic Revolution that have been put forth.
Advent of Syntactic Language
It has been argued that while a primitive form of
language, or proto-language may have existed earlier,
the symbolic and syntactic aspects emerged at this time
(Aiello & Dunbar, 1993; Bickerton, 1990, 1996;
Dunbar, 1993, 1996). Another argument put forth is that
prior to the Upper Paleolithic, language was used
merely in social situations, and thereafter it became
general-purpose, put to use in all kinds of situations
(Aiello & Dunbar, 1993; Dunbar, 1993, 1996).
These arguments lead to the suggestion that the
Upper Paleolithic Revolution was due to onset of more
complex language. As noted by Tomasello (1999),
language has a transformative effect on cognition.
However, to posit that the cultural revolution is due to
the attainment of sophisticated language begs the
question: what cognitive change made possible the kind
432
Gabora, L. (2003). Contextual focus: A cognitive explanation for the cultural revolution of the Middle/Upper Paleolithic. In R. Alterman & D. Hirsch (Eds.), Proceedings of the 25th Annual Meeting of the Cognitive Science Society (pp. 432-437). Austin, TX: Cognitive Science Society.Contextual Focus: A Cognitive Explanation for the Cultural Revolution of theMiddle/Upper PaleolithicLiane Gabora ([email protected])Department of Psychology, University of California, Berkeley3210 Tolman Hall, Berkeley CA, 94720-1650 USAandCenter Leo Apostel for Interdisciplinary Studies (CLEA), Free University of BrusselsKrijgskundestraat 33, B1160 Brussels, Belgium, EUROPE of thought that sophisticated language requires? And
unless there was a underlying cognitive change
involved, why did the cultural situation change so
rapidly? Thus the riddle of the Upper Paleolithic is not
resolved through the attainment of complex language.
Material Culture as Externalized Memory
Donald's (1991, 2001) explanation is that material
culture began to function as an externalized symbol
storage. Certainly artifacts serve the purpose of
anchoring knowledge or desirable states of mind so that
they can be referred back to later. One sees a depiction
of bison on a cave, and there is a re-living of the
experience of watching a stampede of bison. One looks
at notches in a log and knows how many days have
gone by. The result is not so different from retrieving
knowledge from memory, though the source of the
knowledge is external rather than internal.
However, the external world functioned as a form of
memory long before there were symbolic artifacts. A
look of disapproval on a mother's face could remind a
child not to eat a poisonous mushroom as readily as
retrieval of a memory of doing this and getting sick.
The look on the mother's face is not a material artifact,
yet it functions for the child as an external memory
source, in much the same way as a bison painting or
notches in a log. Moreover, since material objects are
manifestations of ideas, which begin in minds, it seems
reasonable that we look to the mind, not the outside
world, for the root cause of the cultural revolution.
Exploration of Conceptual Spaces
Another possibility is that it reflected an enhanced
ability to blend concepts (Fauconnier & Turner, 2002)
or to map, explore, and transform conceptual spaces
(Mithen, 1998). Mithen refers to Boden's (1990)
definition of a conceptual space as a 'style of
thinking -- in music, sculpture, choreography,
chemistry, etc.' As for why hominids suddenly became
good at exploring and transforming conceptual spaces,
he is somewhat vague:
There is unlikely to be one single change in the human mind
that enabled conceptual spaces to become explored and
transformed. Although creative thinking seems to appear
suddenly in human evolution, its cognitive basis had a long
evolutionary history during which the three foundations
evolved on largely an independent basis: a theory of mind, a
capacity for language, and a complex material culture. After
50,000 years ago, these came to form the potent ingredients of
a cognitive/social/material mix that did indeed lead to a
creative explosion (p. 186).
Mithen may be on to something with the notion that the
cultural revolution is related to the capacity to explore
and transform conceptual spaces. However, although
the capacity for a theory of mind, language, and
complex artifacts may have, in their most primitive
forms, arisen at different times, it is hard to imagine
how they could have evolved independent of one
another. Furthermore, if there is anything that science
has established in the last decade or two it is that a
single, small change in initial conditions can have
enormous consequences (e.g. Bak, Tang, & Weisenfeld,
1988; Kauffman, 1993). Thus the possibility that the
Upper Paleolithic revolution can be explained by a
single change in cognitive functioning is not only the
simplest explanation, it is also consistent with the
sudden transition in the archeological record, and with
our understanding of phase transitions across the
scientific disciplines.
Connection of Domain-Specific Modules
Many suggest that modern cognition arose through the
connecting of domain-specific brain modules, stressing
in particular that this would allow for the production of
analogies and metaphors (Fodor, 1983; Gardner, 1983,
1993; Karmiloff-Smith, 1992; Rozin, 1976; Sperber,
1994). Mithen (1996) suggests that in the Upper
Paleolithic, modules specialized to cope with domains
such as natural history, technology, and social processes
became connected. However, this would require that
there be space enough for not just the modules but the
new connections amongst them, and as we have seen,
this cultural transition does not coincide with an
increase in size or change of shape of the cranium.
Moreover, the logistics of physically connecting these
modules (whose positions in the brain evolved without
foreknowledge that they should one day become
connected) would be formidable. Sperber's (1994)
solution is that the modules got connected not directly,
but indirectly, by way of a special module, the 'module
of metarepresentation' or MMR, which contains
'concepts of concepts'. However, to invent an artifact
that combines information from different domains, such
as an axe with a bison engraved on it, it is not necessary
that the module that deals with tools be connected to the
one that deals with animals, nor even that the concepts
'axe' and 'buffalo' spend time together in a meta-
module. All that is necessary, as explained shortly, is
that these concepts be simultaneously accessible.
Psychological Basis of Creativity
We have looked at several hypotheses to account for
the cultural revolution of the Upper Paleolithic. Each
has merit, but none provides a satisfying explanation for
this burst of creativity. To determine more precisely
what gave rise to the modern human mind, it is useful
to examine the psychological basis of creativity.
433
Gabora Attributes of Creative Individuals
Martindale (1999) identified a cluster of attributes
associated with high creativity. A first one is defocused
attention: the tendency not to focus exclusively on the
relevant aspects of a situation, but notice also seemingly
irrelevant aspects (Dewing & Battye, 1971; Dykes &
McGhie, 1976; Mendelsohn, 1976). A related attribute
is high sensitivity (Martindale, 1977, 1999; Martindale
& Armstrong, 1974), including sensitivity to subliminal
impressions; stimuli that are perceived but of which one
is not conscious of having perceived (Smith & Van de
Meer, 1994).
Creative individuals also tend to have flat associative
hierarchies (Mednick, 1962). The steepness of one's
associative hierarchy is measured by comparing the
number of words generated in response to stimulus
words on a word association test. Those who generate
few words for each stimulus have a steep associative
hierarchy, whereas those who generate many have a flat
associative hierarchy. Thus, once such an individual has
run out of the more usual associations (e.g. 'chair' in
response to 'table'), unusual ones (e.g. 'elbow' in
response to 'table') come to mind. The evidence that
creativity is associated with both defocused attention
and flat associative hierarchies suggests that creative
individuals not only notice details others miss, but these
details get stored in memory and are available later on.
However, a considerable body of research suggests
that creativity involves not just the ability to defocus
and free-associate, but also the ability to focus and
concentrate (Barron, 1963; Eysenck, 1995; Feist, 1999;
Fodor, 1995; Richards et al. 1988; Russ, 1993)1. As
Feist (1999) puts it: "It is not unbridled psychoticism
that is most strongly associated with creativity, but
psychoticism tempered by high ego strength or ego
control. Paradoxically, creative people appear to be
simultaneously very labile and mutable and yet can be
rather controlled and stable" (p. 288). He notes that, as
Barron (1963) put it: "The creative genius may be at
once naïve and knowledgeable, being at home equally
to primitive symbolism and rigorous logic. He is both
more primitive and more cultured, more destructive and
more constructive, occasionally crazier yet adamantly
saner than the average person" (p. 224).
Phases of the Creative Process
How do we make sense of this seemingly paradoxical
description of the creative individual? The evidence that
creativity is associated with both defocused free-
association and focused concentration is in fact
1There is also evidence of an association between creativity
and high variability in physiological measures of arousal such
as heart rate (Bowers & Keeling, 1971), spontaneous galvanic
skin response (Martindale, 1977), and EEG alpha amplitude
(Martindale & Hasenfus, 1978; Martindale, 1999).
consistent with the idea that the creative process
consists of a generative phase followed by an
evaluative phase (Boden, 1991; Dennett, 1978). Indeed
there is an enduring notion that there are two kinds of
thought, or that thought varies along a continuum
between two extremes (Ashby & Ell, 2002; James,
1890/1950, Johnson-Laird, 1983; Neisser, 1963; Piaget,
1926; Rips, 2001; Sloman, 1996). Although the issue is
still a subject of hot debate, the general picture
emerging is as follows. At one end of the continuum is
an intuitive, associative mode conducive to finding
remote or subtle connections between items that are
correlated but not necessarily causally related. This
mode may yield an idea or problem solution, though
perhaps in a vague, unpolished form. At the other end
of the continuum is a rule-based, analytic mode of
thought, conducive to analyzing relationships of cause
and effect. This mode facilitates fine-tuning and
manifestation of the creative work.
Contextual Focus Hypothesis
Let us now now look at a tentative explanation of the
cognitive mechanisms underlying the creative process
(Gabora, 2000, 2002a, 2002b; Gabora & Aerts, 2002).
We take as a starting point some fairly well-established
features of memory. According to the doctrine of neural
re-entrance, the same memory locations get used again
and again (Edelman, 1987). Each memory location is
sensitive to a range of subsymbolic microfeatures
(Smolensky, 1988), or values of them (Churchland &
Sejnowski, 1992). Location A may respond
preferentially to lines of a certain angle (say 90
degrees), neighboring location B respond preferentially
to lines of a slightly different angle (say 91 degrees),
and so forth. However, although location A responds
maximally to lines of 90 degrees, it responds to a lesser
degree to lines of 91 degrees. This kind of organization
is referred to as coarse coding. The upshot is that
storage of an item is distributed across a cell assembly
that contains many locations, and likewise, each
location participates in the storage of many items
(Hinton, McClelland, & Rummelhart, 1986). Items
stored in overlapping regions are correlated, or share
features. Therefore memory is content addressable;
there is a systematic relationship between the state of an
input and the place it gets stored. Thus episodes stored
in memory can thereafter be evoked by stimuli that are
similar or 'resonant' (Hebb, 1949; Marr, 1969).
Let us consider the significance of this memory
architecture for creativity. To be constantly in a state of
defocused attention, in which relevant dimensions of a
situation do not stand out strongly from irrelevant ones,
would be clearly impractical. It is only when one does
not yet know what are the relevant dimensions -- or
when those assumed to be relevant turn out not to
be -- that defocused attention is of use. After the
434
Gabora relevant dimensions have been found it is most efficient
to focus on them exclusively. Indeed it has been shown
that in stimulus classification tasks, psychological space
is stretched along dimensions that are useful for
distinguishing members of different categories, and
shrunk along nonpredictive dimensions (Nosofsky,
1987; Kruschke, 1993). In ALCOVE, a computer
model of category learning, only when activation of
each input unit was multiplied by an attentional gain
factor did the output match the behavior of human
subjects (Kruschke, 1992; Nosofsky & Kruschke,
1992). Thus learning and problem solving involve both
(1) associating stimuli with outcomes, and (2) shifts in
attention that determine how one 'parses' the space. Let
us refer to the situation in which many stimulus
dimensions or aspects of a situation activate memory to
an almost equal degree as a flat activation function, and
the situation in one focuses exclusively on one stimulus
dimension, or aspect of a situation, as a spiky activation
function. We can refer to the ability to spontaneously
adjust the shape of the activation function in response to
the situation at hand as the capacity for contextual
focus.
Let us now explore the possibility that creative
individuals are not always in a state of defocused
attention, but that they can enter this state when useful
(such as in a word association test). Thus when one
encounters a problem or inconsistency, or seeks self-
expression, one enters a state of defocused attention
conducive to associative thought through a flattening of
the activation function. More diverse memory locations
get activated and provide ingredients for the next
thought. Because of the distributed, content-addressable
structure of memory, a seemingly irrelevant element of
the situation may evoke an episode from memory that
shares this element. The connection between them may
inspire a new idea.
The vague idea generated in a defocused state is
clarified by focusing attention on the new connection
and shifting to a more analytic mode. The activation
function becomes spikier, and the region searched and
retrieved from narrower. This continues until, to use
Posner's (1964) terms, one has filtered out the irrelevant
dimensions and condensed the relevant ones.
In sum, it is proposed that the capacity for contextual
focus is the distinguishing feature of the modern human
mind, and the reason for the cultural revolution of the
Upper Paleolithic.
Why Connected Modules is Not Necessary
Let us see why to blend items from different modules
together it is not necessary that they be connected. Once
modules start encoding more of the richness of their
respective domains, to the extent that these domains
have elements in common, an individual will start to
have experiences that activate multiple modules
simultaneously. Let us say, for instance, that Mithen is
right about natural history and social situations being
the domain of different modules. Both these modules
will come to contain memory locations that respond to
the chaotic, volatile elements of a situation. So, for
example, an individual in a volatile mood might remind
one of stormy weather. Thus the modules themselves
need not be connected for a blend to occur so long as a
situation can simultaneously activate them.
Possible Explanation for Lag Between Brain
Expansion and Cultural Revolution
An increase in brain size provides more storage space,
thus memories can be laid down in richer detail. But it
doesn't follow that this increased space could
immediately be navigated in the most efficient way.
This is particularly the case for situations involving
simultaneous activation of multiple modules; there is no
reason to think their contents would be compatible
enough to coexist in a stream of thought. The stormy
individual might prompt one to take shelter without
leading to the without realization that ones' own mood
could take on this stormy quality. A situation where
three apples reminds one of ones' three children could
prompt one to bring home the right number of apples,
without leading to the realization that 'three-ness' can be
independent of apples or children. It seems reasonable
that it took time to fine-tune the cognitive system such
that items from different domains could be blended
together and recursively redescribed in a coordinated
manner. Only then could the full potential of this large
brain be realized.
Summary and Discussion
The period of history that exhibits the most impressive
cultural transition is the Upper Paleolithic. To gain
insight into what caused this unprecedented explosion
of creativity, it is useful to examine the creative
process. Creative individuals are prone to states of
defocused attention, and tend toward flat associative
hierarchies, suggesting a proclivity for associative
thought. However, creativity involves not just an
intuitive, associative mode of thought, but also an
analytic, evaluative mode. This suggests that creativity
requires the ability to shift between these modes.
Thus it is tentatively proposed that the arrival of art,
science, religion, and likely also complex language and
a restructuring of social relationships in the Upper
Paleolithic was due to the onset of contextual focus: the
capacity to focus or defocus attention in response to the
situation, thereby shifting between analytic and
associative modes of thought. New ideas germinate in a
defocused state in which one is receptive to the possible
relevance of many dimensions of a situation. They are
435
Gabora refined in a focused state, conducive to filtering out
irrelevant dimensions and condensing relevant ones.
Contextual focus is not a matter of more memory, but
of a more sophisticated way of using memory; thus the
proposal is consistent with there being no increase in
brain size at this time. It is consistent with a point made
by Bickerton (1990) and Leakey (1984) that brain size
cannot be equated with intelligence. However it
deviates from Bickerton's perspective in that it is not
language per se that made the difference, but rather a
kind of cognitive functioning that made not only
language possible, but all aspects of group survival that
can benefit from being considered from different
perspectives and at different degrees of abstraction. It
does not require that modules be connected, but merely
simultaneously accessible.
Acknowledgments
I would like to thank Ellen Van Keer for comments.
This research was supported by Grant G.0339.02 of the
Flemish Fund for Scientific Research.
References
Aiello, L. & Dunbar, R. (1993). Neocortex size, group
size, and the evolution of language. C u r r e n t
Anthropology, 34, 184-193.
Ashby, F. G., & Ell, S. W. (2002). Single versus
multiple systems of learning and memory. In J.
Wixted & H. Pashler (Eds.) Stevens' handbook of
experimental psychology: Volume 4 Methodology in
experimental psychology. New York: Wiley.
Bak, P., Tang, C., & Weisenfeld, K. (1988). Self-
organized criticality. Physical Review, A 38, 364.
Barron, F. (1963). Creativity and psychological health.
Princeton, NJ: Van Nostrand.
Bar-Yosef, O. (1994). The contribution of southwest
Asia to the study of the origin of modern humans. In
M. Nitecki & D. Nitecki (Eds.) Origins of
anatomically modern humans, Plenum Press.
Bickerton, D. (1990). Language and species, Chicago:
Chicago University Press.
Boden, M. (1990). The creative mind: Myths and
mechanisms. London: Weidenfeld & Nicolson.
Bowers, K. S. & Keeling, K. R. (1971). Heart-rate
variability in creative functioning. Psychological
Reports, 29, 160-162.
Churchland, P. S. & Sejnowski, T. (1992). The
Computational Brain. Cambridge MA: MIT Press.
Davidson, I. (1991). The archeology of language
origins: A review. Antiquity, 65, 39-48.
Dennett, D. (1978). Brainstorms. Cambridge MA: MIT
Press.
Dewing, K. & Battye, G. (1971). Attentional
deployment and non-verbal fluency. Journal of
Personality and Social Psychology, 17, 214-218.
Donald, M. (1991). Origins of the modern mind,
Cambridge, MA: Harvard University Press.
Donald, M. (2001). A mind so rare, Norton Press.
Dunbar, R. (1993). Coevolution of neocortical size,
group size, and language in humans. Behavioral and
Brain Sciences, 16(4), 681-735.
Dunbar, R. (1996). Grooming, gossip, and the evolution
of language. London: Faber & Faber.
Dykes, M. & McGhie, A. (1976). A comparative study
of attentional strategies in schizophrenics and highly
creative normal subjects. British Journal of
Psychiatry, 128, 50-56.
Edelman, G. (1987). Neural darwinism: The theory of
neuronal group selection. New York: Basic Books.
Eysenck, H. (1995). Genius: The natural history of
creativity. Cambridge: Cambridge University Press.
Fauconnier, G. & Turner, M. (2002).The way we think:
the mind's hidden
Conceptual blending and
complexities. New York: Basic Books.
Feist, G. J. (1999). The influence of personality on
artistic and scientific creativity. In R. J. Sternberg
(Ed.) Handbook of creativity. Cambridge UK:
Cambridge University Press.
Fodor, E. M. (1995). Subclinical manifestations of
psychosis-proneness, ego-strength, and creativity.
Personality and Individual Differences, 18, 635-642.
Fodor, J. (1983). The modularity of mind. Cambridge
MA: MIT Press.
Gabora, L. (2000). Toward a theory of creative
inklings. In R. Ascott (Ed.) Art, Technology, and
Consciousness, Bristol, UK: Intellect Press.
Gabora, L. (2002a). Cognitive mechanisms underlying
the creative process. Proceedings of Fourth
International Conference on Creativity and
Cognition (pp. 126-133), Loughborough Univ., UK.
Gabora, L. (2002b). The beer can theory of creativity.
In P. Bentley & D. Corne (Eds.) C r e a t i v e
evolutionary systems, San Francisco: Morgan
Kaufmann.
Gabora, L. & Aerts, D. (2002). Contextualizing
concepts using a mathematical generalization of the
quantum formalism. Journal of Experimental and
Theoretical Artificial Intelligence, 14(4), 327-358.
Gardner, M. (1983). Frames of mind: The theory of
multiple intelligences. New York: Basic Books.
Hebb, D. (1949). The Organization of behavior. New
York: Wiley.
Hinton, G., McClelland, J. & Rummelhart, D. (1986).
Distributed representations. In D. Rummelhart, &
J.McClelland (Eds.) Parallel distributed processing.
Cambridge MA: MIT Press.
James, W. (1890/1950). The principles of psychology.
New York: Dover.
Johnson-Laird, P.N. (1983). Mental models. Cambridge
MA: Harvard University Press.
Karmiloff-Smith, A. (1992). Beyond modularity: A
developmental perspective on cognitive science.
Cambridge, MA: MIT Press.
Kauffman, S. (1993). Origins of order. Oxford: Oxford
University Press.
436
Gabora Posner, M. (1964). Information reduction in the
analysis of sequential tasks. Psychological Review,
71, 491-504.
Richards, R. Kinney, D., Lunde, I., Benet, M., &
Merzel, A. (1988). Creativity in manic depressives,
cyclothymes, their normal relatives, and control
subjects. Journal of Abnormal Psychology, 97, 281-
289.
Richerson, P. & Boyd, R. (2000). Climate, culture, and
the evolution of cognition. In C. Heyes & L. Huber
(Eds.) The evolution of cognition. Cambridge, MA:
MIT Press.
Rips, L. (2001). Necessity and natural categories.
Psychological Bulletin, 127(6), 827-852.
Rozin, P. (1976). The evolution of intelligence and
access to the cognitive unconscious. In J. M. Sprague
& A. N. Epstein (Eds.) Progress in psychobiology
and physiological psychology. New York: Academic
Press.
Ruff, C., Trinkaus, E. & Holliday, T. (1997). Body
mass and encephalization in Pleistocene Homo.
Nature, 387, 173-176.
Russ, S. W. (1993). Affect and creativity. Hillsdale NJ:
Erlbaum.
Sloman, S. (1996). The empirical case for two systems
of Reasoning. Psychological Bulletin, 9(1), 3-22.
Smith, G. J. W. & Van de Meer, G. (1994). Creativity
through psychosomatics. Creativity Research
Journal, 7, 159-170.
Smolensky, P. (1988). On the proper treatment of
connectionism. Behavioral and Brain Sciences, 11, 1-
43.
Soffer, O. (1994). Ancestral lifeways in Eurasia -- The
Middle and Upper Paleolithic records. In M. Nitecki
& D. Nitecki, (Eds.) Origins of anatomically modern
humans. New York: Plenum Press.
Sperber, D. (1994). The modularity of thought and the
epidemiology of representations. In L. A. Hirshfield
& S. A. Gelman (Eds.) Mapping the mind: Domain
specificity in cognition and culture, Cambridge UK:
Cambridge University Press.
Stringer, C. & Gamble, C. (1993). In search of the
Neanderthals. London: Thames & Hudson.
Tomasello, M. (1999). The Cultural Origins of Human
Cognition. Cambridge: Harvard University Press.
White, R. (1993) Technological and social dimensions
of 'Aurignacian-age' body ornaments across Europe.
In H. Knecht, A. Pike-Tay, & R. White (Eds.) Before
Lascaux: The complex record of the Early Upper
Paleolithc. New York: CRC Press.
White, R. (1982) Rethinking the Middle/Upper
Paleolithic Transition. Current Anthropology, 23,
169-189.
Klein, R. (1989). Biological and behavioral
perspectives on modern human origins in South
Africa. In P. Mellars & C. Stringe (Eds.) The human
revolution. Edinburgh: Edinburgh University Press.
Kruschke, J. (1992). ALCOVE: an exemplar-based
learning.
connectionist model of category
Psychological Review, 99, 22-44.
Leakey, R. (1984). The origins of humankind. New
York: Science Masters Basic Books.
Marr, D. A. (1969). theory of the cerebellar cortex.
Journal of Physiology, 202, 437-470.
Martindale, C. (1999). Biological bases of creativity. In
R. J. Sternberg (Ed.) Handbook of creativity.
Cambridge UK: Cambridge University Press.
Martindale, C. & Armstrong, J. (1974). The relationship
of creativity to cortical activation and its operant
control. Journal of Genetic Psychology, 124, 311-
320.
Martindale, C. & Hasenfus, N. (1978). EEG differences
as a function of creativity, stage of the creative
process, and effort to be original. B i o l o g i c a l
Psychology, 6, 157-167.
Mednick, S. (1962). The associative basis of the
creative process. Psychological Review, 69, 220-232.
Mendelsohn, G. A. (1976). Associative and attentional
processes in creative performance. Journal of
Personality, 44, 341-369.
Mellars, P. (1973). The character of the middle-upper
transition in south-west France. In C. Renfrew (Ed.)
The explanation of culture change. London:
Duckworth.
Mellars, P. (1989a). Technological changes in the
Middle-Upper Paleolithic transition: Economic,
social, and cognitive perspectives. In P. Mellars & C.
Stringer (Eds.) The human revolution. Edinburgh:
Edinburgh University Press.
Mellars, P. (1989b). Major issues in the emergence of
modern humans. Current Anthropology, 30, 349-385.
Mithen, S. (1996). The prehistory of the mind: A search
for the origins of art, science, and religion. London:
Thames & Hudson.
Mithen, S. (1998). A creative explosion? Theory of
mind, language, and the disembodied mind of the
Upper Paleolithic. In S. Mithen (Ed.) Creativity in
human evolution and prehistory. New York:
Routledge.
Neisser, U. (1963). The multiplicity of thought. British
Journal of Psychology, 54, 1-14.
Nosofsky, R. (1987). Attention, similarity, and the
identification-categorization relationship. Journal of
Experimental Psychlogy: Learning, Memory and
Cognition, 13, 87-108.
Nosofsky, R. & Kruschke, J. (1992). Investigations of
an exemplar-based connectionist model of category
learning. In D. L. Medin (ed.) The psychology of
learning and motivation, 28, 207-250.
437
Gabora |
1903.12155 | 1 | 1903 | 2019-03-28T17:42:05 | Theta-gamma cross-frequency coupling enables covariance between distant brain regions | [
"q-bio.NC",
"nlin.PS"
] | Cross-frequency coupling (CFC) is thought to play an important role in communication across distant brain regions. However, neither the mechanism of its generation nor the influence on the underlying spiking dynamics is well understood. Here, we investigate the dynamics of two interacting distant neuronal modules coupled by inter-regional long-range connections. Each neuronal module comprises an excitatory and inhibitory population of quadratic integrate-and-fire neurons connected locally with conductance-based synapses. The two modules are coupled reciprocally with delays that represent the long-range conduction time. We applied the Ott-Antonsen ansatz to reduce the spiking dynamics to the corresponding mean field equations as a small set of delay differential equations. Bifurcation analysis on these mean field equations shows inter-regional conduction delay is sufficient to produce CFC via a torus bifurcation. Spike correlation analysis during the CFC revealed that several local clusters exhibit synchronized firing in gamma-band frequencies. These clusters exhibit locally decorrelated firings between the cluster pairs within the same population. In contrast, the clusters exhibit long-range gamma-band cross-covariance between the distant clusters that have similar firing frequency. The interactions of the different gamma frequencies produce a beat leading to population-level CFC. We analyzed spike counts in relation to the phases of the macroscopic fast and slow oscillations and found population spike counts vary with respect to macroscopic phases. Such firing phase preference accompanies a phase window with high spike count and low Fano factor, which is suitable for a population rate code. Our work suggests the inter-regional conduction delay plays a significant role in the emergence of CFC and the underlying spiking dynamics may support long-range communication and neural coding. | q-bio.NC | q-bio |
Theta-gamma cross-frequency coupling enables covariance between distant brain
regions
Akihiko Akao
Graduate School of Engineering, The University of Tokyo,
7-3-1 Hongo, Bunkyo-Ku, Tokyo 113-0033, Japan
Graduate School of Information Science and Technology, 1-5 Yamadaoka, Suita, Osaka 565-0871, Japan
Sho Shirasaka
Yasuhiko Jimbo
Graduate School of Engineering, The University of Tokyo,
7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan
Department of Mathematics, University of Pittsburgh, Pittsburgh, Pennsylvania 15260, USA
Bard Ermentrout
Research Center for Advanced Science and Technology, The University of Tokyo,
Kiyoshi Kotani
4-6-1 Komaba, Meguro-ku, Tokyo 153-8904, Japan and
JST, PRESTO, 4-1-8 Honcho, Kawaguchi-shi, Saitama 332-0012, Japan
Cross-frequency coupling (CFC), where the amplitude of a fast neuronal oscillation is modulated
by a second slower frequency, is thought to play an important role in long-range communication
across distant brain regions. However, neither the mechanism of its generation nor the influence on
spiking dynamics is well understood. Here, we investigate the multiscale dynamics of two interacting
distant neuronal modules coupled by inter-regional long-range connections. Each neuronal module
comprises an excitatory and inhibitory population of quadratic integrate-and-fire neurons connected
locally with conductance-based synapses. The two modules are then coupled reciprocally with delays
that represent the conduction times over the long distance between them. By assuming a Lorentzian
distribution to the probability density function of the membrane potential, we are able to apply the
Ott-Antonsen ansatz to reduce the corresponding mean field equations of the spiking dynamics to
a small set of delay differential equations. Bifurcation analysis on these mean field equations shows
that inter-regional conduction delay is sufficient to produce CFC via a torus bifurcation, as well
as a gamma oscillation via a Hopf bifurcation. Spike correlation and covariance analysis during
the CFC revealed that several local clusters in excitatory population exhibit synchronized firing
in gamma-band frequencies. These clusters exhibit locally decorrelated firings between the cluster
pairs within the same population because of their different firing frequencies. In contrast, the same
clusters exhibit long-range gamma-band cross-covariance between the corresponding cluster in the
distant populations that have similar firing frequency. The interactions of the different gamma
frequencies in each module produce a beat leading to population-level CFC. In order to investigate
the impact of CFC on neuronal spike timings, we analyzed spike counts in relation to the phases of
the macroscopic fast and slow oscillations of the mean membrane potential. We found population
spike counts vary with respect to macroscopic phases. Such firing phase preference accompanies a
phase window with high spike count and low Fano factor, which is suitable for a population rate
code. In addition, we analyzed the firing phase preference of the local clusters. We found these
clusters also exhibit firing phase preference that differs between the clusters, similar to experimental
findings. Our work suggests that the inter-regional conduction delay plays a significant role in the
emergence of CFC and the underlying spiking dynamics may support long-range communication
and neural coding.
I.
INTRODUCTION
Rhythms and neuronal oscillations are ubiquitous in
the central nervous system (CNS) where they are be-
lieved to play a role in many functions ranging from
cognition to motor function [1, 2]. These macroscopic
rhythms (seen,
in electroencephalogram
recordings or in extracellular recordings of the local field
potential) emerge from the interactions of populations of
for example,
excitatory and inhibitory neurons that comprise the cor-
tex and other CNS regions. Thus, there have been many
papers written on the biophysical mechanisms that un-
derlie these ubiquitous rhythms [3].
Cross-frequency coupling(CFC), which is a phenom-
ena where a low frequency oscillation modulates the am-
plitude of a high frequency oscillation, has received a
good deal of recent attention experimentally [4 -- 7] and
theoretically[8 -- 12].
It has been suggested that high
frequency neuronal oscillations may contribute to lo-
cal computations, while low frequency oscillations are
employed in long-range communication across different
brain regions[8, 13]. Therefore, CFC is assumed to play
some role in integrating local computations distributed
across the distant brain regions and thus facilitating
higher cognitive function. This assumption,
in which
CFC contributes to long-range integration, is supported
by some recent experimental studies[14 -- 17]. However,
the dynamical mechanism for the emergence of CFC dur-
ing long-range neuronal interactions and the dynamical
characteristics of such CFC are still unclear.
To elucidate the mechanisms for macroscopic neuronal
dynamics, large scale mathematical models of neuronal
populations have been frequently utilized [18]. Conven-
tional firing rate models are popular to describe macro-
scopic rhythms[19, 20]. However, they assume that un-
derlying spiking dynamics is asynchronous, which is not
suitable for high frequency oscillations [20 -- 22]. Recently,
the derivation of mean field equations for a population
of theta neurons (or the equivalent quadratic integrate-
and-fire neurons) via the Ott-Antonsen ansatz[23] has
been applied [24]. This approach allows one to cap-
ture the synchronized spiking dynamics as well as high
frequency oscillations[21].
In previous work, pulse-
coupling[25], gap junctions[26],
locally connected in-
hibitory population[27], a large population [28], exci-
tatory and inhibitory (E-I) populations[29] and time-
varying modulation[30] have all been considered using
this approach, and behavior including, in particular, the
emergence of macroscopic oscillations and chaos has been
investigated. However, this prior research focused on the
local dynamics without long-range interactions and CFC
was also not described in these studies. In addition, these
works were performed with relatively simple models,
where the biological factors which are known to affect the
fast macroscopic oscillation such as synaptic conductance
dynamics[31] and synaptic reversal potential[32, 33] have
not been fully incorporated.
In our previous work, we introduced a modified
quadratic integrate-and-fire neuron and analyzed the
emergence of (fast) gamma oscillations and analyzed
their macroscopic phase response curves[34, 35]. How-
ever, because our focus was on the local emergence of
high frequency oscillations, the dynamics during long-
range interactions between distant brain regions was not
analyzed.
Thus, here, we analyze the dynamics of two delay-
coupled E-I modules composed of modified quadratic
integrate-and-fire (QIF) neurons. We adopt the Ott-
Antonsen ansatz for each population and extract their
mean field dynamics. We introduce the long range in-
teraction through a synaptic conduction delay between
the two distant E-I modules and derive the mean field
dynamics of the whole system as a set of delay differen-
tial equations (DDEs). We perform bifurcation analysis
on the mean field DDEs and investigate how the macro-
scopic oscillations are facilitated from these long-range
2
interactions.
We first introduce the model and the resulting Ott-
Antonsen reduction. We show that a conduction delay
between the regions is sufficient to produce CFC and
then show that the CFC arises in the mean field model
as a torus bifurcation near the intersection of two dis-
tinct Hopf bifurcation branches. We analyze the spiking
model in the parameter regions where there is CFC and
show that it induces long-range multi-cluster gamma-
band cross-covariance.
In addition, we find spike tim-
ings of individual neurons have a preference for specific
phases of both the theta and the gamma oscillations, thus
generating a suitable temporal window for a population
rate code. Our results suggest that CFC is one of the
consequences of inter-regional coupling and that the un-
derlying spiking dynamics may support long-range infor-
mation integration and neural coding.
II. METHODS
A. Dynamics of the spiking neuronal population
VT − VR
gY
Y
We consider two pairs of E-I modules as shown in
Fig. 1(a), namely module 1 and 2.
In each module,
we consider E and I populations of the modified QIF
neurons[34, 35], namely E1, I1, E2 and I2.
i (t) , of
i .
(1)
dV X
i
dt
C
=gLX
− (cid:61)
Parameters are listed in Tab. I. V X
The dynamics of the membrane potential V X
the i-th neuron in population X is written as
i (t) − VT)
(V X
i (t) − VR)(V X
syn) + I X
i (t) − V Y
X(t)(V X
i (t) follows a re-
i (t) ≥ Vpeak, then V X
i (t) ← Vreset.
i (t) represents the action potential
X(t) , synaptic conductance from Y to X, is written
X) . (2)
We assume that the synaptic connections from popu-
X. The dynamics
lation Y to X occur with probability pY
of gY
as
setting rule as if V X
Such resetting of V X
or firing of neuron i.
δt −(tY(k)) + τ Y
X(t)+¯gY
gY
dgY
X
dt
= − 1
τ Y
d
X⋅pY
X⋅
NY(cid:61)
k=1
Parameters are listed in Tab. II and Tab. III. The con-
duction delay τ Y
X is τdelay for the cases where coupling is
inter-regional. τ Y
X is zero for the other cases where cou-
pling is local. It has been experimentally shown that the
conduction delay between distant brain regions can be up
to 40 (ms) [36]. Therefore, we set τdelay = 31 (ms) as the
nominal default value. For inter-module connections, the
synaptic strength pY
X is set as pdelay. For local connec-
X is set as plocal except pI2
tions , the synaptic strength pY
I2
TABLE I. Parameters used Eq.1.
3
Default value (Unit)
X ={I1, I2, E1, E2}
Y ={I1, I2, E1, E2}
C = 1(µF/cm2)
= NI2
= NE2
= 100
= 400
NI1
NE1
1 ≤ i ≤ NX
= 0.1(mS/cm2)
= 0.08(mS/cm2)
= gLI2
= gLE2
VR = −62(mV)
VT = −55(mV)
gLI1
gLE1
Distributed among population X
following the Lorentz distribution
fX(IX) = 1
π
∆X
I− ¯IX2+∆2
= 1(A/cm2)
= 0.1(A/cm2)
= ¯II2
= ∆I2
X
¯IE1
∆E1
= ¯IE2
= ¯II1
= ∆E2
= ∆I1
syn = V I2
V I1
syn = V E2
V E1
syn = −70 (mV)
syn = 0 (mV)
Symbol
Property
X
Y
NX
C
i
gLX
VR
VT
I X
i
¯IX
∆IX
V Y
syn
post-synaptic population
pre-synaptic population
number of neurons
membrane capacitance
index of neuron
leak conductance
resting potential
threshold potential
quenched current
median of I X
i
HWHM of I X
i
synaptic reversal potential
which is the recurrent inhibition in the I2 population. pI2
I2
is set as plocal×0.9 in order to introduce a slight difference
between module 1 and module 2 (to break any symme-
try). We set the area of a neuron as 2.9 × 10
to match the physiological plausible value[37, 38]. Peak
conductances ¯gX
Y are also set as to match physiological
plausible values [31, 39, 40].
−4(cm2)
The description of the whole system (spiking neuronal
population and synaptic dynamics) is obtained by a set
of 1000 of Eq. 1 (400 E and 100 I cells for each area) and
12 of Eq. 2.
For the numerical simulation of the spiking population,
we transformed Eq. 1 to the form of the theta neuron
to avoid numerical delicacies near the spiking threshold.
We introduce the following transformation:
θX
i
2
.
i = VR + VT
V X
+ VT − VR
2
2
tan
i (t) = −gLX cos θX
(3)
We then take the limit as Vpeak =−Vreset = +∞, which
naturally allows us to capture the neuronal spike as well
as the refractory period that is evoked by the spike.
Then, Eq.1 can be transformed into
i (t))I X
i (t)) − sin θX
i (t) ,
2V Y
syn − VR − VT/(VT − VR). Note that V = ±∞ in
i (t) + h(1 + cos θX
Y(t)qY(1 + cos θX
2/(VT − VR)
+ (cid:61)
(4)
=
where
d
dt
and
gX
θX
qY
=
C
h
Y
i
Eq. 1 corresponds to θ = ±π in Eq.4.
We used a set of 1000 of Eq. 4 and 12 of Eq. 2 to
numerically simulate the dynamics of the spiking popu-
lation.
B. Dynamics of mean field equations
(5)
d
dt
d
dt
π ∆IX ,
We adopt the Ott-Antonsen ansatz[23], which allows us
to obtain a mean field description of the spiking model
in the limit as N → ∞. Following the previous work[24],
we get mean field equations for Eq. 1 and Eq.2 as:
rX(t)2 + bX(t)vX(t) + cX(t) + ¯IX ,
rX = 2aX rX(t)vX(t) + bX(t)rX(t) + aX
vX = aX vX(t)2 − π2
X(t) + ¯gY
−gLX(VT + VR)/(VT − VR)
cX(t)
X ⋅ NY ⋅ rY(t − τ Y
X),
bX(t)
gLX/(VT − VR),
X(t)
= −gLX VT VR/(VT − VR) + ∑
X(t)V Y
=
and
syn.
The details of the derivation are in Appendix A. We will
use this mean field equation to investigate the emergence
of macroscopic oscillations by bifurcation analysis.
= − 1
τ Y
d
− ∑
dgY
X
dt
X ⋅ pY
where
Y gY
Y gY
(6)
(7)
aX
aX
gY
=
4
Symbol
Property
TABLE II. Parameters used Eq.2.
τ Y
d
¯gY
X
kY
tY(k)
τ Y
X
τdelay
pY
X
plocal
pdelay
pI2
I2
decay time constant
peak conductance
index of pre-synaptic neuron
spike time of k-th neuron in Y
conduction delay from Y to X
conduction delay between modules
synaptic strength from Y to X
synaptic strength within a module (except pI2
I2)
synaptic strength between modules
recurrent synaptic strength in population I2
Default value (Unit)
d = 5 (ms)
d = τ I2
τ I1
d = τ E2
d = 2 (ms)
τ E1
Listed in Tab. III
(ms)
1 ≤ k ≤ NY
0(ms) : coupling within a module
τdelay(ms) : coupling between modules
τdelay = 31 (ms)
plocal: coupling within a module (except pI2
I2)
pdelay: coupling between modules
pI2
I2: coupling from I2 to I2
plocal = 0.15
pdelay = 0.125
= 0.135
pI2
I2
Y .
TABLE III. Peak conductance ¯gX
Default value (Unit)
Property
¯gE1
E1
¯gI1
E1
E2
E1
Symbol
= ¯gE2
= ¯gI2
= ¯gE2
= ¯gI2
= ¯gE2
= ¯gI2
E2
¯gE1
I1
¯gI1
I1
E1
I2
I2
= ¯gE1
= ¯gI1
E2 AMPA on pyramidal cell 4.069×10
E2 AMPA on interneuron 3.276×10
GABA on pyramidal cell 2.672×10
GABA on interneuron 2.138×10
−3(mS/cm2)
−3(mS/cm2)
−2(mS/cm2)
−2(mS/cm2)
C. Bifurcation analysis of the mean field equations
We use DDE-BIFTOOL to analyze the existence and
stability of the solutions to Eqs. 5, 6 and 7. The nu-
merical methods used in DDE-BIFTOOL are detailed in
[41 -- 46].
III. RESULTS
A. Conduction delay induces cross-frequency
coupling
where τdelay has values of 0(ms), 12.5(ms) and 31(ms) to
We performed three sets of numerical simulations
investigate the effect of the inter-module conduction de-
lay on the dynamics. Parameters except τdelay are fixed
at the default values shown in Tabs. I - III. Simulation
was performed in both the spiking population (Eqs. 2
and 4) and the mean-field equations (Eqs. 5, 6 and 7).
The two corresponding results were compared to con-
firm the validity of the mean field equations in the pres-
ence of the delay. (We note that the validity of the Ott-
Antonsen approach has not been formally proven in the
case where there are delays, although it has been empiri-
cally shown to work for delayed cases [27, 28, 47].) When
spikes are asynchronous [Fig. 1(b)] and the synaptic con-
ductances are constant values, although we can see fluc-
tuations due to the finite size effect in the spiking popula-
τdelay = 0(ms), no macroscopic rhythm is observed. The
tion [Fig. 1(c)]. When τdelay = 12.5(ms), a gamma oscil-
as a gamma oscillation. When τdelay = 31(ms), the spikes
lation is observed. The spikes are partially synchronous
[Fig. 1(d)] and the synaptic conductances exhibits oscil-
lations at around 50Hz [Fig. 1(e)], which can be regarded
are partially synchronous [Fig. 1(f)]. The synaptic con-
ductances are oscillating in the gamma range [Fig. 1(g)].
Moreover, these dynamics can be regarded as a form of
CFC because the amplitude of the gamma oscillation is
modulated by a slow rhythm. Also, this CFC can be
called "theta-gamma coupling" because the fast oscilla-
tion is about 50Hz which is in the gamma range and the
slow modulation is about 10 Hz which is in the theta
range. These simulations show that the inter-module de-
lay τdelay drastically affects the dynamics, especially on
the emergence of the rhythm. We also note that both the
gamma oscillation and CFC emerge just by increasing the
time-delay in our model.
We found that the gamma oscillation was a limit cycle
and the CFC was a torus. This is indicated in a Lorenz
plot of gI1
I1
plane where gI1
, which is a plot of(gI1
I1[n + 1]) in x-y
I1[n] is the n-th local maximum value of gI1
I1[n], gI1
I1
[Fig. 1(h)] [49]. Here, gI1
I1
equations (Eqs. 5, 6 and 7). When τdelay = 12.5(ms), the
limit cycle. When τdelay = 31(ms), the Lorenz plot is a
Lorenz plot is a point, which indicates the dynamics is a
is simulated with the mean field
circle which indicates the dynamics is a torus[48].The bi-
furcation analysis of these attractors will be investigated
in the next session. We also show the corresponding mean
field (Eqs. 5, 6 and 7) behavior along with that of the
spiking model in [Figs. 1(c), (e) and (g)]. The dynamics
of the mean field equations are generally consistent with
the corresponding spiking population (Eqs. 4 and 2) de-
spite the fluctuations due to the finite size of the spiking
model.
FIG. 1. Population dynamics of delay-coupled modified theta
neurons. (a) Schematics. The coupling between the two mod-
ules is delayed because of long-range conduction time.
(b-
g) Result of simulations. plocal = 0.15, pdelay = 0.125 and
= 0.135. (b,c) τdelay = 0 (ms). (d,e) τdelay = 12.5 (ms).
pI2
I2
(f,g) τdelay = 31 (ms). (b,d,f) Rastergram. (c, e, g) Time se-
ries of gI1
E1 . The dynamics of spiking population and
mean field equations are in good agreement. (h) Lorenz map
of the peak of gI1
I1 and gE1
I1(t).
B. Cross-frequency coupling by delay-induced
torus bifurcation
We performed bifurcation analysis on the mean field
equations to reveal the mechanism for the qualitative dif-
ference of the dynamics that we found in Fig. 1. We show
three bifurcation diagrams in Figs. 2 (a-c).
5
Fig.
2 (a) is a one-parameter bifurcation diagram
where the horizontal axis is τdelay, and the vertical axis is
the value of the equilibrium point or the envelope of the
oscillation for the dynamical variable gI1
. Parameters
I1
used in Fig. 1 (b,c), Fig. 1 (d,e) and Fig. 1 (f,g) corre-
sponds to circle, triangle and diamond, respectively. We
can see limit cycles emerge via Hopf bifurcations (HB),
and a torus emerges in a parameter region where two
distinct limit cycle orbits coexist. For τdelay = 0(ms),
the black dotted line indicates the equilibrium point is
stable and the light blue dotted lines indicate unstable
equilibrium. As τdelay increases, a limit cycle oscillation
emerges from a HB point at τdelay = 0.39(ms). The limit
cycle ends with another HB at τdelay = 3.4(ms). For
larger τdelay, another HB is induced at τdelay = 4.0(ms)
and ends with τdelay = 6.4(ms). As τdelay increases, limit
cycles with large and small amplitudes appear regularly
in turn with different intervals. As the consequence of
their regular appearance, around τdelay = 30(ms), two
limit cycles coexist. In this region, we can see the limit
cycles become unstable, and a torus emerges via a torus
bifurcation. Unlike the limit cycles and fixed point, which
are computed via continuation with DDE-BIFTOOL, the
red curve showing the tori is computed by direct forward
integration of the mean field equations.
Fig. 2(b) and (c) are two-parameter bifurcation di-
agrams which show the region of equilibrium solutions,
limit cycles, and tori in a 2D parameter space. The hor-
izontal axis is for the parameter τdelay. The vertical axis
is for pdelay(b) and pI1
I1
(c).
In Fig.
In Fig. 2(b), we found CFC emerges as a torus in
the area surrounded by two torus branches emanate from
the double-Hopf bifurcation (DHB) point where two HB
boundaries overlap.
2(b), the green line of
pdelay = 0.135 corresponds to the parameters shown
in Fig.
2(a). We can also see the HB region peri-
odically emerges. These HB regions become wider as
pdelay becomes large, resulting in U-shaped HB bound-
aries. We can see two distinct HB boundaries overlap at
some points, which are the DHB points where the com-
plex eigenvalues of the two sets (total of 4) exist on the
imaginary axis. As shown in a previous analysis of this
codimension 2 bifurcation, a two-torus branches emanate
from a DHB point [51, 52]. The area surrounded by the
torus branch shown above is coincident with the area
where the torus occurs in Fig. 2 (a). From these facts,
we have shown that the generation mechanism of CFC is
due to the torus bifurcations and it emanates from the
DHB point where two HB branches overlap.
In Fig. 2 (c), gamma oscillations emerge when pI2
I2
> 0.2), regardless of the values of
is large enough (pI2
I2
τdelay . The emergence of gamma oscillations due to
strong local inhibition has also been reported in previ-
≤ 0.2, stable gamma oscillations
ous studies. When pI2
I2
and CFC coexist depending on pI2
and τdelay. We note
I2
CFC emerges in networks even with symmetric coupling
(b)(d)(c)(e)(f)(g)(h)050100time (ms)05001000# neuron050100time (ms)05001000# neuron050100time (ms)00.050.10.15g (mS/cm2)050100time (ms)00.050.10.15g (mS/cm2)050100150200250300time (ms)05001000# neuron050100150200250300time (ms)00.050.10.15g (mS/cm2)0.050.10.15gI1I1[n]0.050.10.15gI1I1[n+1]delay= 12.5delay= 31delay= 0delay= 12.5delay= 31gI1I1gE1E1spiking modelmean field(a) 6
= plocal = 0.15). For example,
between modules (i.e. pI2
I2
when τdelay = 30 (ms) as indicated by a star, it is inside
the torus bifurcation region, and the emergence of the
torus is confirmed also from the numerical simulation.
In Fig. 2 (d-i), the numerical simulations of the repre-
sentative parameters are shown. Since the qualitative
features of them agree with the prediction of the bifur-
cation diagram, this validates the bifurcation analysis of
the mean field model.
C.
Inter-population spike correlations
3
i
2
k
, t
, t
To understand how the cross-frequency coupling af-
fects the underlying spiking dynamics, we simulated the
spiking dynamics (Eq. 2 and Eq. 4) in the case of CFC
and analyzed the spike correlations. Parameters are the
same as in Fig. 1 (f,g). We obtain the spike times of
the i th neuron in X population as t
, ...
1
from the simulation. The spike trains are defined as
yX
tion analysis, we basically followed previous work [50].
). For the method of spike correla-
i (t) = Σkδ(t − t
X(i)
X(i)
X(i)
X(i)
The rastergram of the E1 population is shown in Fig.
3(a). Neurons are sorted in ascending order of I E1
. We
can see two lines running periodically in the rastergram
(colored as red and blue). The firing rate of each neuron
in E1 is shown in Fig. 3(b). Note that the firing rate is
the same within red and blue clusters, respectively (Red
cluster: 34 ≤ i ≤ 93, Blue cluster: 375 ≤ i ≤ 384). In
Fig. 3(c) and 3(d), the same analysis is performed for
E2 . Similarly, we found two lines running periodically
(colored as green and purple) and the firing rates are the
same within each cluster (Green cluster: 431 ≤ i ≤ 470,
Purple cluster: 774 ≤ i ≤ 781).
To quantify the relationship between the spike trains,
, which is correlation coefficient be-
we computed ρXY
ij
tween the ith neuron in X and the jth neuron in Y as
i (t, t + ∆t) , N Y
i (t, t + ∆t) VarN Y
j (t, t + ∆t)
CovN X
VarN X
j (t, t + ∆t) ,
i (t, t + ∆t) is the number of spikes emitted
i (t, t + ∆t) = (cid:69) t+∆t
i (t
′)dt
N X
(8)
(9)
yX
.
′
by the ith neuron in X population during a time bin ∆t,
which is given by
where N X
ij =
ρXY
t
In our study, the time bin ∆t was set as 10 (ms). Fig.
3(e) displays the pairwise correlations for all of the pairs
within and between E1 and E2. We can visually con-
firm high correlation within the red and green clusters,
as well as negative correlation between the red and green
clusters.
FIG. 2. Bifurcation diagrams (a-c) and simulations (d-i). (a)
One parameter bifurcation diagram with τdelay and gI1
I1 . Black
dotted line indicates stable steady state, light blue dotted line
is one pair of eigenvalues with positive real parts and dark
blue dotted line is four positive real part eigenvalues. Solid
blue line indicates stable limit cycle orbit and solid gray line
indicates unstable limit cycle orbit (maximum and minimum
values). Red line indicates the torus orbit (maximum and
minimum values, computed by integration of the system). In
the following bifurcation diagrams, parameters used in Fig. 1
(b,c), Fig. 1 (d,e) and Fig. 1 (f,g) are marked as , and
੦, respectively. (b) Two parameter diagram with τdelay and
pdelay. The Hopf bifurcation and torus bifurcation are plot-
ted. Also, the number of positive eigenvalues are indicated
by shaded blue area. The region of a stable torus solution
is indicated by red shaded area. (c) Two parameter bifurca-
tion diagram with τdelay and pI2
I2. Similarly, the bifurcation
boundaries and the number of positive eigenvalues are plot-
ted. (d-i) Result of numerical simulation with parameter sets
E2 . (d) :
marked in Fig. 2 (a-c). Blue: gE1
τdelay = 29 (ms) (e) ੦: τdelay = 31 (ms) (f) : τdelay = 33
(ms) (g)
: τdelay = 19 (ms), pI2
= 0.2 (h) : τdelay = 22
(ms), pI2
I2
= 0.2 (i) : τdelay = 30 (ms), pI2
E1 . Orange: gE2
= 0.15.
I2
I2
We further plotted histograms of correlation coeffi-
cients as shown in Figs. 3(f) and 3(g).
In Fig. 3(f),
we plotted histograms of correlations within red clusters
(red), within blue clusters (blue), and between red and
blue clusters (gray) in E1 population. We can see the
firing within red and blue clusters is highly correlated,
while low correlation is shown between the red and blue
clusters, which depicts the local decorrelation of the clus-
ters. We also plotted the histogram of the correlation
coefficients between the clusters located in different pop-
ulations (E1 and E2) in Fig. 3(g). Red-green pairs and
blue-purple pairs are highly correlated, while red-purple
and blue-green pairs exhibit low correlations. The results
indicate that highly correlated spiking activity, which is
known as a typical index of spike transmission, emerges
across regions in individual or sub-cluster level via the
time-delay in communication.
ij (τ) =(cid:61)
Although some cluster pairs are shown to be decorre-
lated from the histogram, one might overlook temporally-
localized correlated activity between the clusters. To fur-
ther investigate the temporal structure of the correlation,
we derived C XY
between the ith neuron in X and the jth neuron in Y as
ij (τ) which is a cross-covariance function
j (tn − τ)
i (tn) Y Y
i (t, t + ∆T) 1
i (tn) = ∫ tn+δt
− 1
∆T
(10)
Y X
yX
n
tn
∆T
N Y
N X
C XY
j (t, t + ∆T),
i (t
′)dt
ij (T) over the corresponding neuron pairs.
′ is the
time binned spike train. The time bin δt was set as 1(ms)
and ∆T was 2(sec). Then, the averaged cross-covariance
between the colored clusters is derived as the averaged
C XY
where tn = nδt and Y X
In Fig. 3(h), we can see the cross-covariance between
red and blue clusters is relatively small (gray), in rela-
tion to the pairs within clusters (red and blue), which
shows the two clusters are decorrelated although they
belong to the same excitatory population (E1). We note
that the both red and blue clusters show oscillatory firing
in gamma band with slightly different frequencies (red:
around 40Hz, blue: around 50Hz). We also evaluated
cross-covariance functions for the inter-regional pairs as
shown in Fig. 3(i). Red-green pairs are oscillating with
the same frequency (around 50 Hz) and therefore highly
correlated. Similarly, blue-purple pairs are oscillating
with the same frequency (around 40 Hz). We note that
blue-purple pairs are positively correlated while red-green
pairs are negatively correlated in Fig. 3(g). This is the
consequence of the phase lag between them as seen in
Fig. 3(i).
These results indicate that the long range delayed cou-
pling entrains the neurons in different regions to fire
in the same frequencies in the gamma-band and gen-
erate cross-covariance across the distant brain regions.
Also, the gamma-band entrainment can occur with sev-
eral gamma-band pairs at the same time, which results in
the beat between the gamma-bands and resulting CFC.
7
FIG. 3. CFC exibits long-range multi-cluster gamma-band
cross-covariance (a) Rastergram of E1. Two clusters are col-
ored as red and blue. (b) Firing rate of each neuron. Note
that the firing rate is the same within each cluster. (c) Raster-
gram of E2. Two clusters are colored as green and purple.
(d) Firing rate of each neuron. Similarly, the firing rate is
the same within each clusters. (e) Correlation coefficients for
each neuron pair. Time bin is 10(ms). (f) Histogram of cor-
relation coefficients for pairs within red clusters (red), pairs
within blue clusters (Blue) and pairs between red and blue
clusters (Gray) in E1. We can see firings within each clusters
are highly correlated, while there is low correlation between
the clusters. (g) Histogram of correlation coefficients between
clusters belonging to different populations. Red-green pair
and blue purple pairs are highly correlated, while red-purple
and blue-green pairs exhibit low correlations. (h) Averaged
cross-covariance function for corresponding pairs in (f). Time
bin is 1(ms). Both red and blue clusters exhibit oscillatory
firing. Note that the frequency is slightly different, which re-
sults in low cross-covariance between red and blue clusters. (i)
Averaged cross-covariance function for corresponding pairs in
(g). Time bin is 1(ms). Red-green pairs and blue-purple pairs,
which are located in distant neuronal modules, are oscillating
with the same frequency and exhibit high correlation.
D. Suitable temporal window for population rate
code
It is experimentally reported that neurons tend to fire
in specific phases of ongoing background oscillations[53],
and this relation between phase of oscillation and spike
timing is known to encode information[54]. To evaluate
any relationship between the phase of the background os-
cillation and spike timing, we introduced the mean mem-
brane potential to evaluate the phase of the background
oscillation and investigated the spike counts in relation
to these phases in E1 population. There are two possible
measures to evaluate the mean excitability of a neuronal
population: population spike count and mean membrane
voltage. In this study, we used the mean membrane po-
tential because it is a smoother function of time than pop-
ulation spike counts. (It can be defined as a continuous
function, while the population spike count is discrete.)
In order to avoid the effect of physiologically implausible
blow-up of membrane potential, we obtained the mem-
brane potential of each neuron as: V X
i =(VR + VT)/2 +
i + )(VT − VR)/2. A small con-
stant = 2.0⋅10
−4 is introduced in the denominator in or-
der to avoid the divergence to infinity at θ = ±π [34, 55].
Then, from the time course of individual neuron data
given by Eq.2 and Eq.4, the mean membrane potential
sin θX
i /(1 + cos θX
V(t) is evaluated as
Based onV(t) , we derived the envelopes and defined
V(t) and its envelopes are shown. Envelopes are cap-
tured as the amplitude of the analytic signal ofV(t),
The zero phase for φθ(t) and φγ(t) is defined as the nega-
which was derived via Hilbert filters. To capture the
theta and gamma phase, we applied two Hilbert filters
with different filter length (250ms for θ, 50ms for γ) [56].
macroscopic phase for the theta oscillation (φθ) and the
gamma oscillation (φγ). In Fig. 4 (a), the time series of
V(t) = 1
NE1(cid:61)
V E1
i
(11)
NE1
i=1
plotted whole
tive peak of the envelope as shown in Fig. 4 (b). Because
the macroscopic dynamics is a torus, where the frequency
components are rationally independent generically, then
the trajectory covers the whole area of the φθ-φγ- 2D
space as shown in Fig 4 (c).
on the 2D space as shown in Fig. 4 (d). We can see the
spike count is modulated by both φθ and φγ.
(t − δt/2, t + δt/2)
sE1φθ(t) , φγ(t)
φm = 2mπ/M in φθ-φγ- 2D space and derived averaged
spike counts for each bin ¯sE1φi
γ
To further investigate how the spike count is mod-
ulated by the phases, we introduced M-binned phases
γ = EsE1φi
population
i N E1
= ∑NE1
counts
spike
θ, φj
θ, φj
where φi ≤ φi
shown in Fig.
M = 40.
θ < φi+1 and φj ≤ φj
γ < φj+1 as
4 (e). Here, the number of bins is
4 (e), there is a region of high
In Fig.
We
.
i
8
γ = VarsE1φi
θ, φj
θ, φj
F E1φi
¯sE1.
In addition, to evaluate the variability of the
spike count, we derived Fano factors for the spike count
γ/¯sE1φi
γ [50], for
θ, φj
each bin as shown in Fig. 4 (f). From these calculations,
we can see that the region with high ¯sE1 in Fig.4(e) is
interposed by the two regions with high F E1 as shown
in Fig. 4(f). Between the region with high F E1, we can
see a narrow area in which ¯sE1 is high and F E1 is low.
We note that for a population rate code, we would like a
high spike count average and a low Fano factor.
We also investigated how the temporal firing activity of
the sub-clusters, which was found in Fig. 3 (a), is mod-
ulated by the macroscopic phases. The averaged spike
red : 34 ≤ i ≤ 93), the blue
count for the red cluster ( ¯sE1
blue : 375 ≤ i ≤ 384) and the asynchronous
cluster (¯sE1
neurons between the red and blue clusters (¯sE1
async :
94 ≤ i ≤ 374) are shown in Figs. 4 (g) ,(h) and (i).
We can see these clusters exhibit several distinct types of
phase-specific firing modulated by both φθ and φγ
FIG. 4. θ -γ oscillations and spiking dynamics. (a) Time se-
ries ofV(t), the mean of the membrane potential of neurons
in E1(Black) and the envelopes ofV(t) for theta oscillations
(red) and gamma oscillations (blue). (b) Introducing φθ and
φγ. (c) Time evolution of φθ and φγ. (d) Population spike
count plotted over φθ - φγ space. (e) Averaged spike counts
for each bin in φθ-φγ space. (f) Fano factors of spike count.
(g-i) Averaged spike counts of the sub-clusters found in Fig.
3. (g) The red cluster (h) The blue cluster. (i) The neurons
between the red and blue clusters.
IV. DISCUSSION
We employed numerical and theoretical analyses for
inter-regionally coupled neuronal populations in a multi-
scale point of view. We found the emergence of nontrivial
CFC, induced by time-delay, with correlated spike trains
between regions. We introduced quenched variability to
individual neurons, instead of individual noise, in order
to analyze the effect of time-delayed interactions thus
avoiding the non-Markov state where noise and delay co-
exist.
Another key technique is that we consider conductance
based synapses and voltage dependent neuronal dynam-
ics. By the Lorentzian ansatz [24] (it is associated with
the Ott-Antonsen ansatz [23]), the dynamics of neuronal
populations reduce to a set of macroscopic DDEs. These
enable us to unveil the significant impact of the delay on
the rhythmogenesis. As the derived equations are valid
regardless of the frequencies, we successfully investigate
theta-gamma interactions, which is not possible by con-
ventional firing-rate models [20, 22].
In addition, the
mean (Eq. 11) and individual voltage (Eq. 1) of neu-
rons are appropriately evaluated under biologically plau-
sible values for the conductances and reversal potentials
of GABA and AMPA [31].
By the analyses of the macroscopic equations, we found
that time-delay destabilizes the asynchronous firing state
via a Hopf bifurcation (HB). Further increase of the time-
delay leads to a destabilization of the oscillation (that
emerged from the HB) via a torus bifurcation which pro-
duces the CFC. Fig. 2 (b) shows that the torus bifurca-
tion emerges at the intersection of two curves of HBs. [51]
performed a partial analysis of the case in which there is
a double HB in a delay model in the non-resonant case
(when the frequencies of the HBs were not rationally re-
lated); in a broad range of parameters, it is possible to
find a stable torus that emerges.
(See also [57] for a
neuronal example.) This appears to be the case in our
system. Under such bifurcations, diverse interactions be-
tween the macroscopic oscillation and individual neurons
can be realized. By analyses of the microscopic spike
trains under CFC, we found that the inter-population
correlation can be much larger than that of the intra-
population. We also found that there is a firing prefer-
ence of sub-populations as a function of the phase of the
slow oscillation.
In previous studies, the gamma oscillation were found
to emerge from local inhibitions [3, 34]. Our findings here
show that a gamma oscillation can emerge from inter-
regional excitatory time-delayed interactions (e.g. Fig. 2
(a)). This delay-induced mechanism is different from the
conventional mechanisms (ING and PING), thus implies
a novel generation mechanism for the gamma oscillation.
Pairwise spike correlation has been widely analyzed in
physiological experiments. It is reported to increase or
decrease with respect to task demands[58]. Although the
relation of correlations to brain functions has been ex-
perimentally suggested, it is still hard to interpret them
9
and understand the origin of such correlations. As for
the dynamical mechanism, Litwin-Kumar et al. pro-
posed that a slight heterogeneity of neural interaction
increased pairwise correlation in the same cluster [50].
In our model, there are two clusters in each population
that show the same firing rate. High correlations in abso-
lute value emerge between neurons within different pop-
ulations, that are located in distant regions, while low
correlation occur between populations that have different
firing rates. CFC induced by time delay can thus serve
multi-band information transfer by using sub-cluster cor-
relations and may support neuronal multiplexing [59].
Regarding the spike preferences of macroscopic phase,
Georgiou et al. show that the spike timing of neurons
in V4 and the frontal eye fields occurs at specific phases
of LFP in V4 during visual attentions [60]. Sellers et
al. also found that neurons in Prefrontal cortex fire at
specific phases of not only local theta and gamma oscil-
lation, but also phase of theta oscillation of anatomically
connected distant region [61]. These studies suggest that
the spike preference of macroscopic phase could encode
and transfer information. In Fig. 4, our model also ex-
hibits spike preference of the macroscopic phase of theta
and gamma oscillations.
Senior et al. evaluated firing timing of hippocampal
CA1 pyramidal cells and found phase preferences in re-
lation to theta and gamma oscillations [53]. They also
found distinct two types neuron groups that have a dif-
ferent phase preference (Fig. 4 B in [53]) and firing rate
(Fig. 2 B in [53]). Such clusters, which exhibit distinct
phase preference and firing rates, were also found in our
model during CFC. Moreover, we found these clusters
exhibit long-range gamma-band cross-covariance across
distant regions, suggesting the impact on long-range in-
formation transfer. We note the emergence of such sub-
clusters in a population, which were characterized by dif-
ferent firing rates, are also reported even under collective
chaos [27, 62]. Further studies are needed to figure out
whether sub-clusters can also contribute to long-range
communication in a synchronized manner in chaotic re-
gions, where population level signals are hard to correlate
[22].
The theta-gamma neural code hypothesis, proposed
by Lisman et al., assumes that the nested oscillation,
where the phase of theta oscillation and the amplitude of
gamma oscillation are coupled, encode multiple memory
items in an ordered way[63]. Although this hypothesis is
supported by recent experimental studies [64], it is less
clear whether the dynamical properties of such nested
oscillations have the ability to store information. We
showed, in Figs. 4(e) and 4(f), that nested oscillations
accompany a phase-specific time window with high-spike
counts and low-Fano factor. Within such a temporal win-
dow, stimuli are likely to be robustly encoded by changes
of spike counts.
In real brains, interactions between multiple regions
would exhibit diverse time-delayed interactions [65]. Fur-
ther investigation about delayed interactions and infor-
mation transmission will be important in order to under-
stand how spike dynamics is affected by the macroscopic
phase of multiple brain regions [66].
ACKNOWLEDGMENTS
This study was supported in part by Grant-in-Aid for
JSPS Research Fellow Grant No. 16J04952 to A.A. K.K.
was supported by JST PRESTO (JPMJPR14E2) and
KAKENHI (18H04122) B.E. was supported by the US
NSF DMS-1712922. We would like to thank Brent Do-
iron for helpful discussions.
Appendix A: Derivation of mean field equations
In this section, the details of the derivation of mean
field equations are described.
In the derivation, we
mostly follow the previous work by Montbrió et al [24].
1. Spiking description to reduce by Ott-Antonsen
ansatz
i (t) , i-th
The dynamics of the membrane potential V X
Y
C
i ,
gY
(A1)
dV X
i
dt
VT − VR
=gLX
− (cid:61)
which can be written as
neuron in population X, is
i (t) − VT)
(V X
i (t) − VR)(V X
syn) + I X
i (t) − V Y
X(t)(V X
i (t) + cX(t) + I X
i (t)2 + bX(t)V X
bX(t)
gLX/(VT − VR),
−gLX(VT + VR)/(VT − VR) − ∑
X(t)
X(t)V Y
cX(t) = −gLX VT VR/(VT − VR) + ∑
X(t) is written as
X)). (A3)
δ(t−(t
(kY) +τ Y
=aXV X
X(t)+ ¯gY
i ,
(A2)
=
and
Also, the dynamics of gY
dV X
i
dt
X ⋅pY
X ⋅
where
Y gY
Y gY
NY(cid:61)
syn.
aX
gY
=
dgY
X
dt
= − 1
τ Y
d
k=1
Here, the N-body description of the whole system is
and 12 of gY
X.
1012-dimensional system with 1000 of V X
i
ρX(VXࢯIX , t)
2.
Introducing probability density function
the probability density function ρX(VXࢯIX , t) for the
Taking the continuum limit as NX → ∞, we introduce
ν
10
ρX(VXࢯIX , t) dVX describes
population X, where ∫ ν+∆ν
the probability of neurons in the population whose mem-
conserved, ρX(VXࢯIX , t) follows the continuity equation:
brane potentials VX are between ν and ν + ∆ν and cur-
rent is IX at time t. Because the number of neuron is
X + bX(t)VX + cX(t) + IX
ρX(VXࢯIX , t)
∂
∂t
= − ∂
∂VX
aX V 2
ρX(VXࢯIX , t)]
(A4)
3. Adopting the "Ott-Antonsen ansatz"
,
π
d
dt
(A6)
xX(IX , t)
Here, we start to derive the mean field dynamics of
the system by applying so-called "Ott-Antonsen ansatz
(OAA)". In the previous study[24], the ansatz is extended
to Quadratic integrate-and-fire neurons as:
(A5)
which is a Lorentzian distribution with dynamical vari-
probability density function ρX. Adopting the ansatz,
we obtain the low dimensional behavior as
d
dt
ρX(VXࢯIX , t) = f(IX)
[VX − yX(IX , t)]2 + xX(IX , t)2
ables xX(IX , t) and yX(IX , t). Here, xX(IX , t) and
yX(IX , t) represent the low dimensional behavior of the
xX(IX , t) = 2aX xX(IX , t)y(IX , t) + bX(t)xX(IX , t),
yX(IX , t) = − aX(xX(IX , t))2 + aX(yX(IX , t))2
+ bX(t)yX(IX , t) + cX(t) + IX .
complex variable wX(IX , t) =
xX(IX , t) + iyX(IX , t), the two coupled equations can
wX(IX , t) = iaX wX(IX , t)2+bX(t)wX(IX , t)+cX(t)+IX .
4. Description of mean field dynamics: rX(t) and
rate rX(t) and the mean voltage vX(t). The firing rate
of the population rX(t) is obtained by summing up the
threshold. Taking the firing threshold Vpeak → ∞, rX(t)
X + bX(t)VX + cX(t) + IX)
rX(t) =(cid:69) ∞
flux for all IX at VX = Vpeak , where Vpeak is the firing
We introduce two macroscopic observables: the firing
be written in complex form as
vX(t)
can be defined as
Introducing the
(A7)
(A8)
d
dt
lim
VX →∞(aX V 2
ρX(VX , IX , t)dIX .
−∞
(A9)
Substituting the ansatz, we get
rX(t) = aX
Following that fX(IX) is now given as the Lorentzian
xX(IX , t)fX(IX)dIX .
(cid:69) ∞
(A10)
−∞
π
function, this improper integration can be evaluated us-
ing an analytic continuation and the residue theorem,
then we get
rX(t) = aX
π xX( ¯IX − i∆IX , t).
(A11)
p.v.(cid:69) ∞
the VX and IX values, so that
Next, the mean voltage of the population v(t) can be
defined by integrating the VX-weighted ρX(V, I, t) for all
ρX(VX , IX , t)VX dVX dIX . (A12)
vX(t) = (cid:69) ∞
−R h(x)dx in order to avoid
−∞ h(x)dx = limR→∞∫ R
Note that we resort to the Cauchy principal value
p.v.∫ ∞
indeterminacy of the improper integral. Substituting the
ansatz, we get
vX(t) = (cid:69) ∞
yX(IX , t)fX(IX)dIX .
As the same as rx(t) case, following that fX(I) is now
given as the Lorentzian function, this improper integra-
tion can be evaluated using the residue theorem to get
(A13)
−∞
−∞
−∞
vX(t) = yX( ¯IX − i∆IX , t).
11
(A14)
Finally, substituting Eq. A11 and Eq. A14 into Eq.A8,
the population dynamics is obtained as
d
dt
rX(t) = 2aX rX(t)vX(t) + bX(t)rX(t) + aX
vX(t) = − π2
rX(t)2 + aX vX(t)2 + bX(t)vX(t) + cX(t) + ¯IX .
π ∆IX ,
(A15)
aX
d
dt
(A16)
5. Description of mean field synaptic dynamics:
X(t)
gY
rate rX(t), Eq. A3 can be written using rX(t) instead of
Since we have the dynamics of the population firing
using delta function as
X(t) + ¯gY
gY
X ⋅ NY ⋅ rY(t − τ Y
X).
X ⋅ pY
(A17)
dgY
X
dt
= − 1
τ Y
d
[1] Buzsáki, György. Rhythms of the Brain. Oxford Univer-
sity Press, 2006.
[2] Wang, Xiao-Jing. " Neurophysiological and computa-
tional principles of cortical rhythms in cognition." Phys-
iological reviews 90.3 (2010): 1195-1268.
[3] Buzáki,György and Xiao-jing Wang, Mechanisms of
gamma oscillations. Annual review of neuroscience 35
(2012): 203-225.
[4] Jensen, Ole, and Laura L. Colgin. " Cross-frequency cou-
pling between neuronal oscillations." Trends in cognitive
sciences 11.7 (2007): 267-269.
[5] Canolty, Ryan T., and Robert T. Knight. " The func-
tional role of cross-frequency coupling." Trends in cogni-
tive sciences 14.11 (2010): 506-515.
[6] Aru, Juhan, et al. "Untangling cross-frequency coupling
in neuroscience." Current opinion in neurobiology 31
(2015): 51-61.
[7] Belluscio, Mariano A., et al. " Cross-frequency phase --
phase coupling between theta and gamma oscillations in
the hippocampus." Journal of Neuroscience 32.2 (2012):
423-435.
[8] Hyafil, Alexandre, et al. "Neural cross-frequency cou-
pling: connecting architectures, mechanisms, and func-
tions." Trends in neurosciences 38.11 (2015): 725-740.
[9] Sase, Takumi, et al. "Bifurcation analysis on phase-
amplitude cross-frequency coupling in neural networks
with dynamic synapses." Frontiers in computational neu-
roscience 11 (2017): 18.
[10] Hyafil, Alexandre, et al. "Speech encoding by coupled
cortical theta and gamma oscillations." Elife 4 (2015):
e06213.
[11] Onslow, Angela CE, Matthew W. Jones, and Rafal Bo-
gacz. "A canonical circuit for generating phase-amplitude
coupling." PLoS One 9.8 (2014): e102591.
[12] Chehelcheraghi, Mojtaba, et al. "A neural mass model
of cross frequency coupling." PloS one 12.4 (2017):
e0173776.
[13] Varela, Francisco, et al. " The brainweb: phase synchro-
nization and large-scale integration." Nature reviews neu-
roscience 2.4 (2001): 229.
[14] Fontolan, Lorenzo, et al. " The contribution of frequency-
specific activity to hierarchical information processing in
the human auditory cortex." Nature communications 5
(2014): 4694.
[15] Doesburg, Sam M., et al. " Theta modulation of inter-
regional gamma synchronization during auditory atten-
tion control." Brain research 1431 (2012): 77-85.
[16] Dynamic cross-frequency couplings of local field poten-
tial oscillations in rat striatum and hippocampus during
performance of a T-maze task Adriano B. L. Tort, Mark
A. Kramer, Catherine Thorn, Daniel J. Gibson, Yasuo
Kubota, Ann M. Graybiel, and Nancy J. Kopell
[17] Canolty, Ryan T., et al. "High gamma power is phase-
locked to theta oscillations in human neocortex." science
313.5793 (2006): 1626-1628.
[18] Breakspear, Michael. " Dynamic models of large-scale
brain activity." Nature neuroscience 20.3 (2017): 340.
[19] Wilson, Hugh R., and Jack D. Cowan. " Excitatory and
inhibitory interactions in localized populations of model
neurons." Biophysical journal 12.1 (1972): 1-24.
[20] Ermentrout, G. Bard, and David H. Terman. Mathemati-
cal foundations of neuroscience. Vol. 35. Springer Science
& Business Media, 2010.
[21] Devalle, Federico, Alex Roxin, and Ernest Montbrió. "
Firing rate equations require a spike synchrony mecha-
nism to correctly describe fast oscillations in inhibitory
networks." PLoS computational biology 13.12 (2017):
e1005881.
[22] Battaglia, Demian, Nicolas Brunel, and David Hansel.
"Temporal decorrelation of collective oscillations in neu-
ral networks with local inhibition and long-range excita-
tion." Physical review letters 99.23 (2007): 238106.
[23] Ott, Edward, and Thomas M. Antonsen. " Low dimen-
sional behavior of large systems of globally coupled oscil-
lators." Chaos: An Interdisciplinary Journal of Nonlinear
Science 18.3 (2008): 037113.
[24] Montbrió, Ernest, Diego Pazó, and Alex Roxin. " Macro-
scopic description for networks of spiking neurons." Phys-
ical Review X 5.2 (2015): 021028.
[25] Luke, Tanushree B., Ernest Barreto, and Paul So. " Com-
plete classification of the macroscopic behavior of a het-
erogeneous network of theta neurons." Neural computa-
tion 25.12 (2013): 3207-3234.
[26] Laing, Carlo R. " Exact neural fields incorporating gap
junctions." SIAM Journal on Applied Dynamical Sys-
tems 14.4 (2015): 1899-1929.
[27] Pazó, Diego, and Ernest Montbrió. " From quasiperiodic
partial synchronization to collective chaos in populations
of inhibitory neurons with delay." Physical review letters
116.23 (2016): 238101.
[28] Devalle, Federico, Ernest Montbrió, and Diego Pazó.
"Dynamics of a large system of spiking neurons with
synaptic delay." Physical Review E 98.4 (2018): 042214.
[29] Dumont, Grégory, G. Bard Ermentrout, and Boris
Gutkin. " Macroscopic phase-resetting curves for spik-
ing neural networks." Physical Review E 96.4 (2017):
042311.
[30] So, Paul, Tanushree B. Luke, and Ernest Barreto. " Net-
works of theta neurons with time-varying excitability:
Macroscopic chaos, multistability, and final-state uncer-
tainty." Physica D: Nonlinear Phenomena 267 (2014): 16-
26.
[31] Brunel, Nicolas, and Xiao-Jing Wang. " What deter-
mines the frequency of fast network oscillations with
irregular neural discharges? I. Synaptic dynamics and
excitation-inhibition balance." Journal of neurophysiol-
ogy 90.1 (2003): 415-430.
[32] Stiefel, Klaus M., et al. " Phase dependent sign changes
of GABAergic synaptic input explored in-silicio and
in-vitro." Journal of computational neuroscience 19.1
(2005): 71-85.
[33] Vida, Imre, Marlene Bartos, and Peter Jonas. " Shunt-
ing inhibition improves robustness of gamma oscillations
12
in hippocampal interneuron networks by homogenizing
firing rates." Neuron 49.1 (2006): 107-117.
[34] Kotani, Kiyoshi, et al. " Population dynamics of the mod-
ified theta model: macroscopic phase reduction and bi-
furcation analysis link microscopic neuronal interactions
to macroscopic gamma oscillation." Journal of The Royal
Society Interface 11.95 (2014): 20140058.
[35] Akao, Akihiko, et al. " Relationship between the mech-
anisms of gamma rhythm generation and the magnitude
of the macroscopic phase response function in a popu-
lation of excitatory and inhibitory modified quadratic
integrate-and-fire neurons." Physical Review E 97.1
(2018): 012209.
[36] Caminiti, Roberto, et al. " Evolution amplified process-
ing with temporally dispersed slow neuronal connectiv-
ity in primates." Proceedings of the National Academy
of Sciences 106.46 (2009): 19551-19556.
[37] Bloomfield, S. A., J. E. Hamos, and S. M. Sherman. "
Passive cable properties and morphological correlates of
neurones in the lateral geniculate nucleus of the cat." The
Journal of physiology 383.1 (1987): 653-692.
[38] D. A. McCormick and J. R. Huguenard, J. Neurophysiol.
68, 1384 (1992).
[39] M. Bartos, I. Vida, M. Frotscher, J. R. P. Geiger, and P.
Jonas, J. Neurosci. 21, 2687 (2001).
[40] A. Gupta, Y. Wang, and H. Markram,
[41] K. Engelborghs, T. Luzyanina, and D. Roose, Numerical
bifurcation analysis of delay differential equations using
DDE-BIFTOOL, ACM Trans. Math. Softw. 28 (1), pp.
1-21, 2002.
[42] K. Engelborghs, T. Luzyanina, G. Samaey. DDE-
BIFTOOL v. 2.00: a Matlab package for bifurcation
analysis of delay differential equations. Technical Report
TW-330, Department of Computer Science, K.U.Leuven,
Leuven, Belgium, 2001.
[43] J. Sieber, K. Engelborghs, T. Luzyanina, G. Samaey, D.
Roose: DDE-BIFTOOL Manual - Bifurcation analysis of
delay differential equations. arxiv.org/abs/1406.7144.
[44] Sebastiaan Janssens: On a Normalization Technique for
Codimension Two Bifurcations of Equilibria of Delay Dif-
ferential Equations. Master Thesis, Utrecht University
(NL), supervised by Yu.A. Kuznetsov and O. Diekmann,
dspace.library.uu.nl/handle/1874/312252, 2010.
[45] Bram Wage: Normal form computations for Delay Dif-
ferential Equations in DDE-BIFTOOL. Master Thesis,
Utrecht University (NL), supervised by Y.A. Kuznetsov,
dspace.library.uu.nl/handle/1874/296912, 2014.
[46] M. M. Bosschaert: Switching from codimension 2 bifurca-
tions of equilibria in delay differential equations. Master
Thesis, Utrecht University (NL), supervised by Y.A.
Kuznetsov,
dspace.library.uu.nl/handle/1874/334792,
2016.
[47] Lee, Wai Shing, Edward Ott, and Thomas M. Antonsen.
"Large coupled oscillator systems with heterogeneous in-
teraction delays." Physical review letters 103.4 (2009):
044101.
[48] Candaten, Matteo, and Sergio Rinaldi. " Peak-to-peak
dynamics: A critical survey." International Journal of
Bifurcation and Chaos 10.08 (2000): 1805-1819.
[49] Strogatz, Steven H. Nonlinear dynamics and chaos: with
applications to physics, biology, chemistry, and engineer-
ing. CRC Press, 2018.
[50] Litwin-Kumar, Ashok, and Brent Doiron. " Slow dynam-
ics and high variability in balanced cortical networks
with clustered connections." Nature neuroscience 15.11
(2012): 1498.
[51] Buono, Pietro-Luciano, and Jacques Belair. Restrictions
and unfolding of double Hopf bifurcation in functional
differential equations. Journal of Differential Equations
189.1 (2003): 234-266.
[52] Knobloch, E. " Normal form coefficients for the nonres-
onant double Hopf bifurcation." Physics Letters A 116.8
(1986): 365-369.
[53] Senior, Timothy J., et al. "Gamma oscillatory firing re-
veals distinct populations of pyramidal cells in the CA1
region of the hippocampus." Journal of Neuroscience 28.9
(2008): 2274-2286.
[54] O'Keefe, John, and Michael L. Recce. "Phase relation-
ship between hippocampal place units and the EEG theta
rhythm." Hippocampus 3.3 (1993): 317-330.
[55] Ermentrout, Bard. "Gap junctions destroy persistent
states in excitatory networks." Physical Review E 74.3
(2006): 031918.
[56] Although we arbitrary controlled the filter length as to
capture the theta and gamma envelopes, we also con-
firmed these result are robust against the minor changes
of the filter length.
[57] Ermentrout, G. B., and J. D. Cowan. Secondary bifurca-
tion in neuronal nets. SIAM Journal on Applied Mathe-
matics 39.2 (1980): 323-340.
[58] Doiron, Brent, et al. "The mechanics of state-dependent
neural correlations." Nature neuroscience 19.3 (2016):
13
383.
[59] Akam, Thomas, and Dimitri M. Kullmann. "Oscillatory
multiplexing of population codes for selective communi-
cation in the mammalian brain." Nature Reviews Neuro-
science 15.2 (2014): 111.
[60] Gregoriou, Georgia G., et al. "High-frequency, long-range
coupling between prefrontal and visual cortex during at-
tention." science 324.5931 (2009): 1207-1210.
[61] Sellers, Kristin K., et al. "Oscillatory dynamics in the
frontoparietal attention network during sustained atten-
tion in the ferret." Cell reports 16.11 (2016): 2864-2874.
[62] Luccioli, Stefano, and Antonio Politi. "Irregular collec-
tive behavior of heterogeneous neural networks." Physi-
cal review letters 105.15 (2010): 158104.
[63] Lisman, John E., and Ole Jensen. " The theta-gamma
neural code." Neuron 77.6 (2013): 1002-1016.
[64] Heusser, Andrew C., et al. " Episodic sequence mem-
ory is supported by a theta -- gamma phase code." Nature
neuroscience 19.10 (2016): 1374.
[65] Deco, G., Jirsa, V., McIntosh, A. R., Sporns, O., & Köt-
ter, R. (2009). Key role of coupling, delay, and noise in
resting brain fluctuations. Proceedings of the National
Academy of Sciences, pnas-0901831106.
[66] Canolty, Ryan T., et al. "Oscillatory phase coupling co-
ordinates anatomically dispersed functional cell assem-
blies." Proceedings of the National Academy of Sciences
(2010): 201008306.
|
1505.00041 | 3 | 1505 | 2015-09-03T20:14:21 | Modeling neural activity at the ensemble level | [
"q-bio.NC"
] | Here we demonstrate that the activity of neural ensembles can be quantitatively modeled. We first show that an ensemble dynamical model (EDM) accurately approximates the distribution of voltages and average firing rate per neuron of a population of simulated integrate-and-fire neurons. EDMs are high-dimensional nonlinear dynamical models. To faciliate the estimation of their parameters we present a dimensionality reduction method and study its performance with simulated data. We then introduce and evaluate a maximum-likelihood method to estimate connectivity parameters in networks of EDMS. Finally, we show that this model an methods accurately approximate the high-gamma power evoked by pure tones in the auditory cortex of rodents. Overall, this article demonstrates that quantitatively modeling brain activity at the ensemble level is indeed possible, and opens the way to understanding the computations performed by neural ensembles, which could revolutionarize our understanding of brain function. | q-bio.NC | q-bio |
Modeling neural activity at the ensemble level
Joaqu´ın Rapela∗1,2, Mark Kostuk3, Peter F. Rowat4, Tim Mullen1,
Edward F. Chang5, and Kristofer Bouchard6
1Swartz Center for Computational Neuroscience, UCSD
2Instituto de Investigacin en Luz Ambiente y Visin, UNT and
CONICET, Argentina
3Department of Physics, UCSD
4Institute for Neural Computation, UCSD
5Department of Neurological Surgery, UCSF
6Computational Research Division, LBNL
Contents
1 Introduction
2 Building EDMs
2.1 One Population . . . . . . . . . . . . . . . . . . . . . . . . . . . .
2.1.1 One Population of Independent Neurons . . . . . . . . . .
2.1.2 One Population with Feedback . . . . . . . . . . . . . . .
2.2 A Network of Populations . . . . . . . . . . . . . . . . . . . . . .
2.3 Partial Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .
3 Reducing dimensionality in EDMs
3.1 Method to find low-dimensional approximations of EDMs . . . .
3.1.1 Representing the ensemble pdf in a new basis . . . . . . .
3.1.2 Reducing dimensionality in the new basis . . . . . . . . .
3.2 Evaluation of the method . . . . . . . . . . . . . . . . . . . . . .
3.2.1 Firing Rates
. . . . . . . . . . . . . . . . . . . . . . . . .
3.2.2 Ensemble Proability Density Functions . . . . . . . . . . .
3.3 Partial conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .
4 Estimating parameters of EDMs
4.1 Noise Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
4.2 Optimization Surface . . . . . . . . . . . . . . . . . . . . . . . . .
4.3 Gradient of Log-Likelihood Function . . . . . . . . . . . . . . . .
∗[email protected]
1
4
5
6
6
6
6
14
17
17
17
18
20
20
20
23
39
39
41
41
5 Modeling evoked auditory activity in rodents with EDMs
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5.1 Recordings
5.2 Qualitative analysis
. . . . . . . . . . . . . . . . . . . . . . . . .
5.3 Quantitative analysis . . . . . . . . . . . . . . . . . . . . . . . . .
5.4 Partial Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Conclusions
43
43
43
43
43
52
2
Abstract
Here we demonstrate that the activity of neural ensembles can be
quantitatively modeled. We first show that an ensemble dynamical model
(EDM) accurately approximates the distribution of voltages and aver-
age firing rate per neuron of a population of simulated integrate-and-fire
neurons. EDMs are high-dimensional nonlinear dynamical models. To
faciliate the estimation of their parameters we present a dimensionality
reduction method and study its performance with simulated data. We
then introduce and evaluate a maximum-likelihood method to estimate
connectivity parameters in networks of EDMS. Finally, we show that this
model an methods accurately approximate the high-gamma power evoked
by pure tones in the auditory cortex of rodents. Overall, this article
demonstrates that quantitatively modeling brain activity at the ensemble
level is indeed possible, and opens the way to understanding the compu-
tations performed by neural ensembles, which could revolutionarize our
understanding of brain function.
3
1
Introduction
If we observe a fluid at the molecular level we see random motions, but
if we look at it macroscopically we may see a smooth flow. An intriguing
possibility is that by analyzing brain activity at a macroscopic level, i.e.,
at the level of neural ensembles, we may discover patterns not apparent
at the single-neuron level, that are as useful as velocity or temperature
are to understand, and predict, the motion of fluids.
Technology frequently drives science. For instance, thanks to the de-
velopment of microelectordes in the 1930's, we now know with exquiste
detail computations performed by single neurons. We are now experienc-
ing a dramatic increase in our capacity to monitor the activity of larger
and larger populations of neurons with higher and higher spatial and
temporal resolution. These new ensemble recordings may soon allow us
to uncover crucial computations performed by neural ensembles.
Here we present results of developing and evaluating a mathematical
model and estimation methods to characterize the activity of ensembles
of neurons from electrophysiological data.
Section 2 reports the evaluation of an ensemble dyanmical model (EDMs).
And EDMs is a high-dimensional nonlinear dynamical models. To esti-
mate its parameters it is convenient to reduce the number of parameters.
We escribe a dimensionality reduction method for EDMs in Section 3.
Section 4 presents a maximum-likelihood method to estimate parameters
of EDMs. Finally, in Section 5 we show that EDMs can accurately ap-
proximate high-gamma electroencephalographic activity evoked by pure
tones in the auditory cortex of rodents.
4
2 Building EDMs
We wanted to learn how to build Ensemble Density Models (EDMs), dy-
namical models of the state variables (e.g., trans-membrane potential,
time since last spike, etc.) of a population of identical neurons, starting
from a dynamical model of the state variables of a single neuron. For In-
tegrate and Fire (IF) model of single neurons, an EDM should provide the
ensemble probability density function (pdf) ρ(υ, t), from which to com-
pute the probability of finding a neuron in the ensemble with with a given
trans-membrane voltage υ at time t (i.e., P (υ, t) = ρ(υ, t)dt). It should
also provide the average firing rate per neuron in the population, r(t). To
construct EDMs we chose the methodology described in Omurtag et al.
[2000].
To evaluate the EDM we compared its outputs (ρ(υ, t) and r(t)) with
those derived from the direct simulation of a population of 9,000 IF neu-
rons. The value of the density function ρ(υ, t) derived from the direct
simulation was the proportion of IF neurons having a voltage υ at time
t, and the average firing rate per neuron r(t) derived from the the direct
simulation was the proportion of cells in the population at time t with
voltage at threshold.
An EDM is driven by an external excitatory current and modulated
by an external inhibitory current.
In addition, every cell in the EDM
receives inputs from G other neurons in the population. A fraction f
of these G inputs is excitatory, and the remainder are inhibitory. These
intra-population inputs act as feedback mechanisms to the EDM. The
mathematical representation of an EDM for a population of IF neurons
is given in Equations 1 -- 3, modified from Equations 26, 39 and 47 in
Omurtag et al. [2000]. External excitatory and inhibitory currents ap-
pears as σ0
i (t), respectively, in Equation 3, and excitatory and
inhibitory feedback are given by the terms Gf r(t) and G(1 − f )r(t), re-
spectively, in Equation 3.
e (t) and σ0
∂ρ
∂t
(υ, t) = −
∂J
∂υ
(υ, t)
r(t) = J(υ = 1, t)
J(υ, t) = −γυρ(υ, t)
(1)
(2)
ρ(υ ′, t)dυ ′
(3)
υ
+[σ0
−[σ0
e (t) + Gf r(t)]Z υ
i (t) + G(1 − f )r(t)]Z υ/(1−κ)
ρ(υ ′, t)dυ ′
υ−h
We first verified that for a single population the outputs of the EDM
matched those of the direct simulations (Section 2.1). We next built a
network of excitatory and inhibitory populations, and again compared
the outputs of the EDM and those of the direct stimulation (Section 2.2).
5
2.1 One Population
2.1.1 One Population of Independent Neurons
The top panel in Figure 1 shows the ensemble probability density func-
tion, ρ(υ, t), calculated by integrating the differential equation of an EDM,
Equation 1. The bottom panel shows and approximation to this pdf ob-
tained from the histogram of voltages of a direct simulation of a population
of 9,000 IF neurons. Note the large similarity of the pdfs at all time points.
The black line in Figure 2 shows the average firing rate per neuron,
r(t), calculated using the EDM, Equation 2. The grey line shows the
average firing rate per neuron calculated from a direct simulation of a
population of 9,000 IF neurons, as the proportion of cells with voltage at
threshold. Note the almost perfect match between these firing rates.
Figure 3 is as Figure 2 but for a step input current that jumps from 0 to
800 impulses per second at time t = 0. Note that by 0.4 seconds after the
step in the input current the average firing rate per neuron has reached
a new steady state around 10 impulses per second, as revealed by the
EDM (black line in Figure 3) and by the direct simulation of a population
of 9,000 IF neurons (grey line, in Figure 3)). Figure 4 shows the pdf
of voltages at this new steady state calculated by the EDM, Equation 2
(black line in Figure 4) and approximated using the histogram of voltages
at 0.4 ms of a direct simulation of 9,000 IF neurons (grey line in Figure 4).
2.1.2 One Population with Feedback
The previous Figures showed results from the simulation of a single pop-
ulation of independent neurons. Figure 5 is as Figure 2, but shows the
average firing rate per neuron in a population where each neuron receives
excitatory inputs from ten other neurons in the population (i.e., G=10 and
f=1 in Equation 3). For comparison, the dashed line shows the average
firing rate per neuron from the population without feedback.
Figure 6 demonstrates the effect of inhibitory feedback in the popu-
lation of Figure 5 by changing the fraction of inhibitory input neurons
from 0% to 80% (by changing f=1-0 to f=1-0.8, and still using G=10 in
Equation 3).
2.2 A Network of Populations
The previous sections evaluated EDMs in a single population of neurons.
Here we evaluate EDMs of excitatory and inhibitory populations combined
in the network of populations shown in Figure 7. The network is driven
by an excitatory input to the excitatory population. This population has
excitatory feedback (i.e., G=5 and f=1 in Equation 3), and its average
firing rate per neuron output, scaled by a constant Wei = 50, drives the
inhibitory population. The inhibitory population has inhibitory feedback
(i.e., G=5 and f=0 in Equation 3), and its average firing rate per neuron
output , scaled by a constant Wie = 15, modulates the activity of the
excitatory population.
Figures 8 and 9 show the activity in the excitatory and inhibitory
populations, respectively in the network of Figure 7. The upper panels in
6
Figure 1: Ensemble pdf, ρ(υ, t), for a population of IF neurons in response
to a sinusoidal input, calculated by integrating the differential equation of an
EDM (Equation 1, top panel) and approximated by direct simulation of a pop-
ulation of 9,000 IF neurons (bottom panel). Neurons in this populations were
independent of each other (i.e., they had no feedback; G=0 in Equation 3).
7
EDM
Direct simulation
40
35
30
25
20
15
10
5
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
0
0.70
0.75
0.80
Time (sec)
0.85
0.90
Figure 2: Firing rate of a population of neurons, in response to a sinusoidal
input, calculated by direct simulation of a population of 9,000 IF neurons (grey
line) and by integrating a population equation (black line, Equation 3). Neurons
in this populations were independent of each other (i.e., they had no feedback;
G=0 in Equation 3).
8
EDM
Direct simulation
20
15
10
5
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
0
0.00
0.05
0.10
0.15
0.20
Time (sec)
0.25
0.30
0.35
0.40
Figure 3: Firing rate of a population of 9,000 IF neurons, in response to a step
input at time zero. Same format as in Figure 2
9
EDM
Direct simulation
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
4
.
0
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 4: Ensemble pdf, ρ(υ, t) at 0.4 seconds in the neurons of the population
of Figure 3 in response of a step input at time zero. This pdf was calculated
by integrating the differential equation of an EDM (Equation 1, black line) and
approximated using the histogram of the voltages of the simulated neurons (grey
line).
10
EDM
Direct simulation
120
100
80
60
40
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
0
0.70
0.75
0.80
0.85
Time (sec)
0.90
0.95
1.00
Figure 5: Firing rate of a population of 9,000 neurons, in response to a sinusoidal
input, as in Figure 2, but each neuron in this population is connected with ten
presynaptic neurons, and all of these neurons are excitatory (i.e., G=10 and
f=1.0 in Equation 3).
11
EDM
Direct simulation
Direct simulation (no inhibition)
120
100
80
60
40
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
0
0.70
0.75
0.80
0.85
Time (sec)
0.90
0.95
1.00
Figure 6: Same as Figure 5 but for a population of neurons where 80% of the
presynaptic neurons to a given neuron are inhibitory (i.e., G=10 and f=1-0.8 in
Equation 3). For comparison, the dashed line shows the average firing rate per
neuron in the population with feedback but not inhibition of Figure 4.
12
Figure 7: Simulated network of two populations of IF neurons. A sinusoidal
input, σ0(t), was applied to the excitatory population. This population con-
tained excitatory feedback (each neuron in the population received excitatory
input from five other neurons in the population; G=5, f=1 in Equation 3).
The average firing rate per neuron of the excitatory population, scaled by the
coefficient Wei = 50, was the excitatory input for the inhibitory population.
This population contained inhibitory feedback (each neuron in the population
received inhibitory input from five other neurons in the population; G=5, f=0
in Equation 3). The average firing rate per neuron of the inhibitory population,
scaled by a coefficient Wie = 15, was the inhibitory input for the excitatory
population. Excitatory/inhibitory connections are shown by solid/dashed lines.
13
these figures show the population activities computed by the EDMs and
the bottom panels the activity derived from the direct simulation of 9,000
IF neurons. The blue curves, scaled along the left axis, show the average
spike rate per neuron in the population, the magenta and yellow curves,
scaled along the right axis, show the excitatory and inhibitory external
inputs to the population, respectively, and the magenta and yellow dashed
curves, also scaled along the right axis, show the excitatory and inhibitory
feedback inputs to the population. We should have performed the direct
simulations with a larger number of IF neurons to obtain smoother spike
rates and currents. Nevertheless, the activities computed by the EDMs
are in close agreement with those derived from the direct simulation.
2.3 Partial Conclusions
Given a dynamical model of a neuron, we now know how to derive an EDM
for a population of such neurons. For an IF model of a neuron, here we
have shown that EDMs accurately approximate population activity (i.e.,
the pdf of the trans-membrane voltage, ρ(υ, t), and the average firing
rate per neuron, r(t)). The next step in this sub project is to estimate
connectivity parameters (e.g., Wie and Wei in Figure 7) from simulated
data.
14
25
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
15
10
5
0
0.6
25
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
15
10
5
0
0.6
1400
1200
1000
)
s
p
i
(
t
n
e
r
r
u
C
800
600
400
200
0.7
0.8
0.9
Time (s)
1.0
1.1
0
1.2
1400
1200
1000
800
600
)
s
p
i
(
t
n
e
r
r
u
C
400
200
0.7
0.8
0.9
Time (s)
1.0
1.1
0
1.2
Figure 8: The top and bottom panels show the activities of the excitatory
population represented by the population equation, and by the simulation of
9,000 IF neurons. The blue line, scaled on the left axis, plots the firing rate of the
population. The magenta and yellow lines represent excitatory and inhibitory
currents, respectively; and the solid and dashed lines represent external and
feedback currents, respectively. The currents are scaled on the right axis. Note
that the similarity between the average firing rate per neuron obtained from the
population equation and from the direct simulation (blue lines in the top and
bottom panels).
15
25
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
15
10
5
0
0.6
25
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
15
10
5
0
0.6
1400
1200
1000
800
600
)
s
p
i
(
t
n
e
r
r
u
C
400
200
0.7
0.8
0.9
Time (s)
1.0
1.1
0
1.2
1400
1200
1000
800
600
)
s
p
i
(
t
n
e
r
r
u
C
400
200
0.7
0.8
0.9
Time (s)
1.0
1.1
0
1.2
Figure 9: Same as Figure 8 but for the inhibitory population in the two-
populations model.
16
3 Reducing dimensionality in EDMs
In the previous section we showed that Ensemble Density Models (EDMs)
accurately approximated the average firing rate per neuron and the proba-
bility density function (pdf) of direct simulations of ensembles of integrate-
and-fire (IF) neurons. In networks of EDMs, we want to estimate con-
nectivity parameters and state variables (i.e., the pdfs of the different
ensembles) from recorded ensemble firing rates. The state space of each
previously reported EDM contained 210 variables. To make the estimation
of parameters and state variables in networks of EDMs feasible/efficient it
would be helpful to find low-dimensional approximations of EDMs. Here
we report the approximation power of one such low-dimensional approxi-
mation method. This method was inspired by the moving basis technique
in Knight [2000].
3.1 Method to find low-dimensional approxima-
tions of EDMs
The evolution of the ensemble pdf, ρ(υ, t), is given by:
ρ(υ, t) = Q(s(t))ρ(υ, t)
(4)
where Q(s(t)) is a differential operator that depends on the stimulus s(t).
The normalized voltage υ in Equation 4 ranges in the unit interval. To
numerically solve this equation, we discretize υ, {υi = i/N : 1 ≤ i ≤ N }
and ρ, {ρi(t) = ρ(υi − ∆/2, t) : 1 ≤ i ≤ N }, giving a discretization of
Equation 4:
ρ(t) = Q(s(t))ρ(t)
(5)
where Q(s(t)) ∈ RN ×N is the matrix representation of the differential op-
erator Q(s(t). Equation 5 is a system of N differential equations. The
objective of the dimensionality reduction method described below is to
approximate the evolution of ρ(t) using a system of M differential equa-
tions, where M ≪ N . For this, the ensemble pdf ρ(t) is represented in a
basis of eigenvectors of the differential matrix Q(s(t), where many coef-
ficients in the new representation can be discarded without much loss in
approximation power.
3.1.1 Representing the ensemble pdf in a new basis
Let {φn(s(t)) : 0 ≤ n < N } and {λn(s(t)) : 0 ≤ n < N } be the eigenvec-
tors and eigenvalues, respectively, of Q(s(t)):
Q(s(t))φn(s(t)) = λn(s(t))φn(s(t))
or in matrix notation:
Q(s(t))Φ(s(t)) = Φ(s(t))Λ(s(t))
17
(6)
(7)
where φn(s(t)) is the nth column of the matrix Φ(s(t)) and λn(s(t)) is
the nth diagonal element of the diagonal matrix Λ(s(t)).
Assuming the eigenvectors are linearly independent, we represent ρ(t)
as:
ρ(t) =
N
Xn=1
an(t)φn(s(t))
or in matrix notation:
ρ(t) = Φ(s(t)) a(t)
From Equations 9 and 7 it follows:
(8)
(9)
Q(s(t))ρ(t) = Q(s(t))Φ(s(t)) a(t) = Φ(s(t))Λ(s(t)) a(t)
(10)
Using a backward difference to approximate the time derivative of ρ(t):
ρ(t) =
ρ(t) − ρ(t − ∆t)
∆t
in Equation 5 we obtain:
ρ(t) − ∆t Q(s(t))ρ(t) = ρ(t − ∆t)
Now applying Equations 9 and 10 to Equation 12 we get:
(11)
(12)
[Φ(s(t)) (I − ∆tΛ(s(t)))] a(t) = Φ(s(t − ∆t)a(t − ∆t)
(13)
or:
a(t) = (cid:2)(I − ∆tΛ(s(t)))−1 Φ(s(t))−1Φ(s(t − ∆t)(cid:3) a(t − ∆t)
(14)
Equation 14 provides the evolution of the coefficients a(t) in Equa-
tion 9. It just expresses the evolution of the ensemble pdf in Equation 5
in another basis. It is an exact formula (i.e., it is not an approximation)
and has the same dimensionality as Equation 5. The importance of Equa-
tion 14 is that one can approximate the evolution of the ensemble pdf by
discarding many components of a(t), as we explain next.
3.1.2 Reducing dimensionality in the new basis
To understand why one can discard many coefficients in the representation
of the ensemble pdf of Equation 9 without much loss in approximation
power, consider the case where the stimulus, s(t), is constant. In such a
case, the eigenvectors and eigenvalues will neither depend on the stimulus;
i.e., Φ(s(t)) = Φ and Λ(s(t)) = Λ. Then Equations 5 and 9 reduce to:
18
and:
ρ(t) = Qρ(t)
ρ(t) = Φ a(t)
Taking derivatives in Equation 16 we obtain:
ρ(t) = Φ a(t)
(15)
(16)
(17)
and substituing Equation 16 in Equation 15, and applying Equation 7, we
get:
ρ(t) = QΦ a(t) = ΦΛ a(t)
(18)
Equating the right hand sides in Equations 17 and 18, and pre multi-
plying by Φ−1, gives:
or:
with solution:
a(t) = Λ a(t)
ai(t) = λi ai(t)
ai(t) = exp(λit) ai(0)
(19)
(20)
(21)
Thus, the absolute value of a low dimensional coefficient ai evolves as:
ai(t) = exp(ℜ(λi)t) ai(0)
(22)
Because all eigenvalues have negative real part (with the exception of
the zero eigenvalue), the coefficients associated with non-zero eigenvalues
will decay to zero, and the speed of this decay will be proportional to the
absolute value of the real part of the corresponding eigenvalue. Therefore,
to achieve dimensionality reduction in Equation 14 we may discard those
coefficients associated with eigenvalues with larger absolute value of their
real part, since these coefficients will rapidly decay to zero.
19
3.2 Evaluation of the method
We first study how well low-dimensional EDMs approximate the average
firing rate per neuron in direct stimulations (Section 3.2.1) and then how
they approximate the ensemble pdf (Section 3.2.2).
These studies were peformed in data simulated from the network of
EDMs illustrated in Figure 7. The network was driven by an excitatory
sinusoidal input to the excitatory population (shown by the dotted curve
and scaled along the right axis in the top panel of Figure 10). The average
firing per neuron in this population, scaled by a constant Wei = 50, drove
the inhibitory population (shown by the dotted curve and scaled along the
right axis in the top panel of Figure 11). In turn, the average firing rate
per neuron in the inhibitory population, scaled by a constant Wie = 15,
inhibited the excitatory population. Both populations contained feedback
(i.e., each neuron received 10 inputs from ten other cells in the same
population and 80% of these inputs were inhibitory).
3.2.1 Firing Rates
The top panels of Figures 10 and 11 show the average firing rate per neu-
ron obtained from direct simulation (grey curve), from a full-dimensional
EDM (red curve), and from low-dimensional EDMs (the blue, cyan, and
red curves correspond to 17, 5, and 1 moving basis, respectively). The
full dimensional EDM and its low dimensinal approximations with 17 and
5 moving basis almost perfectly approximate the average firing rate per
neuron of the direct simulation. The approximation power of the EDM
with only one moving basis is not as good, but it look reasonable.
3.2.2 Ensemble Proability Density Functions
We compared the normalized histogram of number of directly simulated
neuron per voltage bin (i.e., the ensemble pdf from direct simulation) with
the pdfs calculated with EDMs (i.e., Equation 5). For this we computed at
every time step the Kullback-Leibler (KL) divergence (in bits) between the
pdfs obtained by direct stimulation and those obtained from EDMs. These
KL divergences are shown in the bottom panels of Figures 10 and 11.
We see that the pdfs obtained from EDMs were good approximation of
the pdfs from the direct simulation at times of large average firing rate per
neuron (between 0.73 and 0.85 seconds and between 1.03 and 1.2 seconds
in the bottom panel of Figure 10, and between 0.72 and 0.82 seconds and
between 1.03 and 1.18 seconds in the bottom panel of Figure 11).
At times of low averaged firing rate per neuron the difference between
the pdfs obtained from direct simulation and those obtained from EDMs
were an order of magnitude larger for the inhibitory than for the excitatory
ensemble. This is probably because the excitatory ensemble was driven
by a large and smooth sinusoidal input while the inhibitory ensemble
was driven by the weaker and non-smooth average firing rate per neuron
of the excitatory ensemble. Also, at most time points, the larger the
number of moving basis in low-dimensional EDMs, the better the EDM
pdf approximated the pdf obtained from direct simulation, as can be seen
more clearly in the bottom panel of Figure 11.
20
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
30
25
20
15
10
5
0
0.6
)
M
D
E
.
i
m
S
t
c
e
r
i
D
(
L
K
0.45
0.40
0.35
0.30
0.25
0.20
0.15
0.10
0.05
0.00
0.6
Direct Sim.
EDM (full)
(05 mb)
(03 mb)
(01 mb)
Exc. Input
0.7
0.8
0.9
1.0
1.1
1200
1100
1000
900
800
700
600
500
400
)
s
p
i
(
t
n
e
r
r
u
C
0.7
0.8
0.9
Time (sec)
1.0
1.1
Figure 10: Average firing rate per neuron (top) and KL divergence between
the ensemble pdf calculated by direct simulation and that caculated by EDMs
(bottom) for the excitatory ensemble.
21
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
)
M
D
E
.
i
m
S
t
c
e
r
i
D
(
L
K
30
25
20
15
10
5
0
0.6
35
30
25
20
15
10
5
0
0.6
Direct Sim.
EDM (full)
(05 mb)
(03 mb)
(01 mb)
Exc. Input
0.7
0.8
0.9
1.0
1.1
1000
)
s
p
i
(
t
n
e
r
r
u
C
800
600
400
200
0
0.7
0.8
0.9
Time (sec)
1.0
1.1
Figure 11: Average firing rate per neuron (top) and KL divergence between
the ensemble pdf calculated by direct simulation and that caculated by EDMs
(bottom) for the inhibitory ensemble.
22
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
1
8
.
0
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 12: Ensemble pdfs for the inhibitory ensemble at 0.81s.
To try to understand why the pdfs obtained from direct simulation
were different from those obtained from low-dimensional EDMs, we plot-
ted these pdfs for the inhibitory ensemble in an interval of low average
firing rate per neuron (from 0.81 seconds in Figure 12 to 1.09 seconds in
Figure 26).
In the transition between weak to zero average firing rate
per neuron (from 0.81 seconds in Figure 12 to 0.93 seconds in Figure 18)
the low-dimensional pdfs moved faster towards lowers voltage than the
pdfs from the full-dimensional EDM and those from direct simulation.
Similarly, in the transition between zero to weak average firing rate per
neuron (from 0.95 seconds in Figure 19 to 1.09 seconds in Figure 26) the
low-dimensional pdfs moved faster towards higher voltages than the pdfs
from the full-dimensional EDM and those from direct simulation. This
suggests that the moving basis discarded in the low-dimensional approxi-
mations of EDMs help prevent the EDM pdf to transition too fast to and
away from the pdf corresponding to large average firing rate per neuron.
3.3 Partial conclusions
We have developed a method to reduce the dimensionality of the state
space of EDMs. We observed that the pdfs from low-dimensional EDMs
23
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
3
8
.
0
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 13: Ensemble pdfs for the inhibitory ensemble at 0.83s.
24
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
5
4
3
2
1
0
)
s
5
8
.
0
,
υ
(
ρ
−1
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 14: Ensemble pdfs for the inhibitory ensemble 0.85s.
25
)
s
7
8
.
0
,
υ
(
ρ
10
8
6
4
2
0
−2
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 15: Ensemble pdfs for the inhibitory ensemble 0.87s.
26
)
s
9
8
.
0
,
υ
(
ρ
20
15
10
5
0
−5
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 16: Ensemble pdfs for the inhibitory ensemble 0.89s.
27
)
s
1
9
.
0
,
υ
(
ρ
250
200
150
100
50
0
−50
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 17: Ensemble pdfs for the inhibitory ensemble 0.91s.
28
)
s
3
9
.
0
,
υ
(
ρ
250
200
150
100
50
0
−50
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 18: Ensemble pdfs for the inhibitory ensemble 0.93s.
29
)
s
5
9
.
0
,
υ
(
ρ
200
150
100
50
0
−50
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 19: Ensemble pdfs for the inhibitory ensemble 0.95s.
30
)
s
7
9
.
0
,
υ
(
ρ
100
80
60
40
20
0
−20
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 20: Ensemble pdfs for the inhibitory ensemble 0.97s.
31
)
s
9
9
.
0
,
υ
(
ρ
12
10
8
6
4
2
0
−2
0.0
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
0.2
0.4
υ
0.6
0.8
1.0
Figure 21: Ensemble pdfs for the inhibitory ensemble 0.99s.
32
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
6
5
4
3
2
1
0
)
s
1
0
.
1
,
υ
(
ρ
−1
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 22: Ensemble pdfs for the inhibitory ensemble 1.01s.
33
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
3
0
.
1
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 23: Ensemble pdfs for the inhibitory ensemble 1.03s.
34
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
5
0
.
1
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 24: Ensemble pdfs for the inhibitory ensemble 1.05s.
35
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
7
0
.
1
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 25: Ensemble pdfs for the inhibitory ensemble 1.07s.
36
Direct Sim.
EDM (full)
(17 mb)
(05 mb)
(01 mb)
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
)
s
9
0
.
1
,
υ
(
ρ
−0.5
0.0
0.2
0.4
υ
0.6
0.8
1.0
Figure 26: Ensemble pdfs for the inhibitory ensemble 1.09s.
37
move faster to and away from the high-average-firing-rate-per-neuron pdf
than the pdf derived from direct simulation or that from full EDMS.
However, these differences did not have a large impact on the average
firing rate per neuron produced by low-dimensional EDMs, which almost
perfectly approximates the average firing rate per neuron computed by
direct simulation.
Our next step is to estimate connectivity parameters and state space
variables in networks of EDMs from recorded spike rates. We will first do
this with simulated data.
38
Figure 27: Network of EDMs.
4 Estimating parameters of EDMs
Here we evaluate a maximum likelihood method to estimate the connec-
tivity parameters θ = [Wei, Wie] in the EDMs network in Figure 27. We
seek the connectivity parameters for which a set of N pairs measurements,
YN = [y(0), . . . , y(N − 1)], are most probable; ie:
θml = arg max
P (YN θ)
θ
The pair of measurements at time n, y(n) comprises measurements from
the excitatory and inhibitory ensembles; i.e., y(n) = [ye(n), yi(n)].
4.1 Noise Model
As a first approximation we make the following three assumptions on the
noise of the measurements. These are the assumptions made in previous
work by our collaborators [Kostuk et al., 2012].
1. Noise is independent in time; i.e., P (YN θ) = ΠN
2. Noise is independent in the different populations; i.e., P (y(n)θ) =
n=1P (y(n)θ).
P (ye(n)θ)P (yi(n)θ).
3. Gaussian noise with a known variance, σ2, in each population; i.e.,
P (y.(n)θ) = N (y.(n)r.(nθ), σ2), where r.(nθ) is the activity gen-
erated by the ensemble at time n.
With these assumptions the log likelihood function of the model pa-
rameters reduces to:
log P (YN θ) = K − ΣN
n=0
(ye(n) − re(nθ))2 + (ye(n) − re(nθ))2
2σ2
The red lines in Figure 28 plots the activity generated by the excita-
tory and inhibitory ensemble with connectivity parameters Wei = 50 and
Wie = 15. The blue lines plot the noisy measurements (σ = 2) that we
use below for parameter estimation.
39
Excitatory Ensemble
35
30
25
20
15
10
5
0
−5
)
s
p
i
(
e
t
a
R
g
n
i
r
i
F
Inhibitory Ensemble
Mes rements
EDM spike rates
30
25
20
15
10
5
0
−5
−10
0.0
0.1
0.2
Time (s)
0.3
−10
0.0
0.1
0.2
Time (s)
0.3
Figure 28: EDMs spike rates (red) and noisy measurements (blue, σ = 2) used
to estimate the connectivity parameters in the network of Figure 27.
40
4.2 Optimization Surface
The levelplot in Figure 29 shows the optimization surface for the set of
parameters show in the axes. We see that it has a convex shape with peak
at the true parameters. Thus, an iterative gradient-ascent optimization
procedure should climb to the maximum-likelihood parameters from any
starting set of parameters.
4.3 Gradient of Log-Likelihood Function
To compute the gradient of the log-likelihood function at a given set of
parameters θ = [wei, wie] one needs to integrate the EDMs to generate
at each time step the activities of both ensembles, re(n, θ) and ri(n, θ),
as well as the ensembles pdfs, ρe(nθ) and ρi(nθ). With these quanti-
ties in hand, the gradient of the log-likelihood function can be computed
recursively in a second integration step (Equation 23).
∂ log P (Ymθ)
∂Wei
∂re[n; θ]
∂Wei
∂ρe[nθ]
∂Wei
∂Qe[n; θ]
∂Wei
(ye[n] − re[n; θ]) ∂re[n;θ]
∂Wei
= Σm
n=0
+ (yi[n] − ri[n; θ]) ∂ri[n;θ]
∂Wei
σ2
= ∆v σE
e [n] (qr,
∂ρe[nθ]
∂Wei
)
∂Qe[n − 1; θ]
= ∆t
ρe[n − 1θ] + [I + ∆t Qe[n − 1; θ]]
= −Wei
∂Wei
∂ri[n − 1; θ]
∂Wei
A(2)
∂ρe[n − 1θ]
∂Wei
(23)
The white arrows in Figure 29 point in the direction of the gradi-
ent, and are scaled according to the gradient magnitude. Note that, as
expected, the gradient direction is perpendicular to the level lines.
With a convex log-likelihood function, for which we can compute its
gradient, we can now use an iterative gradient ascent procedure to max-
imize this function. The green line in Figure 29 shows a gradient-ascent
trajectory that in only three steps accurately approximated the maximum
likelihood parameters.
41
65
60
55
i
e
W
50
45
40
35
0
σ=2.00
0
−120000
−240000
−360000
−480000
−600000
d
o
o
h
i
l
i
e
k
L
g
o
L
−720000
−840000
−960000
5
10
15
Wie
20
25
30
Figure 29: Log-likelihood function, its gradient, and a sample gradient ascent
path. The contour plot plots the log-likelihood function for the parameter values
shown on the axes for the network in Figure 27. The white arrows show the
log-likelihood gradient computed analytically. The green curve is a gradient
ascent trajectory starting at Wei = 37 and Wie = 3.
42
5 Modeling evoked auditory activity in
rodents with EDMs
Here we show that the average firing-rate per neuron predicted by EDMs
well approximates the evoked high-gamma power (HGP) from auditory
neurons in response to stimulation with pure-tones.
5.1 Recordings
We used very high-resolution simultaneous surface and laminar recordings
(Figure 30) in anesthetized rodents stimulated with pure tones at different
frequencies and amplitudes. The blue trace in Figure 31 shows the high-
gamma power (HGP) evoked by the presentation of tones in one sample
electrode of the array. Each vertical dotted color line marks the onset of
a tone of a corresponding frequency.
We see that these evoked waveforms are very stereotypical. Some
waveforms have a large peak followed by a bump. Other waveforms have
two peaks, with a smaller peak preceding a larger one. We also see inf
Figure 32 waveforms that have more than one peak preceding a larger one
and waveforms having a larger peak followed by a smaller one.
5.2 Qualitative analysis
The recorded HGP waveforms appear remarkably similar to those pro-
duced by EDMs when stimulated by sinusoids. To confirm this similar-
ity we manually chose parameters for EDMs to reproduce the shape of
recorded waveforms (Figures 33, 34, 35, and 36).
5.3 Quantitative analysis
Using a similar method as described above to learn connectivity param-
eters in networks of EDMs, we etimated parameters so that an EDM
approximated as close as possible the HGP evoked by the presentation of
pure tones to rodents. In total we estimated 13 parameters: three param-
eters of an EDM, two parameters for its initial condition, four parameters
for the sinusoidal excitatory input, and four parameters for the inhibitory
input. The recorded and approximated HGP are given by the blue and
red solid curves, respectively, in Figure 37.
5.4 Partial Conclusions
These preliminary results show that EDMs, besides accurately approxi-
mating the average firing rate of ensembles of simulated IF neurons, well
approximate the HGP in the auditory cortex of rats evoked by auditory
tones.
43
Figure 30: Electrodes used for simultaneous surface and laminar recordings.
Surface recordings were obtained from an 8 × 8 grid of subdurally implanted
electrodes covering an area of 1.6 mm2. Laminar recordings were obtained from
a 32-channel politrodes of length 650 µm. These rercordings were combined
with optogenetic manipulations.
44
801
807
809
813
817
822
828
850
851
737 Hz
853
1743 Hz
4121 Hz
9743 Hz
23035 Hz
0.000040
0.000035
0.000030
0.000025
0.000020
0.000015
0.000010
0.000005
r
e
w
o
P
a
m
m
a
G
-
h
g
H
i
0
500
1000
1500
2000
Figure 31: High-gamma power evoked by the presentation of pure tones at
different frequencies and amplitudes. Blue traces show the HGP evoked by the
presentation of tones in one sample electrode of the array. Each vertical dotted
color line marks the onset of a tone of a corresponding frequency.
45
853
855
859
860
873
875
876
877
884
882
737 Hz
1743 Hz
4121 Hz
9743 Hz
23035 Hz
0.000040
0.000035
0.000030
0.000025
0.000020
0.000015
0.000010
0.000005
r
e
w
o
P
a
m
m
a
G
-
h
g
H
i
2000
2500
3000
3500
4000
Figure 32: More examples of HGP evoked by the presentation of pure tones.
The format is as in Figure 31.
46
Figure 33: The blue curve in the inset shows the average firing-rate simulated
by an EDM with manually chosen parameters. The magenta and yellow solid
curves plot its excitatory and inhibitory inputs, respectively. The first waveform
is qualitatively similar to the ECoG waveforms with a bump following a large
peak.
47
Figure 34: The blue curve in the inset shows the average firing-rate simulated by
an EDM with manually chosen parameters. The magenta and yellow solid curves
plot its excitatory and inhibitory inputs, respectively. The second waveform is
qualitatively similar to the ECoG waveforms with a lower peak preceding a large
one.
48
Figure 35: The blue curve in the inset shows the average firing-rate simulated
by an EDM with manually chosen parameters. The magenta and yellow solid
curves plot its excitatory and inhibitory inputs, respectively. EDMs can generate
waveforms with multiple lower peaks preceding a larger one.
49
Figure 36: The blue curve in the inset shows the average firing-rate simulated
by an EDM with manually chosen parameters. The magenta and yellow solid
curves plot its excitatory and inhibitory inputs, respectively. EDMs can generate
waveforms with a larger peak preceding a smaller one.
50
[ 10.10648747 21.14308782 0.80066418 12.04027179 11.95724926
722.23861385 279.88373368 4.54564969 -1.16267511 634.38337831
327.55274887 6.58685284 -2.53546525]
Data
Estimate
eInput
iInput
0.05
0.10
Time (s c)
0.15
1100
1000
900
800
700
600
500
400
300
0.20
)
s
p
i
(
e
t
a
R
e
k
p
S
t
u
p
n
i
I
r
e
w
o
P
a
m
m
a
G
i
-
h
g
H
d
e
a
c
S
l
y
t
i
s
n
e
D
y
t
i
l
i
b
a
b
o
r
P
30
25
20
15
10
5
0
−5
0.00
4.0
3.5
3.0
2.5
2.0
1.5
1.0
0.5
0.0
0.0
0.2
0.4
0.6
Normalized Votage
0.8
1.0
Figure 37: Learning ensemble properties to approximate physiological record-
ings. We estimated 13 parameters so that the average firing rate per neuron
predicted by an EDM approximates as close as possible the recorded HGP in
response of a pure tone. The title shows the values of these parameters. The
blue and red curves in the top panel plot the recorded HGP and the predicted
average firing rate per neuron, respectively. The dotted green and magenta
curves in the top panel shows the estimated excitatory and inhibitory inputs to
the EDM. The curve in the bottom panel plot the estimated initial condition
probability density function of the EDM. The average firing rate predicted by
the EDM is a good approximation of the recorded HGP.
51
6 Conclusions
We have shown that an ensemble model accurately reproduce the probabil-
ity density function of the transmembrane voltage, as well as the average-
firing rate per neurona, in a large ensemble of integrate-and-fire simulated
neurons (Section 2). We developed and evaluated methods to reduce the
dimensionality (Section 3) and estimate parameters (Section 4). Finally
we demonstrated the feasibility of EDMs to model the high-gamma power
evoked by pure tones in the auditory cortex of rodents.
The possiblity of quantitatively model the activity of ensemble of neu-
rons may allow us to uncover fundamental computations performed by
neural ensembles.
References
B.W. Knight. Dynamics of encoding in neuron populations: some general
mathematical features. Neural Computation, 12:473 -- 518, 2000.
M. Kostuk, B.A. Toth, C.D. Meliza, D. Margoliash, and H.D.I. Abarbanel.
Dynamical estimation of neuron and network properties II: path inte-
gral monte carlo methods. Biological Cybernetics, 106:155 -- 167, 2012.
A. Omurtag, B.W. Knight, and L. Sirovich. On the simulation of large
populations of neurons. Journal of Computational Neuroscience, 8:51 --
63, 2000.
52
|
1907.10879 | 1 | 1907 | 2019-07-25T07:46:22 | Synaptic Time-Dependent Plasticity Leads to Efficient Coding of Predictions | [
"q-bio.NC"
] | Latency reduction of postsynaptic spikes is a well-known effect of Synaptic Time-Dependent Plasticity. We expand this notion for long postsynaptic spike trains, showing that, for a fixed input spike train, STDP reduces the number of postsynaptic spikes and concentrates the remaining ones. Then we study the consequences of this phenomena in terms of coding, finding that this mechanism improves the neural code by increasing the signal-to-noise ratio and lowering the metabolic costs of frequent stimuli. Finally, we illustrate that the reduction of postsynaptic latencies can lead to the emergence of predictions. | q-bio.NC | q-bio |
Synaptic Time-Dependent Plasticity Leads to
Efficient Coding of Predictions
Pau Vilimelis Aceituno1,2, Masud Ehsani1, and Jurgen Jost1,3
1Max Planck Institute for Mathematics in the Sciences, Inselstraβe 22,
04103 Leipzig, Germany
2Max Planck School of Cognition, Stephanstraβe 1a, 04103 Leipzig
3Santa Fe Institute, 1399 Hyde Park Road, 87501 Santa Fe, New
Mexico, USA
Abstract
Latency reduction of postsynaptic spikes is a well-known effect of Synaptic
Time-Dependent Plasticity. We expand this notion for long postsynaptic spike
trains, showing that, for a fixed input spike train, STDP reduces the number of
postsynaptic spikes and concentrates the remaining ones. Then we study the con-
sequences of this phenomena in terms of coding, finding that this mechanism im-
proves the neural code by increasing the signal-to-noise ratio and lowering the
metabolic costs of frequent stimuli. Finally, we illustrate that the reduction of
postsynaptic latencies can lead to the emergence of predictions.
1
Introduction
Living organisms need to make accurate predictions in order to survive [BVCS10,
Hoh13], posing the question of how do brains learn to make those predictions. Early
general models based on classical conditioning [RW+72, MBG95], as well as mech-
anistic models explaining the neural substrate for those predictions [SDM97, Hee17]
assume that prediction performance is feed back into the predicting neural popula-
tion, similarly to supervised or reinforcement learning paradigms that are common in
machine learning. However, recent studies have found that sensory neurons without
feedback from higher brain areas encode predictive information [PMBB15], a finding
that has been supported by simulation studies [SMP18]. This implies that a bottom-
up process without explicit feedback -- similar to unsupervised learning -- should also
generate predictions.
In this paper we present such a mechanism by focusing on postsynaptic latency
reduction. This is a well-known effect of Synaptic Time-Dependent Plasticity (STDP)
first mentioned by Song et. al [SMA00] for a single post-synaptic neuron driven by
an specific excitatory input pattern. This effect was explored in detail in a simulation
1
study by Guyonneau et. al [GVT05] who showed that this effect is robust to noise
in the form of jitter and Poissonian spontaneous activity. They further analyze STDP
on a single neuron receiving fixed (among trials) Poissonian spike trains from each
pre-synaptic neuron and showed that by STDP weights of the earliest afferents will
be increased, regardless of the firing rate and level of synchrony of the corresponding
neurons. Masquelier et. al [MGT08] showed how a single post-synaptic neuron by
STDP would learn a single frequent excitatory pattern of spikes even when there is
a strong background noise presented and associate this with how postsynaptic firing
latency in response to pattern presentation would decrease. In this article we revisit this
phenomenon at the micro level with plastic inhibitory neurons added to the previous
setups and analyze the effect of latency reduction at network level.
The gist of our argument is that latency reduction implies that neurons fire as early
as possible for a given input spike train that is repeated very often; as neurons do not
differentiate between a specific stimulus and an early clue of such a stimulus -- both be-
ing part of seemingly the same input spike train -- the neurons can, by STDP, fire earlier
than the stimulus itself. Furthermore, we expand on the previous studies focused on
excitatory neurons to include inhibition and illustrate the parameter regime in which
inhibitory plasticity is compatible with latency reduction. However, the latency reduc-
tion mechanism has other uses in terms of neural code. First, as neurons fire as early
as possible when a stimulus is presented, their spikes will concentrate in a small time
window, and thus they are easier to decode. Second, we show that the latency reduction
can also lead to a reduction of the number of spikes, which translates as a reduction of
metabolic costs for encoding frequent stimuli.
We develop our argument by studying simple models of neurons subject to fixed
input spike trains. We use a combination of simulations and mathematical analysis to
derive our results, starting from the evolution of a single postsynaptic spike at very short
timescales we expand to larger scales that conclude in the emergence of predictions and
efficient code at the level of populations of neurons in large timescales.
The rest of this paper is organized as follows. First, we present the models of
neurons and STDP in Sec. 2. Second, study the effects of STDP in a single postsynaptic
spike in very small timescales ∼ 10ms, focusing on latency reduction and the reduction
of the number of postsynaptic spikes in Sec. 3. Third, we expand those results to
long postsynaptic spike trains in Sec. 4, finding that STDP forces postsynaptic neurons
to fire only once at the onset of the presynaptic spike train. Third, we provide an
interpretation of this concentration in terms of neural code performance, showing that
it leads to lower metabolic costs and lower the decoding errors in Sec. 5. We finalize by
illustrating that the same mechanism of latency reduction leads to encoding predictions
in Sec. 6.
2 Models
2.1 Leaky integrate-and-fire neuron
Neurons are cells that constitute the basic computational units in the nervous system.
Their main feature is the capacity to receive information through electrical impulses,
2
combine this information and send impulses to other neurons. In this paper we model
them as Integrate-and-Fire neurons with a refractory period [Lap07]. In this model, the
state of a neuron at a given time is described by its membrane potential v(t), which
evolves according to the equation
dv(t)
dt
= −(v(t) − v0) + i(t),
τm
(1)
where τm = 10ms, v0 = −70mV . i(t) is the input to the neuron at time t. When the
membrane potential reaches a certain threshold vth = −50mV , the neuron "fires" or
"spikes", meaning that it emits a pulse of current. After firing, the membrane potential
is reset to its resting state v0 and kept frozen at this value for a fixed period of time
called the refractory period tref = 4ms.
The firing of a neuron generates pulses of current that arrive at other neurons, which
in turn update their membrane potentials. If neuron a receives the spikes of neuron b we
will say that there is a synapse going from the second to the first. The receiving neuron
is called postsynaptic and the sending neuron is the presynaptic one. This synapse is
characterized by a weight wab and a delay dab which correspond, respectively, to the
gain and the latency that the pulse of neuron a goes through before arriving at b.
Input spike trains
2.2
Neurons communicate mainly through action potentials or spikes, which are typically
modeled as Diracs, hence the input to a neuron can be described as
(cid:88)
i(t) =
wnδ(t − tn),
(2)
n
where wn is the weight of the spike, which corresponds to the strength ot the synapse
from which the spike comes, and tn is the arrival time of the spike. The weights of
the synapses can be positive, if the presynaptic neuron is excitatory, or negative, if it
is inhibitory. Through this paper we will assume that every neuron gets an input that
will be repeated, meaning that a neuron will always get spikes from different synapses,
and although the weights of the synapses might change, the times tn of the spikes will
remain the same in every input repetition.
2.3 Synaptic Time-Dependent Plasticity
Networks of neurons learn by modifying the strength of the connections between them
-- previously denoted by w There is a rich literature on what rules those weights follow
in biological neurons and their respective implications [DA01]. For the purposes of
this paper, the biological neurons that we will analyze and simulate will adapt their
connections according to the Synaptic Time-Dependent Plasticity (STDP) paradigm
[SG10, GKvHW96].
In STDP the weight of a connection is modified depending on the time interval
between pairs of pre- and post-synaptic spikes. For every pair the weight of the synapse
3
is modified according to the equations
(cid:40)
∆w(∆t) =
A+(w)e
−A−(w)e
− ∆t
τs
− ∆t
τs
if ∆t ≥ 0
if ∆t < 0
(3)
(4)
where ∆t = tpost − tpre is the time difference between the postsynaptic spike and the
presynaptic one, τs = 20ms. Based on previous works [KH00, VRBT00], we define
A+ and A− as
A+(w) = η+(wmax − w),
A−(w) = η−(w − wmin)
where η− = 0.015, η+ = 0.01, we
follow the same rules as their excitatory counterparts but with parameters η− = 0.045, η+ =
0.03 and wi
with probability 0.8 or inhibitory with probability 0.2.
max = 20mV . Each synapse comes from a presynaptic excitatory neuron
max = 10mV and wmin = 0. Inhibitory synapses
2.4 Model limitations and required features
We must note that the models used here are heavy simplifications of real neurons. LIF
neurons do not exhibit the rich range of dynamics that real neurons can have [Izh04],
ion channel kinetics are more complicated than simple Diracs [CBC95] and the STDP
model used here cannot account for the evolution of synaptic weights when the fre-
quency of postsynaptic or presynaptic spikes is high [PG06]. However, those models
contain the main biologically realistic features that we need for the rest of this study.
First, the time constants of the neuron membrane potentials [GKNP14] and the STDP
interactions [BP98] are at least an order of magnitude smaller than the duration of the
input spike trains associated to biologically realistic stimuli-evoked spatiotemporal pat-
terns [RWP07, PVB+98]. Second, the sparse firing of the typical neuron [RBH+11].
Third, the potentiation of presynaptic spikes preceding postsynaptic ones, which is
consistent with most experiments [PG06]. Finally, the homeostatic consideration that
neurons should not widely increase their firing rate, which is a natural requirement
on metabolic grounds [TN04] can easily be incorporated by the depressive term A−.
Thus we will keep these well-known models [GKNP14] on the grounds that they are
analytically tractable and qualitatively plausible.
3 Evolution of a single postsynaptic spike
In this section we show that STDP can change individual postsynaptic spikes by reduc-
ing their latencies and their number. We will start by presenting simple scenarios with
excitatory inputs in which both effects are easy to illustrate, then show how inhibitory
synapses can be added to the model, and finally show that those effects can appear in
random input spike trains by presenting simulations. It is worth noticing that the time
windows in this section are on the order of τs and the number of repetitions of each
input pattern will be small.
4
3.1 Latency Reduction
If a fixed train of presynaptic spikes is repeated very often, then the spikes that ar-
rive before the postsynaptic spike get reinforced. This implies that the postsynaptic
spike might then be triggered earlier [SMA00, GKvHW96]. When this happens, the
refractory period of the postsynaptic neuron would prevent a second spike on the orig-
inal spiking site. However, when the postsynaptic spike happens earlier and earlier, it
might lead to a proliferation of spikes by having a new spike appear at the time of the
original postsynaptic spike. Following previous literature [SMA00,AN00,KGH01], to
prevent this effect, we assume that long term depression -- the weakening of synaptic
weights -- is stronger than long term potentiation -- the strengthening of postsynaptic
weights.
This is easy to understand in a simple scenario: Considering a very long, excitatory
presynaptic spike train which generates a single postsynaptic spike at some time t0.
The postsynaptic spike will advance through the spike train, and after some repetitions
it will be triggered one presynaptic spike earlier. After this advancement is repeated
many times, the postsynaptic spike is triggered at time t∞, very far (in time) from the
place where it was first triggered, so that
t∞ (cid:28) t0.
(5)
The membrane potential decays exponentially, meaning that the effect of the postsy-
τm ), which
naptic spike at time t∞ on the membrane potential are of order O(e
is negligible. Thus, the membrane potential at time t0 is now only dependent on the
presynaptic spikes that are close. If those presynaptic spikes have been left as they
where by the passage of the postsynaptic spike, then a new postsynaptic spike will be
generated at time t0. To not have this postsynaptic spike appear, it is therefore nec-
essary that the passage of the postsynaptic spike weakens the presynaptic ones. We
illustrate this point in Fig. 1 with the functions and parameters that we will use in
subsequent sessions.
− t0−t∞
Note that the argument that we give here is qualitative in nature, in the sense that
we simply state that LTD should dominate LTP through the constant η, but we have
not studied how to find that ratio. As this would depend on the exact parameters of
the regular spike train -- and thus would not be directly generalizable -- , we will simply
assume that the brain operates in a parameter regime in which spikes do not proliferate.
3.2 Late spike disappearance through synaptic noise
If latencies might be reduced, then two postsynaptic spikes that are triggered at distant
points in time might become close as time progresses. We must then ask what happens
to a pair of postsynaptic spikes that occur very close in time. In this section we show
that in the absence of synaptic noise the two spikes can coexist, but random modifica-
tions of the presynaptic weights -- induced, for instance, by other presynaptic inputs --
can lead to the disappearance of the second postsynaptic spike.
There are many possible scenarios that we might consider when we have pairs
of postsynaptic spikes in the same neuron: we must consider the time between the
5
Figure 1: Latency reduction and spike proliferation: We plot the membrane poten-
tial (left) and firing times (right) of a postsynaptic neuron that receives a constant train
of spikes with inter-spike interval of 3.5ms and strength 5.5mV , from time t = 0ms
to t = 150ms, and we add an extra spike at t = 150ms with potential 2mV . The
neuron generates a single postsynaptic spike at the original input presentation (Repeti-
tion 0). The upper plots reflect the case η+ = η−, while for the lower ones we picked
2 η+ = η−. After an initialization period, the postsynaptic spike moves forward in
3
time at a constant rate. As this happens, a single presynaptic spike will get reinforced
proportionally to the η+ and dampened proportionally to η−. If LTP is equal to LTD,
after the postsynaptic spike happens much earlier than before, the membrane potential
of the postsynaptic neuron will reach the threshold again. This second postsynaptic
spike would move forward in time at the same speed as the strengths of the spikes
are left unchanged by the compensation of LTD and LTP (upper plots). In the case
where η+ < η−, the depression compensates the potentiation, so there is no second
postsynaptic spike.
6
050100150200250time (ms)70656055504540Membrane Potential (mV)Single neuron membrane potentialRepetition 100Repetition 70Repetition 00102030405060708090Input Repetitons020406080100120140160Firing Time (ms)Single neuron firing time050100150200250time (ms)70656055504540Membrane Potential (mV)Single neuron membrane potentialRepetition 100Repetition 70Repetition 00102030405060708090Input Repetitons020406080100120140160Firing Time (ms)Single neuron firing time two spikes, the movements in time of both of them and the possibility of synaptic
noise. The case when two postsynaptic spikes happen originally very close in time
is extremely rare -- because postsynaptic spikes are sparse. The case where the first
postsynaptic spike also moves is not interesting, because the spike will move forward
in time, increasing the distance between the two postsynaptic spikes and thus reducing
the LTD effect on the second spike. Therefore we will consider the case where there is
an early postsynaptic spike at some fixed time that will remain in place, and a second
postsynaptic spike that will initially be triggered very far in time.
The intuition here is that there is a time interval for the second postsynaptic spike,
in which the LTD of the first postsynaptic spike would lead to a decrease in the mem-
brane potential of the postsynaptic neuron at the time of the second postsynaptic spike,
which could lead to the irreversible disappearance of the second postsynaptic spike or
its recession. Outside of this time interval, the second postsynaptic spike will reduce
its latency, approaching the early postsynaptic spike and the dangerous zone. In the
remaining of this section we will show that this interval is never reached in a determin-
istic system but that the addition of noise can enforce this disappearance.
We start by showing that repeating always the same input spike train without noise
cannot lead to the reduction of the number of postsynaptic spikes. Consider a long
presynaptic spike train with presynaptic spikes arriving at t0, t1, ...tN , which generates
two postsynaptic spikes, one at time t0, which is fixed and will appear at every pre-
sentation of the spike train, and another one that is originally triggered at tN . For the
second spike to disappear, it can either do so at tN or first advance through the spike
train -- that means, being triggered at tN−1, then at tN−2 and so on -- and eventually
die. For now, we assume that tN −t0 (cid:28) τs, so that initially the spike at time tN evolves
independently of the spike at time t0, and it would not disappear at tN . Consider now
that the input has been repeated long enough so that the second postsynaptic spike is
now triggered at ti, and the effects of the STDP generated by the spike at t0 are not
negligible to the presynaptic weight ti−1, which is associated to the presynaptic spike
at ti−1. If the postsynaptic spike is originally triggered at ti, then it would move to
ti−1 only if, after repeating the same input many times,
i−1(cid:88)
v(ti−1) =
− tk−ti−1
τm ≥ vth.
wke
(6)
k=1
After v(ti−1) crosses the vth threshold, the postsynaptic spike at ti moves to ti−1,
and thus the time difference between every presynaptic spike at t ≤ ti−1 and the
postsynaptic spike is reduced. This naturally implies that the synaptic weights wk
for all k ≤ i − 1 increase, thus the postsynaptic spike cannot disappear because the
membrane potential at v(ti−1) can not decrease unless the postsynaptic spike moves to
ti−2.
This argument assumes that presynaptic spike trains are always repeated with fixed
spike timings but with weights that are affected by LTP and LTD. This is generally
not true, as there are many factors that can introduce stochasticity on the evolution of
the weights, such as jitter, the stochastic nature of molecular dynamics on the synaptic
cleft and on the neuron membrane.
If we now consider the stability of both postsynaptic spikes with respect to that
7
where ξt is the contribution of the random evolution of the weights to v(t) given by
ξti =
− tk−ti
τm
δwke
(8)
i(cid:88)
k=1
i(cid:88)
k=1
i(cid:88)
noise, we easily realize that they are not equal: while the presynaptic spikes that gener-
ate the first postsynaptic spike are only subject to LTP and noise, the presynaptic spikes
that generate the second spike -- which happen necessarily between postsynaptic spikes
-- are subject to both LTP -- from the late postsynaptic spike -- and LTD -- from the earlier
postsynaptic spike -- on top of the noise.
This difference implies that the noise can make a postsynaptic spike disappear or
recede, either by directly weakening the associated presynaptic weights or strengthen-
ing them, so that the postsynaptic spike moves into a region where LTD dominates and
it would be later erased or receded.
To explain this in the setting that we used before, consider a neuron with a postsy-
naptic spike at time ti that would not move to ti−1 in the previous deterministic system.
However, now the weights evolve by the combined effects of that spike, an earlier post-
synaptic spike at time t0 and some noise. The membrane potential at time ti and after
r repetitions of the input spike train follows
v(ti) =
− tk−ti
τm + ξti,
wke
(7)
where δwk is the deviation of weight wk from its deterministic evolution; in the case of
Gaussian noise, for instance, it would lead to an Ornstein -- Uhlenbeck process for the
evolution of v(ti, r) on the repetition variable r. Note that the noise is not a variable
reinitialized at every repetition, as it has some momentum as weights evolve slowly.
If this postsynaptic spike train is repeated very often, the deterministic part of the
weights goes to a fixed value, which is zero for k > i and thus v(tk) ∼ ξtk for all
k > i. Thus, under the assumption
ξti <
− tk−ti
τm − vth
wke
(9)
k=1
for a few repetitions of the input spike train, the seconds spike vanishes. Therefore
among the subsequent input repetitions subjected to the ever present postsynaptic spike
at t0,the weights wk will decrease for all values of k, hence it is possible that v(ti) <
vth thereafter. This will result in the irreversible disappearance of the postsynaptic
spike at ti or its delay. This is illustrated in Fig. 2.
3.3 Generalization to inhibitory plasticity
Until now we have considered only excitatory neurons. However, in biological sys-
tems, inhibitory synapses are also present and have plasticity [VFD+13]. Naturally,
this might compromise the effects described in the previous section, as an inhibitory
8
Figure 2: Noise deletes late spike in a regular presynaptic spike train: We plot
the membrane potential (left) and firing times (right) of a postsynaptic neuron that
receives a constant train of spikes with inter-spike interval of 5ms and strength 7.5mV ,
from time t = 0ms to t = 150ms, and we add an extra spike at t = 80ms with
potential 5mV , with a forces postsynaptic spike at time 0.5ms. Note also that, during
its existence, the latency of the postsynaptic spike subject to noise decreases faster than
its noiseless counterpart.
synapse that gets potentiated could counteract the effects of excitatory STDP. For in-
stance, it might decrease the membrane potential and thus increase the latency of the
postsynaptic neuron [EJL15]. Our goal in this section is to find the parameter regime in
which the presence of inhibitory plasticity does not compromise the latency decrease
and, by extension, the disappearance of postsynaptic spikes.
Intuitively, as long as the STDP in inhibitory synapses is weaker than the STDP
in excitatory ones, the latency of postsynaptic spikes would still be decreased. The
question is then to find a way of measuring "how much weaker" it has to be. To
address this issue, we must find a boundary parameter set for inhibitory synapses that
guarantees that latency would be reduced, and then we can simply take any parameter
set that is between this boundary parameter set and the only-excitatory STDP.
To identify the parameter regime in which latency reduction for a single spike ap-
pears, we assume that the STDP keeps the balance between excitation and inhibition, in
the sense that the average input to a single neuron is maintained [Bru00]. To maintain
this balance, the potentiation of excitatory synapses is compensated by the potentia-
tion of inhibitory synapses. Potentiating all synapses but maintaining the average input
leads to the increase in fluctuations of the membrane potential, meaning that the mem-
9
020406080100120time (ms)70656055504540Membrane Potential (mV)Single neuron membrane potentialRepetition 300Repetition 70Repetition 0050100150200250300Input Repetitons01020304050607080Firing Time (ms)Single neuron firing time020406080100120time (ms)70656055504540Membrane Potential (mV)Single neuron membrane potentialRepetition 300Repetition 10Repetition 0050100150200250300Input Repetitons01020304050607080Firing Time (ms)Single neuron firing time brane potential preceding a postsynaptic spike would change more around the average,
and thus it can still lead to an earlier postsynaptic spike.
Consider a single postsynaptic spike at time tpos. For t < tpost,
v(t) =
− t−tk
τm ,
wke
(10)
(cid:88)
tk<t
and initially v(t) < vth. Now we wonder what happens when the weights wk change,
specifically if the postsynaptic spike would advance, recede or disappear. This depends
on the exact values of wk and tk, so to make more generic statements we are interested
in the value
Er [∆tpost] = E(cid:2)tr
post − tpost
(cid:3) =(cid:0)E(cid:2)tr
post
(cid:3) − tpost
(cid:1) Pr [∃s]
(11)
where r accounts for the number of times that the spike train has been repeated, and
Pr [∃s] is the probability that a postsynaptic spike still exists, and the expectations are
taken over the presynaptic spike trains -- a list of tuples (wk, tk) sampled from some
predefined distribution -- that generate a postsynaptic spike at time tpost. In simpler
words, we are simply trying to calculate whether the postsynaptic spike is expected
to move forward (E [∆tpost] (r) < 0) or backward (E [∆tpost] (r) < 0), ignoring the
ones that disappeared, if we only have some information about the distribution from
which the list of (wk, tk) was sampled.
We know that increasing the input excitatory weights can only lead to an earlier
postsynaptic spike, because v(t) can only increase and thus it might reach vth ear-
lier. We will take this a step further and assume that this statement is also true about
the average weights, meaning that when the expected input increases, the expected
postsynaptic firing time also increases. In more formal terms, we are assuming that
Er [∆tpost] is a function that decreases monotonically with
E [∆rv(t)] = E
∆ri(t)e
− t−x
τm dx
=
E [∆ri(t)] e
− t−x
τm dx,
(12)
(cid:20)(cid:90) t
−∞
(cid:21)
(cid:90) t
−∞
for all t < tpost, meaning that if the expected value of ∆v(t) averaged over all realiza-
tions of the input spike train producing a spike at tpost is positive, then Er [∆tpost] will
be negative.
This assumption, albeit natural, requires some careful consideration. Specifically,
we must clarify the distribution over which the expectations are taken, which corre-
sponds to all possible presynaptic spike trains shortly preceding a postsynaptic spike.
Those spike trains have fixed timings for every postsynaptic spike under consideration,
but are updated systematically because postsynaptic spikes evolve with the input rep-
etitions and the noise. Thus, this distribution considers samples in which a new spike
has just appeared or samples where a postsynaptic spike has recently been displaced
by a short time.
The subsequent step is to find the conditions that guarantee that E [∆rv(t)] in-
creases. A sufficient condition for this to happen is to have
E [∆ri(t)] = ∆rE [ie(t)] − ∆rE [ii(t)] > 0, ∀t < tpost
(13)
10
where E [ie(t)] is the expected input to the neuron at time t, and E [ii(t)], E [ii(t)] is
simply its decomposition in inhibitory and excitatory inputs, which gives us
(cid:90) ∞
(cid:90) ∞
0
0
E [ie(t)] = ρe
E [ii(t)] = ρi
µwe(w, t)dw
µwi(w, t)dw
(14)
(cid:90) ∞
0
(cid:90) ∞
(cid:90) ∞
0
0
+, we
(cid:90) ∞
ρe
0
(cid:90) ∞
(cid:90) ∞
0
ρe
0
where ρe, ρi are the rates of incoming spikes and µwe(w, t), µwi(w, t) the probabilities
of the weights associated to time t.
Thus, to maintain the condition from Eq. 13 we must ensure that the parameters
µwe, µwi, ηe
+, ηi
min, wi
min are such that
∆we(r)µwe(w, t)dw > ρi
∆wi(r)µwi(w, t)dw,
(15)
where ∆w(r) are given by the STDP Eq. 3 over many repetitions -- counted by r --
of the input spike train. We will now find a parameter regime in which this holds by
finding its boundary. In other words, we are interested on the parameter set in which
ρe
∆we(r)µwe(w, t)dw = ρi
∆wi(r)µwi(w, t)dw.
(16)
Note that it is not enough to find two weight distributions µwe, µwi where
Ae
+(we)µwe(w, t)dw = ρi
Ai
+(wi)µwi(w, t)dw,
(17)
because this would only work for the first input repetition. We have to ensure that even
after STDP changes the distribution, the equality holds. A simple way to achieve this
is to set the inhibitory and excitatory parameters to be equal. It is obvious that if the
probability distributions of weights, input rates and STDP parameters are the same,
then the change in input will affect the inhibitory and excitatory synapses in the same
way. However, we know that this is not the case, as there are typically fewer inhibitory
synapses than excitatory ones. Thus, we modify this symmetry to include rescaling,
meaning that we have the ratio
α =
ρi
ρe
(cid:90) ∞
that is also intrinsic to the probability distributions
αµwi (αx, t) = µwe (x, t) ∀x, t.
and the STDP parameters
αAi
+(αx) = Ae
+(x) ∀x.
By a simple change of variable we can show that, if those properties are satisfied,
Ae
+(x)µwe (x, t)dx =
1
α
ρi
0
αAi
+(αx)αµwi (αx, t)
1
α
d(αx)
(cid:90) ∞
ρe
0
(cid:90) ∞
0
= ρi
(18)
(19)
(20)
(21)
Ai
+(y)µwi (y, t)dy.
11
wi
Furthermore, if we take a pair of inhibitory and excitatory weights such that we = αwi
we have that after STDP,
αwi → α(wi + Ai
+(wi)) = αwi + αAi
+(we) ← we,
(22)
+(αwe) = we + Ae
meaning that the weight probability changes in such a way that
+(x), t(cid:1) = µwe (x, t) = αµwi (αx, t) = αµ(cid:48)
(cid:0)x + Ae
µ(cid:48)
(cid:0)α(cid:0)x + Ai
+(x)(cid:1) , t(cid:1) ,
we
we and µ(cid:48)
(23)
where µ(cid:48)
wi are the weight distributions after STDP has acted once. Thus, if
Eq. 19 holds at some point, it will also hold for all subsequent iterations of the input
spike pattern.
Thus, we have found a set of conditions that satisfy Eq. 19 at r = 0 and for any
subsequent r > 0 for the case where the postsynaptic spike does not change during
the r repetitions. Notice that the self-consistency of this condition does not make any
assumptions about the learning constant or ∆t dependent term on STDP, or even its
sign, it only requires that the expected increase -- or decrease -- in excitatory input is
matched by the expected increase -- or decrease -- in inhibitory input. In particular,
this symmetry does not change if the postsynaptic spike advances places, because the
STDP kernel has the same ratio of potentiating inhibitory and excitatory synapses. In
other words, when a postsynaptic spike changes places before the rth input repetitions,
the variance of the input before the postsynaptic spike still increases and, converselly,
it still decreases after the postsynaptic spike.
Now we have a large set of parameters in which latency reduction is expected to
happen. Any STDP parameters for which αAi
+(x) combined with Eq. 19,
or distribution of weights with αµwi (αx, t) < µwe (x, t) with Eq. 20, or both cases
combined.
+(αx) < Ae
It is worth noticing that the case when all the equalities Eq. 19 and Eq. 20 are met
we would still expect the latency to decrease. The reason is that even if
the variance of v(t) increases. More explicitly,
∆rVar [v(t)] = ∆r
Var [i(t)] dt = ∆r
(cid:90) t
E(cid:2)i2(t)(cid:3) dt =
−∞
(cid:90) t
= ∆r
−∞
E [∆rv(t)] = E [∆ri(t)] = 0,
(cid:90) t
e(t)(cid:3) dt +
−∞
(cid:16)E(cid:2)i2(t)(cid:3) − E [i(t)]2(cid:17)
(cid:90) t
i (t)(cid:3) dt
∆rE(cid:2)i2
−∞
(cid:90) t
−∞
∆rE(cid:2)i2
(24)
dt
(25)
where the term E [i(t)]2 = 0 by the symmetry of the weights and it is maintained at
zero by the symmetry of the STDP. Since we are only concerned with t < tpost, STDP
potentiates both inhibitory and excitatory synapses, so
∆rE(cid:2)i2
i (t)(cid:3) , ∆rE(cid:2)i2
e(t)(cid:3) > 0
(26)
and therefore the variance increases.
Naturally, if the variance of a certain distribution increases, then the probability of
reaching a value higher than some threshold -- vth -- also increases. Notice that even with
12
an increase on variance of v(t) we would not expect the postsynaptic spike to recede,
as the inputs i(t) at t (cid:46) tpost are excitatory -- because that v(tpost) > vth -- and thus by
Eq. 26, the membrane potential at v(tpost) should increase.
The approach outlined here can be also used for other STDP kernels. While the
symmetry in the excitatory and inhibitory STDP kernels might not exist for some
choices of inhibitory and excitatory plasticity, the approach can, in principle, still be
used by simply adding a bias on the number or weight of excitatory synapses that would
compensate the asymmetry in the STDP kernel. In the interest of simplicity we will
also skip this.
The non-proliferation of spikes can be derived by a similar argument, although in
this case the mean or the variance (or both) of the presynaptic input to the postsynaptic
neuron will decrease due to the depressive nature of STDP for t > tpost. In general,
the idea of having the depression stronger than the potentiation would still work, as
long as the depression of inhibitory synapses is weaker or equal than that of excitatory
synapses. As this calculation is essentially the same as the one we just presented, we
will omit it.
3.4 Numerical verification for random input spike trains
The examples presented to illustrate the latency reduction and the disappearance of
late postsynaptic spikes were simple, so we must now extend them to a more general
case. To do so, we simulated spike trains where the times of the presynaptic spikes
are random, including only excitatory or excitatory and inhibitory STDP, noise and
the presence of an earlier postsynaptic spike. The results are presented in Table 1 and
agree with our previous conclusions: a single postsynaptic spike tends to reduce its
latency, if there are multiple postsynaptic spikes in a short time window the later one
tends to disappear, and the presence of noise increases those effects. Note that we have
not included jitter or probabilistic presynaptic spikes, choosing instead to have noise
directly on the weight evolutions. As both cases have been addressed before [GVT05]
with similar conclusions, we shall not repeat them here.
So far we have only considered effects in small time scales, meaning that there
were only a few spikes on a time interval of the order of 10ms, and the postsynaptic
spike train would evolve over a few repetitions, on the order of 20. This leads us to
the conclusion that, with plausible assumptions on the parameters of our model, an
individual postsynaptic neuron will fire a given postsynaptic spike earlier after many
repetitions of the same presynaptic spike train and that if two postsynaptic spikes are
close in time, then the later one could disappear.
4 Postsynaptic Spike Train
Now we study the effects of the previously described phenomena, which act on small
temporal scales and affect only one or two postsynaptic spikes, for a population of
postsynaptic neurons, each one receiving many presynaptic spike trains happening over
time scales much larger than τm or τs. Specifically, we will explore the latency reduc-
tion and suppression or delaying of late postsynaptic spikes change the postsynaptic
13
STDP
Type
Noise
Var.
Spike
at t=0
E & I
E & I
E
E
E & I
E & I
E
E
0
0.2
0
0.2
0
0.2
0
0.2
No
No
No
No
Yes
Yes
Yes
Yes
Spike
Count
Increase
0.04 %
0.2 %
1.1 %
4.7 %
0.0 %
0.2 %
2.8 %
2.8 %
Spike
Count
Decrease
11.6 %
23.6 %
0.0 %
4.1 %
11.3 %
24.6 %
4.1 %
3.96 %
Spike
Latency
Increase
6.0 %
16.2 %
1.0 %
7.5 %
6.17 %
16.4 %
7.5 %
7.2 %
Spike
Latency
Decrease
19.7 %
14.8 %
46.6 %
39.5 %
20.1 %
14.7 %
39.7 %
39.8 %
Average
Latency
Change
-1.33 ms
-1.58 ms
-1.27 ms
-0.89 ms
-1.27 ms
-1.65 ms
-0.89 ms
-0.91 ms
Table 1: Effects of STDP on short random spike trains: We explored the effects of
STDP on the postsynaptic spike train of a neuron receiving 8 excitatory and 2 inhibitory
presynaptic spikes arriving at uniformly sampled times on the interval [0, 30ms] and
the stimulus is repeated 20 times. The first three columns determine the set-up: the
STDP Type indicates if STDP was active for excitatory presynaptic neurons only (E)
or for inhibitory as well as excitatory (E & I) with the inhibitory STDP having the
parameters to exactly compensate the excitatory one as presented in Section 3.3, the
second column indicates the variance of the Gaussian noise added to every weight at
every stimulus repetition, and the third column indicates whether we added a postsy-
naptic spike at the begining of the time window. The remaining columns explain the
results: the fourth one indicates the percentage of spike trains in which new postsy-
naptic spikes appeared, the fifth one the percentage of spike trains in which a spike
disappeared, the sixth one the percentage of spike trains in which a single postsynaptic
spike (not counting the imposed one at t = 0) happened later after learning, the sev-
enth one corresponds to the postsynaptic spike happening earlier, and the last one is
the average latency change of the postsynaptic spikes (here we only accounted for the
cases where there was a single postsynaptic spike at the beginning and at the end of the
learning). We calculated the percentages and averages from 10000 randomly generated
spike trains in which a single postsynaptic spike was triggered at the beginning of the
training. The results clearly show that in all cases spike latencies tend to decrease as
well as the number of spikes. Naturally, if STDP affects only excitatory synapses and
there is no noise, spikes do not disappear nor do they happen later. Also, as expected
adding a postsynaptic spike at t = 0 reduces the strength of the subsequent postsynap-
tic spikes, so they tend to disappear more and have stronger latency increases.
14
spike distribution.
Before studying the previous effects, we must validate some of the assumptions that
we made in the previous section. Specifically, we assume that all the input spikes came
from different synapses, which allowed us to treat the weights of all presynaptic spike
as independent. This is a valid assumption when we are considering a short time inter-
vals, as the sparsity of presynaptic firing and the existence of refractory periods implies
that a single synapse would typically not fire more than once during a short presynap-
tic spike train. However, when there is a long presynaptic spike train, a presynaptic
neuron might contribute to that spike train more than once, thus our assumption might
be invalid and the phenomena described in the previous section might not appear. To
ensure that the phenomena of latency reduction and late spike disappearance are still
present in long spike trains we use a combinatorial argument and count the number of
synapses that might evolve in a non-trivial fashion, which we present in Appendix A.
We can now consider the first time an input presynaptic spike train is presented.
Every neuron starts at v(0) = 0 and then its membrane potential will change depending
on its inputs. As the input spike train consists of independent spikes with independent
weights, the times of the first spike have a probability distribution f 1
0 (t) with support
on t > 0, which depends on the parameters of the input spike train. After spiking every
neuron resets its membrane potential to zero, and thus the distribution of inter-spike
intervals f ISI
(t) follows
f ISI
0
(t) = f 1
(27)
After the input has been repeated many times, the distribution of postsynaptic
spikes changes to f 1∞ and f ISI∞ respectively. Specifically, the first spikes reduce their
latency and thus move closer to t = 0, while the inter-spike intervals increase, due to
the depressive effect of postsynaptic spikes that repels or eliminates late postsynaptic
spikes. Therefore,
0 (t − tref ).
0
(cid:90) t
(cid:90) t
0
0
(cid:90) t
(cid:90) t
0
0
F 1∞(t) =
F ISI∞ (t) =
f 1∞(t)dt ≥
f ISI∞ (t)dt ≤
f 1
0 (t) = F 1
0 (t)
f ISI
0
(t) = F ISI
0
(t)
(28)
(29)
0
and F F
0 are the cumulative probability distributions of the
where F F∞, F ISI∞ , F ISI
inter-spike intervals and first spikes respectively. This is illustrated in Fig. 3 show-
ing that indeed the first spikes move forward through STDP and the later spikes are
more separated, which is consistent with the results from previous sections.
For the next section it will be convenient to look at the instantaneous firing rate,
which is obtained by accumulating the times of all spikes.
∞(cid:88)
Pr [tk ∈ [t, t + ∆t]]
s(t) = lim
∆t→0
k=1
∆t
where tk is the time of the kth spike. Since the time of the kth spike is the sum of the
inter-spike intervals of the first k − 1 spikes and the first spike, and the probability of
a sum is given by the convolution of the probability distributions, we can rewrite the
15
Figure 3: Cumulative probability distribution of the inter-spike interval : We plot
the cumulative probability distribution of the time of the first spike and the inter-spike
interval when a presynaptic spike train is presented for the first time (left) and after
many repetitions (right). We simulate 2000 neurons each receiving a presynaptic spike
train lasting 600ms with 200 presynaptic spikes, both inhibitory and excitatory, and
whose arrival time is uniformly sampled. Every synapse evolves through STDP and
being subject to both the fixed spike train with probability 0.33 or a random pair of
pre- and post-synaptic spikes with probability 0.66. We plot the time of the first spike
(blue) and the inter-spike interval for second third and fourth spikes, but subtracting
the refractory period to have a pertinent comparison with the first spiking time. We can
see that initially the first spike time is the same as the inter-spike interval for all the
spikes, but after STDP is applied the time of the first spike reduces, implying that the
blue line moves to the left with respect to the time before learning (in the black dotted
line) while the other inter-spike intervals increase, thus moving the curves to the right.
previous function as
s(t) =(cid:0)f 1 + f 1 ∗ f ISI + f 1 ∗ f ISI ∗ f ISI + ...(cid:1) (t) =
(cid:32)
f 1 ∗
(cid:33)
∞(cid:88)
(cid:0)f ISI(cid:1)∗k
(30)
where ∗ is the convolution operator, ∗k is the convolution power. Note that f 1 and f ISI
depend on how many times the input has been repeated. We will refer to the subindex
0 and ∞ to refer respectively to the cases where the presynaptic spike train is presented
for the first time or when it has been presented many times.
k=0
5 Implications for the Neural Code
The postsynaptic spike trains generated by neural populations are analogous to codes
that transmit information about presynaptic spikes to other neurons. As STDP is a
learning mechanism that modifies the postsynaptic spike train, we expect that it should
improve this encoding. Each input stimulus triggers spikes in a certain neural popula-
tion, and every neuron in that population has a certain performance associated to it, the
two most common ones being the energy consumption and resistance to noise [R+96].
16
020406080100Inter-Spike Interval0.00.20.40.60.81.0Cumulative Probability1st spike2nd spike3rd spike4th spike020406080100Inter-Spike Interval0.00.20.40.60.81.0Cumulative Probability1st spike2nd spike3rd spike4th spikeBefore learning (cid:90) T
0
(cid:90) ∞
5.1 Metabolic Efficiency
The energy consumption for a particular stimulus can be translated to our set-up as
the number of postsynaptic spikes triggered when the stimulus is present. That is, we
consider the integral
S =
s(t)dt.
(31)
When the spike train is long, we can ignore the initial distribution of spikes and use the
approximation
S ≈ Npost
T
tISI
(32)
where Npost is the number of postsynaptic neurons and tISI is the average inter-spike
interval given by
And by noting that
f ISI∞ (t)dt ≤
(cid:90) t
0
therefore,
tISI =
tf ISI dt.
(33)
(cid:90) t
0
(t) ⇒
f ISI
0
(cid:90) T
0
tref
tf ISI∞ (t)dt ≤
(cid:90) T
0
tf ISI
0
(t) ⇒ 1
tISI
0
≤ 1
tISI∞
,
(34)
S∞ = N
T
tISI∞
≤ N
T
tISI
0
= S0,
(35)
and thus the number of spikes decreases when the input is repeated, as illustrated in
Fig. 4.
5.2 Decoding Accuracy
The activity presented in a neural population must be decoded in subsequent layers
of the neural hierarchy. Here we show that, under some plausible assumptions on the
decoding populations, the increased concentration of spikes can make the decoding
more efficient in terms of signal strength.
The first step is to consider the decoder architecture. A very generic set-up consists
of a filter with a threshold detector described by the equations
y(t) = (h ∗ s) (t)
(cid:40)1 ⇐⇒ max
d =
t
0 otherwise,
y(t) ≥ θ
(36)
where d is the decoder output, h is the filter and θ is the threshold. Note that this
threshold must be selected in such a way that the detection happens when the input
is present -- thus low enough to be reached even in the presence of noise -- and at the
same time it should not fire when the input is not present, even if there is spurious
activity -- thus high enough to avoid false positives. Regardless of the choice of θ, the
17
performance of the decoder would change depending on the maximum value of y(t),
with higher values imply8ing better decoding performance. Our goal is then to identify
regimes in which this might happen.
For simplicity, we will assume that the filter h is a simple flat window given by
h(t) = Θ(t) − Θ(t − L)
with Θ is the step function and L is the window length. Therefore,
(cid:90) t
yM = max
t
y(t) =
s(t)dt.
t−L
(37)
(38)
The main parameter is thus the window length. When L ∼ T , then ymax ∝ S, and it
decreases when STDP is applied. Conversely, when L is small y(t) ∝ s(t), thus the
maximum is proportional to the maximum concentration of spikes, which is always at
the beginning of the spike train and increases by STDP.
Thus, the final question is which value of L is relevant. We consider L = τm, which
accounts for the fact that the decoder is also built from neurons which have roughly the
same timescales as the encoding neurons. This implies that as long as the concentration
of spikes at the beginning of the spike train is on that order of magnitude we will have
an increase in y(t) and thus the decoding would become better after STDP is applied,
and it is indeed what we observe in Fig. 4.
6 The Emergence of Predictions
When a group of neurons encodes a stimulus we mean that those neurons fire when the
stimulus is presented. However, the neurons themselves are not aware agents and do
not know anything about that stimulus; they simply receive a spike train that is strong
enough to trigger their spiking. From the point of view of an encoding neurons, there is
no difference between the stimulus-induced presynaptic spike train and any other input
spike train that always precedes the stimulus.
Combining this observation with the results from previous sections showing that
neurons will fire at the onset of a frequent input spike train, we can conclude that a
neuron that "encodes" a stimulus can start firing before the stimulus is presented if
another stimulus appears before it. As an illustrative example, imagine listening to a
melody. Different parts of the melody trigger the activity of different groups of neurons
in the same area of the brain. If the melody is repeated very often, the neurons P 1 that
react to an early part of the melody will systematically fire before the neurons P 2 that
react to a later part. As the melody is repeated, neurons in P 2 will always fire after
receiving spikes from neurons in P 1 and thus the synapses from P 1 to P 2 will be
reinforced. Eventually, the reinforced synapses might trigger spikes in P 2 before the
late part of the melody sounds. This can be extended to more populations encoding
more stimuli, and thus the whole melody is encoded through simultaneous activity of
all the neurons which originally encode only separate notes. This is illustrated and
simulated in Fig. 5.
18
Figure 4: Instantaneous firing rate and its effects on coding: Number of spikes
per bins of 4ms on the first 50ms of a spike train (left), evolution of the number of
postsynaptic spikes per input repetition (center) and maximum signal yM per input
repetition with window length of L = τm. We simulate 1000 neurons each receiving a
presynaptic spike train lasting 600ms with 200 presynaptic spikes, both inhibitory and
excitatory, and whose arrival time is uniformly sampled. Every synapse is initialized
by applying 500 pairs of pre- and postsynaptic spikes with a ∆t sampled uniformly
at random in the interval [−20ms, 20ms] and letting STDP modify it. Then, every
synapse evolves through STDP by being subject to both the fixed spike train with prob-
ability 0.33 or a random pair of pre- and post-synaptic spikes with probability 0.66. As
discussed in the previous sections, after the input pattern is repeated many times, the
first spikes arrive early, creating an initial burst of spikes (left), but the overall number
of postsynaptic spikes decreases when the input is repeated very often (center) and at
the same time the initial burst of activity increases the value of yM (right) for small
time windows -- L = 1, 10, 20 -- but not for long ones -- L = 50, 100. Notice that the
reduction on the number of spikes is relatively small with respect to the original val-
ues. This should be taken as a qualitative result that depends on the parameters of the
simulation, but it will increase if we had a larger time window.
7 Discussion
In this paper we start by analyzing and expanding previous findings on latency re-
duction [SMA00, GVT05]. Then we interpret them in communication terms: those
mechanisms lead to encoding the more common inputs with less spikes while con-
centrating the remaining spikes in smaller time windows. This leads us to the con-
clusion that STDP reduces the amount of spikes used to encode frequent stimuli, in
line with the idea that metabolic efficiency is one of the guiding principles of the
brain [HOCS10, Lau01]. The same phenomena also improves decoding performance
of the neural code by concentrating encoding spikes on small time windows, in line
with the notion that synchronization is a learned behavior used to improve communi-
cation between neuronal assemblies [Sin11, Fri05, VDM94]. Finally, we show that the
latency reduction can explain how the nervous system learns to forecast even without
any feedback.
This study is another example of how simple, well-known plasticity rules that are
present at synaptic level lead to modifications that are advantageous at the organism
level. Furthermore, the fact that the same mechanism improves the neural code and cre-
ates predictions might explain how the ability of the brain to make predictions -- which
19
01020304050t020406080100120s(t)Before STDPAfter STDP02004006008001000Input Repetition304003060030800310003120031400Spike Count02004006008001000Input Repetition707580859095100Percentage of Maximum SignalL = 1L = 10L = 20L = 50L = 100 Figure 5: Encoding Predictions: Schema for the emergence of predictions (left) and
firing latencies of neurons in encoding population (right): An external event creates
three stimulus that trigger all the neurons in corresponding distinct neural populations
P 1, P 2, P 3 with delays randomly sampled on the intervals dS1P 1 ∈ [0ms, 1000ms]
and dS2P 2 ∈ [1000ms, 2000ms] and dS3P 3 ∈ [2000ms, 3000ms] respectively. The
three populations, with N = 50 neurons each one, also have synapses that between
them with delays sampled from a uniform distribution between dP iP j ∈ [5ms, 45ms].
Originally, almost all neurons in each population fire only after receiving inputs from
their respective stimuli, but after the external event is repeated very often, the inter-
population connections become strong enough to trigger some spikes before the stim-
ulus is received. Notice that even though the connections between populations are
originally symmetric -- P i is connected to P j in the same way as P j is to P i -- after
synapses have adapted they follow the temporal order.
20
External EventStimulus 2Stimulus 1P 2P 1dS1P1 = 1sStimulus 3P 3dS2P2 = 2sdS3P3 = 3s0255075100125150175200Stimulus repetition50010001500200025003000LatencyEvolution of Median LatencyP 1P 2P 3 is one of the core problems in cognitive science -- could have emerged as a consequence
of evolutionary pressures on metabolic cost and information transmission.
Naturally, our work is also interesting for researchers in machine learning, as it
shows that Hebbian learning rules, which are classically used to infer or reinforce cor-
relations [DA01], can be used to find predictions by adding a temporal asymmetry in
the synaptic plasticity kernel. Furthermore, the fact that the same mechanism gives
rise to predictions and coding efficiency is another example of the intimate relation-
ship between machine learning and coding [MMK03], thus it might be interesting for
information theorists.
The results exposed here also open new questions. The effects of latency reduc-
tion in networks of neurons -- specially recurrent ones -- , or the potential use of this
prediction capabilities of STDP for machine learning require further study but could
be useful extensions. However, the most immediate question is whether this unsuper-
vised process is used in the nervous system. An experimental study should identify the
neurons that encode two temporally correlated stimuli and follow the evolution of la-
tencies as the stimuli are repeated, while simultaneously ensuring that this process was
due to STDP alone without interference of reward systems that have been previously
proposed.
21
A Postsynaptic spikes evolve independently
The problem with having multiple spikes per presynaptic neuron is that all of the presy-
naptic spikes coming from the same synapse have the same weight, and therefore when
a postsynaptic spike is close to one of those presynaptic spikes, all of the presynaptic
spikes that come from that synapse will undergo the same weight modifications. There
are two scenarios when this would be a problem:
1. A single synapse undergoes STDP from two or more different spikes: If there are
two postsynaptic spikes, affected by their respective presynaptic spikes, but some
of those presynaptic spikes come from the same synapse, the resulting weight
change from STDP would be a combination of the effects of both postsynaptic
spikes. This is undesirable as the effects could be opposite: one postsynaptic
spike could induce depression while the other potentiation, and thus the evolution
of one of the presynaptic spikes would not evolve as our STDP rule predicts.
2. A new postsynaptic spike appears spontaneously from STDP: Typically, STDP
applies only when there exists a postsynaptic spike. However, if some synapses
are very strong due to STDP, and those synapses have spikes that are close to-
gether, they could generate a new postsynaptic spike. This would automatically
generate pairs of presynaptic spikes that are affected by two postsynaptic spikes
simultaneously (thus we would be in the previous case). Furthermore, the spon-
taneous generation of new postsynaptic spikes is itself problematic.
Consider M presynaptic neurons which fire with a rate λ, and a postsynaptic neuron
that receives them with a rate ρ = M λ during a time interval of length T , generating
spost postsynaptic spikes. Furthermore, each one of those postsynaptic spikes imposes
STDP that affects the presynaptic spikes that are close to it. For simplicity, we will as-
sume that the noticeable effect on the presynaptic spikes is restricted to a time window
of size lτS where l is a small integer number.
We start by studying case (1). If we have spost postsynaptic spikes, then the effects
of STDP are noticeable for
ta = lτSspost
(39)
milliseconds in which all presynaptic spikes should come from different synapses.
Given that the arrival times of each spike are uniform of the whole interval, the ex-
pected number of presynaptic neurons that fire in that interval more than once is given
by
(λta)ke−λta
= N(cid:0)1 − e−λta − λtae−λta(cid:1) ,
and by a Taylor expansion to order two,
1 − λta +
E [#1] ≈ N
k!
(cid:18)
(cid:19)
= N
λ2t2
a
2
.
+ λta − λ2t2
a
λ2t2
a
2
∞(cid:88)
k=2
N
(40)
(41)
To get an intuition of the magnitude of these numbers, consider, for instance, an input
spike train lasting 1s with presynaptic spike rate of 0.5Hz which generates two post-
synaptic spikes and we pick the relevant time window to be twice τS, so l = 2 and
22
spost = 3. Then, the expected number of events of type (1) would be
E [#1] ≈ M
400
.
(42)
Furthermore, not all of those events would actually be problematic; ifall of them are
potentiating or depressing, then this does not change our analysis.
For case (2) we argue that in order to spontaneously generate new spikes, the
synapses affected STDP must be very strong and excitatory, and a few of those strong
excitatory synapses must coincide within a small time window of order τm.
The synapses that can be very strong are those in the ta time, meaning that we
expect
(43)
independent synapses to be close to wmax. Each one of those synapses can fire within
the remaining T − ta time at a rate λ, so we would expect to have presynaptic a rate of
STDP-affected spikes of
na = ρta = ρlτSspost,
(44)
Now we must compute the probability that enough of them coincide to generate a
postsynaptic spike.
λa = naλ.
We denote this number by k and we will compute the number of spontaneous post-
synaptic spikes that would appear for every k. We start by considering k = 2 of those
presynaptic spikes (although for some choices of wmax we have to start at k > 2), and
note that in order to have the postsynaptic spike, we must have
wmax + wmaxe
− ∆tk=2
τm + σv > vth
(45)
where σv is a term that accounts for the presence of other spikes that could be driving
the membrane potential higher, and ∆tk=2 is the time interval between the two spikes.
By rearranging,
where ϑ = vth−σv
time interval between spikes is given by an exponential distribution, so
(46)
. Since the spikes follow a Poisson distribution, the probability of a
∆tk=2 < i2 = τm ln (ϑ − 1) ,
wmax
Pr [#2k = 2] = Pr [∆tk=2 < i2] = 1 − e−λai2,
and the number of those intervals tends to λaT for large T , so
E [#2]k=2 = λaT(cid:0)1 − e−λai2(cid:1) ,
(47)
(48)
For k > 2, the estimation can be done by applying the fact that two contiguous spikes
are independent, and therefore the inter-spike intervals are also independent, so we
can multiply their probabilities. Furthermore, we should not have any two spikes at a
distance closer than i2, so
(cid:90) ∞
(cid:90) ∞
Pr [#2k = 3] <
λae−λax
i2
i2
− x
τm + e
τm − ϑ
− y
1 + e
(cid:105)
dydx,
(49)
(cid:104)
λae−λayΘ
23
where the inequality comes because we let the interval time go to infinity, while T is
finite. We can ignore the Θ
term and we obtain
τm + e
1 + e
− x
τm − ϑ
− y
(cid:104)
(cid:105)
Pr [#2k = 3] <(cid:0)1 − e−λai2(cid:1)2
(cid:0)1 − e−λai2(cid:1)2
∞(cid:88)
E [#2]k=3 ≤ T λa
∞(cid:88)
E [#2]k=j ≤ T λa
.
.
j=2
j=2
= T λaeλai2 = T λa (ϑ − 1)λa .
(cid:0)1 − e−λai2(cid:1)j−1
(50)
(51)
(52)
And here the number of pairs of contiguous time intervals is also lower than T λa,
which gives us
Naturally, the same upper bound can be computed for any k, so
E [#2]∀k =
Which will be low as long as λa is low. If we have, for instance, M = 50, l = 2,
spost = 3 and λ = 0.5 Hz, we obtain λa = 6 · 10−3. Then, if we take σv = wmax/2,
ϑ = 1.5,
E [#2]∀k ≈ 3T
1000
,
(53)
with T being in milliseconds, this means that for an input spike train lasting half a
second, generating 3 postsynaptic spikes, there would be one expected spontaneous
postsynaptic spike.
The estimates from Eq. 42 and 53 give a relatively low number of coupled postsy-
naptic spikes or spontaneous spikes. We will therefore assume, from now on, that the
effects described in Sec.3 are valid and happen in every postsynaptic spike on every
neuron independently of the presence of other postsynaptic spikes.
24
References
[AN00]
[BP98]
[Bru00]
[BVCS10]
[CBC95]
[DA01]
[EJL15]
[Fri05]
Larry F Abbott and Sacha B Nelson. Synaptic plasticity: taming the
beast. Nature neuroscience, 3(11s):1178, 2000.
Guo-qiang Bi and Mu-ming Poo. Synaptic modifications in cultured hip-
pocampal neurons: dependence on spike timing, synaptic strength, and
postsynaptic cell type. Journal of neuroscience, 18(24):10464 -- 10472,
1998.
Nicolas Brunel. Dynamics of sparsely connected networks of excitatory
and inhibitory spiking neurons. Journal of computational neuroscience,
8(3):183 -- 208, 2000.
Andreja Bubic, D Yves Von Cramon, and Ricarda I Schubotz. Predic-
tion, cognition and the brain. Frontiers in human neuroscience, 4:25,
2010.
Franc¸ois Chapeau-Blondeau and Nicolas Chambet. Synapse models
for neural networks: From ion channel kinetics to multiplicative coef-
ficientw ij. Neural computation, 7(4):713 -- 734, 1995.
Peter Dayan and Laurence F Abbott. Theoretical neuroscience, volume
806. Cambridge, MA: MIT Press, 2001.
Felix Effenberger, Jurgen Jost, and Anna Levina. Self-organization in
balanced state networks by stdp and homeostatic plasticity. PLoS com-
putational biology, 11(9):e1004420, 2015.
Pascal Fries. A mechanism for cognitive dynamics: neuronal com-
munication through neuronal coherence. Trends in cognitive sciences,
9(10):474 -- 480, 2005.
[GKNP14] Wulfram Gerstner, Werner M Kistler, Richard Naud, and Liam Paninski.
Neuronal dynamics: From single neurons to networks and models of
cognition. Cambridge University Press, 2014.
[GKvHW96] Wulfram Gerstner, Richard Kempter, J Leo van Hemmen, and Hermann
Wagner. A neuronal learning rule for sub-millisecond temporal coding.
Nature, 383(6595):76, 1996.
[GVT05]
Rudy Guyonneau, Rufin VanRullen, and Simon J Thorpe. Neurons tune
to the earliest spikes through stdp. Neural Computation, 17(4):859 -- 879,
2005.
[Hee17]
David J Heeger. Theory of cortical function. Proceedings of the National
Academy of Sciences, 114(8):1773 -- 1782, 2017.
25
[HOCS10] Andrea Hasenstaub, Stephani Otte, Edward Callaway, and Terrence J
Sejnowski. Metabolic cost as a unifying principle governing neu-
ronal biophysics. Proceedings of the National Academy of Sciences,
107(27):12329 -- 12334, 2010.
[Hoh13]
Jakob Hohwy. The predictive mind. Oxford University Press, 2013.
[Izh04]
[KGH01]
[KH00]
[Lap07]
[Lau01]
[MBG95]
[MGT08]
Eugene M Izhikevich. Which model to use for cortical spiking neurons?
IEEE transactions on neural networks, 15(5):1063 -- 1070, 2004.
Richard Kempter, Wulfram Gerstner, and J Leo van Hemmen. Intrinsic
stabilization of output rates by spike-based hebbian learning. Neural
computation, 13(12):2709 -- 2741, 2001.
Werner M Kistler and J Leo van Hemmen. Modeling synaptic plasticity
in conjunction with the timing of pre-and postsynaptic action potentials.
Neural Computation, 12(2):385 -- 405, 2000.
L. Lapique. Recherches quantitatives sur lexcitation electrique des nerfs
traite comme une polarization. Journal de Physiologie Pathologie Gene-
tique, 1907.
Simon B Laughlin. Energy as a constraint on the coding and processing
of sensory information. Current opinion in neurobiology, 11(4):475 --
480, 2001.
Ralph R Miller, Robert C Barnet, and Nicholas J Grahame. Assessment
of the rescorla-wagner model. Psychological bulletin, 117(3):363, 1995.
Timoth´ee Masquelier, Rudy Guyonneau, and Simon J Thorpe. Spike
timing dependent plasticity finds the start of repeating patterns in con-
tinuous spike trains. PloS one, 3(1):e1377, 2008.
[MMK03]
David JC MacKay and David JC Mac Kay. Information theory, inference
and learning algorithms. Cambridge university press, 2003.
[PG06]
[PMBB15]
[PVB+98]
Jean-Pascal Pfister and Wulfram Gerstner. Triplets of spikes in a
model of spike timing-dependent plasticity. Journal of Neuroscience,
26(38):9673 -- 9682, 2006.
Stephanie E Palmer, Olivier Marre, Michael J Berry, and William Bialek.
Predictive information in a sensory population. Proceedings of the Na-
tional Academy of Sciences, 112(22):6908 -- 6913, 2015.
Yifat Prut, Eilon Vaadia, Hagai Bergman, Iris Haalman, Hamutal Slovin,
and Moshe Abeles. Spatiotemporal structure of cortical activity: proper-
ties and behavioral relevance. Journal of neurophysiology, 79(6):2857 --
2874, 1998.
[R+96]
Theodore S Rappaport et al. Wireless communications: principles and
practice, volume 2. prentice hall PTR New Jersey, 1996.
26
[RBH+11] Alex Roxin, Nicolas Brunel, David Hansel, Gianluigi Mongillo, and
Carl van Vreeswijk. On the distribution of firing rates in networks of
cortical neurons. Journal of Neuroscience, 31(45):16217 -- 16226, 2011.
[RW+72]
Robert A Rescorla, Allan R Wagner, et al. A theory of pavlovian con-
ditioning: Variations in the effectiveness of reinforcement and nonre-
inforcement. Classical conditioning II: Current research and theory,
2:64 -- 99, 1972.
[RWP07]
John D Rolston, Daniel A Wagenaar, and Steve M Potter. Precisely
timed spatiotemporal patterns of neural activity in dissociated cortical
cultures. Neuroscience, 148(1):294 -- 303, 2007.
[SDM97]
Wolfram Schultz, Peter Dayan, and P Read Montague. A neural sub-
strate of prediction and reward. Science, 275(5306):1593 -- 1599, 1997.
[SG10]
[Sin11]
[SMA00]
[SMP18]
[TN04]
[VDM94]
[VFD+13]
J. Sjstrm and W. Gerstner. Spike-timing dependent plasticity. Scholar-
pedia, 5(2):1362, 2010. revision #184913.
Wolf Singer. Dynamic formation of functional networks by synchro-
nization. Neuron, 69(2):191 -- 193, 2011.
Sen Song, Kenneth D Miller, and Larry F Abbott. Competitive heb-
bian learning through spike-timing-dependent synaptic plasticity. Na-
ture neuroscience, 3(9):919, 2000.
Audrey J Sederberg, Jason N MacLean, and Stephanie E Palmer. Learn-
ing to make external sensory stimulus predictions using internal correla-
tions in populations of neurons. Proceedings of the National Academy
of Sciences, page 201710779, 2018.
Gina G Turrigiano and Sacha B Nelson. Homeostatic plasticity in the de-
veloping nervous system. Nature Reviews Neuroscience, 5(2):97, 2004.
Christoph Von Der Malsburg. The correlation theory of brain function.
In Models of neural networks, pages 95 -- 119. Springer, 1994.
Tim P Vogels, Robert C Froemke, Nicolas Doyon, Matthieu Gilson,
Julie S Haas, Robert Liu, Arianna Maffei, Paul Miller, Corette Wierenga,
Melanie A Woodin, et al. Inhibitory synaptic plasticity: spike timing-
dependence and putative network function. Frontiers in neural circuits,
7:119, 2013.
[VRBT00] Mark CW Van Rossum, Guo Qiang Bi, and Gina G Turrigiano. Sta-
ble hebbian learning from spike timing-dependent plasticity. Journal of
neuroscience, 20(23):8812 -- 8821, 2000.
27
|
1711.01629 | 2 | 1711 | 2018-03-15T05:47:37 | Mutual Information in Frequency and its Application to Measure Cross-Frequency Coupling in Epilepsy | [
"q-bio.NC",
"cs.IT",
"eess.SP",
"cs.IT"
] | We define a metric, mutual information in frequency (MI-in-frequency), to detect and quantify the statistical dependence between different frequency components in the data, referred to as cross-frequency coupling and apply it to electrophysiological recordings from the brain to infer cross-frequency coupling. The current metrics used to quantify the cross-frequency coupling in neuroscience cannot detect if two frequency components in non-Gaussian brain recordings are statistically independent or not. Our MI-in-frequency metric, based on Shannon's mutual information between the Cramer's representation of stochastic processes, overcomes this shortcoming and can detect statistical dependence in frequency between non-Gaussian signals. We then describe two data-driven estimators of MI-in-frequency: one based on kernel density estimation and the other based on the nearest neighbor algorithm and validate their performance on simulated data. We then use MI-in-frequency to estimate mutual information between two data streams that are dependent across time, without making any parametric model assumptions. Finally, we use the MI-in- frequency metric to investigate the cross-frequency coupling in seizure onset zone from electrocorticographic recordings during seizures. The inferred cross-frequency coupling characteristics are essential to optimize the spatial and spectral parameters of electrical stimulation based treatments of epilepsy. | q-bio.NC | q-bio | Mutual Information in Frequency and its
Application to Measure Cross-Frequency Coupling
in Epilepsy
Rakesh Malladi, Member, IEEE, Don H Johnson, Fellow, IEEE, Giridhar P Kalamangalam,
Nitin Tandon, and Behnaam Aazhang, Fellow, IEEE
8
1
0
2
r
a
M
5
1
]
.
C
N
o
i
b
-
q
[
2
v
9
2
6
1
0
.
1
1
7
1
:
v
i
X
r
a
Abstract-We define a metric, mutual
information in fre-
quency (MI-in-frequency), to detect and quantify the statistical
dependence between different
frequency components in the
data, referred to as cross-frequency coupling and apply it to
electrophysiological recordings from the brain to infer cross-
frequency coupling. The current metrics used to quantify the
cross-frequency coupling in neuroscience cannot detect if two
frequency components in non-Gaussian brain recordings are
statistically independent or not. Our MI-in-frequency metric,
based on Shannon's mutual information between the Cramér's
representation of stochastic processes, overcomes this shortcom-
ing and can detect statistical dependence in frequency between
non-Gaussian signals. We then describe two data-driven estima-
tors of MI-in-frequency: one based on kernel density estimation
and the other based on the nearest neighbor algorithm and
validate their performance on simulated data. We then use
MI-in-frequency to estimate mutual
information between two
data streams that are dependent across time, without making
any parametric model assumptions. Finally, we use the MI-in-
frequency metric to investigate the cross-frequency coupling in
seizure onset zone from electrocorticographic recordings during
seizures. The inferred cross-frequency coupling characteristics
are essential to optimize the spatial and spectral parameters of
electrical stimulation based treatments of epilepsy.
Index Terms-Mutual
information in frequency; dependent
data; Cramér's spectral representation; cross-frequency coupling;
epilepsy; seizure onset zone.
I. INTRODUCTION
Epilepsy is a very common neurological disorder affecting
nearly 1% of the world's population. Epilepsy is characterized
by repeated, unprovoked seizures. Nearly a third of all epilepsy
patients have medically refractory epilepsy (medication is not
effective in these patients). For these patients, surgical resection
of the seizure onset zone (SOZ) (the regions of the brain
responsible for generating and sustaining seizure activity [3]) or
electrical stimulation are possible treatment options. However,
the efficacy of these treatments is variable and almost always
never results in a cure [4], [5]. There is tremendous interest
in leveraging the recent advances in electrical stimulation
This work is funded in part by grant 1406447 from National Science
Foundation and Texas Instruments and was done at Rice University. A portion
of this work was presented at Cosyne [1] and Asilomar [2].
Rakesh Malladi is with LinkedIn Corporation, Sunnyvale, CA. Don H John-
son and Behnaam Aazhang are with the Department of Electrical and Computer
Engineering, Rice University, Houston, TX. Giridhar P Kalamangalam is with
Department of Neurology at University of Florida, Gainesville, FL. Nitin
Tandon is with Department of Neurosurgery at University of Texas Health
Center, Houston, TX. E-mail: [email protected], {dhj, aaz}@rice.edu,
[email protected], [email protected].
[6] and optogenetics [7] to develop spatiotemporally specific
approaches to treat epilepsy. A crucial step in this endeavor is
to develop an understanding of the coupling between neuronal
oscillations in different frequency bands during seizures. This
coupling or statistical dependence across frequency components
between signals is referred to as cross-frequency coupling
(CFC) [8], [9]. Our main objective is to learn the dynamics of
cross-frequency coupling during seizures in epilepsy patients
from the electrocorticographic (ECoG) data.
Elaborating the characteristics of epileptic seizures using
cross-frequency coupling between ECoG data has been the
focus of many papers. CFC has been used to predict the onset
of seizure in [10] and detect epileptic seizures in [11]. CFC
has also been used to localize the area for surgical resection
in epilepsy patients [12]–[14]. Variations in CFC from preictal
(before a seizure) to ictal (during a seizure) to postictal (after
a seizures) in epilepsy patients have been analyzed in [15],
[16]. In addition, the CFC in interictal stages is compared with
that around seizures in [11], [17]–[19]. In this paper, we study
CFC within and between various regions inside the seizure
onset zone to determine the dominant frequencies involved in
seizures and to learn the variations in coupling strength between
various spatial regions inside SOZ. The results from this study
are crucial to optimize the spectral and spatial parameters of
next generation epilepsy treatments.
Cross-frequency coupling or dependence across frequencies
in the data could be in a single recording or between recordings,
not necessarily at the same frequency. Coherence can identify
if two frequency components are statistically independent
or not and quantify the dependence for linear, Gaussian
processes [20]. There is no such equivalent metric for non-
Gaussian signals. Since the time-series data recorded from the
brain are neither linearly related nor Gaussian, neuroscientists
typically use heuristic metrics that cannot identify if two
frequency components are statistically independent or not and
can only capture second-order dependencies. Some of the
popular heuristics estimate the phase-amplitude, amplitude-
amplitude, phase-phase coupling between the low and high
frequency components in the electrophysiological recordings
from brain [8], [21]–[23]. In fact, a recent review article on
CFC metrics suggests the use of cross-frequency 'correlation'
instead of 'coupling' to describe these heuristic CFC metrics
[22]. Furthermore, a list of confounds affecting the current
CFC metrics is provided in [22]. A more comprehensive metric
that detects statistical independence and thereby, capture both
2
linear and nonlinear dependencies, would be invaluable in
determining how neuronal oscillations at various frequencies
are involved in the computation, communication, and learning
in the brain. Here we propose a new methodology or metric to
estimate the cross-frequency coupling (CFC) in neuroscience
that overcomes the challenges of the existing approaches and
as a proof-of-concept, we infer CFC characteristics of epileptic
seizures using our metric.
Mutual information in frequency (MI-in-frequency), defined
for linear Gaussian processes using coherence in [24], [25],
can indeed be further developed into a general technique to
estimate CFC. Inspired by prior work [26], we define MI-
in-frequency between two frequencies in a signal (or two
signals) as the Shannon's mutual information (MI) between
the Cramér's spectral representations [27], [28] of the two
signals at the corresponding frequencies. Cramér's spectral
representation transforms a time-domain stochastic process
into a stochastic process in the frequency domain, the samples
of which can be estimated at each frequency from the time-
domain data samples [29]. MI-in-frequency metric is equivalent
to coherence measures for linear, Gaussian signals and can be
thought of as 'coherence' for non-Gaussian signals. The MI-in-
frequency metric is one of the three mutual information based
metrics used in [26] to analyze linear relationships between
seismic data and [26] is not focussed on defining a single
metric to capture the statistical dependence across frequency.
We extend this approach to define a single metric, MI-in-
frequency, to capture statistical dependencies across frequency
for both linear and nonlinear data and use it measure CFC in the
brain. We then describe two data-driven algorithms – one based
on kernel density estimation (KDMIF) and the other based
on nearest neighbor estimation (NNMIF) – to estimate MI-in-
frequency without assuming any parametric model of the data.
We considered these two approaches since they outperformed
other approaches in estimating MI from i.i.d. data and there is
no clear winner between them [30], [31]. We also demonstrate
the superiority of MI-in-frequency over existing CFC metrics
by comparing against modulation index [8], [21], a commonly
used CFC metric, on simulated data.
In addition to estimating CFC between ECoG data, we use
MI-in-frequency to develop a data-driven estimator for mutual
information (MI). Note that MI estimation is a solved problem
if the data samples are i.i.d. [32] or are sampled from linear,
Gaussian processes [24], [25], [33], [34]. As mentioned earlier,
real-world data is neither independent across time nor Gaussian
and the underlying model is often unknown. Our data-driven
MI estimation algorithm applies to dependent data, without
making any parametric model assumptions. The key idea is to
make the problem computationally tractable by focussing only
on those frequencies in the two data streams that are statistically
dependent, which are identified by MI-in-frequency metric. Our
MI estimator converges to the true value for Gaussian models
and we validate its performance on nonlinear models.
Finally, we apply the MI-in-frequency estimators to infer the
cross-frequency coupling in the seizure onset zone (SOZ), by
analyzing electrocorticographic (ECoG) data from the SOZ of
9 patients with medial temporal lobe epilepsy in whom a total
of 25 seizures were recorded. We investigate the dynamics of
CFC in preictal, ictal and postictal periods within one SOZ
electrode and between electrodes in different regions in the
SOZ. We observe an increase in coupling in gamma and ripple
high-frequency oscillations during seizures, with the largest
increase within a SOZ electrode and a very small increase
between electrodes in different regions inside SOZ. In addition,
low-frequency coupling and linear interactions between SOZ
electrodes also increase during the postictal state.
II. CRAMÉR'S SPECTRAL REPRESENTATION OF
STOCHASTIC PROCESSES
Consider a stochastic processes X (t) , t ∈ R. Let SX (ν) for
ν ∈ R be the spectral distribution function of X and sX (ν),
its power spectral density, if it exists. Two basic spectral
representations are associated with the stochastic process
X (t) - power spectral distribution and Cramér's representation
[27], [28]. The Cramér's representation of X (t) and its key
properties are stated in the following theorem.
Theorem 1. (page 380 in [28]) Let X (t) be a second order
stationary, mean-square continuous and zero mean stochastic
process. Then there exists a complex-valued, finite-variance,
ν ∈ R, such that
orthogonal increment process (cid:101)X (ν) in the frequency domain
with E(cid:104)
The process (cid:101)X (ν) = (cid:101)XR (ν) + j(cid:101)XI (ν) satisfying the above
of X (t). d(cid:101)X (ν) is the complex random variable representing
Furthermore, if the X (t) is real-valued, then (cid:101)X(cid:0) − ν(cid:1) =
(cid:101)X (cid:63)(cid:0)ν(cid:1), E(cid:2)d(cid:101)XR
the amplitude of oscillation in the interval from ν to ν + dν
in X (t). The integral in Theorem 1 is a Fourier-Stieltjes
integral. Intuitively, Theorem 1 decomposes X (t) into an
orthogonal increment complex process in the frequency domain.
theorem is the spectral process or the Cramér's representation
ej2πνtd(cid:101)X (ν) ,
= 0, and E(cid:104)d(cid:101)X (ν)2(cid:105)
−∞
E(cid:2)(cid:0)d(cid:101)XR (ν)(cid:1)2(cid:3) = E(cid:2)(cid:0)d(cid:101)XI (ν)(cid:1)2(cid:3) = 1
(cid:0)ν(cid:1)d(cid:101)XI
(cid:0)ν(cid:1)(cid:3) = 0, and
= dSX (ν) .
(cid:105)
d(cid:101)X (ν)
2 dSX (ν) .
(1)
∞(cid:82)
X (t) =
We have the following theorem for the special case of a real-
valued Gaussian process X (t).
Theorem 2. (page 385 in [28]) Let X (t) be a real-valued
stationary, mean-square continuous Gaussian process with zero
mean and power spectral distribution function SX (ν) , ν ∈ R.
Then the real and imaginary parts of its spectral process (cid:101)XR (ν)
and (cid:101)XI (ν) are zero mean, mutually independent, identically
distributed Gaussian processes satisfying (1).
Example: Consider the zero mean stationary Gaussian
process X (t) = A cos (2πν0t + Θ), where A is Rayleigh
random variable with parameter σA that is independent of
Θ, which is uniform in [0, 2π). The increments of the spectral
process of X (t) are all zero, except at ν = ±ν0, where the
path of the real part of spectral process (cid:101)X (ν) has two jumps
2 exp (±jΘ) [28]. This implies that the sample
increment is A
of same magnitude and direction at frequencies ±ν0, while
that of the imaginary part has two jumps of same magnitude,
2 cos Θ and A
but opposite directions at ±ν0. The magnitude of the jump
at ν0 in the real and imaginary parts is A
2 sin Θ
respectively, both of which are Gaussian random variables with
mean zero and variance 1
A. This spectral process is intuitive
because we know X (t) has all its energy only at frequencies
±ν0 and the variance of the increments of the spectral process
of X (t) which is nonzero only at ±ν0. We therefore expect
probability to be constant, except for jumps at ±ν0.
d(cid:101)X (ν) is equal to the differential power spectral distribution
all sample paths of the random process (cid:101)X (ν) with non-zero
2 σ2
Note that
if the process is wide sense-stationary and
Gaussian, then power spectral distribution would have all the
information about the process and its relationship with Cramér's
representation is given by Theorem 2. Otherwise, power spectral
distribution only captures the second-order dependencies in
the process. Since ECoG signals are not Gaussian, we use
Cramér's representation to transform a time-domain stochastic
process into a stochastic process in the frequency domain.
III. MUTUAL INFORMATION IN FREQUENCY
We first define MI between frequencies within a pro-
cess and between two processes in continuous time. We
then extend this definition to discrete-time stochastic pro-
the increments of
spectral processes or the Cramér's representation of X(t)
and Y (t) at frequencies νi and νj respectively. Let the joint
probability density of the four dimensional random vector
cesses. Consider d(cid:101)X (νi) and d(cid:101)Y (νj),
of the real and imaginary parts of d(cid:101)X (νi) and d(cid:101)Y (νj)
be denoted by P(cid:0)d(cid:101)XR (νi) , d(cid:101)XI (νi) , d(cid:101)YR (νj) , d(cid:101)YI (νj)(cid:1).
P(cid:0)d(cid:101)XR (νi) , d(cid:101)XI (νi)(cid:1), P(cid:0)d(cid:101)YR (νj) , d(cid:101)YI (νj)(cid:1). The MI-in-
= I(cid:0)(cid:8)d(cid:101)XR (νi) , d(cid:101)XI (νi)(cid:9);(cid:8)d(cid:101)YR (νj) , d(cid:101)YI (νj)(cid:9)(cid:1),
P(cid:0)d(cid:101)XR(νi),d(cid:101)XI (νi),d(cid:101)YR(νj ),d(cid:101)YI (νj )(cid:1)
P(cid:0)d(cid:101)XR(νi),d(cid:101)XI (νi)(cid:1)P(cid:0)d(cid:101)YR(νj ),d(cid:101)YI (νj )(cid:1)(cid:27)
(cid:26)
frequency between X (t) at νi and Y (t) at νj is defined as
The corresponding two-dimensional marginal densities are
MIXY (νi, νj)
= E
(2)
log
,
(3)
(cid:0)νi, νj
where I ({·,·} ;{·,·}) is the standard mutual
information
between two pairs of two dimensional real-valued random
vectors [34]. The MI between two different frequencies νi, νj
in the same process Y (t) is similarly defined as
(cid:0)νj
(cid:1)(cid:9)(cid:1).
that(cid:2)d(cid:101)YR (ν) , d(cid:101)YI (ν)(cid:3) is a continuous-valued random vector
MIY Y
The MI between the components of Y at frequencies νi =
is ∞, a consequence of the fact
νj = ν, MIY Y (ν, ν),
(cid:0)νi
(cid:1)(cid:9);(cid:8)d(cid:101)YR
(cid:1)=I(cid:0)(cid:8)d(cid:101)YR
(cid:0)νj
(cid:1),d(cid:101)YI
(cid:0)νi
(cid:1),d(cid:101)YI
whose conditional differential entropy is not lower bounded.
MI-in-frequency defined in (2), (3) is a non-negative number. If
MI-in-frequency between two frequencies is zero, then they are
independent and if not, MI-in-frequency is a measure of the sta-
tistical dependence between the two frequency components. MI-
in-frequency between two processes is not symmetric in general,
i.e., MIXY (νi, νj) (cid:54)= MIXY (νj, νi). However, it is symmetric
within a process, i.e., MIY Y (νi, νj) = MIY Y (νj, νi).
X (t) = A cos (2πν0t + Θ) and Y (t) = X (t)2. Then d(cid:101)Y (ν)
Example: Continuing with our example in section II, let
3
is zero except at ν = 0, where the spectral increment is A2
2 ,
4 exp (±j2Θ).
and at ν = ±2ν0, where the increment is A2
As a result, the frequency components at ±ν0 in X and at
frequencies {0,±2ν0} in Y are statistically dependent and
hence the MI-in-frequency obtained from (2) at these frequency
pairs will be positive. In addition, the frequency components
in Y at ν ∈ {0,±2ν0} are dependent and hence the MI-in-
frequency within Y at these frequencies will also be positive.
A. Gaussian Inputs to LTI Filters
Let's now consider the special case where X (t), a Gaussian
process with power spectral density sX (ν) serves as the input
to a linear, time-invariant (LTI) filter with transfer function
H1 (ν) and Y (t) is output observed in additive colored noise
(white noise W (t) passed through a LTI filter with transfer
function H2 (ν)). The processes X (t) and Y (t) are related
by
y (t) = h1 (t) ∗ x (t) + h2 (t) ∗ w (t) ,
(4)
where ∗ denotes convolution operation, x(t), y(t) and w(t) are
sample paths of X (t), Y (t) and W (t) respectively. W is
a Gaussian process with power spectral density sW (ν) and
independent of X. h1(t) and h2(t) are continuous-time impulse
responses of LTI filters, whose transfer functions are H1 (ν)
and H2 (ν) respectively. Let d(cid:101)X (ν), d(cid:102)W (ν) and d(cid:101)Y (ν) be
the spectral process increments of the Gaussian processes X,
W and Y . We have from Theorem 2,
(cid:2)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:3)∼N(cid:0)0, 1
(cid:2)d(cid:102)WR (ν) , d(cid:102)WI (ν)(cid:3)∼N(cid:0)0, 1
2 sX (ν) I(cid:1) ,
2 sW (ν) I(cid:1) ,
(5)
where N (µ, Σ) represents Gaussian distribution with mean µ
and covariance Σ, 0 is the two element zero vector and I is
the 2 × 2 identity matrix. In addition, we can show for the
model in (4) that
d(cid:101)Y (ν) = H1 (ν) d(cid:101)X (ν) + H2 (ν) d(cid:102)W (ν) .
(6)
The proof of (6) is in the appendix. The MI-in-frequency
defined in (2) is further simplified for the model in (4) using
(5), (6) and stated in the following theorem.
Theorem 3. For the model given in (4), the MI between X (t)
at frequency νi and Y (t) at frequency νj is zero, when νi (cid:54)= νj
and the MI between X (t) and Y (t) at frequency νi = νj =
ν (cid:54)= 0 is
MIXY (ν, ν) = 2 × I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1)
equal to I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1), which is just half
The proof of the above theorem is in the appendix. Note
that at ν = 0, the MI-in-frequency between X and Y is
= log(cid:0)1 +
H1(ν)2sX (ν)
H2(ν)2sW (ν)
(cid:1).
of the right hand side of (7). We intuitively expect different
frequency components in the Gaussian input and its output
from a linear system to be independent and Theorem 3 confirms
that the proposed definition of MI-in-frequency agrees with
this intuition. In addition, the MI between X and Y is ∞
when H2 (ν) = 0, since the components of X and Y at
(7)
4
such ν are linearly related. The MI between two different
frequencies in Y (t), generated from (4),
is zero due to
the linearity of the filters and Gaussian inputs. Furthermore,
we can also show for the Gaussian processes X and Y
related by (4) that MI-in-frequency is related to coherence
CXY (ν) ∈ [0, 1], by MIXY (ν, ν) = − log (1 − CXY (ν)).
The proof is in the appendix. This result implies MI-in-
frequency between Gaussian processes related by (4) can
be estimated with the coherence. In addition, Theorem 3
also shows that MI-in-frequency between Gaussian processes
related by (4) can be estimated by estimating the mutual
information between (cid:2)d(cid:101)XR (νi) , d(cid:101)XI (νi)(cid:3) and d(cid:101)YR (νj), a
three dimensional estimate as opposed to a four dimensional
estimate in general.
B. Discrete-time Stochastic Processes
We now extend the definition of MI-in-frequency between
continuous-time stochastic processes in (2), (3) to discrete-
time stochastic processes. In practice, we only have access
to data samples from a real-valued, discrete-time stochastic
process, sampled at a given Nyquist sampling frequency Fs.
Sampled signals have periodic spectra, with a period equalling
Fs. In addition, components in the process with frequencies in
the range [Fs/2, Fs] correspond to negative frequencies [35].
Therefore, the actual frequency content in the signal is confined
∈ [0, 0.5] to
to [0, Fs/2]. We use normalized frequency λ = ν
Fs
describe the frequency axis in case of discrete-time stochastic
processes, instead of ν which was used for continuous-time
stochastic processes. The MI-in-frequency between discrete-
time processes is therefore obtained by replacing νi, νj by
the normalized frequencies λ1, λ2 ∈ [0, 0.5]
in (2), (3).
Multivariate autoregressive models, commonly used to model
electro-physiological signals recorded from brain [20], [36],
are a special case of the discrete-time equivalent of (4). The
analytic expression for MI at frequency λ for such discrete-
time Gaussian processes is therefore similarly obtained by
replacing the frequencies ν by λ in (7), which is also equal
to − log (1 − CXY (λ)). This shows that for the special case
of discrete-time Gaussian processes, MI-in-frequency metric is
equivalent to coherence and the definitions in [24], [25].
IV. DATA-DRIVEN ESTIMATION OF MI-IN-FREQUENCY
We describe two data-driven estimators–a kernel density
based (KDMIF) and a nearest neighbor based (NNMIF)
λi component of X and λj component of Y . The input to
both these algorithms are the N samples of X and Y . The
first step in both KDMIF and NNMIF estimators involves
estimator to estimate MI-in-frequency, (cid:99)MIXY (λi, λj), between
estimating the samples of spectral process increments d(cid:101)X (λi)
and d(cid:101)Y (λj), of X at λi and of Y at λj respectively. In the
MI from the samples of spectral process increments, d(cid:101)X (λi)
and d(cid:101)Y (λj).
second step, the KDMIF estimator uses the kernel density based
MI estimator [32], [37], whereas NNMIF estimator uses the
k-nearest neighbor based MI estimator [32], [38] to estimate
A. Kernel Density Based MI-in-frequency (KDMIF) Estimator
1) Estimation of Samples of Spectral Process Increments:
The first step of the algorithm is estimating the samples of
spectral process increments of X and Y from N dependent
data samples. We assume there is a finite memory in both
these processes and choose a value for a parameter Nf , which
is much larger than the length of dependence or memory in
the data and determines the frequency resolution of our MI-in-
frequency estimates. We assume data in different windows are
independent of each other. Ideally, consecutive windows should
be separated to ensure no dependence across windows and
avoid the dependence across the window boundaries, but our
simulation results demonstrate that not separating the windows
doesn't affect performance significantly. N samples of X are
split into Ns non-overlapping windows with Nf = N
data
Ns
points in each window. Let us denote the samples in lth window
of X and Y respectively by two Nf element one-dimensional
vectors, xl and yl, for l = 1, 2,··· , Ns.
Fourier transform (DTFT) of xl at normalized frequency α.
For λi = i
Nf
variable d(cid:101)X (λi). Let F(cid:8)xl(cid:9) (α) denote the discrete-time
∈ [0, 1] and i ∈ [0, Nf − 1], let us define d(cid:101)xl (λi)
and integrated Fourier spectrum, (cid:101)xl (λi), by
d(cid:101)xl (λi) = F(cid:8)xl(cid:9) (λi) and(cid:101)xl (λi) =
F(cid:8)xl(cid:9) (λm) . (8)
i(cid:80)
It is stated in [29] that the random variable for which (cid:101)xl (λi)
Let us now focus on estimating samples of the random
m=0
R (λi) + id(cid:101)xl
is just one realization, tends to the spectral process of X
at λi in mean of order γ, for any γ > 0, as the number
of samples goes to infinity and assuming the underlying
distribution is stationary and satisfies a mixing assumption.
to obtain the Ns samples of the spectral process increments of
, j ∈ [0, Nf − 1] and the resulting samples are
Y at λj = j
Nf
the spectral process increments of X at λi = i
. The lth
Nf
I (λi), is a
and λi + dλ, is just the DTFT of the samples in window
l. Calculating the DTFT with the FFT for each of the Ns
windows separately yields an Nf × Ns matrix, whose ith
Also, d(cid:101)xl (λi), which is the increment in (cid:101)xl (λi) between λi
row, d(cid:101)x (λi) = (cid:2)d(cid:101)x1 (λi) , d(cid:101)x2 (λi) ,··· , d(cid:101)xNs (λi)(cid:3) is the
complex-valued vector containing Ns samples of d(cid:101)X (λi),
element of d(cid:101)x (λi), d(cid:101)xl (λi) = d(cid:101)xl
particular realization of d(cid:101)X (λi). A similar procedure is used
denoted by d(cid:101)y (λj) =(cid:2)d(cid:101)y1 (λj) , d(cid:101)y2 (λj) ,··· , d(cid:101)yNs (λj)(cid:3).
(cid:0)λi
(cid:0)d(cid:101)xl
P(cid:0)d(cid:101)XR (λi) , d(cid:101)XI (λi) , d(cid:101)YR (λj) , d(cid:101)YI (λj)(cid:1). The density is
the two-dimensional densities, P(cid:0)d(cid:101)XR (λi) , d(cid:101)XI (λi)(cid:1) and
2) Estimating MI-in-frequency: The MI-in-frequency
estimate
samples,
for
l = 1, 2,··· , Ns, using a kernel density based plug-in
nonparametric estimator [32]. The Ns data samples are split
into Ntr training and Nts test samples. The training data is
used to estimate the four-dimensional joint probability density
estimated using a kernel density estimator with Gaussian
kernels, the optimal bandwidth matrix selected using smoothed
cross-validation criterion [37] and implemented using 'ks'
package in R [39]. The joint density is marginalized to estimate
(cid:1), d(cid:101)xl
(cid:1), d(cid:101)yl
I
(cid:0)λj
(cid:1)(cid:1),
(cid:1)(cid:1)
(cid:0)λi
from the Ns
(cid:0)d(cid:101)yl
R
(cid:0)λj
is
now obtained
R
I
and
P(cid:0)d(cid:101)YR (λj) , d(cid:101)YI (λj)(cid:1), by recognizing that the bandwidth
matrix for the two-dimensional marginal is the appropriate
2 × 2 sub-matrix from the 4 × 4 bandwidth matrix of the joint
density. The estimates of the joint and the marginal densities
at the Nts test samples are plugged into the following equation
(9) to estimate MI-in-frequency.
(cid:99)MIXY (λi, λj)
(cid:98)P(cid:0)d(cid:101)xl
(cid:98)P(cid:0)d(cid:101)xl
R(λi),d(cid:101)xl
R(λi),d(cid:101)xl
(cid:80)
= 1
Nts
log
l
I (λj )(cid:1)
I (λj )(cid:1) .
R(λj ),d(cid:101)yl
R(λj ),d(cid:101)yl
I (λi)(cid:1)(cid:98)P(cid:0)d(cid:101)yl
I (λi),d(cid:101)yl
B. Nearest Neighbor Based MI-in-frequency (NNMIF) Estima-
tor
1) Estimation of Samples of Spectral Process Increments:
to two-dimensional
process increments of X at λi and Y at λj respectively.
The first step in the nearest neighbor based MI-in-frequency es-
timator is exactly same as that of KDMIF estimator. Following
between two data points with indices l1, l2 ∈ [1, Ns]
according to
is
2) Estimating MI-in-frequency: MIXY (λi, λj) is now
l = 1, 2,··· , Ns using nearest neighbor based MI estimator
the algorithm in
[38]. We apply the first version of
[38]
the steps described in section IV-A1, we estimate d(cid:101)xl (λi) and
d(cid:101)yl (λj), for l = 1, 2,··· , Ns, the Ns samples of the spectral
estimated from d(cid:101)xl (λi) ∈ R2 and d(cid:101)yl (λj) ∈ R2, for
random variables d(cid:101)X (λi) and
d(cid:101)Y (λj) to compute (cid:99)MIXY (λi, λj). Consider the joint four
dimensional space (cid:0)d(cid:101)X (λi) , d(cid:101)Y (λj)(cid:1) ∈ R4. The distance
max(cid:8)(cid:107)d(cid:101)xl1 (λi) − d(cid:101)xl2 (λi)(cid:107),(cid:107)d(cid:101)yl1 (λj) − d(cid:101)yl2 (λj)(cid:107)(cid:9).
(cid:0)d(cid:101)xl (λi) , d(cid:101)yi (λj)(cid:1) and its K th nearest neighbor,
y denote the number of samples of d(cid:101)X (λi) and
d(cid:101)Y (λj) within an infinity norm ball of radius less than l
centered at d(cid:101)xl (λi) and d(cid:101)yi (λj) respectively. From [38], the
(cid:99)MIXY (λi, λj) = ψ (K) + ψ (Ns)
(cid:0)ψ(cid:0)nl
l denote the distance between the data sample
Let
for
l = 1, 2,··· , Ns. We used K = 3 in this paper [30]. Let
x and nl
nl
MI-in-frequency between X and Y at normalized frequencies
λi and λj is given by
x + 1(cid:1) + ψ(cid:0)nl
y + 1(cid:1)(cid:1) ,
calculated using the
infinity norm,
Ns(cid:80)
− 1
(10)
Ns
l=1
where ψ (·) is the Digamma function.
C. Significance Testing
The statistical significance of the MI-in-frequency estimates
obtained from both KDMIF and NNMIF estimators is now
tested using the following procedure. We permute the samples
in the vector d(cid:101)x (λi) randomly and estimate the MI-in-
of d(cid:101)Y (λj). Instead of adding random phase or permuting
frequency between the permuted vector and the Ns samples
the phase time series, which are typically used to test the
statistical significance of phase-amplitude coupling metrics
[40], we permute the samples of spectral process increments
since our metric can detect coupling across phase and amplitude
jointly. This process is repeated Np times to obtain Np
permuted MI-in-frequency estimates, under the null hypothesis
5
of independence. The permuted MI estimates will be almost
zero since the permutations make the spectral processes almost
independent. If the actual MI estimate, (cid:99)MIXY (λi, λj), is
judged larger than all the permuted Np estimates, then there
is a statistically significant dependence between the processes
at these frequencies.
(9)
V. MI BETWEEN DATA WITH TEMPORAL DEPENDENCIES
We now use MI-in-frequency to estimate mutual information
between dependent data. The data-driven MI estimator, summa-
rized in Algorithm 1, takes in N samples of X and Y as input
and outputs the mutual information between X and Y , I (X; Y ),
,
∀ (i, j) such that i, j ∈ [0, Nf − 1].
by estimating (cid:99)MIXY (λi, λj), where λi = i
, λj = j
Nf
Nf
Algorithm 1: Mutual Information Estimator
Data: (x [n] , y [n]), for x [n] , y [n] ∈ R, n ∈ [0, N − 1].
Result: I (X; Y )
Algorithm:
A) Estimate (cid:99)MIXY (λi, λj) at all possible pairs (λi,λj),
(cid:0)λip , λjq
(cid:1) is
using either the KDMIF or the NNMIF estimator.
there exists a λjq ∈ Λy such that (cid:99)MIXY
Identify the sets Λx, Λy, such that for each λip ∈ Λx
B) Let d(cid:101)X (Λx) =(cid:2)d(cid:101)X (λi1 ) ,··· , d(cid:101)X (λiP )(cid:3) ∈ R2P ,
d(cid:101)Y (Λy) =(cid:2)d(cid:101)Y (λj1) ,··· , d(cid:101)Y (cid:0)λjQ
(cid:1)(cid:3) ∈ R2Q. The
(cid:17)
d(cid:101)X (Λx) ; d(cid:101)Y (Λy)
statistically significant and vice-versa. Let P, Q
respectively denote the cardinality of Λx, Λy.
mutual information between X and Y is given by
I (X; Y ) =
(cid:16)
I
1
,
max(P,Q)
where the right hand side is estimated from Ns i.i.d.
samples using any nonparametric MI estimator [32].
A. Identifying Coupled Frequencies
Nf
MI-in-frequency, (cid:99)MIXY (λi, λj), between λi = i
The first step in our MI estimator involves estimating the
frequency
component in X and λj = j
component in Y , for all
(i, j) such that i, j ∈ [0, Nf − 1] using either the KDMIF
Nf
(section IV-A) or the NNMIF (section IV-B) algorithms.
Statistical significance of the resulting estimates is assessed
using the procedure described in section IV-C. The resultant
MI-in-frequency estimates across all frequency pairs can be
graphically visualized by plotting the statistically significant
MI-in-frequency estimates on a two-dimensional image grid,
whose rows and columns correspond to frequencies of X and
Y respectively. Let Λx and Λy respectively denote the set
of frequency components of X and Y , such that for each
(cid:99)MIXY
λip ∈ Λx, there exists at least one λjq ∈ Λy for which
(cid:1) is statistically significant and vice-versa.
(cid:0)λip , λjq
6
B. Estimating Mutual Information
The final step in our algorithm estimates MI between the
spectral process increments of X and Y at frequencies in Λx
and Λy respectively. With P, Q denoting the cardinality of
Λx, Λy respectively, let d(cid:101)X (Λx) and d(cid:101)Y (Λy) denote the 2P
and 2Q-dimensional random vector comprising the spectral
process increments of X, Y at all frequencies in Λx and Λy
respectively. We already computed Ns i.i.d. samples of these
two random vectors to estimate MI-in-frequency estimates in
the previous step of this algorithm. The desired MI estimate
is computed from the mutual information between d(cid:101)X (Λx)
and d(cid:101)Y (Λy), which is estimated using the k-nearest neighbor
based estimator developed in [38], according to
(cid:16)
(cid:17)
d(cid:101)X (Λx) ; d(cid:101)Y (Λy)
.
(11)
I (X; Y ) =
1
max (P, Q)
I
The MI estimator in (11) can be further simplified for
discrete-time Gaussian processes. Without loss of generality,
consider two Gaussian processes X and Y , related by
y[n] = h1[n] ∗ x[n] + h2[n] ∗ w[n],
(12)
where h1[n], h2[n] are linear time-invariant (LTI) filters and
W is white Gaussian noise independent of X. For the model
in (12), which is the discrete-time equivalent of (4), the data-
driven estimation in (11) can be further simplified to
(cid:99)MIXY (λi; λi) , where λi = i
.
(13)
Nf
I (X; Y ) = 1
Nf
Nf /2(cid:80)
i=0
This result is obtained because linear models do not introduce
cross-frequency dependencies and because negative frequencies
do not carry any extra information. Furthermore, the relation-
ship between the MI and the MI-in-frequency for two processes
related by (12) is stated in the following theorem.
0.5(cid:82)
Theorem 4. Consider two discrete-time Gaussian stochastic
processes X and Y related by (12). The mutual information
between these processes, a scalar, is given by
I (X; Y ) =
MIXY (λ, λ) dλ.
(14)
0
The proof of the above theorem is in the appendix. This
theorem means that MI between two Gaussian processes over
the entire time can be obtained by integrating the contribution
from each frequency component. It is easy to see that the right
hand side of (13) is just the Riemann sum of the integral on the
right hand side of (14), which converges to the true value as
Nf tends to infinity. This implies our MI estimator converges
to the true value for discrete-time Gaussian processes.
Note that the MI estimation algorithm does not make any
parametric assumptions on the underlying model between
X and Y . The computation of MI via (11) can be greatly
simplified by clustering the frequencies in Λx and Λy into
groups such that there are no significant dependencies across
groups and using the chain rule of mutual information. In
addition, if we observe after the first step that significant MI-
in-frequency estimates occur only at (λi, λi) ,∀i∈(cid:2)0, Nf − 1(cid:3),
then the MI can be estimated using (13).
Finally, as we mentioned earlier, MI estimation between
Gaussian processes is a solved problem in the sense that we
can analytically compute it if the covariance of the Gaussian
processes is known [3] and there are several estimators whose
performance is thoroughly analyzed [4]. MI in frequency for
Gaussian processes is analyzed by Brillinger [24]. In this paper,
we extended Brillinger's work to define MI-in-frequency for
any process. In the following section, we use simulated data to
validate that the extensions we proposed to any process in this
paper are still in agreement with the prior work on Gaussian
processes and also work for non-Gaussian processes.
VI. PERFORMANCE EVALUATION ON SIMULATED DATA
The performance of the data-driven MI-in-frequency and
mutual information estimators described in section IV and
section V respectively is validated on simulated data. The
statistical significance of the estimates was assessed using the
procedure described in section IV-C. In addition, we compare
the performance of the MI-in-frequency estimators against
modulation index [8], [9], [22], a commonly used phase-
amplitude coupling metric in neuroscience.
A. Comparing the KDMIF and NNMIF Estimators
Consider two stochastic processes X and Y , where X is a
white Gaussian process with standard deviation σx and Y is
obtained by
y[n] = h[n] ∗ x[n] + w[n],
(15)
where W is a white Gaussian process with standard deviation
σw that is independent of X and h[n] is a linear time-invariant
filter. We compared the performance of the kernel density
based and nearest neighbor based estimators by benchmarking
the estimates against the true value of MI-in-frequency and
the mutual information between X and Y for the model in
(15). We used two different filers: a two-tap low pass filter,
h[n] = [β, 1 − β] , for β ∈ [0, 1] and a 33-tap bandpass filter
with passband in [0.15, 0.35] normalized frequency range. We
observed that modulation index, a popular CFC metric, was
unable to correctly detect and quantify the strength of cross-
frequency coupling for both these models.
1) Lowpass Filter: The samples of X and Y are generated
from (15) with σx = σw = 1 and a lowpass filter with unit-
impulse response [β, 1 − β], for various values of β ∈ [0, 1].
The true value of MI-in-frequency at normalized frequency
λ ∈ [0, 0.5] is obtained substituting the parameters of this
model in (7) and is plotted in Fig. 1a for β = 0.5. In addition,
the MI-in-frequency estimated by the KDMIF and NNMIF
algorithms from N = 64 × 104 data samples, with Nf =
64, Ns = 104 is also plotted in Fig. 1a. It is seen that the
estimates from both algorithms follow the true value closely,
without the knowledge of the underlying model. In addition,
we evaluate the bias and the rate of convergence of both these
algorithms as a function of Ns, with Nf = 64 in Fig. 1b.
The bias is defined as the average value of the ratio of MI-in-
frequency estimate and its true value in the passband of the
lowpass filter. We observe that the NNMIF algorithm converges
faster and has lower bias than the KDMIF algorithm. We now
7
(a) MI-in-frequency
(b) Bias of the Estimators
(c) Mutual Information
Fig. 1. Comparing the performance of the kernel density based and nearest neighbor based estimators, KDMIF and NNMIF respectively, on simulated
generated from (15) using a two-tap lowpass filter. In Fig. 1a, the MI-in-frequency estimates obtained from KDMIF and NNMIF estimators along with the true
value of MI-in-frequency are plotted against the normalized frequency λ for β = 0.5. Fig. 1b plots the bias (mean of the ratio of the estimate and the true
value in the filter passband) against the number of data samples used for estimation for β = 0.5. Fig. 1c plots the MI estimate between X and Y obtained
from kernel density and nearest neighbor algorithms along with the true value of MI for β ∈ [0, 1].
(a) MI-in-frequency
(b) Bias of the Estimators
(c) Mutual Information
Fig. 2. Comparing the performance of the kernel density based and nearest neighbor based estimators, KDMIF and NNMIF respectively, on simulated generated
from (15) using a 33-tap bandpass filter with passband in [0.15, 0.35] normalized frequency. In Fig. 2a, the MI-in-frequency estimates obtained from KDMIF
and NNMIF estimators along with the true value of MI-in-frequency are plotted against the normalized frequency λ for σw = 1. Fig. 2b plots the bias (mean
of the ratio of the estimate and the true value in the filter passband) against the number of data samples used for estimation for σw = 1. Fig. 2c plots the plots
the MI estimate between X and Y from kernel density and nearest neighbor algorithms along with the true value of MI for different values of σw ∈ [0.5, 2].
use both these algorithms to estimate the mutual information
between X and Y for β ∈ [0, 1]. The analytical expression for
the true value of MI1 for this model is derived in [41]. It is
evident from Fig. 1c that the MI estimates obtained from the
nearest neighbor based estimator is closer to the true value
than those from the kernel density based estimator.
2) Bandpass Filter: The samples of X are generated from
a standard white Gaussian random process with σx = 1 and
those of Y are generated from (15) using a 33-tap finite-
impulse-response bandpass filter with passband in [0.15, 0.35]
normalized frequency range for different values of noise
standard deviation, σw ∈ [0.5, 2]. We used the kernel density
and the nearest neighbor based algorithms to estimate the MI-
in-frequency and the mutual information between X and Y .
The true value of MI-in-frequency is obtained from (7) and
of mutual information is numerically calculated using power
spectral density (chapter 10 in [34]). It is clear from Fig. 2b
that the nearest neighbor based algorithm converges to the true
value faster than the kernel density based algorithm. The nearest
neighbor based algorithm also provides more accurate estimates
of both MI-in-frequency and mutual information between X
1Note that for this particular model, mutual information is equal to the
directed information from X to Y and the analytical expression is given in
equation (18) in [41].
and Y , as evident from Fig. 2a, Fig 2c respectively. In addition,
nearest neighbor based MI-in-frequency algorithm runs faster
than kernel density based algorithm. We, therefore, conclude
that the nearest neighbor based MI-in-frequency algorithm
outperforms kernel density based algorithms and only depict
the results obtained from nearest neighbor based algorithm in
the remainder of the paper.
B. Comparison with Modulation Index
We now compare the effectiveness of MI-in-frequency
against modulation index in detecting cross-frequency coupling,
using the simulated model commonly used to validate CFC
metrics [21], [22], [42]. Modulation index quantifies the rela-
tionship between the phase and amplitude envelopes extracted
by the Hilbert transform [8]. Consider two random cosine
waves, sl[n] and sh[n], at frequencies fl and fh respectively.
Let fs denote the sampling frequency. The samples of time-
series X and Y are generated from the following model:
n + θ
, sh[n] = A cos
sl[n] = A cos
x[n] = sl[n] + w1[n], y[n] = (1 + sl[n]) sh[n] + w2[n], (16)
where A is a Rayleigh random variable with parameter 1 and
θ is a uniformly distributed random variable between 0 and
2π fh
fs
n + θ
(cid:16)
2π fl
fs
(cid:17)
(cid:16)
(cid:17)
600.250.5MIbetweenXandYat600.350.7TrueValueKDMIFNNMIFNumberofSamples,Ns102103104Bias0.60.811.1KDMIFNNMIF-0 0.51 Mutual Information0.160.260.36TrueValueKernelDensityBasedNearestNeighborBased600.250.5MIbetweenXandYat600.40.8TrueValueKDMIFNNMIFNumberofSamples,Ns102103104Bias0.60.81KDMIFNNMIF<w0.51 1.52 Mutual Information0 0.20.4TrueValueKernelDensityBasedNearestNeighborBased 8
(a) MI-in-frequency
(b) Modulation Index
(c) MI-in-frequency
(d) Modulation Index
Fig. 3. Comparing the performance of MI-in-frequency against modulation index in detecting cross-frequency coupling in data generated from (16). In Fig. 3a
and Fig. 3b, MI-in-frequency estimates obtained from nearest neighbor algorithm and modulation index are plotted respectively, when fl = 5 Hz and fh = 60
Hz in (16). Fig. 3c and Fig. 3d respectively plot the MI-in-frequency estimates and modulation index estimates, when fl = 15 Hz and fh = 60 Hz in (16).
(a) (cid:99)MIY Y (λi, λj )
(b) (cid:99)MIXY (λi, λj )
(c) MI between X and Y
Fig. 4.
(a) MI-in-frequency estimates from the nearest neighbor based algorithm between the frequency components within the random processes Y , obtained
from the single cosine data-generation model, (18) with σw = 1. Note that the MI-in-frequency estimates along the principal diagonal are not plotted, since
they are equal to ∞. (b) MI-in-frequency estimates between random processes X and Y related by the single cosine data-generation model with σw = 1.
It is clear that MI-in-frequency estimator correctly identifies the pairwise frequency dependencies. (c) MI-in-frequency between X at λ0 and Y at 2λ0,
(cid:99)MIXY (λ0, 2λ0), obtained from (10) along with the MI estimate between X and Y , I (X; Y ), obtained from Algorithm 1 for various values of the noise
standard deviation, σw.
2π that is independent of A. w1[n], w2[n] are samples of i.i.d
white Gaussian noise process with standard deviation 1. We
generated samples from this model with fl = 5 Hz, fh = 60
Hz and fs = 200 Hz. MI-in-frequency between X and Y is
estimated using the nearest neighbor based algorithm from
N = 40 × 104 samples with Ns = 104 and plotted in Fig. 3a.
Modulation index between X and Y estimated by using the
Matlab toolbox [21], with the amplitude envelope estimated
by the Hilbert transform and is plotted in Fig. 3b. It is clear
that both MI-in-frequency and modulation index successfully
detect the cross-frequency coupling between 5 Hz component
of X and {55, 60, 65} Hz components of Y for these parameter
values. We then generated X and Y from (16) with fl = 15
Hz and all other parameter values unchanged. Fig. 3c plots the
MI-in-frequency estimates obtained via NNMIF algorithm and
as expected, we detect the CFC between 15 Hz component of
X and {45, 60, 75} Hz components of Y . However, modulation
index, depicted in Fig. 3d, was not able to correctly detect the
CFC between X and Y for these parameter values. In addition,
the strength of the modulation index decreased from around
0.5 when fl = 5 Hz in Fig. 3b to 0.05 when fl = 15 Hz in
Fig. 3d. This is because metrics like modulation index can only
detect the CFC correctly with good frequency resolution only
when one of the frequencies involved is very small compared
to the other frequency. Otherwise, the bandwidth of the filter
used to extract the phase and the amplitude envelope should
be larger, which will reduce the frequency resolution in the
estimated CFC (note the smearing in Fig. 3d, when compared to
Fig. 3b) [22], [42]. In addition, we tested modulation index on
data generated from (15) and (17) and found that modulation
index is unable to detect the cross-frequency coupling for these
relationships. This is not surprising since the modulation index
like metrics are tuned to detect CFC when the underlying
coupling is of the form in (16), whereas the MI-in-frequency
defined in this paper overcomes this shortcoming, as evident
from its performance on various simulated models.
C. Nonlinear Models
We now consider square nonlinearity, where the random
processes X and Y are related by
y[n] = x[n]2 + w[n],
(17)
where w[n] is white Gaussian noise with standard deviation
σw. Modulation index was not able to detect and quantify
the cross-frequency coupling for this model. We estimated
the MI-in-frequency between frequency components within Y ,
(cid:99)MIY Y (λi, λj), between the frequency components of X and
Y , (cid:99)MIXY (λi, λj), and the mutual information between X and
Y , I (X; Y ), from N = 32 × 104 samples of X and Y with
Ns = 104, for different values of noise standard deviation,
σw ∈ [0, 10]. Computing the true value of MI-in-frequency and
mutual information is nontrivial because of the nonlinearity. The
performance of the algorithms is assessed by checking if they
detect the cross-frequency coupling at expected frequency pairs
Frequency of X (Hz)5 153045607590Frequency of Y (Hz)5 15304560759001.22.4Frequency of X (Hz)5 153045607590Frequency of Y (Hz)5 15304560759000.240.48Frequency of X (Hz)5 153045607590Frequency of Y (Hz)5 15304560759001.22.4Frequency of X (Hz)5 153045607590Frequency of Y (Hz)5 15304560759000.0240.0486j(#2!5cycles=sample)0 4 8 12166i(#2!5cycles=sample)0 4 8 121601.12.26j(#2!5cycles=sample)0 4 8 12166i(#2!5cycles=sample)0 4 8 121602.24.4<w0 5 10Mutual Information0.41.62.812cMIXY(60;260)^I(X;Y) 9
(a) (cid:99)MIY Y (λi, λj )
(b) (cid:99)MIXY (λi, λj )
(c) MI between X and Y
Fig. 5.
(a) MI-in-frequency estimates from the nearest neighbor based algorithm between the frequency components within the random processes Y ,
obtained from the two cosine data-generation model, (19). The MI-in-frequency estimates are not plotted along the diagonal, since they are equal to ∞. (b)
MI-in-frequency estimates between random processes X and Y related by the two cosine data-generation model. It is clear that MI-in-frequency estimator
correctly identifies the pairwise frequency dependencies between X and Y . (c) I (X; Y ), the MI estimate between X and Y obtained from Algorithm 1 for
various values of the noise standard deviation, σw.
and by checking if the mutual information estimates decrease
with increasing noise power as expected. We considered two
different models for the stochastic process X, such that its
samples are dependent across time.
1) Random Cosine with Squared Nonlinearity: The samples
of X are generated from a random cosine wave,
x[n] = A cos (2πλ0n + θ) ,
(18)
where A is a Rayleigh random variable with parameter 1,
θ is a uniform random variable between 0 and 2π that is
independent of A and λ0 = 4
32. It is easy to see that frequency
components of X are statistically independent and this is
confirmed by the NNMIF estimator. However, because of
the square nonlinearity in (17), the DC component of Y and
the 2λ0 component of Y will be statistically dependent and
this is confirmed by Fig. 4a, which plots the MI-in-frequency
between components of Y generated with σw = 1 using
the NNMIF algorithm. The common information between
these two processes will be present between λ0 component of
X and the {0, 2λ0} components of Y . This cross-frequency
dependence is confirmed by Fig. 4b, which plots the estimates
of MI-in-frequency between X and Y obtained by the NNMIF
algorithm from (10): we observe that significant dependencies
occur only at (λ0, 0) and (λ0, 2λ0) frequency pairs. As a
result, P = 1, Q = 2. The MI estimate from Algorithm 1,
I (X; Y ) = 1
is plotted in
2
Fig. 4c. The MI estimate decreases with increasing σw as
expected. In addition, we note for this model that the DC
component of Y does not contain any extra information about
X, given the 2λ0 component of Y . Therefore, we expect
I
1
2
sult verified in Fig. 4c, since the two curves are very close.
d(cid:101)X(λ0);(cid:8)d(cid:101)Y (0), d(cid:101)Y (2λ0)(cid:9)(cid:17)
2(cid:99)MIXY (λ0; 2λ0), a re-
(cid:16)
d(cid:101)X(λ0);(cid:8)d(cid:101)Y (0), d(cid:101)Y (2λ0)(cid:9)(cid:17)
(cid:16)
I
= 1
2) Two Random Cosines with Squared Nonlinearity: The
samples of random process X are generated according to
x[n] = A1 cos (2πλ1n + θ1) + A2 cos (2πλ2n + θ2) ,
(19)
where A1, A2 are independent Rayleigh random variables
with parameter 1, θ1, θ2 are independent uniformly distributed
random variables between 0 and 2π that are independent of
32 , λ2 = 6
A1, A2, and λ1 = 4
32. As before, the frequency
components of X are statistically independent. However, after
some basic algebra, it is easy to see that the all possible pairs of
frequency components of Y in {0, λ2 − λ1, 2λ1, λ2 + λ1, 2λ2}
are statistically dependent, except for (2λ1, 2λ2) frequency pair,
and we expect to see statistically significant MI-in-frequency
estimates between these frequency components. This is con-
firmed by Fig. 5a, which plots the MI-in-frequency estimates
within Y , generated with σw = 1 and obtained by the NNMIF
algorithm. The pairwise frequency dependencies between X
and Y occur at (λ1, 0), (λ1, λ2 − λ1), (λ1, 2λ1), (λ1, λ2 + λ1),
(λ2, 0), (λ2, λ2 − λ1), (λ2, λ2 + λ1) and (λ2, 2λ2). Fig. 5b
plots the estimates of pairwise MI-in-frequency between X
and Y generated with σw = 1 and obtained by the data-
driven NNMIF algorithm using (10). The algorithm correctly
identifies all the dependent frequency pairs and P = 2, Q = 5.
We then apply the algorithm described in section V and plot
the estimates the MI for different values of noise standard
deviation σw in Fig. 5c. Again, the MI decreases with increasing
noise power, as expected. These different models validate
the superiority of MI-in-frequency over other existing metrics
to detect cross-frequency coupling and also demonstrate the
performance and accuracy of the data-driven MI-in-frequency
and MI estimators.
VII. CFC IN SEIZURE ONSET ZONE
Epilepsy is a common neurological disorder characterized
by repeated, unprovoked seizures. The seizure onset zone
(SOZ) comprises regions of the brain that are responsible for
generating and sustaining seizures [3]. Surgical resection of the
seizure onset zone is the prescribed treatment for a large portion
of medically refractory epilepsy patients with focal epilepsy.
However, surgical resection risks damage to critical functional
zones that are frequently adjacent or even overlapping with
the seizure focus, depending on location of the focus [45]. An
ideal solution might be a closed-loop neuromodulation strategy
that stimulates the epileptic [41], [46] and other networks [47]
at the optimal frequency with spatial and temporal specificity
[6], [7]. In this paper, we focus on learning more about the
characteristic frequencies and the spatial specificity of epileptic
6j(#2!5cycles=sample)0 4 8 12166i(#2!5cycles=sample)0 4 8 121600.91.86j(#2!5cycles=sample)0 4 8 12166i(#2!5cycles=sample)0 4 8 121602.24.4<w0 5 10Mutual Information0.20.61 ^I(X;Y) 10
CLINICAL DETAILS OF THE PATIENTS ANALYZED.
TABLE I
Patient
Number of Seizures
Analyzed
Age/Sex
Seizure Onset Zone
P1
P2
P3
P4
P5
P6
P7
P8
P9
3
3
2
3
3
3
3
3
2
22/M
61/M
29/M
21/F
24/M
35/M
26/M
41/M
18/F
RAH 1-2, RPH 4,
RAMY 2-3
LAH 2-4, LPH 2
PD 4, 5
LF 28, LP 4
MST 1, TP 1, HD 1
LPH 5, 6, LPSM 8, LMH 5,
RMH 4, 5, RPSM 7
AH 3-5, PH 4
AMY 2, 3
AH 1, 2, 5, PH 5
TOP 3, 4
LAH 5, LAMY 3
RAH 3-5, LPH 6,
RPH 5-7
Outcome
of Surgery
Class IA
Class IIIA
Class IA
Class IA
Class IB
Class IA
Class IIB
N/A
Class IB
The full forms of the electrodes in seizure onset zone column in Table I: RAH - right anterior hippocampus, RPH - right posterior hippocampus,
RAMY - right amygdala, PD - posterior hippocampal depth, MST - mid-subtemporal lobe, TP - temporopolar, HD - hippocampal depth and
AST - anterior sub-temporal lobe, LMH - left mid hippocampus, AH - anterior hippocampus, PH - posterior hippocampus, AMY - amygdala,
TOP - temporo-occipito-parietal. The outcomes are in Engel epilepsy surgery outcome scale [43], [44]: "Class IA - completely seizure free
since surgery, class IB - non disabling simple partial seizures only since surgery, class IIB - rare disabling seizures since surgery ('almost
seizure-free'), class IIIA - worthwhile seizure reduction, class IV - no worthwhile improvement".
networks. Specifically, we investigate cross-frequency coupling
between various regions in the seizure onset zone during the
evolution of seizures and identify the frequencies with strong
coupling. We estimate the cross-frequency coupling (CFC)
from ECoG data recorded from the SOZ electrodes using our
nearest neighbor based MI-in-frequency estimator. We infer
the characteristics of CFC within and between various regions
inside the seizure onset zone.
We analyzed ECoG data, sampled at Fs = 1 kHz, from a
total of 25 seizures recorded from nine patients with medial
temporal lobe epilepsy. Clinical details of the patients, along
with the seizure onset zone identified from ECoG data [41],
are summarized in Table I. The seizure start and end time were
marked by the neurologist. We analyzed ECoG recordings
from SOZ electrodes during preictal (window spanning up to
3 minutes immediately before the seizure starts), ictal (during
seizures) and postictal (window spanning up to 3 minutes
immediately after the seizure ends) periods. We only focussed
on the oscillations in alpha (7.5-12.5 Hz), beta (12.5 - 30
Hz), gamma (30-80 Hz) and ripples (80-200 Hz), excluding
60 Hz line noise and its harmonics. The CFC oscillations are
analyzed at spectral resolution of 10 Hz by choosing Nf = 100,
and the exact frequencies considered are {10, 20,··· , 200} Hz,
excluding {60, 120, 180} Hz. The resulting 17×17 CFC matrix
from each ECoG electrode and between all pairs of ECoG
electrodes in the SOZ is estimated using nearest neighbor based
estimator (section IV-B) during preictal, ictal and postictal
periods during all the twenty five seizures.
We then grouped the ECoG electrodes into distinct anatom-
ical regions based on their label and analyzed the average
Fig. 6. Binary mask plotting the frequency pairs with statistically significant
differences across ahypotheses tested after applying false discovery rate
correction. White and black colored regions represent frequency pairs with
and without statistically significant variation respectively.
CFC within a SOZ electrode, between two electrodes in the
same anatomical region and between electrodes in different
anatomical regions. For instance, consider patient P1. ECoG
electrodes in the SOZ of patient P1 are grouped into three
different anatomical regions–RAH, RPH, and RAMY (Table I).
We estimated 5 CFC matrices, one per SOZ electrode, to infer
the average CFC within an electrode in SOZ in this patient.
We estimated 20 CFC matrices between all pairs of electrodes
in the SOZ. Of these, 4 CFC matrices (2 to learn the CFC
between the 2 SOZ electrodes in RAH and 2 to learn the CFC
between the 2 SOZ electrodes in RAMY regions) are grouped
to learn the average CFC between electrodes in the same
anatomical region in SOZ. The remaining 16 CFC matrices are
grouped to learn the CFC between different regions inside the
1 5 7 11131719201 5 7 1113171920 11
(a) Preictal period
(c) Difference between postictal and ictal period
Fig. 7. Cross-frequency coupling within an electrode inside the seizure onset zone. In Fig. 7a, MI-in-frequency estimates over the frequencies {10, 20, · · · , 200}
Hz excluding {60, 120, 180} Hz are obtained from each electrode in SOZ during preictal period and the median of the resulting CFC estimates from all the
SOZ electrodes in the twenty five seizures from the nine temporal lobe epilepsy patients analyzed is plotted. In Fig. 7b, MI-in-frequency estimates are obtained
from each electrode in SOZ in the ictal period and the difference between the median CFC estimate in ictal and preictal period is plotted. Similarly, Fig. 7c
plots the difference in the median CFC between postictal and ictal periods.
(b) Difference between ictal and preictal period
(a) Preictal period
(c) Difference between postictal and ictal period
Fig. 8. Cross-frequency coupling between electrodes in different regions inside the seizure onset zone. In Fig. 8a, MI-in-frequency estimates over the frequencies
{10, 20, · · · , 200} Hz excluding {60, 120, 180} Hz are obtained between electrodes in different SOZ regions during the preictal period and the median of the
resulting CFC estimates from the twenty five seizures in the nine temporal lobe epilepsy patients analyzed is plotted. In Fig. 8b, MI-in-frequency estimates are
obtained between electrodes in different SOZ regions from the ictal period and the difference between the median CFC estimate from the ictal and preictal
period is plotted. Similarly, Fig. 8c plots the difference in the median CFC between postictal and ictal periods.
(b) Difference between ictal and preictal period
SOZ. The estimated CFC matrices are grouped into these three
spatial categories for all the nine patients during preictal, ictal
and postictal periods. We only presented the results for CFC
within a SOZ electrode and between electrodes in different
SOZ regions during preictal, ictal and postictal periods.
We used the permutation procedure outlined in section IV-C
to estimate the CFC under the null hypothesis and assess
the significance of the estimated CFC values across the six
conditions considered (CFC during preictal, ictal, postictal
periods within a SOZ electrode and between electrodes in
different SOZ regions) using Wilcoxon signed-rank test [48].
We also used the Wilcoxon signed-rank test to identify the
frequency pairs with significant variation in CFC between
preictal and ictal periods and between ictal and postictal periods,
both within a SOZ electrode and between electrodes in different
SOZ regions (four hypotheses in total). In addition, we used
the Mann-Whitney U-test [48] to identify frequency pairs
with significant changes in CFC within a SOZ electrode and
between electrodes in different SOZ regions across preictal,
ictal and postictal periods (three hypotheses in total). We
estimated 3621 p-values in total (13 × 17 × 16 + 5 × 17)
and applied false discovery rate correction at a significance
level of 0.01 to account for multiple comparisons [49]. The
frequency pairs with significant statistical variation across
all the hypotheses considered are depicted using a binary
mask in Fig. 6, in which black and white colored regions
respectively represent frequency pairs without statistically
significant variation and with statistically significant variation.
Lack of statistical significance at the black regions in Fig. 6
could be because of insufficient data or could be due to a
neuronal transition mechanism as the brain moves from preictal
to ictal to postictal state. It is important to note that if we
tested only a subset of the thirteen hypotheses, then some of
the frequency pairs in black colored regions in Fig. 6 could
become statistically significant.
The median CFC within an electrode in SOZ during preictal,
ictal and postictal periods grouped across all twenty five
seizures in nine patients analyzed is plotted in Fig. 7. In
Fig. 7a, median CFC in the preictal period is plotted, while the
difference between median CFC in the ictal and preictal period,
and between postictal and ictal period is plotted in Fig. 7b
and Fig. 7c respectively. We need to multiply the binary mask
in Fig. 6 with the plots in Fig.7 to obtain frequency pairs
with significant statistical variation. The (i, j)th element in
1 5 7 11131719201 5 7 11131719200.211.91 5 7 11131719201 5 7 1113171920-0.20.30.71 5 7 11131719201 5 7 1113171920-0.8-0.20.41 5 7 11131719201 5 7 111317192000.050.11 5 7 11131719201 5 7 111317192000.10.21 5 7 11131719201 5 7 111317192000.130.26 12
the matrix in Fig. 7a is the median MI-in-frequency between
the 10i and 10j Hz frequency components during preictal
period across all SOZ electrodes in the twenty five seizures
analyzed. The principal diagonal in the three CFC matrices is
not plotted since MI-in-frequency between same frequencies
in a signal is infinity. It is clear from this figure that ripple
frequencies are heavily synchronized during preictal stage
within an electrode in SOZ. The synchronization between
all frequency pairs, particularly in gamma and ripples, seemed
to increase during the seizure when compared to just before
the seizure. And finally, the synchronization between high-
frequency bands decreased, and low frequencies become more
synchronized amongst themselves and with high-frequencies
in the postictal period compared to the ictal period within an
electrode in SOZ.
The median CFC between electrodes in different SOZ regions
grouped across all twenty five seizures in nine patients analyzed
is plotted in Fig. 8. We need to multiply the binary mask in
Fig. 6 with the plots in Fig. 8 to obtain frequency pairs with
significant statistical variation. The median CFC during the
preictal period is plotted in Fig. 8a. It is clear from the principal
diagonal that neighboring regions in SOZ have weak linear
interactions (possibly due to their spatial proximity) just before
a seizure starts. From Fig. 7a and Fig. 8a, it is clear that the
CFC strength is much lower between regions when compared
to within an electrode. From Fig. 8b, we observe a small
increase in CFC between regions as the brain transitions to
seizure state. However, the increase is much smaller between
regions when compared to the increase observed in Fig. 7b,
which suggests that different SOZ regions potentially drive
the rest of the brain into a seizure state independently, which
implies any non-surgical treatment should target these different
regions simultaneously to disrupt the epileptic network. As
the brain transitions to postictal state, we observe a sharp
increase in linear coupling between electrodes in different
SOZ regions, which suggests that postictal periods, unlike ictal
periods, are characterized by an increase in linear interactions.
These results highlight the role of gamma and ripple high-
frequency oscillations (HFOs) during seizures and the dynamic
reorganization of synchronization between neuronal oscillations
inside the seizure onset zone during the course of a seizure.
These results also suggest that multiple regions inside the
seizure onset zone might have to be targeted simultaneously
using neuromodulation techniques to control seizure activity.
VIII. DISCUSSION AND CONCLUSIONS
Detecting and quantifying relationships between multiple
data streams recorded from a physical system is of interest in
many science and engineering disciplines. However, since the
underlying model is often unknown and nonlinear, detecting
and quantifying the relationships in data is very challenging
in most real-world applications. Brownian distance covariance
[50], maximal information coefficient [51] are some of the
recent works that attempt to overcome this challenge in the
most general case. Furthermore,
in neuroscience, we are
also interested in decomposing the relationships in frequency
domain and estimating cross-frequency coupling (CFC) from
electrophysiological recordings. Motivated to understand non-
linear frequency coupling in electrophysiological recordings
from the brain and inspired by [26], we defined MI-in-frequency
between stochastic processes that are not necessarily Gaussian
and estimated it using data-driven estimators. We found that the
nearest neighbor based MI-in-frequency estimator outperforms
the kernel-based MI-in-frequency estimator. MI-in-frequency
can be thought of as 'coherence' for non-Gaussian signals. At
a first glance, CFC could be estimated by first filtering the
data into appropriate frequency bands and then applying the
techniques in [21], [50], [51]. However, [22] summarizes all the
caveats and confounds in estimating CFC using this approach.
In contrast, the MI-in-frequency metric estimates CFC without
explicitly band-pass filtering the data into appropriate frequency
bands.
We then compared the performance of MI-in-frequency
against modulation index [8], [21], a popular CFC metric used
to measure phase-amplitude coupling that involves bandpass
filtering, on simulated data and observed that MI-in-frequency
outperforms the existing metrics used to estimate CFC. The
main advantages of the MI-in-frequency approach over existing
methods to estimate CFC are that it detects statistical inde-
pendence, detects dependencies across phase and amplitude
jointly, applies to linear and nonlinear dependencies, and is not
dependent on parameters like the filter bandwidth. Our approach
will need more data when compared with coherence since MI-
in-frequency detects both linear and nonlinear dependencies
in frequency. From the simulation results on linear models,
we need about 103 samples to be within 10% of the true
value. For the ECoG data sampled at 1 kHz and a desired
spectral resolution of 10 Hz, this implies the total number of
data samples is of the order of 100 seconds or a couple of
minutes, which is roughly the size of preictal, ictal and postictal
windows used in section VII. In summary, we developed a
metric to detect statistical independence in frequency which
outperforms existing CFC metrics and for the first time, utilized
frequency domain to estimate mutual information over time
between dependent data.
The MI-in-frequency metric can be further extended along
several directions and some of them are outlined here. We
can move to wavelet based analysis to improve the fixed
time-frequency resolution of our Fourier-based approach in
future work. The assumption of data stationarity in observation
window (also assumed by most CFC metrics) can be potentially
relaxed by utilizing time-frequency distributions and developing
heuristics to measure the dependencies across frequency.
However, the inherent trade-off involved is that we are not
guaranteed to detect statistical independence. It is also possible
to define and estimate conditional MI-in-frequency to eliminate
indirect coupling estimated between two signals because of a
third signal which is coupled to both.
We then apply the MI-in-frequency estimators to infer the
coupling between neuronal oscillations before, during and
after seizures in the seizure onset zone. Spatially, we used
the electrode labels to identify the different regions in the SOZ.
This is just one possible way to analyze the spatial variation
in CFC. Some of the other possible options include using the
distance between electrodes or using the underlying neuronal
13
seizure activity and will serve as the first step towards the
development of a patient-specific, closed-loop, non-surgical
treatment for epilepsy.
IX. ACKNOWLEDGMENTS
The authors wish to thank Suganya Karunakaran for the
helpful discussions on statistical hypothesis testing and proof-
reading the manuscript.
REFERENCES
[1] R. Malladi, D. H. Johnson, G. Kalamangalam, N. Tandon, and
B. Aazhang, "Measuring cross-frequency coupling using mutual infor-
mation and its application to epilepsy," in Cosyne Abstracts, Salt Lake
City, USA, 2017.
[2] --, "Data-driven estimation of mutual information using frequency
domain and its application to epilepsy," in Asilomar Conference on
Signals, Systems and Computers, 2017.
[3] H. O. Lüders, I. Najm, D. Nair, P. Widdess-Walsh, and W. Bingman,
"The epileptogenic zone: general principles," Epileptic Disorders, vol. 8,
no. 2, pp. 1–9, 2006.
[4] F. Rosenow and H. Lüders, "Presurgical evaluation of epilepsy," Brain,
vol. 124, no. 9, pp. 1683–1700, 2001.
[5] G. K. Bergey, M. J. Morrell, E. M. Mizrahi, A. Goldman, D. King-
Stephens, D. Nair, S. Srinivasan, B. Jobst, R. E. Gross, D. C. Shields
et al., "Long-term treatment with responsive brain stimulation in adults
with refractory partial seizures," Neurology, vol. 84, no. 8, pp. 810–817,
2015.
[6] S. Sunderam, B. Gluckman, D. Reato, and M. Bikson, "Toward rational
design of electrical stimulation strategies for epilepsy control," Epilepsy
& Behavior, vol. 17, no. 1, pp. 6–22, 2010.
[7] E. Krook-Magnuson and I. Soltesz, "Beyond the hammer and the scalpel:
selective circuit control for the epilepsies," Nature neuroscience, vol. 18,
no. 3, pp. 331–338, 2015.
[8] R. T. Canolty, E. Edwards, S. S. Dalal, M. Soltani, S. S. Nagarajan, H. E.
Kirsch, M. S. Berger, N. M. Barbaro, and R. T. Knight, "High gamma
power is phase-locked to theta oscillations in human neocortex," Science,
vol. 313, no. 5793, pp. 1626–1628, 2006.
[9] R. T. Canolty and R. T. Knight, "The functional role of cross-frequency
coupling," Trends in Cognitive Sciences, vol. 14, no. 11, pp. 506–515,
2010.
[10] C. Alvarado-Rojas, M. Valderrama, A. Fouad-Ahmed, H. Feldwisch-
Drentrup, M. Ihle, C. Teixeira, F. Sales, A. Schulze-Bonhage, C. Adam,
A. Dourado et al., "Slow modulations of high-frequency activity (40–140
Hz) discriminate preictal changes in human focal epilepsy," Scientific
reports, vol. 4, 2014.
[11] K. Edakawa, T. Yanagisawa, H. Kishima, R. Fukuma, S. Oshino,
H. M. Khoo, M. Kobayashi, M. Tanaka, and T. Yoshimine, "Detection
of epileptic seizures using phase–amplitude coupling in intracranial
electroencephalography," Scientific reports, vol. 6, 2016.
[12] M. Guirgis, Y. Chinvarun, M. del Campo, P. L. Carlen, and B. L.
Bardakjian, "Defining regions of interest using cross-frequency coupling
in extratemporal lobe epilepsy patients," Journal of Neural Engineering,
vol. 12, no. 2, p. 026011, 2015.
[13] S. A. Weiss, A. Lemesiou, R. Connors, G. P. Banks, G. M. McKhann,
R. R. Goodman, B. Zhao, C. G. Filippi, M. Nowell, R. Rodionov et al.,
"Seizure localization using ictal phase-locked high gamma a retrospective
surgical outcome study," Neurology, vol. 84, no. 23, pp. 2320–2328,
2015.
[14] S. Liu, Z. Sha, A. Sencer, A. Aydoseli, N. Bebek, A. Abosch, T. Henry,
C. Gurses, and N. F. Ince, "Exploring the time–frequency content of high
frequency oscillations for automated identification of seizure onset zone
in epilepsy," Journal of Neural Engineering, vol. 13, no. 2, p. 026026,
2016.
[15] H. Zhou, Y. Li, Y.-L. Hsin, and W. Liu, "Phase-amplitude coupling
analysis for seizure evolvement using Hilbert Huang transform," in IEEE
38th Annual International Conference of the Engineering in Medicine
and Biology Society (EMBC).
IEEE, 2016, pp. 1022–1025.
[16] R. Zhang, Y. Ren, C. Liu, N. Xu, X. Li, F. Cong, T. Ristaniemi, and
Y. Wang, "Temporal-spatial characteristics of phase-amplitude coupling
lobe epilepsy," Clinical
in electrocorticogram for human temporal
Neurophysiology, vol. 128, no. 9, pp. 1707–1718, 2017.
(a) Within a SOZ electrode
(b) Between electrodes in different SOZ
regions
Fig. 9. Cross-frequency coupling during interictal periods. In Fig. 9a, MI-
in-frequency estimates over the frequencies {10, 20, · · · , 200} Hz excluding
{60, 120, 180} Hz are obtained from each electrode in SOZ during interictal
period and the median of the resulting CFC estimates from all the SOZ
electrodes in patients P1 and P2 is plotted. In Fig. 9b, MI-in-frequency estimates
over the frequencies {10, 20, · · · , 200} Hz excluding {60, 120, 180} Hz are
obtained between electrodes in different SOZ regions during the interictal
period and the median of the resulting CFC estimates in patients P1 and P2
is plotted.
cell types to split the electrodes into different regions in SOZ.
Our MI-in-frequency metric provides a framework that can
be utilized to learn the CFC characteristics for any desired
spatial grouping. In addition, the frequency resolution of our
estimated CFC was constant and wavelet transform, instead of
Fourier transform, can be utilized to provide greater resolution
at lower frequencies.
We observed that the high-frequency synchronization within
an ECoG electrode in SOZ increases during seizures and
decreases immediately after the seizure, which is accompanied
by an increase in low-frequency coupling. However,
the
coupling between different anatomical regions in SOZ does
not increase noticeably during seizures and is also followed
by a large increase in linear interactions immediately after
a seizure. These observations suggest that seizure activity is
characterized by nonlinear interactions and is potentially due
to the independent efforts by various regions within SOZ,
which implies that all these regions are potential spatial targets
for electrical stimulation. Furthermore, we did a preliminary
investigation to learn if there are the differences in CFC between
interictal periods and seizure periods. Fig. 9 plots the CFC
within an ECoG electrode and between ECoG electrodes in
different regions in SOZ during interictal period in two patients
(P1 and P2). Comparing Fig. 9 with Fig. 7a and Fig. 8a, it looks
like the CFC within a SOZ electrodes at higher frequencies
slightly increases, while CFC between electrodes in different
regions across the diagonal (or equivalently, linear interactions)
slightly decreases as the brain transitions from interictal to
preitctal periods. We plan to extend this analysis to a larger
patient cohort. Building a real-time seizure prediction system
utilizing the variations in CFC between interictal and seizure
periods is the focus of our current [52] and future work. In
addition, the CFC characteristics were patient-specific and we
presented the median CFC across all the patients considered.
Going forward, the MI-in-frequency metric should be applied
to infer the CFC between channels in SOZ and outside SOZ to
learn how SOZ drives the rest of the brain into a seizure state
in each epilepsy patient. The results from such an analysis will
improve our understanding of the CFC mechanisms underlying
1 5 7 11131719201 5 7 11131719200.10.81.61 5 7 11131719201 5 7 111317192000.080.16 14
[17] M. Amiri, B. Frauscher, and J. Gotman, "Phase-amplitude coupling is
elevated in deep sleep and in the onset zone of focal epileptic seizures,"
Frontiers in human neuroscience, vol. 10, 2016.
[18] M. Cotic, Y. Chinvarun, M. del Campo, P. L. Carlen, and B. L. Bardakjian,
"Spatial coherence profiles of ictal high-frequency oscillations correspond
to those of interictal low-frequency oscillations in the ecog of epileptic
patients," IEEE Transactions on Biomedical Engineering, vol. 63, no. 1,
pp. 76–85, 2016.
[19] B. Frauscher, N. von Ellenrieder, F. Dubeau, and J. Gotman, "Different
seizure-onset patterns in mesiotemporal lobe epilepsy have a distinct
interictal signature," Clinical Neurophysiology, vol. 128, no. 7, pp. 1282–
1289, 2017.
[20] L. Faes and G. Nollo, Multivariate frequency domain analysis of causal
INTECH Open Access
interactions in physiological time series.
Publisher, 2011.
[21] A. C. Onslow, R. Bogacz, and M. W. Jones, "Quantifying phase–
amplitude coupling in neuronal network oscillations," Progress in
Biophysics and Molecular Biology, vol. 105, no. 1, pp. 49–57, 2011.
[22] J. Aru, J. Aru, V. Priesemann, M. Wibral, L. Lana, G. Pipa, W. Singer,
and R. Vicente, "Untangling cross-frequency coupling in neuroscience,"
Current Opinion in Neurobiology, vol. 31, pp. 51–61, 2015.
[23] R. Pascual-Marqui, P. Faber, T. Kinoshita, Y. Kitaura, K. Kochi, P. Milz,
K. Nishida, and M. Yoshimura, "The dual frequency RV-coupling
coefficient: a novel measure for quantifying cross-frequency information
transactions in the brain," arXiv preprint arXiv:1603.05343, 2016.
[24] D. R. Brillinger, "Second-order moments and mutual information in the
analysis of time series," Recent Advances in Statistical Methods, pp.
64–76, 2002.
[25] R. Salvador, A. Martinez, E. Pomarol-Clotet, J. Gomar, F. Vila, S. Sarro,
A. Capdevila, and E. Bullmore, "A simple view of the brain through a
frequency-specific functional connectivity measure," Neuroimage, vol. 39,
no. 1, pp. 279–289, 2008.
[26] D. R. Brillinger and A. Guha, "Mutual information in the frequency
domain," Journal of Statistical Planning and Inference, vol. 137, no. 3,
pp. 1076–1084, 2007.
[27] H. Cramér and M. Leadbetter, Stationary and related stochastic processes:
sample function properties and their applications, ser. Wiley series in
probability and mathematical statistics. Tracts on probability and statistics.
Wiley, 1967.
[28] H. J. Larson and B. O. Shubert, Probabilistic models in engineering
sciences. Wiley, 1979, vol. 2.
[29] D. R. Brillinger, Time Series: Data Analysis and Theory. Philadelphia,
PA, USA: Society for Industrial and Applied Mathematics, 2001.
[30] S. Khan, S. Bandyopadhyay, A. R. Ganguly, S. Saigal, D. J. Erickson III,
V. Protopopescu, and G. Ostrouchov, "Relative performance of mutual
information estimation methods for quantifying the dependence among
short and noisy data," Physical Review E, vol. 76, no. 2, p. 026209,
2007.
[31] E. Schaffernicht, R. Kaltenhaeuser, S. S. Verma, and H.-M. Gross, "On
estimating mutual information for feature selection," in International
Conference on Artificial Neural Networks. Springer, 2010, pp. 362–367.
[32] Q. Wang, S. R. Kulkarni, and S. Verdú, "Universal estimation of
information measures for analog sources," Foundations and Trends in
Communications and Information Theory, vol. 5, no. 3, pp. 265–353,
2009.
[33] M. Pinsker, Information and information stability of random variables
and processes, ser. Holden-Day series in time series analysis. Holden-
Day, 1964.
[34] T. M. Cover and J. A. Thomas, Elements of information theory.
Wiley & Sons, 2012.
John
[35] A. V. Oppenheim, R. W. Schafer, J. R. Buck et al., Discrete-time signal
processing. Prentice Hall Englewood Cliffs, NJ, 1989, vol. 2.
[36] E. Pereda, R. Q. Quiroga, and J. Bhattacharya, "Nonlinear multivariate
analysis of neurophysiological signals," Progress in Neurobiology, vol. 77,
no. 1, pp. 1–37, 2005.
[37] D. W. Scott, Multivariate density estimation: theory, practice, and
visualization.
John Wiley & Sons, 2015.
[38] A. Kraskov, H. Stögbauer, and P. Grassberger, "Estimating mutual
information," Physical Review E, vol. 69, no. 6, p. 066138, 2004.
[39] T. Duong et al., "ks: Kernel density estimation and kernel discriminant
analysis for multivariate data in R," Journal of Statistical Software,
vol. 21, no. 7, pp. 1–16, 2007.
[40] D. Dvorak and A. A. Fenton, "Toward a proper estimation of phase–
amplitude coupling in neural oscillations," Journal of Neuroscience
Methods, vol. 225, pp. 42–56, 2014.
[41] R. Malladi, G. Kalamangalam, N. Tandon, and B. Aazhang, "Identifying
seizure onset zone from the causal connectivity inferred using directed
information," IEEE Journal of Selected Topics in Signal Processing,
vol. 10, no. 7, pp. 1267–1283, Oct 2016.
[42] J. I. Berman, J. McDaniel, S. Liu, L. Cornew, W. Gaetz, T. P. Roberts,
and J. C. Edgar, "Variable bandwidth filtering for improved sensitivity
of cross-frequency coupling metrics," Brain Connectivity, vol. 2, no. 3,
pp. 155–163, 2012.
[43] J. Engel, "Update on surgical treatment of the epilepsies summary of the
second international palm desert conference on the surgical treatment of
the epilepsies (1992)," Neurology, vol. 43, no. 8, pp. 1612–1612, 1993.
[44] C. Tonini, E. Beghi, A. T. Berg, G. Bogliun, L. Giordano, R. W. Newton,
A. Tetto, E. Vitelli, D. Vitezic, and S. Wiebe, "Predictors of epilepsy
surgery outcome: a meta-analysis," Epilepsy Research, vol. 62, no. 1, pp.
75–87, 2004.
[45] U. Gleissner, R. Sassen, M. Lendt, H. Clusmann, C. Elger, and
C. Helmstaedter, "Pre-and postoperative verbal memory in pediatric
patients with temporal lobe epilepsy," Epilepsy research, vol. 51, no. 3,
pp. 287–296, 2002.
[46] S. Karunakaran, M. J. Rollo, K. Kim, J. A. Johnson, G. P. Kalamangalam,
B. Aazhang, and N. Tandon, "The interictal mesial temporal lobe epilepsy
network," Epilepsia, 2017.
[47] K. Kim, A. D. Ekstrom, and N. Tandon, "A network approach for
modulating memory processes via direct and indirect brain stimulation:
toward a causal approach for the neural basis of memory," Neurobiology
of learning and memory, vol. 134, pp. 162–177, 2016.
[48] G. W. Corder and D. I. Foreman, Nonparametric statistics: A step-by-step
approach.
John Wiley & Sons, 2014.
[49] Y. Benjamini and Y. Hochberg, "Controlling the false discovery rate:
a practical and powerful approach to multiple testing," Journal of the
Royal Statistical Society. Series B (Methodological), pp. 289–300, 1995.
[50] G. J. Székely, M. L. Rizzo et al., "Brownian distance covariance," The
annals of applied statistics, vol. 3, no. 4, pp. 1236–1265, 2009.
[51] D. N. Reshef, Y. A. Reshef, H. K. Finucane, S. R. Grossman, G. McVean,
P. J. Turnbaugh, E. S. Lander, M. Mitzenmacher, and P. C. Sabeti,
"Detecting novel associations in large data sets," Science, vol. 334, no.
6062, pp. 1518–1524, 2011.
[52] S. Hooper, E. Biegert, M. Levy, J. Pensock, L. V. D. Spoel, X. Zhang,
T. Zhang, N. Tandon, and B. Aazhang, "On developing an FPGA based
system for real time seizure prediction," in Asilomar Conference on
Signals, Systems and Computers, 2017.
X. APPENDIX
a) Proof of (6): We have from (4),
∞(cid:82)
−∞
∞(cid:82)
−∞
−∞
−∞
+
−∞
−∞
y(t) =
∞(cid:82)
h2(t − τ )
h1(t − τ )x(τ )dτ+
h2(t − τ )w(τ )dτ (20)
−∞
h1(t − τ )
∞(cid:82)
⇒ ∞(cid:82)
ej2πντ d(cid:101)X (ν) dτ
ej2πνtd(cid:101)Y (ν) =
∞(cid:82)
∞(cid:82)
ej2πντ d(cid:102)W (ν) dτ (from Theorem 1)
∞(cid:82)
∞(cid:82)
h1(t − τ )e−j2πν(t−τ )dτ d(cid:101)X (ν)+
∞(cid:82)
∞(cid:82)
h2(t − τ )e−j2πν(t−τ )dτ d(cid:102)W (ν)
ej2πνt(cid:110)
(cid:111)
∞(cid:82)
H1 (ν) d(cid:101)X (ν)+H2 (ν) d(cid:102)W (ν)
=⇒ d(cid:101)Y (ν) = H1 (ν) d(cid:101)X (ν) + H2 (ν) d(cid:102)W (ν) .
ej2πνt
ej2πνt
(22)
(21)
−∞
=
−∞
−∞
−∞
−∞
=
.
b) Proof of Theorem 3: We will first prove that
MIXY (ν1, ν2) is zero, when X and Y are related by (4) and
ν1 (cid:54)= ν2. Since the processes X (t) and W (t) are independent,
their spectral processes are also independent. In addition, we
also know from Theorem 2 that the spectral increments of
the Gaussian process X (t) are independent. It is clear from
(6) that given H1 (ν) and H2 (ν), (cid:2)d(cid:101)YR (ν2) , d(cid:101)YI (ν2)(cid:3) is
15
I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1), since the imaginary part of
Y is zero.
c) Relationship between MI in frequency and coherence:
The coherence CXY (ν) ∈ [0, 1] between two processes X and
Y related by (4) is given by
CXY (ν) =
sXY (ν)2
sX (ν)sY (ν) =
⇒ − log (1 − CXY (ν)) = log(cid:0)1 +
H1(ν)2sX (ν)
sX (ν)H1(ν)2+sW (ν)H2(ν)2 .
H1(ν)2sX (ν)
H2(ν)2sW (ν)
(cid:1)
= MIXY (ν, ν) .
(30)
d) Proof of Theorem 4: Now we consider two discrete-
time Gaussian stochastic processes X [n] and Y [n] that are
related by
y[n] = h1[n] ∗ x[n] + h2[n] ∗ w[n],
(31)
where h1[n] and h2[n] are the impulse responses of two
discrete-time linear, time-invariant filters. (31) is the discrete-
time equivalent of (4). It was shown in chapter 10 in [33]
that mutual information between the discrete-time Gaussian
stochastic processes X [n] and Y [n] is related to coherence
according to
I (X; Y ) = − 0.5(cid:82)
0.5(cid:82)
I (X; Y ) =
0
0
From (30) and (32), we have
log (1 − CXY (λ)) dλ.
MIXY (λ, λ) dλ.
(32)
(33)
completely determined by the two-dimensional random vec-
tors (cid:2)d(cid:101)XR (ν2) , d(cid:101)XI (ν2)(cid:3) and (cid:2)d(cid:102)WR (ν2) , d(cid:102)WI (ν2)(cid:3), both
vector(cid:2)d(cid:101)XR (ν1) , d(cid:101)XI (ν1)(cid:3) when ν1 (cid:54)= ν2. This implies the
MI between(cid:2)d(cid:101)YR (ν2) , d(cid:101)YI (ν2)(cid:3) and(cid:2)d(cid:101)XR (ν1) , d(cid:101)XI (ν1)(cid:3),
of which are independent of the two-dimensional random
which is defined as MIXY (ν1, ν2), is zero.
1
0,
, (24)
Σ21 Σ22
where Σ11 = 1
(cid:2)d(cid:101)YR (ν) , d(cid:101)YI (ν)(cid:3)∼N(cid:0)0,(cid:0) 1
We will now derive the analytical expression for MIXY (ν,ν),
for ν (cid:54)= 0. Let H1(ν)=H1R(ν)+jH1I (ν) and H2(ν)=H2R(ν)+
jH2I (ν). We can see from (5), (6) that
2 sX (ν)H1 (ν)2 +
(23)
where N represents Gaussian distribution, 0 is a two element
zero vector and I is the 2 × 2 identity matrix. In addition,
2 sW (ν)H2 (ν)2(cid:1) I(cid:1) ,
(cid:20)Σ11 Σ12
(cid:18)
(cid:21)(cid:19)
(cid:2)d(cid:101)XR (ν),d(cid:101)XI (ν),d(cid:101)YR (ν),d(cid:101)YI (ν)(cid:3)∼N
(cid:0)sX (ν)H1 (ν)2 + sW (ν)H2 (ν)2(cid:1), I is the 2× 2 identity
(ν) I, σ2(cid:101)Y
2 σ2(cid:101)Y
2 sX (ν) I, Σ22 = 1
(cid:21)
MIXY (ν, ν) = I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9);(cid:8)d(cid:101)YR (ν) , d(cid:101)YI (ν)(cid:9)(cid:1)
= I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1)+
= I(cid:0)(cid:8)d(cid:101)XR (ν) ,d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1)+
I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YI (ν)d(cid:101)YR (ν)(cid:1)
I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YI (ν)(cid:1),
matrix and 0 is a four element zero vector. In addition,
1
2 sX (ν) H1I (ν)
1
2 sX (ν) H1R(ν)
Now, the MI between X and Y at frequency ν is given by
2 sX (ν) H1R(ν)
− 1
2 sX (ν) H1I (ν)
Σ12 = ΣT
(cid:20) 1
(ν) =
21 =
(26)
(25)
.
where (25) follows from the chain rule of mutual information
[34] and (26) follows because the real and imaginary parts
of the spectral process of a Gaussian process are independent
from Theorem 2. In addition, (cid:2)d(cid:101)XR (ν) , d(cid:101)XI (ν) , d(cid:101)YR (ν)(cid:3)
is a Gaussian distributed random vector with zero mean and
covariance matrix Σ(cid:48), which is easily obtained from (24). Since
the mutual information between components of a Gaussian
random vector depends only on the determinants of the joint
distribution's covariance matrices and that of marginals [34],
we can easily show that
2 σ2(cid:101)Y )
Σ11( 1
2 log
Σ(cid:48)
H1(ν)2sX (ν)
H2(ν)2sW (ν)
I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1) = 1
2 log(cid:0)1 +
(cid:1).
I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9);d(cid:101)YI (ν)(cid:1) =
2 log(cid:0)1 +
(cid:1).
MIXY (ν, ν) = 2 × I(cid:0)(cid:8)d(cid:101)XR (ν) , d(cid:101)XI (ν)(cid:9); d(cid:101)YR (ν)(cid:1)
H1(ν)2sX (ν)
H2(ν)2sW (ν)
From (26), (27) and (28), we have
Similarly, we can also show that
(27)
(28)
= 1
1
= log(cid:0)1 +
(cid:1).
H1(ν)2sX (ν)
H2(ν)2sW (ν)
(29)
At ν = 0, MI-in-frequency between X and Y is equal to
|
1310.6934 | 2 | 1310 | 2013-10-30T03:17:29 | Generalized activity equations for spiking neural network dynamics | [
"q-bio.NC"
] | Much progress has been made in uncovering the computational capabilities of spiking neural networks. However, spiking neurons will always be more expensive to simulate compared to rate neurons because of the inherent disparity in time scales - the spike duration time is much shorter than the inter-spike time, which is much shorter than any learning time scale. In numerical analysis, this is a classic stiff problem. Spiking neurons are also much more difficult to study analytically. One possible approach to making spiking networks more tractable is to augment mean field activity models with some information about spiking correlations. For example, such a generalized activity model could carry information about spiking rates and correlations between spikes self-consistently. Here, we will show how this can be accomplished by constructing a complete formal probabilistic description of the network and then expanding around a small parameter such as the inverse of the number of neurons in the network. The mean field theory of the system gives a rate-like description. The first order terms in the perturbation expansion keep track of covariances. | q-bio.NC | q-bio | Generalized activity equations for spiking neural network
dynamics
Michael A. Buice1 and Carson C. Chow2
1Allen Institute for Brain Science
2Laboratory of Biological Modeling,
NIDDK, NIH, Bethesda, MD
(Dated: June 23, 2021)
Abstract
Much progress has been made in uncovering the computational capabilities of spiking neural
networks. However, spiking neurons will always be more expensive to simulate compared to rate
neurons because of the inherent disparity in time scales - the spike duration time is much shorter
than the inter-spike time, which is much shorter than any learning time scale. In numerical analysis,
this is a classic stiff problem. Spiking neurons are also much more difficult to study analytically.
One possible approach to making spiking networks more tractable is to augment mean field activity
models with some information about spiking correlations. For example, such a generalized activity
model could carry information about spiking rates and correlations between spikes self-consistently.
Here, we will show how this can be accomplished by constructing a complete formal probabilistic
description of the network and then expanding around a small parameter such as the inverse of the
number of neurons in the network. The mean field theory of the system gives a rate-like description.
The first order terms in the perturbation expansion keep track of covariances.
3
1
0
2
t
c
O
0
3
]
.
C
N
o
i
b
-
q
[
2
v
4
3
9
6
.
0
1
3
1
:
v
i
X
r
a
1
Introduction
Even with the rapid increase in computing power due to Moore's law and proposals to
simulate the entire human brain notwithstanding [1], a realistic simulation of a functioning
human brain performing nontrivial tasks is remote. While it is plausible that a network the
size of the human brain could be simulated in real time [2, 3] there are no systematic ways to
explore the parameter space. Technology to experimentally determine all the parameters in
a single brain simultaneously does not exist and any attempt to infer parameters by fitting to
data would require exponentially more computing power than a single simulation. We also
have no idea how much detail is required. Is it sufficient to simulate a large number of single
compartment neurons or do we need multiple-compartments? How much molecular detail
is required? Do we even know all the important biochemical and biophysical mechanisms?
There are an exponential number of ways a simulation would not work and figuring out which
remains computationally intractable. Hence, an alternative means to provide appropriate
prior distributions for parameter values and model detail is desirable. Current theoretical
explorations of the brain utilize either abstract mean field models or small numbers of more
biophysical spiking models. The regime of large but finite numbers of spiking neurons
remains largely unexplored. It is not fully known what role spike time correlations play in
the brain. It would thus be very useful if mean field models could be augmented with some
spike correlation information.
This paper outlines a scheme to derive generalized activity equations for the mean and
correlation dynamics of a fully deterministic system of coupled spiking neurons. It synthe-
sizes methods we have developed to solve two different types of problems. The first problem
was how to compute finite system size effects in a network of coupled oscillators. We adapted
the methods of the kinetic theory of gases and plasmas [4, 5] to solve this problem. The
method exploits the exchange symmetry of the oscillators and characterizes the phases of all
the oscillators in terms of a phase density function η(θ, t), where each oscillator is represented
as a point mass in this density. We then write down a formal flux conservation equation of
this density, called the Klimontovich equation, which completely characterizes the system.
However, because the density is not differentiable, the Klimontovich equation only exists in
the weak or distributional sense. Previously, e.g [6 -- 9] the equations were made usable by
taking the "mean field limit" of N → ∞ and assuming that the density is differentiable in
2
that limit, resulting in what is called the Vlasov equation. Instead of immediately taking
the mean field limit, we regularize the density by averaging over initial conditions and pa-
rameters and then expand in the inverse system size N−1 around the mean field limit. This
results in a system of coupled moment equations known as the BBGKY moment hierarchy.
In [10], we solved the moment equations for the Kuramoto model perturbatively to compute
the pair correlation function between oscillators. However, the procedure was somewhat
ad hoc and complicated. We then subsequently showed in [11], that the BBGKY moment
hierarchy could be recast in terms of a density functional of the phase density. This density
functional could be written down explicitly as an integral over all possible phase histories,
i.e. a Feynman-Kac path integral. The advantage of using this density functional formalism
is that the moments to arbitrary order in 1/N could be computed as a steepest-descent
expansion of the path integral, which can be expressed in terms of Feynman diagrams. This
made the calculation more systematic and mechanical. We later applied the same formalism
to synaptically coupled spiking models [12].
Concurrently with this line of research, we also explored the question of how to generalize
population activity equations, such as the Wilson-Cowan equations, to include the effects
of correlations. The motivation for this question is that the Wilson-Cowan equations are
mean field equations and do not capture the effects of spike-time correlations. For example,
the gain in the Wilson-Cowan equations is fixed, (which is a valid approximation when the
neurons fire asynchronously), but correlations in the firing times can change the gain [13].
Thus, it would be useful to develop a systematic procedure to augment population activity
equations to include spike correlation effects. The approach we took was to posit plausible
microscopic stochastic dynamics, dubbed the spike model, that reduced to the Wilson-Cowan
equations in the mean field limit and compute the self-consistent moment equations from that
microscopic theory. Buice and Cowan [14] showed that the solution of the master equation
of the spike model could be expressed formally in terms of a path integral over all possible
spiking histories. The random variable in the path integral is a spike count whereas in the
path integral for the deterministic phase model we described above, the random variable is
a phase density. To generate a system of moment equations for the microscopic stochastic
system, we transformed the random spike count variable in the path integral into moment
variables [15]. This is accomplished using the effective action approach of field theory, where
the exponent of the cumulant generating functional, called the action, which is a function
3
of the random variable is Legendre transformed into an effective action of the cumulants.
The desired generalized Wilson-Cowan activity equations are then the equations of motion
of the effective action. This is analogous to the transformation from Lagrangian variables of
position and velocity to Hamiltonian variables of position and momentum. Here, we show
how to apply the effective action approach to a deterministic system of synaptically coupled
spiking neurons to derive a set of moment equations.
Approach
Consider a network of single compartment conductance-based neurons
= − n(cid:88)
N(cid:88)
C
dVi
dt
τ r
i
τj
dxr
i
dt
dsj
dt
dgij
dt
τg
gr(xr
i )(Vi − vr) +
gijsj(t)
r=1
j=1
= f (Vi, xi)
= h(Vj, sj)
= φ(gij, V )
The equations are remarkably stiff with time scales spanning orders of magnitude from
milliseconds for ion channels, to seconds for adaptation, and from hours to years for changes
in synaptic weights and connections. Parameter values must be assigned for 1011 neurons
with 104 connections each. Here, we present a formalism to derive a set of reduced activity
equations directly from a network of deterministic spiking neurons that capture the spike
rate and spike correlation dynamics. The formalism first constructs a density functional
for the firing dynamics of all the neurons in a network. It then systematically marginalizes
the unwanted degrees of freedom to isolate a set of self-consistent equations for the desired
quantities. For heuristic reasons, we derive an example set of generalized activity equations
for the first and second cumulants of the firing dynamics of a simple spiking model but the
method can be applied to any spiking model.
A convenient form to express spiking dynamics is with a phase oscillator. Consider the
quadratic integrate-and-fire neuron
dVi
dt
= Ii + V 2
i + αiu(t)
(1)
4
where I is an external current and u(t) are the synaptic currents with some weight αi. The
spike is said to occur when V goes to infinity whereupon it is reset to minus infinity. The
quadratic nonlinearity ensures that this transit will occur in a finite amount of time. The
substitution V = tan(θ/2) yields the theta model [16]:
= 1 − cos θi + (1 + cos θi)(Ii + αiu)
dθi
dt
(2)
which is the normal form of a Type I neuron near the bifurcation to firing [17]. The phase
neuron is an adequate approximation to spiking dynamics provided the inputs are not overly
strong as to disturb the limit cycle. The phase neuron also includes realistic dynamics such
as not firing when the input is below threshold. Coupled phase models arise naturally in
weakly coupled neural networks [18 -- 20]. They include the Kuramoto model [21], which we
have previously analyzed [10, 11].
Here, we consider the phase dynamics of a set of N coupled phase neurons obeying
θi = F (θ, γi, u(t))
u(t) = −βu(t) + βν(t)
ν(t) =
1
N
δ(t − tl
j)
N(cid:88)
(cid:88)
j=1
l
(3)
(4)
(5)
where each neuron has a phase θi that is indexed by i, u is a global synaptic drive, F (θ, γ, u)
is the phase and synaptic drive dependent frequency, γi represents all the parameters for
neuron i drawn from a distribution with density g(γ), ν is the population firing rate of the
network,tl
j is the lth firing time of neuron j and a neuron fires when its phase crosses π. In
the present paper, we consider all-to-all or global coupling through a synaptic drive variable
u(t). However, our basic approach is not restricted to global coupling.
We can encapsulate the phase information of all the neurons into a neuron density func-
tion [10 -- 12, 22, 23].
η(θ, γ, t) =
(6)
where δ(·) is the Dirac delta functional, and θi(t) is a solution to system (3)-(5). The
neuron density gives a count of the number of neurons with phase θ and synaptic strength
i=1
δ(θ − θi(t))δ(γ − γi)
γ at time t. Using the fact that the Dirac delta functional in (5) can be expressed as
j) = θjδ(π − θj), the population firing rate can be rewritten as
(cid:80)
l δ(t − tl
N(cid:88)
1
N
(cid:90)
ν(t) =
dγ F (π, γ, u(t))η(π, γ, t)
(7)
5
The neuron density formally obeys the conservation equation
∂
∂t
η(θ, γ, t) +
∂
∂θ
[F η(θ, γ, t)] = 0
(8)
with initial condition η(θ, γ, t0) = η0(θ, γ) and u(t0) = u0. Equation (8) is known as the
Klimontovich equation [4, 24]. The Klimontovich equation, the equation for the synaptic
drive (4), and the firing rate expressed in terms of the neuron density (7), fully define
the system. The system is still fully deterministic but is now in a form where various
sets of reduced descriptions can be derived. Here, we will produce an example of a set of
reduced equations or generalized activity equations that capture some aspects of the spiking
dynamics. The path we take towards the end will require the introduction of some formal
machinery that may obscure the intuition around the approximations. However, we feel
that it is useful because it provides a systematic and controlled way of generating averaged
quantities that can be easily generalized.
For finite N , (8) is only valid in the weak or distributional sense since η is not differen-
tiable. In the N → ∞ limit, it has been argued that η will approach a smooth density ρ that
evolves according to the Vlasov equation that has the same form as (8) but with η replaced
by ρ [4 -- 7, 10]. This has been proved rigorously in the case where noise is added using the
theory of coupled diffusions [25 -- 28]. This N → ∞ limit is called mean field theory.
mean field theory, the original microscopic many body neuronal network is represented by
In
a smooth macroscopic density function. In other words, the ensemble of networks prepared
with different microscopic initial conditions is sharply peaked at the mean field solution. For
large but finite N , there will be deviations away from mean field [10 -- 12, 23]. These devia-
tions can be characterized in terms of a distribution over an ensemble of coupled networks
that are all prepared with different initial conditions and parameter values. Here, we show
how a perturbation theory in N−1 can be developed to expand around the mean field solu-
tion. This requires the construction of the probability density functional over the ensemble
of spiking neural networks. We adapt the tools of statistical field theory to perform such a
construction.
6
Formalism
as
The complete description of the system given by equations (4), (7), and (8) can be written
u(t) + βu(t) − β
∂
∂t
η(θ, γ, t) +
∂
∂θ
dγ F (π, γ, u(t))η(π, γ, t) = 0
[F (θ, γ, u(t))η(θ, γ, t)] ≡ Lη = 0
(9)
(10)
(cid:90)
(cid:90)
The probability density functional governing the system specified by the synaptic drive
and Klimontovich equations (9) and (10) given initial conditions (η0, u0) can be written as
Du0(t)Dη0(θ, γ) P [η, uη0, u0] P0[η0, u0, γ]
P [η, u] =
(11)
where P [η, uη0, u0] is the conditional probability density functional of the functions (η, u),
and P0[η0, u0] is the density functional over initial conditions of the system. The integral is
a Feynman-Kac path integral over all allowed initial condition functions. Formally we can
write P [η, uη0, u0] as a point mass (Dirac delta) located at the solutions of (9) and (10)
given the initial conditions:
δ [Lη − η0δ(t − t0)] δ
dγ F (π, γ, u(t))η(π, γ, t) − u0δ(t − t0)
u + βu − β
(cid:90)
(cid:20)
(cid:21)
The probability density functional (11) is then
P [η, u] =
Du0(t)Dη0(θ, γ) δ [Lη − η0δ(t − t0)]
(cid:90)
(cid:20)
(cid:21)
(cid:90)
(cid:90)
× δ
u + βu − β
dγ F (π, γ, u(t))η(π, γ, t) − u0δ(t − t0)
P0[η0, u0, γ]
(12)
delta is given by δ(x) ∝(cid:82) dk eikx. Using the infinite dimensional Fourier functional transform
Equation (12) can be made useful by noting that the Fourier representation of a Dirac
then gives
P [η, u] =
DηDu e−N S[η,η,u,u].
The exponent S[η, u] in the probability density functional is called the action and has the
form
where
(cid:90)
Sϕ =
S = Su + Sϕ + S0
dθdγdt ϕ(x) [∂tϕ(x) + ∂θF (θ, γ, u(t))ϕ(x)]
7
(13)
(14)
represents the contribution of the transformed neuron density to the action,
(cid:90)
(cid:18)
(cid:90)
Su =
1
N
dt u(t)
u(t) + βu(t) − β
dγF (π, γ, u(t))[ ϕ(π, γ, t) + 1]ϕ(π, γ, t)
(15)
(cid:19)
represents the global synaptic drive, S0[ ϕ0(x0), u0(t0)] represents the initial conditions, and
x = (θ, γ, t). For the case where the neurons are considered to be independent in the initial
state, we have
S0[ ϕ0(x0), u0(t0)] = − 1
N u(t0)u0 − ln(cid:0)1 +(cid:82) dθdγ ϕ0(θ, γ, t0)ρ0(θ, γ, t0)(cid:1)
(16)
where u0 is the initial value of the coupling variable and ρ0(θ, γ, t) is the distribution from
which the initial configuration is drawn for each neuron. The action includes two imaginary
auxiliary response fields (indicated with a tilde), which are the infinite dimensional Fourier
transform variables. The factor of 1/N appears to ensure correct scaling between the u and
ϕ variables since u applies to a single neuron while ϕ applies to the entire population. The
full derivation is given in [12] and a review of path integral methods applied to differential
equations is given in [29]. In the course of the derivation we have made a Doi-Peliti-Jannsen
transformation [12, 30], given by
ϕ(x) = η(x)e−η(x)
ϕ(x) = eη(x) − 1
In deriving the action, we have explicitly chosen the Ito convention so that the auxiliary
variables only depend on variables in the past. The action (13) contains all the information
about the statistics of the network.
The moments for this distribution can be obtained by taking functional derivatives of
a moment generating functional. Generally, the moment generating function for a random
variable is given by the expectation value of the exponential of that variable with a single
parameter. Because our goal is to transform to new variables for the first and second
cumulants, we form a "two-field" moment generating functional, which includes a second
parameter for pairs of random variables,
(cid:90)
exp(N W [J, K]) =
Dξ exp
(cid:20)
(cid:90)
−N S[ξ] + N
(cid:90)
N
2
dx J i(x)ξi(x) +
8
(cid:21)
dxdx(cid:48)ξi(x)K ij(x, x(cid:48))ξj(x(cid:48))
(17)
where J and K are moment generating fields, ξ1(x) = u(t), ξ2(x) = u(t), ξ3(x) = ϕ(x),
ξ4(x) = ϕ(x), and x = (θ, γ, t). Einstein summation convention is observed beween upper
and lower indices. Unindexed variables represent vectors. The integration measure dx is
assumed to be dt when involving indices 1 and 2. Covariances between an odd and even index
corresponds to a covariance between a field and an auxiliary field. Based on the structure of
the action S and (17) we see that this represents a linear propagator and by causality and
the choice of the Ito convention is only nonzero if the time of the auxiliary field is evaluated
at an earlier time than the field. Covariances between two even indices correspond to that
between two auxiliary fields and are always zero because of the Ito convention.
The mean and covariances of ξ can be obtained by taking derivatives of the action W [J, K]
in (17), with respect to J and K and setting J and K to zero:
δW
δJ i = (cid:104)ξi(cid:105)J,K=0
(cid:104)ξiξj(cid:105)
δW
δK ij =
1
2
(cid:12)(cid:12)(cid:12)(cid:12)J,K=0
Expressions for these moments can be computed by expanding the path integral in (17)
perturbatively around some mean field solution. However, this can be unwieldy if closed
form expressions for the mean field equations do not exist. Alternatively, the moments at
any order can be expressed as self-consistent dynamical equations that can be analyzed or
simulated numerically. Such equations form a set of generalized activity equations for the
means ai = (cid:104)ξi(cid:105), and covariances Cij = N [(cid:104)ξiξj(cid:105) − aiaj].
We derive the generalized activity equations by Legendre transforming the action W ,
which is a function of J and K, to an effective action Γ that is a function of a and C. Just
as a Fourier transform expresses a function in terms of its frequencies, a Legendre transform
expresses a convex function in terms of its derivatives. This is appropriate for our case
because the moments are derivatives of the action. The Legendre transform of W [J, K] is
Γ[a, C] = −W [J, K] +
which must obey the constraints
(cid:90)
(cid:90)
dxdx(cid:48)(cid:20)
(cid:21)
dxJ iai +
1
2
aiaj +
1
N
Cij
K ij
(18)
(cid:20)
δW
δJ i = ai
1
δW
δK ij =
2
aiaj +
1
N
Cij
(cid:21)
9
and
δΓ
δai
δΓ
δCij
(cid:2)K ij + K ji(cid:3)
1
2
aj
≡ Γi,00 = J i +
≡ Γ0,ij =
1
2N
K ij
(19)
The generalized activity equations are given by the equations of motion of the effective
action, in direct analogy to the Euler-Lagrange equations of classical mechanics, and are
obtained by setting J i = 0 and K ij = 0 in (19).
In essence, what the effective action does is to take a probabilistic (statistical mechani-
cal) system in the variables ξ with action S and transform them to a deterministic (classical
mechanical) system with an action Γ. Our approach here follows that used in [15] to con-
struct generalized activity equations for the Wilson Cowan model. However, there are major
differences between that system and this one. In [15], the microscopic equations were for
the spike counts of an inherently probabilistic model so the effective action and ensuing
generalized activity equations could be constructed directly from the Markovian spike count
dynamics. Here, we start from deterministically firing individual neurons and get to a prob-
abilistic description through the Klimontovich equation.
It would be straightforward to
include stochastic effects into the spiking dynamics.
Using (18) in (17) gives
exp(−N Γ[a, C]) =
(cid:90)
(cid:90)
(cid:20)
dxdx(cid:48)(cid:20)
(cid:90)
Dψ exp
−N S[ξ] + N
dx J i(ξi − ai)
(cid:21)
(cid:21)
+
N
2
ξiξj − aiaj − 1
N
Cij
K ij
(20)
where J and K are constrainted by (19). We cannot compute the effective action explicitly
but we can compute it perturbatively in N−1. We first perform a shift ξi = ai + ψi, expand
the action as S[a + ψ] = S[a] +(cid:82) dx(Li[a]ψi + (1/2)(cid:82) dx(cid:48)Lij[a]ψiψj) +··· and substitute for
J and K with the constraints (19) to obtain
exp(−N Γ[a, C]) = exp(−N S[a] − N Tr Γ0,ijCij)
Dψ exp
−N
dx
Li[a]ψi
(cid:90)
+
1
2
(cid:19)
(cid:90)
where
Tr AijBij =
dxdx(cid:48)Aij(x, x(cid:48))Bij(x, x(cid:48))
10
(cid:90)
(cid:90)
(cid:18)
(cid:20)
(cid:90)
(cid:90)
dx(cid:48)Lij[a]ψiψj
+ N
dx Γi,00ψi + N 2
dxdx(cid:48)ψiψjΓ0,ij
(cid:21)
(21)
(22)
Our goal is to construct an expansion for Γ by collecting terms in successive orders of N−1
in the path integral of (21). Expanding Γ as Γ[a, C] = Γ0 + N−1Γ1 + N−2Γ2 and equating
coefficients of N in (21) immediately leads to the conclusion that Γ0 = S[a], which gives
exp(−N Γ[a, C]) = exp(cid:0)−N S[a] − Tr Γ0,ij
1 Cij
Dψ exp
(cid:1)(cid:90)
(cid:90)
(cid:90)
(cid:20)
−N
2
+ N
(cid:21)
dxLij[a]ψiψj
dx Γ0,ij
1 ψiψj
where higher order terms in N−1 are not included. To lowest nonzero order Γ0,ij = N−1Γ0,ij
since Γ0 is only a function of a and not C. If we set
1
1 = (1/2)Lij − (1/2)Qij,
Γ0,ij
we obtain
(cid:18)
exp(−N Γ[a, C]) = exp
Tr LijCij +
1
2
Tr QijCij
(cid:90)
−N S[a] − 1
2
Dψ exp
×
(cid:20)
(cid:90)
−N
2
dx Qij[a]ψiψj
(cid:19)
(cid:21)
(23)
(24)
to order N−1. Qij is an unknown function of a and C, which we will deduce using self-
consistency. The path integral in (24), which is an infinite dimensional Gaussian that
is proportional to 1/(cid:112)det Qij = exp(−(1/2) ln det Qij) =
(cid:18)
(cid:19)
can be explicitly integrated,
exp(−(1/2) Tr ln Qij), using properties of matrices. Hence, (24) becomes
exp(−N Γ[a, C]) = exp
−N S[a] − 1
2
Tr LijCij − 1
2
Tr QijCij +
Tr ln Qij
and
Γ[a, C] = S[a] +
1
2N
Tr LijCij +
1
2N
Tr ln Qij − 1
2N
Tr QijCij
Taking the derivative of Γ with respect to Cij yields
Γ0,ij =
1
2N
Lij + (Q−1)kl ∂
∂Cij
Qlk − ∂
∂Cij
(QklClk)
(cid:18)
Self consistency with (23) then requires that Qij = (C−1)ij which leads to the effective action
Γ[a, C] = S[a] +
1
2N
Tr ln(C−1)ij +
1
2N
Tr LijCij
(25)
1
2
(cid:19)
where
(cid:90)
dx(cid:48) (C−1)ik(x, x(cid:48))Ckj(x(cid:48), x0) = δijδ(x − x0)
11
and we have dropped the irrelevant constant terms.
The equations of motion to order N−1 are obtained from (19) with J i and K ij set to zero:
+
1
2N
δ
δai
Tr LijCij = 0
[−(C−1)ij + Lij] = 0
δS[a]
δai
1
2N
and (27) can be rewritten as(cid:90)
dx(cid:48)Lik(x, x(cid:48))Ckj(x(cid:48), x0) = δijδ(x − x0)
(26)
(27)
(28)
Hence, given any network of spiking neurons, we can write down the a set of generalized
activity equations for the mean and covariance functions by 1) constructing a neuron density
function, 2) writing down the conservation law (Klimontovich equation), 3) constructing
the action and 4) using formulas (26) and (28). We could have constructed these equations
directly by multiplying the Klimontovich and synaptic drive equations by various factors of u
and η and recombining. However, as we saw in [15] this is not a straightforward calculation.
The effective action approach makes this much more systematic and mechanical.
Phase Model Example
We now present a simple example to demonstrate the concepts and approximations in-
volved in our expansion. Our goal is not to analyze the system per se but only to demon-
strate the application of our method in a heuristic setting. We begin with a simple nonleaky
integrate-and-fire neuron model, which responds to a global coupling variable. This is a
special case of the dynamics given above, with F given by
F [θ, γ, u] = I(t) + γu
The action from (14) and (15) is
S[a] =
dθdγdt a4(x) [∂ta3(x) + ∂θ(I + γa1(t))a3(x)]
(cid:90)
(cid:90)
+
1
N
(cid:18)
(cid:90)
dt a2(t)
a1(t) + βa1(t) − β
dγ (I + γa1(t))[a4(π, γ, t) + 1]a3(π, γ, t)
(29)
(cid:19)
(30)
and we ignore initial conditions for now.
12
(cid:90)
1
N
1
N
(cid:90)
=
+
=
=
In order to construct the generalized activity equations we need to compute the first and
second derivatives of the action Li and Lij. Taking the first derivative of (30) gives
L1[a](x, x(cid:48)) =
δS[a(x)]
δa1(t(cid:48))
L2[a](x, x(cid:48)) =
L3[a](x, x(cid:48)) =
δS[a(x)]
δa2(t(cid:48))
δS[a(x)]
δa3(x(cid:48))
L4[a](x, x(cid:48)) =
δS[a(x)]
δa4(x(cid:48))
(cid:90)
dt a2(t)
d
dt
(cid:90)
δ(t − t(cid:48)) + βa2(t(cid:48)) − a2(t(cid:48))β
dθdγ dtγa4(x)∂θa3(x)δ(t − t(cid:48))
(cid:20)(cid:90)
(cid:20)da1
dt(cid:48) + βa1(t(cid:48)) − β
dt a4(θ(cid:48), γ(cid:48), t)∂tδ(t − t(cid:48)) +
a2(t(cid:48))(I + γ(cid:48)a1(t(cid:48)))(a4(π, γ(cid:48), t(cid:48)) + 1)δ(π − θ(cid:48))
− β
N
= ∂t(cid:48)a3(x(cid:48)) + ∂θ(cid:48)(I + γ(cid:48)a1(t(cid:48)))a3(x(cid:48)) − β
N
dγ γ[a4(π, γ, t(cid:48)) + 1]a3(π, γ, t(cid:48))
(cid:21)
(cid:90)
dγ(I + γa1(t(cid:48)))[a4(π, γ, t(cid:48)) + 1]a3(π, γ, t(cid:48))
dθa4(θ, γ(cid:48), t(cid:48))∂θ(I + γ(cid:48)a1(t(cid:48)))δ(θ − θ(cid:48))
(cid:21)
a2(t(cid:48))(I + γ(cid:48)a1(t(cid:48)))a3(π, γ(cid:48), t(cid:48))δ(π − θ(cid:48))
(31)
The mean field equations are obtained by solving Li = 0 using (31). We immediately see
that a2 = a4 = 0 are solutions, which leaves us with
(cid:90)
a1 + βa1 − β
dγ(I + γa1)a3(π, γ, t) = 0
∂ta3 + (I + γa1)∂θa3 = 0
(32)
(33)
The mean field equations should be compared to those of the spike response model [31, 32].
We can also solve (33) directly to obtain
(cid:18)
(cid:90) t
(cid:19)
a3(x, t) = ρ0
θ −
dt(cid:48) [IΩ(t(cid:48)) + γa1(t(cid:48))] , γ, Ω
where ρ0 is the initial distribution.
If the neurons are initially distributed uniformly in
phase, then ρ0 = g(γ)/2π and the mean field equations reduce to
t0
a1(t) + βa1(t) − β
2π
(I + ¯γa1(t)) = 0
(34)
which has the form of the Wilson-Cowan equation, with (β/2π) (I + ¯γa1) acting as a gain
function. Hence, the Wilson-Cowan equation is a full description of the infinitely large
system limit of a network of globally coupled simple phase oscillators in the asynchronous
state. For all other initial conditions, the one-neuron conservation equation (called the
Vlasov equation in kinetic theory) must be included in mean field theory.
13
To go beyond mean field theory we need to compute Lij(x, x(cid:48), x(cid:48)(cid:48)) = δLi(x, x(cid:48))/δaj(x(cid:48)(cid:48)):
1
N
(cid:20)
(cid:20)
γ(cid:48)(cid:48)(cid:90)
(cid:20)
L11[a] = 0
L12[a] =
L13[a] =
L14[a] =
(cid:90)
(cid:20) d
dt(cid:48) + β − β
(cid:90)
dγ γ[a4(π, γ, t(cid:48)(cid:48)) + 1]a3(π, γ, t(cid:48)(cid:48))]
δ(t(cid:48)(cid:48) − t(cid:48))
− d
dt(cid:48)(cid:48) + β − β
dθ a4(x)δ(γ − γ(cid:48)(cid:48))∂θδ(θ − θ(cid:48)(cid:48)) − β
N
γ(cid:48)(cid:48)a2(t(cid:48))[a4(π, γ(cid:48)(cid:48), t(cid:48)(cid:48)) + 1]δ(π − θ(cid:48)(cid:48))
(cid:21)
δ(t(cid:48) − t(cid:48)(cid:48))
γ(cid:48)(cid:48)∂θ(cid:48)(cid:48)a3(x(cid:48)(cid:48)) − β
N
γ(cid:48)(cid:48)a2(t(cid:48)(cid:48))a3(π, γ(cid:48)(cid:48), t(cid:48)(cid:48))δ(π − θ(cid:48)(cid:48))
δ(t(cid:48) − t(cid:48)(cid:48))
dγ γ[a4(π, γ, t(cid:48)) + 1]a3(π, γ, t(cid:48))
δ(t(cid:48) − t(cid:48)(cid:48))
(cid:21)
(cid:21)
(cid:21)
L21[a] =
1
N
L22[a] = 0
L23[a] = − β
N
L24[a] = − β
N
(cid:20)(cid:90)
(I + γ(cid:48)(cid:48)a1(t(cid:48)))[a4(π, γ(cid:48)(cid:48), t(cid:48))) + 1]δ(π − θ(cid:48))δ(t(cid:48) − t(cid:48)(cid:48))
(I + γ(cid:48)(cid:48)a1(t(cid:48)))a3(π, γ(cid:48)(cid:48), t(cid:48))]δ(π − θ(cid:48)(cid:48))δ(t(cid:48) − t(cid:48)(cid:48))
(cid:21)
δ(t(cid:48) − t(cid:48)(cid:48))
L31[a] =
L32[a] = − β
N
dθ a4(θ, γ(cid:48), t(cid:48))γ(cid:48)∂θδ(θ − θ(cid:48)) − β
N
(I + γ(cid:48)a1(t(cid:48)))(a4(π, γ(cid:48), t(cid:48)) + 1)δ(π − θ(cid:48))δ(t(cid:48) − t(cid:48)(cid:48))
a2(t(cid:48))γ(cid:48)[a4(π, γ(cid:48), t(cid:48)) + 1]δ(π − θ(cid:48))
L33[a] = 0
L34[a] = [δ(θ(cid:48) − θ(cid:48)(cid:48))∂t(cid:48)(cid:48) − ∂θ(cid:48)(cid:48)(I + γ(cid:48)a1(t(cid:48)))δ(θ(cid:48)(cid:48) − θ(cid:48))
(cid:21)
− β
N
a2(t(cid:48))(I + γ(cid:48)a1(t(cid:48)))δ(π − θ(cid:48))δ(π − θ(cid:48)(cid:48))
δ(γ(cid:48) − γ(cid:48)(cid:48))δ(t(cid:48)(cid:48) − t(cid:48))
(cid:20)
(cid:21)
L41[a] =
∂θ(cid:48)γ(cid:48)a3(x(cid:48)) − β
N
a2(t(cid:48))γ(cid:48)a3(π, γ(cid:48), t(cid:48))δ(π − θ(cid:48))
δ(t(cid:48) − t(cid:48)(cid:48))
(I + γ(cid:48)a1(t(cid:48)))a3(π, γ(cid:48), t(cid:48))δ(π − θ(cid:48))δ(t(cid:48) − t(cid:48)(cid:48))
L42[a] = − β
N
L43[a] = ∂t(cid:48)δ(x(cid:48) − x(cid:48)(cid:48)) + ∂θ(cid:48)(I + γ(cid:48)a1(t(cid:48)))δ(x(cid:48) − x(cid:48)(cid:48))
a2(t(cid:48))(I + γa1(t(cid:48)))δ(π − θ(cid:48))δ(π − θ(cid:48)(cid:48))δ(γ(cid:48) − γ(cid:48)(cid:48))δ(t(cid:48) − t(cid:48)(cid:48))
− β
N
L44[a] = 0
The activity equations for the means to order N−1 are given by (26). The only nonzero
14
contributions are given by L13 and L31 resulting in
(cid:90)
(cid:90)
L2 +
L4 +
1
2N
1
2N
δ
δa2
δ
δa4
dxdx(cid:48)(L13C13 + L31C31) = 0
dxdx(cid:48)(L13C13 + L31C31) = 0
since a2 = a4 = 0 and correlations involving response variables (even indices) will be zero
for equal times. The full activity equations for the means are thus
(cid:90)
(cid:90)
a1 + βa1 − β
dγ(I + γa1)a3(π, γ, t) − β
N
∂ta3 + (I + γa1)∂θa3 +
1
N
γ∂θC(θ, γ, t) = 0
where C(θ, γ, t) = C13(t; θ, γ, t) = C31(θ, γ, t; t).
dγ γC(π, γ, t) = 0
(35)
(36)
We can now use the Lij in (28) to obtain activity equations for Cij. There will be sixteen
coupled equations in total but the applicable nonzero ones are
+ β − β
dγ γa3(π, γ, t)]
dγ (I + γa1)C31(π, γ, t; t0)
+ β − β
dγ γa3(π, γ, t)
dγ (I + γa1)C33(π, γ, t; x0)
dγ (I + γa1(t))a3(π, γ, t)C41(π, γ, t; t0) = 0
dγ (I + γa1(t))a3(π, γ, t)C43(π, γ, t; x0) = 0
(37)
(38)
(cid:20) d
(cid:20) d
dt
dt
(cid:90)
(cid:90)
(cid:90)
(cid:90)
(cid:21)
(cid:21)
(cid:90)
(cid:90)
C11(t; t0) − β
− β
C13(t; x0) − β
− β
γ∂θa3(x)C11(t; t0) + [∂t+(I + γa1)∂θ]C31(x; t0)
− β
N
(I + γa1(t))a3(π, γ, t)δ(π − θ)C21(t, t0) = 0
(39)
γ∂θa3(x)C13(t; x0) + [∂t+(I + γa1(t))∂θ]C33(x, x0)
− β
N
(I + γa1(t))a3(π, γ, t)δ(π − θ)C23(t, x0) = 0
(40)
Adding (38) and (39) and taking the limit t0 → t and setting θ0 = θ, γ0 = γ gives
(cid:21)
(cid:90)
(cid:20)
(cid:90)
∂tC(θ, γ, t) +
β − β
dγ(cid:48) γ(cid:48)a3(π, γ(cid:48), t) + (I + γa1)∂θ
C(θ, γ, t) − β
dγ(cid:48) (I + γ(cid:48)a1)C33(π, γ(cid:48), t; x)
− 2β(I + γa1(t))a3(π, γ, t)δ(π − θ) + γ∂θa3(x)C11(t; t) = 0
where we use the fact that C21(t, t(cid:48)) = N and C43(x; x(cid:48)) = δ(θ − θ(cid:48))δ(γ − γ(cid:48)) in the limit of
t(cid:48) approaching t from below and equal to zero when approaching from above. Adding (37)
15
+ 2β − 2β
dγ γa3(π, γ, t)]
C11(t; t) − 2β
dγ (I + γa1)C(π, γ, t) = 0
[∂t + (I + γa1(t))∂θ] C33(x; x) + 2γ[∂θa3(x)]C(x) = 0
because C41(x; t) = 0 and C23(t; x) = 0. Putting this all together, we get the generalized
gives
(cid:20) d
dt
activity equations
(cid:90)
(cid:90)
(cid:21)
(cid:90)
(cid:90)
and (40) to themselves with t and t0 interchanged and taking the limit of t0 approaching t
+ βa1(t) − β
da1
dt
dγ(I + γa1(t))a3(π, γ, t) − β
N
dγ γC(π, γ, t) = 0
∂ta3(θ, γ, t) + (I + γa1)∂θa3(θ, γ, t) +
γ∂θC(θ, γ, t) = 0
(cid:20)
(cid:90)
1
N
(cid:21)
β − β
dγ(cid:48) γ(cid:48)a3(π, γ(cid:48), t) + (I + γa1)∂θ
C(θ, γ, t)
dγ(cid:48) (I + γ(cid:48)a1)C33(π, γ(cid:48), t; θ, γ, t) − 2β(I + γa1(t))a3(θ, γ, t)δ(π − θ)
(cid:90)
∂tC(θ, γ, t) +
− β
(cid:20) d
dt
(cid:90)
+ γ∂θa3(θ, γ, t)C11(t; t) = 0
+ 2β − 2β
dγ γa3(π, γ, t)]
C11(t; t) − 2β
dγ (I + γa1)C(π, γ, t) = 0
[∂t + (I + γa1(t))∂θ] C33(θ, γ, t; θ, γ, t) + 2γ∂θa3(θ, γ, t)C(θ, γ, t) = 0
(cid:21)
(cid:90)
(41)
(42)
(43)
(44)
(45)
Initial conditions, which are specified in the action, are required for each of these equations.
The derivation of these equations using classical means require careful consideration for each
particular model. Our method provides a blanket mechanistic algorithm. We propose that
these equations represent a new scheme for studying neural networks.
Equations (41)-(45) are the complete self-consistent generalized activity equations for
the mean and correlations to order N−1. It is a system of partial differential equations in
t and θ. These equations can be directly analyzed or numerically simulated. Although the
equations seem complicated, one must bear in mind that they represent the dynamics of
the system averaged over initial conditions and unknown parameters. Hence, the solution of
this PDE system replaces multiple simulations of the original system. In previous work, we
required over a million simulations of the original system to obtained adequate statistics [12].
There is also a possibility that simplifying approximations can be applied to such systems.
The system has complete phase memory because the original system was fully deterministic.
However, the inclusion of stochastic effects will shorten the memory and possibly simplify the
16
dynamics. It will pose no problem to include such stochastic effects. In fact, the formalism
is actually more suited for stochastic systems [15].
Discussion
The main goal of this paper was to show how to systematically derive generalized ac-
tivity equations for the ensemble averaged moments of a deterministically coupled network
of spiking neurons. Our method utilizes a path integral formalism that makes the process
algorithmic. The resulting equations could be derived using more conventional perturba-
tive methods although possibly with more calculational difficulty as we found before [15].
For example, for the case of the stochastic spike model, Buice et al. [15] presumed that
the Wilson-Cowan activity variable was the rate of a Poisson process and derived a sys-
tem of generalized activity equations that corresponded to deviations around Poisson firing.
Bressloff [33], on the other hand, assumed that the Wilson-Cowan activity variable was a
mean density and used a system-size expansion to derive an alternative set of generalized
activity equations for the spike model. The classical derivations of these two interpretations
look quite different and the differences and similarities between them are not readily appar-
ent. However, the connections between the two types of expansions are very transparent
using the path integral formalism.
Here, we derived equations for the rate and covariances (first and second cumulants) of a
deterministic synaptically coupled spiking network as a system size expansion to first order.
However, our method is not restricted to these choices. What is particularly advantageous
about the path integral formalism is that it is straightforward to generalize to include higher
order cumulants, extend to higher orders in the inverse system size, or to expand in other
small parameters such as the inverse of a slow time scale. The action fully specifies the
system and all questions regarding the system can be addressed with it.
To give a concrete illustration of the method, we derived the self-consistent generalized
activity equations for the rates and covariances to order N−1 for a simple phase model.
The resulting equations consist of ordinary and partial differential equations. This is to be
expected since the original system was fully deterministic and memory cannot be lost. Even
mean field theory requires the solution of an advective partial differential equation. The
properties of these and similar equations remain to be explored computationally and ana-
17
lytically. The system is possibly simpler near the asynchronous state, which is marginally
stable in mean field theory like the Kuramoto model [7] and like the Kuramoto model, we
conjecture that the finite size effects will stabilize the asynchronous state [10, 11]. The addi-
tion of noise will also stabilize the asynchronous state. Near asynchrony could be exploited
to generate simplified versions of the asynchronous state.
We considered a globally connected network, which allowed us to assume that networks
for different parameter values and initial conditions converge towards a "typical" system in
the large N limit. However, this property may not hold for more realistic networks. While
the formalism describing the ensemble average will hold regardless of this assumption, the
utility of the equations as descriptions of a particular network behavior may suffer. For
example, heterogeneity in the connectivity (as opposed to the global connectivity we con-
sider here) may threaten this assumption. This is the case with so called "chaotic random
networks" [34] in which there is a spin-glass transition owing to the variance of the connec-
tivity crossing a critical threshold. This results in the loss of a "typical" system in the large
N limit requiring an effective stochastic equation which incorporates the noise induced by
the network heterogeneity. Whether the expansion we present here is useful without further
consideration depends upon whether the network heterogeneity induces this sort of effect.
This is an area for future work. A simpler issue arises when there are a small discrete num-
ber of "typical" systems (such as with bistable solutions to the continuity equation). In this
case, there are noise induced transitions between states. While the formalism has a means
of computing this transition [35], we do not consider this case here.
An alternative means to incorporate heterogeneous connections is to consider a network
of coupled systems. In such a network, a set of generalized activity equations, such as those
derived here or simplified versions, would be derived for each local system, together with
equations governing the covariances between the local systems. Correlation based learning
dynamics could then be imposed on the connections between the local systems. Such a
network could serve as a generalization of current rate based neural networks to include the
effects of spike correlations with applications to both neuroscience and machine learning.
18
Acknowledgments
This work was supported by the Intramural Research Program of the NIH, NIDDK.
[1] H. Markram and collaborators, The Human Brain Project: a report to the European commis-
sion (The HBP-PS Consortium, Lausanne, 2012).
[2] E. M. Izhikevich and G. M. Edelman, Proceedings of the national Academy, USA 105, 3593
(2008).
[3] C. Eliasmith, T. C. Stewart, X. Choo, T. Bekolay, T. DeWolf, Y. Tang, and D. Rasmussen,
Science 338, 1202 (2012).
[4] S. Ichimaru, Basic principles of plasma physics, a statistical approach. (Benjamin, New York,
1973).
[5] D. R. Nicholson, Introduction to plasma theory (Krieger Publishing Co, Malabar, FL, 1993).
[6] R. Desai and R. Zwanzig, Journal of Statistical Physics 19, 1 (1978).
[7] S. Strogatz and R. Mirollo, Journal of Statistical Physics 63, 613 (1991).
[8] L. Abbott and C. van Vreeswijk, Physical Review E 48, 1483 (1993).
[9] A. Treves, Network: Computation in Neural Systems 4, 259 (1993).
[10] E. J. Hildebrand, M. A. Buice, and C. C. Chow, Physical Review Letters 98, 054101 (2007).
[11] M. A. Buice and C. C. Chow, Physical Review E 76, 031118 (2007).
[12] M. A. Buice and C. C. Chow, PLoS Computational Biology 9, e1002872 (2013).
[13] E. Salinas and T. J. Sejnowski, The Journal of neuroscience : the official journal of the Society
for Neuroscience 20, 6193 (2000).
[14] M. Buice and J. Cowan, Progress in Biophysics and Molecular Biology 99, 53 (2009).
[15] M. A. Buice, J. D. Cowan, and C. C. Chow, Neural Computation 22, 377 (2010).
[16] G. Ermentrout and N. Kopell, SIAM Journal on Applied Mathematics pp. 233 -- 253 (1986).
[17] B. Ermentrout, Neural computation 8, 979 (1996).
[18] G. Ermentrout and N. Kopell, Journal of Mathematical Biology 29, 195 (1991).
[19] D. Golomb and D. Hansel, Neural computation 12, 1095 (2000).
[20] F. Hoppensteadt and E. Izhikevich, Weakly Connected Neural Networks (Springer-Verlag, New
York, 1997).
19
[21] Y. Kuramoto, Chemical Oscillations, Waves, and Turbulence (Springer, Berlin, 1984).
[22] M. A. Buice and C. C. Chow, Physical Review E 84, 051120 (2011).
[23] M. A. Buice and C. C. Chow, Journal of Statistical Mechanics: Theory and Experiment 2013,
P03003 (2013).
[24] R. L. Liboff, Kinetic Theory (Springer, New York, 2003).
[25] H. McKean Jr, Proceedings of the National Academy of Sciences of the United States of
America 56, 1907 (1966).
[26] O. Faugeras, J. Touboul, and B. Cessac, Frontiers in computational neuroscience 3, 1 (2009).
[27] J. Touboul, Physica D: Nonlinear Phenomena 241, 1223 (2012).
[28] J. Baladron, D. Fasoli, O.
, and J. Touboul, J. Math. Neurosci. 2, 10 (2012).
[29] M. A. Buice and C. C. Chow, arXiv.org p. 1009.5966 (2010).
[30] H. Janssen and U. Tauber, Annals of Physics 315, 147 (2005).
[31] W. Gerstner, Physical Review E 51, 738 (1995).
[32] W. Gerstner, Neural computation 12, 43 (2000).
[33] P. C. Bressloff, SIAM Journal on Applied Mathematics 70, 1488 (2010).
[34] H. Sompolinsky, A. Crisanti, and H. Sommers, Physical Review Letters 61, 259 (1988).
[35] V. Elgart and A. Kamenev, arXiv cond-mat.stat-mech (2004).
20
|
1205.6438 | 1 | 1205 | 2012-05-29T18:03:26 | Stimulus-dependent maximum entropy models of neural population codes | [
"q-bio.NC",
"physics.bio-ph"
] | Neural populations encode information about their stimulus in a collective fashion, by joint activity patterns of spiking and silence. A full account of this mapping from stimulus to neural activity is given by the conditional probability distribution over neural codewords given the sensory input. To be able to infer a model for this distribution from large-scale neural recordings, we introduce a stimulus-dependent maximum entropy (SDME) model---a minimal extension of the canonical linear-nonlinear model of a single neuron, to a pairwise-coupled neural population. The model is able to capture the single-cell response properties as well as the correlations in neural spiking due to shared stimulus and due to effective neuron-to-neuron connections. Here we show that in a population of 100 retinal ganglion cells in the salamander retina responding to temporal white-noise stimuli, dependencies between cells play an important encoding role. As a result, the SDME model gives a more accurate account of single cell responses and in particular outperforms uncoupled models in reproducing the distributions of codewords emitted in response to a stimulus. We show how the SDME model, in conjunction with static maximum entropy models of population vocabulary, can be used to estimate information-theoretic quantities like surprise and information transmission in a neural population. | q-bio.NC | q-bio |
Stimulus-dependent maximum entropy models of neural population codes
Einat Granot-Atedgi=,a, Gasper Tkacik∗=,b, Ronen Segev≡,c, and Elad Schneidman≡,a
aDepartment of Neurobiology, Weizmann Institute of Science, 76100 Rehovot, Israel
bInstitute of Science and Technology Austria, Am Campus 1, A-3400 Klosterneuburg, Austria
cFacutly of Natural Sciences, Department of Life Sciences and Zlotowski Center for Neuroscience,
Ben Gurion University of the Negev, 84105 Be'er Sheva, Israel
(Dated: September 14, 2018)
Neural populations encode information about their stimulus in a collective fashion, by joint activ-
ity patterns of spiking and silence. A full account of this mapping from stimulus to neural activity
is given by the conditional probability distribution over neural codewords given the sensory input.
To be able to infer a model for this distribution from large-scale neural recordings, we introduce a
stimulus-dependent maximum entropy (SDME) model -- a minimal extension of the canonical linear-
nonlinear model of a single neuron, to a pairwise-coupled neural population. The model is able to
capture the single-cell response properties as well as the correlations in neural spiking due to shared
stimulus and due to effective neuron-to-neuron connections. Here we show that in a population
of 100 retinal ganglion cells in the salamander retina responding to temporal white-noise stimuli,
dependencies between cells play an important encoding role. As a result, the SDME model gives
a more accurate account of single cell responses and in particular outperforms uncoupled models
in reproducing the distributions of codewords emitted in response to a stimulus. We show how the
SDME model, in conjunction with static maximum entropy models of population vocabulary, can
be used to estimate information-theoretic quantities like surprise and information transmission in a
neural population.
INTRODUCTION
Neurons represent and transmit information using
temporal sequences of short stereotyped bursts of electri-
cal activity, or spikes [1]. Much of what we know about
this encoding has been learned by studying the mapping
between stimuli and responses at the level of single neu-
rons, and building detailed models of what stimulus fea-
tures drive a single neuron to spike [2 -- 4]. In most of the
nervous system, however, information is represented by
joint activity patterns of spiking and silence over pop-
ulations of cells.
In a sensory context, these patterns
can be thought of as codewords that convey informa-
tion about external stimuli to the central nervous system.
One of the challenges of neuroscience is to understand the
neural codebook -- a map from the stimuli to the neural
codewords -- a task made difficult by the fact that neu-
rons respond to the stimulus neither deterministically nor
independently.
The structure of correlations among the neurons deter-
mines the organization of the code, that is, how different
stimuli are represented by the population activity [5 -- 8].
These correlations also determine what the brain, having
no access to the stimulus apart from the spikes coming
from the sensory periphery, can learn about the outside
world [9 -- 11]. The source of these correlations, which
arise either from the correlated external stimuli to the
neurons, from "shared" local input from other neurons,
∗ [email protected]
=,≡ equal contributions
or from "private" independent noise, have been heavily
debated [12 -- 15].
In many neural systems, the correla-
tion between pairs of (even nearby or functionally simi-
lar) neurons was found to be weak [16, 17, 26]. Similarly,
the redundancy between pairs in terms of the information
they convey about their stimuli was also typically weak
[18 -- 20]. The low correlations and redundancies between
pairs of neurons therefore led to the suggestion that neu-
rons in larger populations might encode information in-
dependently [21], which was echoed by theoretical ideas
of maximally efficient neural codes [22 -- 24].
Recent studies of the neural code in large populations
have, however, revealed that while the typical pairwise
correlations may be weak, larger populations of neurons
can nevertheless be strongly correlated as a whole [25 --
33]. Maximum entropy models of neural populations
have shown that such strong network correlations can be
the result of collective effects of pairwise dependencies
between cells, and, in some cases, of sparse high-order
dependencies [26, 34, 35]. Most of these studies have
characterized the strength of network effects and spik-
ing synchrony at the level of the total vocabulary of the
population, i.e. the distribution of codewords averaged
over all the stimuli. It is not immediately clear how these
findings affect stimulus encoding, where one needs to dis-
tinguish the impact of correlated stimuli that the cells
receive ("stimulus correlations"), from the impact of co-
variance of the cells conditional on the stimulus ("noise
correlations"). For small populations of neurons, it has
been shown that taking into account correlations for de-
coding or reconstructing the stimulus can be beneficial
compared to the case where correlations are neglected
(e.g. [35 -- 39]). Similarly, generalized linear models high-
lighted the importance of dependencies between cells in
accounting for correlations between pairs and triplets of
retinal ganglion cell responses [40].
Here we present a new encoding model that allows us
to study in fine detail the codebook of large neural popu-
lations. Importantly, this model gives a joint probability
distribution over the activity patterns of the whole pop-
ulation for a given stimulus, while capturing both the
stimulus and noise correlations. This new model belongs
to a class of maximum entropy models with strong links
to statistical physics [27, 41 -- 52] and is directly related
to maximum entropy models of neural vocabulary [26 --
32], allowing us estimate the entropy and its derivative
quantities for the neural code. In sum, the maximum en-
tropy framework enables us to progress towards our goal
of focusing attention on the level of joint patterns of ac-
tivity, rather than capturing low-level statistics (e.g., the
individual firing rates) of the neural code alone.
We start by showing that linear-nonlinear (LN) mod-
els of retinal ganglion cells responding to spatially un-
structured stimuli capture a significant part of the single
neuron response, but still miss much of the detail;
in
particular, we show that they fail to capture the correla-
tion structure of firing among the cells. We next present
our new stimulus-dependent maximum entropy (SDME)
model, which is a hybrid between linear-nonlinear mod-
els for single cells and the pairwise maximum entropy
models. Applied to groups of ∼ 100 neurons recorded si-
multaneously, we find that SDME models outperform the
LN models for the stimulus-response mapping of single
cells and, crucially, give a significantly better account of
the distribution of codewords in the neural population.
RESULTS
We recorded the simultaneous spiking activity of ∼ 110
ganglion cells from the salamander retina [53], presented
with repeats of a 10 s long full-field flicker ("Gaussian
FFF") movie, where the light intensity on the screen
was sampled independently from a Gaussian distribution
with a frequency of 30 Hz (Fig. 1a). This "frozen noise"
stimulus was repeated 726 times, for a total of ∼ 2 h of
stimulation. Most of the recorded cells exhibited tempo-
ral OFF-like behaviors (Fig. 1b). We chose for further
analysis N = 100 cells that were reliably sorted, demon-
strated a robust and stable response over repeats, and
generated at least 2500 spikes during the course of the
experiment.
We discretized neural responses into ∆t = 10 ms bins,
and represented the activity of the neurons in response
to the stimulus as binary words in each of the time
bins.
If neuron i = 1, . . . , N was active in time bin t,
we denoted a spike (or more spikes) as xi(t) = 1, and
xi(t) = 0 if it was silent.
In this representation, the
whole experiment yielded a total of about T ∼ 7.3 · 105
2
FIG. 1: Response of a large population of ganglion
cells to a 10 s long repeated visual stimulus. (a) White
noise uncorrelated Gaussian stimulus presented at 30 Hz and
the spiking patterns of 3 cells to repeated presentations of the
stimulus. (b) Spike-trigerred averages of 110 simultaneously
recorded cells; a subset of 100 cells was chosen for further
analysis. (c) The histogram of pairwise correlations between
cells for repeated Gaussian white noise stimulus (green), re-
peated natural pixel movie (red), and non-repeated natural
pixel movie (blue) [35]. (d) Average pairwise correlation co-
efficient between cells as a function of the distance (mean and
std are across pairs of cells at a given distance).
i
binary word samples. Using repeated presentations of the
same movie, we estimated the average response of each
of the cells across repeats, ri(t) = (cid:104)xi(t)(cid:105)rep, or the peri-
stimulus time histogram (PSTH). Following Refs. [4, 54],
we fitted a linear-nonlinear model for each of the cells in
the experiment, such that the predicted rate rLN
(t) =
Ni(ki · s(t)), where ki is a linear filter matched for the
i-th cell, Ni is its point-wise nonlinear function, and s(t)
is the stimulus fragment from time t − τ until t (here
we used τ = 400 ms, making s(t) a vector of light in-
tensities with 40 components). Linear filters were recon-
structed using reverse correlation (spike-triggered aver-
age), and nonlinearities were obtained by histograming
P (ki · s(t)spike) into K = 20 adaptively-sized bins and
(t) = Ni(ki · s) = P (spikeki · s(t)) by in-
obtaining rLN
verting P (ki · s(t)spike) using Bayes' rule. These LN
models captured most of structure of the PSTH, yet as
the example cell in Fig. 2a shows, they often misesti-
mated the exact firing rates of the neuron, or sometimes
even missed parts of the neural response altogether. In
the Gaussian FFF condition, the normalized (Pearson)
correlation between the measured and predicted PSTH,
(t)), was 0.69 ± 0.06 (mean ± std across
Corr(ri(t), rLN
100 cells).
i
i
The performance gap of the canonical LN models in
stimuluscell 1cell 2time (s)cell 301234567891000.20.40.60.8100.10.20.30.40.5C(xi,xj)ProbabilityNaturalPixelGaussianRepNaturalPixelRep05001000150000.10.2Distance between electrodes (μm)<C(xi,xj)>abcd−0.4−0.3−0.2−0.10−101time from spike (s)Normalized stimulus predicting single neuron responses suggests that either
the single-neuron models need to be improved to account
for the observed behavior, or that interactions between
neurons play an important encoding role and need to be
included. While firing rate prediction performance can
be improved for single neurons by models with higher-
dimensional stimulus sensitivity (e.g. [54, 55]) or dynam-
ical aspects of spiking behavior (e.g.
[56, 57]), previous
work, as well as the results below, demonstrated that
even conditionally-independent models which by con-
struction perfectly reproduce the firing rate behavior of
single cells, often fail to capture the measured correla-
tion structure of firing between pairs of cells, as well as
higher-order statistical structure [26].
We find two salient features of the correlations between
pairs of neurons: (i) the pairwise correlations between
cells in response to the Gaussian FFF movie are typi-
cally weak but are not zero (Fig. 1c, consistently with
previous reports [26, 27, 32]); (ii) the correlation in neu-
ral activities shows a fast decay with distance despite the
infinite correlation length of the stimulus, but the decay
does not reach zero correlation even at relatively large
distances (Fig. 1d). This salient structure, along with
any other potential statistical correlation at the pairwise
order, is characterized by the covariance matrix of ac-
tivities, Cij = (cid:104)xixj(cid:105) − (cid:104)xi(cid:105)(cid:104)xj(cid:105), where the averages are
taken across time and repeats.
We would like to find a model of the neural code that
will be able to reproduce these pairwise statistics. Mo-
tivated by the shortcomings of the canonical LN model,
we therefore asked whether a model that combined the
LN (receptive-field based) aspect of single cells with the
interactions between cells, could give a better account
of the neural stimulus-response mapping.
Importantly,
the new model should capture not only the firing rate
of single cells and the full pairwise correlation structure
between them, but should also accurately predict the full
distribution of the joint activity patterns across the whole
population. Because the joint distributions of activity are
high-dimensional (e.g., the distribution over codewords
across the duration of the experiment, P ({xi}), has 2N
components), this is a very demanding benchmark for
any model.
Here we propose the simplest extension to the
conditionally-independent set of LN models for each cell
in the recorded population, by including pairwise cou-
plings between cells, so that the spiking of cell i can
increase or decrease the probability of spiking for cell
j [58, 59].
In contrast to previous proposals, this cou-
pling will be introduced so that the resulting model is a
maximum-entropy model for P ({xi}s), the conditional
distribution over population activity patterns given the
stimulus. We recall that the maximum entropy models
give the most parsimonious probabilistic description of
the joint activity patterns, which perfectly reproduces
a chosen set of measured statistics over these patterns,
3
without making any additional assumptions [60].
We start by introducing the least structured (maxi-
mum entropy) distribution P (x1, x2, . . . , xNt) that re-
produces exactly the observed average firing rate for each
time bin t and for each neuron i, ri(t) = (cid:104)xi(t)(cid:105)data =
(cid:104)xi(t)(cid:105)P , as well as the overall correlation matrix Cij be-
tween all pairs of cells (c.f. [61]). Thus, we seek P ({xi}t)
that maximizes L:
L [P ({xi}t)] = − (cid:88)
(cid:88)
(cid:88)
(cid:88)
i,t
1
2
+
+
ij
+
{xi},t
P ({xi}t) log2 P ({xi}t)
{xi},t
αi(t)[(cid:104)xi(t)(cid:105)P − (cid:104)xi(t)(cid:105)data]
βij[(cid:104)xixj(cid:105)P,t − (cid:104)xixj(cid:105)data]
µ(t)[P ({xi}t) − 1],
(1)
where the subscript to brackets (cid:104)·(cid:105) denotes whether the
averaging is done over the maximum entropy distribu-
tion (P ), or over the recorded data; Lagrange multipliers
µ ensure that the distributions are normalized. This is
an optimization problem for parameters αi(t) and βij,
which has a unique solution since the entropy is convex.
The functional form of the solution to this optimization
problem is well-known; in our case it can be written as
P T DM E({xi}t) =
1
Z(t)
exp
αi(t)xi +
N(cid:88)
i=1
N(cid:88)
i,j=1
1
2
βijxixj
,
(2)
where the individual time-dependent parameters for
each of the cells, αi(t), and the stimulus-independent
pairwise interaction terms βij, are set to match the
measured firing rates ri(t) and the pairwise correla-
tions Cij; Z(t) is a normalization factor or parti-
tion function for each time bin t, given by Z(t) =
(cid:80){xi} exp
(cid:16)(cid:80)
i αi(t)xi + 1
2
ij βijxixj
.
(cid:80)
(cid:17)
time-dependent maximum entropy (TDME)
The
model
in Eq. (2) is equivalent to an Ising model
from physics, where the single-cell parameters are time-
dependent local fields acting on each of the neurons
(spins), and static (stimulus-independent) infinite-range
interaction terms couple each pair of spins.
In the
limit where interactions go to zero, βij = 0, the model
in Eq. (2) becomes the full conditionally-independent
model, itself a maximum entropy model that reproduces
exactly the firing rate of every neuron, ri(t); in this case
the probability distribution factorizes, and the solution
for αi(t) and Z(t) becomes trivially computable from the
firing rates, ri(t). Time-dependent maximum entropy
models are powerful, since they make no assumptions
about how the stimulus drives the response; they can
serve as useful benchmarks for other models (especially
the conditionally independent model with βij = 0). On
the other hand, these models require repeated stimulus
presentations to fit, involve a number of parameters that
grows linearly with the duration of the stimulus, do not
generalize to new stimuli, and do not provide an explicit
map from the stimuli to the responses.
To make a direct link to the stimulus and allow com-
parison with a set of uncoupled LN models, we take
the general time-dependent maximum entropy model of
Eq. (2) and specialize it to a particular form of stimu-
lus dependence. Rather than having an arbitrary time-
dependent parameter for every neuron for each time
bin, αi(t), we assume that this dependence takes place
through the stimulus projection alone,
i.e. αi(t) =
αi(ki · s(t)), much like in an LN model, where the neu-
ral firing depends on the value of the stimulus projec-
tion onto the linear filter ki. This choice is made purely
for the sake of convenience: the model could be gener-
alized to, e.g., neurons that depend on two linear pro-
jections of the stimulus, by making αi depend jointly on
(k1 · s(t), k2 · s(t)), although such models would be pro-
gressively more difficult to infer from data.
Concretely, we estimated the linear filter ki for each
cell i using reverse correlation, and convolved the filter
with the stimulus sequence, s(t), to get the "generator
signal" gi(t) = ki · s(t). We then looked for the max-
imum entropy probability distribution P ({xi}s(t)), by
requiring that the average firing rate of every cell given
the generator signal is the same in the data and under
the model, i.e. (cid:104)xi(gi)(cid:105)data = (cid:104)xi(gi)(cid:105)P (see Methods);
as before, we also required that the model reproduced
the overall correlation between every pair of cells, Cij.
This gives then a stimulus-dependent maximum entropy
(SDME) model, which takes the following form:
P SDM E({xi}s(t)) =
1
Z(s(t))
exp
αi(gi(t))xi +
N(cid:88)
i=1
N(cid:88)
i,j=1
1
2
βijxixj
(3)
.
The parameters of this model are: N × (N − 1)/2 cou-
plings βij, K × N parameters αi, and a linear filter ki
for each cell. We used a Monte Carlo based gradient de-
scent learning procedure to find the model parameters
α, β numerically (see Methods).
By construction, the SDME model exactly reproduces
the covariance in activities, Cij, between all pairs of cells,
and also the LN model properties of every cell: an arbi-
trary nonlinear function N can be encoded by properly
choosing how parameters αi depend on the linear projec-
tions of the stimulus, gi. We can construct a maximum
entropy model with βij = 0 (no constraints on the pair-
wise correlations Cij). The result is a set of uncoupled
(conditionally independent) LN models:
P LN ({xi}s(t)) ≡ N(cid:89)
1
Zi(s(t))
exp (αi(gi(t))xi) (4)
i=1
4
FIG. 2: SDME model predicts the firing rate of single
cells better than LN models. (a) Example of the PSTH
segment for one cell (blue), the best prediction of an LN model
(green) and of a SDME model (red). (b) Correlation between
the true PSTH and SDME model prediction (vertical axis) vs.
the correlation between the true PSTH and the LN model
prediction (horizontal axis); each plot symbol is a separate
cell, dotted line shows equality. The neuron chosen in panel
(a) is shown in orange.
N(cid:89)
i=1
=
Ni(ki · s(t)).
In sum, the time-dependent maximum entropy (TDME)
model of Eq. (2) is an extension of conditionally indepen-
dent model that additionally reproduces the measured
pairwise correlations between cells.
In a directly anal-
ogous way, the stimulus-dependent maximum entropy
(SDME) model of Eq. (3) is an extension to the set of
uncoupled LN models, Eq. (4), that additionally repro-
duces the measured pairwise correlations between cells.
Because P SDM E (Eq. 3) agrees with P LN (Eq. 4) exactly
in all constrained single-neuron statistics, any improve-
ment in prediction of the SDME, be it in the firing rate
or the codeword distributions, can be directly ascribed
to the effect of the interaction terms, βij.
i
We found that the SDME predicted the firing rate of
single cells better than the LN models, with the nor-
malized correlation coefficient between the true and pre-
(t)) being 0.74±0.06
dicted firing rate, Corr(ri(t), rSDM E
(mean ± std across 100 cells), as shown in Fig. 2b.
The differences between the SDME and the LN models
become more striking on the level of the activity pat-
terns of the whole population. Figures 3a,b show the
log-likelihood ratio for the population activity patterns
x = {xi} under the two models, showing that the SDME
can be orders of magnitude better in predicting the pop-
ulation neural response. These differences are large in
particular for those stimuli that elicit a strong response
(Fig. 3c), that is, precisely where the response consists of
synchronous spiking and the structure of the codewords
can be nontrivial. Moreover, the difference between the
models becomes increasingly significant with the size of
the population N , and particularly apparent for groups
of more than 20 cells (Fig. 3d).
0500100015000102030time (ms)Firing rate (spikes/s) rrSDMErLN0.40.50.60.70.80.90.40.50.60.70.80.9Corr (r,r LN)Corr (r,rSDME) ab 5
FIG. 3: SDME model predicts population activity patterns for N = 100 neurons better than conditionally
independent LN models. (a) The log-likelihood ratio of the population firing patterns under the SDME model and under a
collection of LN models, shown as a function of time (red) for an example stimulus repeat. For reference, the average population
firing rate is shown in grey. (b) A scatterplot of the log-likelihoods under the SDME and LN models; each dot represents an
average over activity patterns {xi} (across repeats) at a given time bin; dotted line shows equality. (c) The log-likelihood ratio
under the SDME and LN models as a function of the average firing rate in the population; SDME outperforms LN models in
particular for patterns with more spiking activity. (d) The average likelihood ratio between the SDME and LN models as a
function of the population size, N (error bars = std over 10 randomly chosen groups of neurons at that N ).
We next examined how well various models for the neu-
ral codebook, P ({xi}s), explain the total vocabulary,
that is, the distribution of neural codewords observed
across the whole duration of the experiment, P ({xi}) =
(cid:104)P ({xi}s(t))(cid:105)t. Despite the nominally large space of pos-
sible codewords -- much larger than the total number of
samples in the experiment (2N (cid:29) T ) -- the sparsity of
spikes and the correlations between neurons restrict the
vocabulary to a much smaller set of patterns. Some of
these occur many times during our stimulus presenta-
tion, allowing us to estimate their empirical probability,
P data({xi}), directly from the experiment, and compare
it to the model prediction [35]. The most prominent ex-
ample of such frequently observed codewords is the silent
pattern, xi = 0, which is seen ∼ 72% of the time.
Figure 4 shows the likelihood ratio of the model prob-
ability and empirical probability for various codewords
observed in the experiment, as a function of the rate at
which these codewords appear in the experiment. While
SDME model in Fig. 4a does not predict the frequencies
of all patterns perfectly, it strongly outperforms the un-
coupled set of LN models in Fig. 4b, and has a slightly
better performance than the full conditionally indepen-
dent model (Fig. 4c), despite the fact that the latter is
determined by N × 1000 = 1 · 105 parameters, the fir-
ing rates of every cell in every time bin. On average,
SDME predicts the probabilities of the patterns of ac-
tivity with no bias, and with a standard deviation of
log(P SDM E/P data) of about 1; uncoupled LN models in
comparison are biased and have a spread that is more
than twice as large. Even more striking is the fact that
LN models assign such low probabilities to some code-
words that they are never generated during our Monte
Carlo sampling (and are therefore not even shown in scat-
terplots of Fig. 4), while they are frequently observed in
the experiment. This discrepancy is quantified by enu-
merating the M most probable patterns in the data and
in the model (by sampling; see Methods), and measuring
the size of the intersection of the two sets of patterns;
in other words, we ask if the model is even able to ac-
cess all the patterns that one is likely to record in the
experiment. As shown in the third row of Fig. 4, SDME
does well on this task, with 434 codewords in the inter-
section of the 500 most likely patterns in the data and
the model; this is a much better performance than for
the uncoupled model, and slightly better than for the
conditionally independent model.
The SDME model was constructed to capture ex-
actly the total correlations in neuronal spiking, Cij =
(cid:104)xixj(cid:105)−(cid:104)xi(cid:105)(cid:104)xj(cid:105). With repeated stimulus, this total cor-
relation can be broken down into the signal and noise
components. The signal correlations, C s
ij, are inferred
by applying the same formula as for the total correla-
tion, but on the spiking raster where the repeated trial
indices have been randomly and independently permuted
for each time bin. This removes any correlation due to
interactions between spikes on simultaneously recorded
trials, and only leaves the correlations induced by the
response being locked to the stimulus. The noise cor-
relation, C n
ij, is then defined as the difference between
ij = Cij − C s
the total and the signal components, C n
ij.
We calculated the noise correlations between all pairs
in our N = 100 neuron dataset. By their definition,
the conditionally independent models cannot reproduce
C n
ij, which are always zero. To assess the performance
of the SDME, we drew samples from our model distri-
bution using the Monte Carlo simulation and compared
the noise correlations in the simulated rasters to the true
noise correlations. The model prediction tightly corre-
lates with the measured values, as shown in Fig. 5. We
observe a systematic deviation of ∼ 25%, most likely be-
cause the assumed dependence on the stimulus through
0102030400102030405678905101520time (s)Population rate (spikes/s/neuron)05101520051015051015Population rate (spikes/s/neuron)02040608000.20.40.60.81Population Size abcd 6
FIG. 5: Measured vs predicted noise correlations for
the SDME model. Noise correlation (see text) is estimated
from recorded data for every pair of neurons, and plotted
against the noise correlation predicted by the SDME model
(each pair of neurons = one dot; shown are N (N − 1)/2 dots
for N = 100 neurons). Conditionally independent models
predict zero noise correlation for all pairs.
tion. As such, they represent only functional and not
physical (i.e.
synaptic) connections between the cells.
Fig. 6b shows the pairwise interaction map for 100 cells;
the histogram of their values (in Fig. 6c) reflects that they
can be of both signs, but the distribution has a stronger
positive tail, i.e. a number of cell pairs tend to spike
together or be silent together with a probability than is
higher than expected from their respective LN models.
We can compare these interactions to the interactions
of a static (non-stimulus-dependent) maximum entropy
model for the population vocabulary [26, 28]:
(cid:88)
i
P M E({xi}) =
1
Z0
exp
α0
i xi +
1
2
. (5)
(cid:88)
ij
β0
ijxixj
In this model for the total distribution of codewords,
there is no stimulus dependence, and the parameters α0
i
and β0
ij are chosen to that the distribution is as random as
possible, while reproducing exactly the measured mean
firing rate of every neuron (cid:104)xi(cid:105)data = (cid:104)xi(cid:105)P M E , and every
pairwise correlation, (cid:104)xixj(cid:105)data = (cid:104)xixj(cid:105)P M E , across the
whole duration of the experiment.
Interestingly, we find that the pairwise interaction
terms in the SDME model of Eq. (3) are closely related
to the interactions in the static pairwise maximum en-
tropy model of Eq. (5): SDME interactions, βij, tend to
be smaller in magnitude, but have an equal sign and rel-
ative ordering, as the static ME interactions, β0
ij. Some
degree of correspondence is expected: an interaction be-
tween neurons i and j in the static ME model captures
the combined effect of the stimulus and noise correlations,
FIG. 4: The performance of various models in account-
ing for the total vocabulary of the population, P ({xi}).
The results for the SDME model are shown in (a), the results
for an uncoupled set of LN models in (b), the results for a
full conditionally independent model in (c). The first row
displays the log ratio of model to empirical probabilities for
various codewords (dots), as a function of that codeword's
empirical frequency in the recorded data. The model prob-
abilities were estimated by generating Monte Carlo samples
from the corresponding model distributions (see Methods);
only patterns that were generated in the MC run as well as
found in the recorded data are shown. The second row sum-
marizes this scatterplot by binning codewords according to
their frequency, and showing the average log probability ra-
tio in the bin (solid line), as well as the 1 std scatter across
the codewords in the bin (shaded area). The highly probable
all-silent state, {xi} = 0, is shown separately as a circle. The
third row shows the overlap between 500 most frequent pat-
terns in the data and 500 most likely patterns generated by
the model (see text).
one linear filter per neuron is insufficient to capture the
complete dependence on stimulus, thereby underestimat-
ing the full structure of stimulus correlation and inducing
an excess in the noise correlation. Despite this, the de-
gree of correspondence in noise correlations observed in
Fig. 5 is telling us that SDME has clearly captured a large
amount of noise covariance structure in neural firing.
How should we interpret the inferred parameters of the
SDME model? LN models have a clear "mechanistic" in-
terpretation in terms of the cell's receptive field and the
nonlinear spiking mechanism. Here, similarly, the stim-
ulus dependent part of the model for each cell, αi, is a
nonlinear function of a filtered version of the stimulus
gi(t) = ki · s(t); in the absence of neuron-to-neuron cou-
plings, the nonlinearity of every neuron would correspond
to Ni(gi) ∼ f (αi(gi)), where f (·) = exp(·)/(1 + exp(·)),
according to Eq. (4). The dependence of αi on the stim-
ulus projection gi is similar across the recorded cells as
shown in Fig. 6a; as expected, higher overlaps with the
linear filter induce higher probability of spiking.
The pairwise interaction terms in the model, βij, are
symmetric, static, and stimulus independent by construc-
43427040866662302309292SDME modelLN (β=0) modelcond. ind. modelabc 7
that is, the information can be written as a difference
of the entropy of the neural vocabulary, and the noise
entropy (the average of the entropy of the codebook),
where the entropy is S[p(x)] = −(cid:82) dx p(x) log2 p(x). Be-
cause of the maximum entropy property of our model
for P M E({xi}), the entropy of our static pairwise model
in Eq. (5) is an upper bound on the transmitted infor-
mation; expressed as an entropy rate, this amounts to
s ≡ S[P M E({xi})]/∆t ≈ 730 bit/s.
The brain does not have direct access to the stimulus,
but only receives codewords {xi}, "drawn" from P ({xi}),
by the retina. At every moment in time, − log2 P ({xi})
measures the surprise about the output of the retina,
and thus about the stimulus. We, as experimenters -- but
not the brain -- have access to stimulus repeats and thus
to P ({xi}s(t)), so we can compute the average value of
surprise (per unit time) at every instant t in the stimulus:
S(t) = − 1
∆t
P ({xi}s(t)) log2 P ({xi}).
(7)
(cid:88)
{xi}
This quantity can be expressed using the entropies and
the learned parameters of our maximum entropy models,
and is plotted as a function of time in Fig. 7. Since aver-
aging across time is equal to averaging over the stimulus
ensemble, we see from Eq. (7) that (cid:104)S(t)(cid:105)t would have
to be identically equal to S[P ({x})] under the condition
that (cid:104)P ({xi}s(t))(cid:105)t = P ({xi}) (marginalization). Since
we build models for P ({xi}) (static ME) and P ({xi}s)
(SDME) from data independently, they need not obey
the marginalization condition exactly, but they will do
so if they provide a good account of the data. Indeed, by
using the static ME and SDME distributions in Eq. (7)
for surprise, we find that (cid:104)S(t)(cid:105)t ≈ 740 bit/s, very close
to the entropy rate s of the total vocabulary and within
our estimated 1% error bars on entropy computation.
To estimate the information transmission, we have to
subtract the noise entropy rate from the output entropy
rate s, as dictated by Eq. (6). The entropy of the SDME
model is an upper bound on the noise entropy; since this
is not a lower bound, we cannot put a strict bound on
the information transmission, but can nevertheless esti-
mate it. Figure 7 shows the "instantaneous information",
I(t) = S(t) − S[P SDM E({xi}s(t))]/∆t, as a function of
time; from Eq. (6), the mutual information rate is a time
average of this quantity, R = I({xi}; s)/∆t = (cid:104)I(t)(cid:105)t.
We find R ≈ 130 bit/s. This quantity can be compared
to the total entropy rate of the stimulus itself (which must
be higher than R), which in our case is ≈ 210 bit/s (see
Methods). While our estimates seem to indicate that a
lot of vocabulary bandwidth (730 bit/s) is "lost" to noise
(600 bit/s), the last comparison shows that the Gaussian
FFF stimulus source itself is not very rich, so that the
estimated information transmission takes up more than
half of the actual entropy rate of the source.
FIG. 6: SDME model parameters. (a) Average values of
the LN-like driving term, αi(gi), where gi = ki · s, across all
cells i (error bars = std across cells), for each of the K = 20
adaptive bins for gi (see Methods). (b) Pairwise interaction
map βij of the SDME model, between all N = 100 neurons in
the experiment. (c) Histogram of pairwise interaction values
from (b), and their average value as a function of the distance
between cells (inset). (d) For each pair of cells i and j, we
plot the value of β0
ij under the static maximum entropy model
of Eq. (5) vs. the βij from the stimulus-dependent maximum
entropy model of Eq. (3).
while in the corresponding SDME interaction, (most of)
the stimulus correlation has been factored out into the
correlated dynamics of the inputs to the neurons i and
j, i.e. αi(gi(t)) and αj(gj(t)). The surprisingly high de-
gree of correspondence, however, indicates that even the
interactions learned from static maximum entropy mod-
els can account for, up to a scaling factor, the pairwise
neuron dependencies that are not due to the correlated
stimulus inputs.
The SDME model is an approximation to the neu-
ral codebook, P ({xi}s), while the static ME model de-
scribes the population vocabulary, P ({xi}). With these
two distributions in hand, we can explore how the pop-
ulation jointly encodes the information about the stimu-
lus into neural codewords -- the joint activity patterns of
spiking and silence. We make use of the fact that we can
estimate the entropy of the maximum entropy distribu-
tions using a procedure of heat capacity integration, as
explained in Refs. [27, 32] (see Methods).
Information
(in bits) per codeword is
I({xi}; s) =
ds P (s)
P ({xi}s) log2
P ({xi}s)
P ({xi})
(cid:90)
(cid:88)
{xi}
= S[P ({xi})] − (cid:104)S[P ({xi}s)](cid:105)P (s);
(6)
Lastly, we asked how important is the inclusion of
20406080100−2−101220406080100−2−101234020040060080010001200# interactionsij04008001200−0.200.20.40.60.811.2Distance (μm)<βij>βijβij05101520−20−15−10−50adaptive bin for gi = −202468−202468β (static ME model)βij (SDME model)ijabcd0 8
the single cell parameters α are static, SDME becomes
the static maximum entropy model of the population vo-
cabulary; if the couplings β are 0, SDME becomes a set
of uncoupled LN models. The framework is general, and
could be easily applied to other neural systems.
We applied this modeling framework to the salaman-
der retina presented with Gaussian white noise stimuli,
and found that the interactions between neurons play
an important role in determining the detailed patterns
of population response. In particular, the SDME model
gave better prediction of PSTH of single cells, yielded
orders of magnitude improvement in describing the pop-
ulation patterns, and captured significant aspects of noise
correlations. The deviations between the SDME and the
uncoupled LN model became significant for > 20 cells,
and tended to occur at "interesting" times in the stimu-
lus, precisely when the neural population was not silent.
The responses of the neural system in the maximum
entropy framework are binary codewords of spiking and
silence across the neural population. The choice of
timescale over which these codewords are defined, here
∆t = 10 ms, is short enough such that multiple spikes are
rarely observed in the same time bin, but long enough so
that most of the strong spike-spike interactions (as well
as fine temporal detail, such as spike-timing jitter) occur
within a single bin. This allows us to view successive time
bins as codewords, although some statistical dependence
between them remains, possibly in the conditional SDME
model (due to multiple timescales on which the neurons
respond to stimuli), and certainly in the static ME model
[31].
If we were to make the time scale much shorter,
e.g. by an order of magnitude, we could make the con-
ditional independence assumption of the responses given
the stimuli and previous spiking, which would lead us
to GLM models [40] or nonequilibrium generalizations of
Ising models [47]. GLMs, in particular, are excellent gen-
erating models for precise spiking rasters, are easy to in-
fer, and allow for asymmetric couplings between neurons.
However, the inference in all these cases is tractable be-
cause there are no interactions between the spikes within
the same time bin (as there are in SDME). This neces-
sitates the use of very short time bins and introduces
strong dependencies between successive time bins, mak-
ing the interpretation of the discretized neural responses
in terms of individual codewords difficult. For this rea-
son, GLM and SDME are complementary approaches:
the first allows for a temporally-detailed probabilistic de-
scription of a spiking process, while the second gives an
explicit expression for the probability distribution over
codewords in longer temporal bins.
SDME allowed us to improve over LN models for sala-
mander retinal ganglion cells both in terms of the PSTH
prediction and, especially, in terms of population activity
patterns. Interestingly, for parasol cells in the macaque
retina under flickering checkerboard stimulation, the gen-
eralized linear model did not yield firing rate improve-
FIG. 7: Surprise and information transmission esti-
mated from the SDME model. (a) Surprise rate (blue) is
estimated from the static ME and SDME models assuming in-
dependence of codewords across time bins. The instantaneous
information rate (red) is the difference between the surprise
and the noise entropy rate, estimated from the SDME model
(see text). The information transmission rate is the average
of the instantaneous information across time. (b) Population
firing rate as a function of time shows that bursts of spiking
strongly correlate with the bursts of surprise and information
transmission in the population. (c) The stimulus (normalized
to zero mean and unit variance) is shown for reference as a
function of time.
pairwise interactions, βij, into the SDME model, com-
pared to a set of uncoupled LN models, when account-
ing for information transmission. We therefore esti-
mated the noise entropy rate for a set of uncoupled LN
models, S[P LN ({xi}s(t))]/∆t, which was found to be
≈ 770 bit/s, considerably higher than the noise entropy
of the SDME model. Crucially, this noise entropy rate
is larger than the total entropy rate s estimated above,
which is impossible for consistent models of the neural
codebook and the vocabulary (since it would lead to
negative information rates). This failure is a quantita-
tive demonstration of the inability of the uncoupled LN
models to reproduce the statistics of the population vo-
cabulary, as shown in Fig. 4b, despite a seemingly small
performance difference on the level of single cell PSTH
prediction.
DISCUSSION
We presented a modeling framework for stimulus en-
coding by large populations of neurons, which combines
an individual neuronal receptive field model, with the
ability to include pairwise interactions between neurons.
The result is a stimulus-dependent maximum entropy
(SDME) model, which is the most parsimonious model of
the population response to the stimulus that reproduces
the linear-nonlinear (LN) aspect of single cells, as well as
the correlation structure between neurons. In two limit-
ing cases, the SDME model reduces to known models: if
246810100101102t [s][bits/s/neuron]surprise Sinfo. I01020avg. rate [Hz](cid:239)(cid:22)0(cid:22)s(t)time [s] ment relative to uncoupled LN models (but did improve
higher order statistics of firing) [40]. In both cases, how-
ever, the improvements reflect the role of dependencies
among cells in encoding the stimulus, and their effect be-
comes apparent when we ask questions about information
transmission by a neural population. Maximum entropy
models can only put an upper bounds on the total en-
tropy and the noise entropy of the neural code (and this
statement remains true even if successive codewords are
not independent), and as such cannot set a strict bound,
but only give an estimate, for the information transmis-
sion. Nevertheless, ignoring the inter-neuron dependen-
cies and using an uncoupled set of LN models predicts
the total population responses so badly that the esti-
mated noise entropy is higher than the upper bound on
the total entropy, which is a clear impossibility, while the
SDME model gives transmission rates that appear rea-
sonable (and positive), amounting to about 60% of the
source entropy rate.
Tkacik and colleagues [61] have suggested that one can
interpret βij in an SDME model as a prior over the activ-
ity patterns that the population would use to optimally
encode the stimulus. For low noise level they argued
that the prior should be minimal (and could help decor-
relate the responses), as the population could faithfully
encode the stimulus, whereas in the noisy regime, the
prior should match the statistics of the sensory world and
thus counteract the effects of noise. Similarly, Berkes and
colleagues [62] suggested a similar reason for the similar-
ity of ongoing and induced activity patterns in the visual
cortex. Our results show that interactions are necessary
for capturing the network encoding, and implicitly re-
flect the existence of such a prior. The recovered inter-
actions are strongly correlated with the interaction pa-
rameters of a static, stimulus independent model over
the distribution of patterns, making it possible for the
brain (which only has access to the spikes, not the stim-
ulus) to learn these values. Whether the interactions
are matched to the statistics of the visual inputs as sug-
gested by Ref [61] is the focus of future work. In parallel,
increasingly detailed statistical models of neural codes
going beyond SDME (e.g. by including temporal depen-
dencies as in Ref [48]), and efforts to infer such mod-
els from experimental data, should focus our attention
on population-level statistics and on finding principled
information-theoretic measures for quantifying the code,
like the surprise and instantaneous information suggested
here.
METHODS
Electrophysiology. Experiments were performed on
the adult tiger salamander, Ambystoma tigrinum. All ex-
periments were in accordance with Ben-Gurion Univer-
sity of the Negev and government regulations. Extracted
9
retinas were placed with the ganglion cell layer facing a
multielectrode array with 252 electrodes (Ayanda Biosys-
tems, Switzerland), and superfused with oxygenated
Ringer medium at room temperature. Extracellularly
recorded signals were amplified (MultiChannel Systems,
Germany) and digitized at 10k Samples/s, and spike-
sorted using custom software written in MATLAB.
Visual stimulation. Stimuli were projected onto the
retina from a CRT video monitor (ViewSonic G90fB) at
a frame rate of 60 Hz; each movie frame was presented
twice, using standard optics. Full Field Flicker (FFF)
stimuli were generated by independently sampling spa-
tially uniform gray levels (with a resolution of 8 bits)
from a Gaussian distribution, with mean luminance of
147 lux and the standard deviation of 33 lux. These
data allow us to estimate the entropy rate of the source
(as used in the main text), by multiplying the entropy
of the luminance distribution with the refresh rate. To
estimate the cells' receptive fields, checkerboard stimu-
lus was generated by selecting each checker (∼ 100 µm
on the retina) randomly every 33 ms to be either black
or white. To identify the RF centers, a two-dimensional
Gaussian was fitted to the spatial profile of the response.
The movies were gamma corrected for the computer mon-
itor. In all cases the visual stimulus entirely covered the
retinal patch that was used for the experiment.
Inferring SDME from data. The LN model for
each neuron i consists of the linear filter ki, and the non-
linear function Ni, which is defined pointwise on a set of
binned values for the generator signal, gi = ki · s. We
used binning into K = 20 bins such that initially each
bin contains roughly the same number of values for gi,
but subsequently the binning is adaptively adjusted (sep-
arately for each neuron) to be denser at higher values of
gi, where the firing rates are higher. We fitted LN models
with varying number of K bins, and have chosen K = 20
when the performance of the LN models appeared to sat-
urate [63].
To find the parameters of the stimulus-dependent max-
imum entropy model (αi(gi), βij), we retained the bin-
ning of the generator signal used for LN model construc-
tion. Given trial values for the SDME parameters, we es-
timated the chosen expectation values (covariance matrix
Cij in firing, and the firing rate conditional on gi, ri(gi))
by Monte Carlo sampling from the trial distribution in
Eq. (3); the learning step of the algorithm is computed by
comparing the expectation values in the trial distribution
and the empirical distribution (computed over the train-
ing half of the stimulus repeats). In detail, we used a gra-
dient ascent algorithm, applying a combination of Gibbs
sampling and importance sampling in order to efficiently
estimate the gradient, by using optimizations similar to
those described in Ref. [64]. Sampling was carried out
in parallel on a 16 node cluster with two 2.66GHz Intel
Quad-Core Xeon processors and 16GB of memory per
node. The calculation was terminated when the average
error in firing rates and coincident firing rates reached
below 1% and 5% respectively, which is within the ex-
perimental error.
To compute the single neuron PSTH and compare the
distributions of codewords from the model to the empir-
ical distribution, we used Metropolis Monte Carlo sam-
pling to draw codewords from the model distributions;
we drew 5000 independent samples (to draw uncorre-
lated configurations, a sample was recorded only after
100 "spin-flip" trials) for every timepoint, for a total of
5 · 106 samples; the same procedure was used also to
draw from the uncoupled (β = 0) models. To estimate
the entropies of high dimensional SDME distributions,
we used the "heat capacity integration" method, detailed
in Ref [32]. Briefly, a maximum entropy model P (x) =
Z−1 exp(−E(x)) (where E is the Hamiltonian function
determined by the choice of constrained operators and
the conjugated parameters) is extended by introducing a
new parameter T , much like the temperature in physics,
T exp(−E(x)/T ). The entropy of the
so that PT (x) = Z−1
0 C(T )/T dT , where
E(T )/T 2, and the variance
the heat capacity C(T ) = σ2
in energy can be estimated at each T by Monte Carlo
sampling.
In practice, we run a separate Monte Carlo
sampling for a finely discretized interval of temperatures,
T ∈ [0, 1], estimate C(T ) for each temperature, and nu-
merically integrate to get the entropy S. We have pre-
viously shown that this procedure yields robust entropy
estimates even for large numbers of neurons [27, 32].
distribution is given by S[PT =1] =(cid:82) 1
Acknowledgements. This work was supported by
The Israel Science Foundation and the Human Frontiers
Science Program.
[1] Rieke F, Warland D, de Ruyter van Steveninck RR &
Bialek W (1997) Spikes: Exploring the Neural Code. MIT
Press, Cambridge, MA.
[2] Aguera y Arcas B & Fairhall AL (2003) What causes a
neuron to spike? Neural Comput 15: 1789 -- 1807.
[3] Bialek W & de Ruyter van Steveninck RR (2005) Fea-
tures and dimensions: Motion estimation in fly vision.
arXiv.org:q-bio/0505003.
[4] Schwartz O, Pillow JW, Rust NC & Simoncelli EP (2006)
Spike-triggered neural characterization. J Vis 6: 484 --
507.
[5] Stopfer M, Bhagavan S, Smith BH, Laurent G (1997)
Impaired odour discrimination on desynchronization of
odour-encoding neural assemblies. Nature 390:70 -- 4.
[6] Riehle A, Grun S, Diesmann M, Aertsen A (1997) Spike
synchronization and rate modulation differentially in-
volved in motor cortical function. Science 278:1950 -- 3.
[7] Harris KD, Csicsvari J, Hirase H, Dragoi G, Buzs´aki G
(2003) Organization of cell assemblies in the hippocam-
pus. Nature 424:552 -- 6.
[8] Averbeck BB, Lee D (2004) Coding and transmission
of information by neural ensembles. Trends Neurosci
27:225 -- 30.
10
[9] Brunel N, Nadal JP (1998) Mutual information, Fisher
information, and population coding. Neural Comp
10:1731 -- 1757.
[10] Abbott LF, Dayan P (1998) he Effect of Correlated Vari-
ability on the Accuracy of a Population Code. Neural
Comp. 11:91-102
[11] Sompolinsky H, Yoon H, KAng K, Shamir M (2001) Pop-
ulation coding in neuronal systems with correlated noise.
Phys Rev E 64:8095 -- 8100.
[12] Schneidman E, Bialek W, Berry MJ (2003) Synergy, re-
dundancy, and independence in population codes. J Neu-
rosci 23:11539 -- 53.
[13] Pola, G, Thiele A, Hoffmann K-P, Panzeri, S (2003) An
exact method to quantify the information transmitted
by different mechanisms of correlational coding. Network:
Comput. Neural Syst. 14:35 -- 60.
[14] Nirenberg S, Latham PE (2003) Decoding neuronal spike
trains: How important are correlations?. Proc. Natl.
Acad. Sci. USA 100:7348 -- 7353.
[15] Averbeck B, Latham PR, Pouget A (2006) Neural cor-
relations, population coding and computation. Nat Rev
Neurosci 7:358 -- 366 .
[16] Bair W, Zohary E, Newsome WT (2001) Correlated firing
in macaque visual area mt: time scales and relationship
to behavior. Journal of Neuroscience 21:1676 -- 97.
[17] Ecker AS, Berens P, Keliris GA, Bethge M, Logothetis
NK, Tolias AS (2010) Decorrelated neuronal firing in cor-
tical microcircuits. Science 327: 584 -- 7.
[18] Puchalla JL, Schneidman E, Harris RA & Berry MJ 2nd
(2005) Redundancy in the population code of the retina.
Neuron 46: 493 -- 504.
[19] Narayanan NS, Kimchi EY, Laubach M (2005) Redun-
dancy and synergy of neuronal ensembles in motor cor-
tex. J Neurosci 25:4207 -- 16.
[20] Chechik G, et al. (2006) Reduction of information re-
dundancy in the ascending auditory pathway. Neuron
51:359 -- 68.
[21] Nirenberg S, Carcieri SM, Jacobs AL & Latham PE
(2001) Retinal ganglion cells act largely as independent
encoders. Nature 411: 698 -- 701.
[22] HB Barlow, Possible principles underlying the transfor-
mation of sensory messages. Sensory communication, ed
Rosenblith W (MIT Press, Cambridge, MA), pp 217 -- 234
(1961).
[23] JJ Atick & AN Redlich, Towards a theory of early visual
processing. Neural Comp 2, 308 -- 320 (1990).
[24] Barlow H (2001) Redundancy reduction revisited. Net-
work 12:241 -- 53.
[25] Schnitzer MJ, Meister M (2003) Multineuronal firing pat-
terns in the signal from eye to brain. Neuron 37:499 -- 511.
[26] Schneidman E, Berry MJ 2nd, Segev R & Bialek W
(2006) Weak pairwise correlations imply strongly corre-
lated network states in a neural population. Nature 440:
1007 -- 12.
[27] G Tkacik, E Schneidman, MJ Berry II & W Bialek
(2006) Ising models for networks of real neurons. arXiv:q-
bio/0611072.
[28] J Shlens, GD Field, JL Gaulthier, MI Grivich, D Petr-
usca, A Sher, AM Litke & EJ Chichilnisky (2006)
The structure of multi-neuron firing patterns in primate
retina. J Neurosci 26: 8254-66.
[29] A Tang, D Jackson, J Hobbs, W Chen, JL Smith, H
PAtel, A Prieto, D Petruscam MI Grivich, A Sher, P
Hottowy, W Dabrowski, AM Litke & JM Beggs (2008) A
maximum entropy model applied to spatial and temporal
correlations from cortical networks in vitro. J Neurosci
28: 505 -- 518.
[30] J Shlens, GD Field, JL Gaulthier, M Greschner, A Sher,
AM Litke & EJ Chichilnisky (2009) The structure of
large-scale synchronized firing in primate retina. J Neu-
rosci 29: 5022 -- 31.
[31] O Marre, SE Boustani, Y Fregnac & A Destexhe (2009)
Prediction of spatio -- temporal patterns of neural activity
from pairwise correlations. Phys Rev Lett 102: 138101.
[32] G Tkacik, E Schneidman, MJ Berry II & W Bialek
(2009) Spin-glass models for a network of real neurons.
arXiv:0912.5409 (2009).
[33] Ganmor E, Segev R, Schneidman E (2011) The architec-
ture of functional interaction networks in the retina. J
Neurosci 31:3044 -- 54.
[34] Ohiorhenuan IE, Mechler F, Purpura KP, Schmid AM,
Hu Q & Victor JD (2010) Sparse coding and high-order
correlations in fine-scale cortical networks. Nature 466:
617 -- 21.
[35] Ganmor E, Segev R & Schneidman E (2011) Sparse low-
order interaction network underlies a highly correlated
and learnable neural population code. Proc Nat'l Acad
Sci USA 108: 9679 -- 84.
[36] Warland DK, Reinagel P & Meister M (1997) Decoding
visual information from a population of retinal ganglion
cells. J Neurophys 78: 2336 -- 2350.
[37] Dan Y, Alonso JM, Usrey WM, Reid RC (1998) Coding
of visual information by precisely correlated spikes in the
lateral geniculate nucleus. Nat Neurosci 1:501 -- 7.
[38] Hatsopoulos NG, Ojakangas CL, Paninski L, Donoghue
JP (1998) Information about movement direction ob-
tained from synchronous activity of motor cortical neu-
rons. Proc Natl Acad Sci USA 95:15706 -- 11.
[39] Brown EN, Frank LM, Tang D, Quirk MC, Wilson MA
(1998) A statistical paradigm for neural spike train de-
coding applied to position prediction from ensemble fir-
ing patterns of rat hippocampal place cells. J Neurosci
18:7411 -- 25.
[40] Pillow JW, Shlens J, Paninski L, Shear A, Litke AM,
Chichilnisky EJ & Simoncelli EP (2008) Spatio-temporal
correlations and visual signaling in a complete neural
population. Nature 454: 995 -- 9.
[41] E Schneidman, S Still, MJ Berry II & W Bialek (2003)
Network information and connected correlations. Phys
Rev Lett 91: 238701.
[42] S Cocco, S Leibler & R Monasson (2009) Neuronal cou-
plings between retinal ganglion cells inferred by efficient
inverse statistical physics methods. Proc Nat'l Acad Sci
USA 106: 14058 -- 62.
[43] S Cocco & R Monasson (2011) Adaptive cluster expan-
sion for inferring Boltzmann machines with noisy data.
Phys Rev Lett 106: 090601.
[44] Y Roudi, E Aurell & JA Hertz (2009) Statistical physics
of pairwise probability models. Front Comput Neurosci
3: 22.
[45] Y Roudi, S Nirenberg & PE Latham (2009) Pairwise
maximum entropy models for studying large biological
systems: when they can work and when they can't. PLoS
Comput Biol 5: e1000380.
[46] Y Roudi, J Trycha & J Hertz (2009) The ising model for
11
neural data: model quality and approximate methods
for extracting functional connectivity. Phys Rev E 79:
051915.
[47] Y Roudi & J Hertz (2011) Mean field theory for nonequi-
librium network reconstruction. Phys Rev Lett 106:
048702.
[48] Vasquez JC, Marre O, Palacios AG, Berry MJ 2nd &
Cessac B (2011) Gibbs distribution analysis of temporal
correlation structure in retina ganglion cells. J Physiol
Paris, in press.
[49] Macke JH, Opper M & Bethge M (2011) Common input
explains higher-order correlations and entropy in a simple
model of neural population activity. Phys Rev Lett 106:
208102.
[50] M Mezard & T Mora, Constraint satisfaction problems
and neural networks: a statistical physics perspective
(2009) J Physiol Paris 103: 107 -- 113.
[51] B Cessac, H Rostro, JC Vasques & T Vieville (2009)
How Gibbs distributions may naturally arise from synap-
tic adaptation mechanisms. J Stat Phys 136: 565 -- 602.
[52] V Sessak & R Monasson (2009) Small-correlation expan-
sions for the inverse Ising problem. J Phys A 42, 055001.
[53] Segev R, Goodhouse J, Puchalla J & Berry MJ 2nd
(2004) Recording spikes from a large fraction of the gan-
glion cells in a retinal patch. Nat Neurosci 7: 1154 -- 61.
[54] AL Fairhall, CA Burlingame, R Narasimhan, RA Harris,
JL Puchalla & MJ Berry II (2006) Selectivity for multiple
stimulus features in retinal ganglion cells. J Neurophysiol
96: 2724 -- 2738.
[55] G Tkacik, A Ghosh, E Schneidman & R Segev (2012)
Retinal adaptation and invariance to changes in higher-
order stimulus statistics. arXiv.org:1201.3552.
[56] J Keat, P Reinagel, RC Reid & M Meister (2001) Pre-
dicting every spike: a model for the responses of visual
neurons. Neuron 30: 803 -- 817.
[57] Ozuysal Y & Baccus SA (2012) Linking the computa-
tional structure of variance adaptation to biophysical
mechanisms. Neuron 73: 1002-1015.
[58] Tkacik G, Information flow in biological networks. The-
sis, Princeton University, 2007.
[59] Granot-Atdegi E, Tkacik G, Segev R & Schneidman E
(2010) A stimulus-dependent maximum entropy model of
the retinal population neural code. Front Neurosci Con-
ference Abstract COSYNE 2010.
[60] ET Jaynes (1957) Information theory and statistical me-
chanics. Phys Rev 106: 620 -- 630.
[61] Tkacik G, Prentice JS, Balasubramanian V & Schneid-
man E (2010) Optimal population coding by noisy spik-
ing neurons. Proc Nat'l Acad Sci USA 107: 14419 -- 14424.
[62] Berkes P, Orban G, Lengyel M & Fiser J (2011) Spon-
taneous cortical activity reveals hallmarks of an optimal
internal model of the environment. Science 331: 83 -- 7.
[63] E Granot-Atedgi (2009) Stimulus-dependent maximum
entropy models and decoding of naturalistic movies from
large populations of retinal neurons. Thesis. Weizmann
Institute of Science, Israel.
[64] T Broderick, M Dudik, G Tkacik, RE Schapire & W
Bialek (2007) Faster solutions of the inverse pairwise
Ising problem. arXiv:0712.2437.
|
1905.01933 | 1 | 1905 | 2019-05-06T11:25:24 | Neuronal coupling benefits the encoding of weak periodic signals in symbolic spike patterns | [
"q-bio.NC"
] | A good understanding of how neurons use electrical pulses (i.e, spikes) to encode the signal information remains elusive. Analyzing spike sequences generated by individual neurons and by two coupled neurons (using the stochastic FitzHugh-Nagumo model), recent theoretical studies have found that the relative timing of the spikes can encode the signal information. Using a symbolic method to analyze the spike sequence, preferred and infrequent spike patterns were detected, whose probabilities vary with both, the amplitude and the frequency of the signal. To investigate if this encoding mechanism is plausible also for neuronal ensembles, here we analyze the activity of a group of neurons, when they all perceive a weak periodic signal. We find that, as in the case of one or two coupled neurons, the probabilities of the spike patterns, now computed from the spike sequences of all the neurons, depend on the signal's amplitude and period, and thus, the patterns' probabilities encode the information of the signal. We also find that the resonances with the period of the signal or with the noise level are more pronounced when a group of neurons perceive the signal, in comparison with when only one or two coupled neurons perceive it. Neuronal coupling is beneficial for signal encoding as a group of neurons is able to encode a small-amplitude signal, which could not be encoded when it is perceived by just one or two coupled neurons. Interestingly, we find that for a group of neurons, just a few connections with one another can significantly improve the encoding of small-amplitude signals. Our findings indicate that information encoding in preferred and infrequent spike patterns is a plausible mechanism that can be employed by neuronal populations to encode weak periodic inputs, exploiting the presence of neural noise. | q-bio.NC | q-bio |
Neuronal coupling benefits the encoding of weak
periodic signals in symbolic spike patterns
Maria Masoliver1, Cristina Masoller1,∗
Abstract
The biophysical mechanisms by which an input signal elicits a neuronal response
are well known (sufficiently large inputs change the membrane potential of the
neuron and generate electrical pulses, known as action potentials or spikes),
yet, a good understanding of how neurons use these spikes to encode the signal
information remains elusive. Recent theoretical studies have focused on how
neurons encode a weak periodic signal (that by itself is unable to generate
spikes) in a noisy environment, where stochastic electrical fluctuations that do
not encode any information occur. Analyzing spike sequences generated by
individual neurons and by two coupled neurons (that were simulated with the
stochastic FitzHugh-Nagumo model), it has been found that the relative timing
of the spikes can encode the signal information. Using a symbolic method
to analyze the spike sequence, preferred and infrequent spike patterns were
detected, whose probabilities vary with both, the amplitude and the frequency
of the signal. To investigate if this encoding mechanism is plausible also for
neuronal ensembles, here we analyze the activity of a group of neurons, when
they all perceive a weak periodic signal. We find that, as in the case of one
or two coupled neurons, the probabilities of the spike patterns, now computed
from the spike sequences of all the neurons, depend on the signal's amplitude
and period, and thus, the patterns' probabilities encode the information of the
∗Corresponding author
Email addresses: [email protected] (Maria Masoliver),
[email protected] (Cristina Masoller)
1Universitat Polit`ecnica de Catalunya. Departament de F´ısica.
Edifici Gaia, Rambla de Sant Nebridi 22, 08222, Terrassa, Barcelona, Spain.
Preprint submitted to Journal of LATEX Templates
May 7, 2019
signal. We also find that the resonances with the period of the signal or with the
noise level are more pronounced when a group of neurons perceive the signal, in
comparison with when only one or two coupled neurons perceive it. Neuronal
coupling is beneficial for signal encoding as a group of neurons is able to encode a
small-amplitude signal, which could not be encoded when it is perceived by just
one or two coupled neurons. Interestingly, we find that for a group of neurons,
just a few connections with one another can significantly improve the encoding
of small-amplitude signals. Our findings indicate that information encoding in
preferred and infrequent spike patterns is a plausible mechanism that can be
employed by neuronal populations to encode weak periodic inputs, exploiting
the presence of neural noise.
Keywords: neural coding, excitability, spike train variability, neuronal noise,
FitzHugh-Nagumo model, time series analysis, symbolic analysis, ordinal
analysis
1. Introduction
A mechanical input such as tapping someone's knee elicits a stretch reflex
as a response. The biophysical mechanism is known, the muscle stretches as a
consequence of the tapping to the tendon, which triggers the generation of spikes
by a sensory neuron, which in turn triggers the generation of spikes by a motor
neuron, leading to muscle contraction and causing the lower leg to bounce back
[1, 2, 3]. On the other hand, the signal encoding mechanism is also known, the
frequency at which the sensory neuron fires encodes the information about how
fast the muscle is stretching [4]. In turn, the firing rate of the motor neuron
encodes the information about the muscle force when it contracts [5]. This
is an example of a neural circuit (two neurons interconnected by a synapse),
which uses the firing rate code as a coding scheme for an external input. Yet,
neurons encode information of different types of signals using different encoding
mechanisms, which are not yet fully understood. Neurons can represent external
or internal inputs in the timing of the individual spikes, in the relative timing
2
of the spikes of two or more neurons, in the individual firing rate, in the average
firing rate of a population of neurons, in the spike arrival times, among other
coding schemes [6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 3, 17, 18].
Understanding the neural code is crucial, not only to gain knowledge of the
operation of the central nervous system, but also, to advance artificial intelli-
gence systems based on neural networks that use spike-processing operations for
classification, pattern recognition, logic operations, etc. [19, 20, 21, 22].
Efforts have focused on understanding the role, on neural coding, of neural
noise (stochastic electrical fluctuations that do not encode any information [23])
and spike temporal correlations, in particular, for encoding and processing weak
sensory signals. By analyzing the coefficient of variation of the inter-spike inter-
val (ISI) distribution (the standard deviation of the ISI distribution divided by
the mean), the well-known phenomena of stochastic resonance and coherence
resonance have been found. While stochastic resonance [24, 25] refers to the en-
hancement of weak signal detection, coherence resonance [26, 27, 28, 29, 30, 31]
refers to the regularization of the spike train, for an optimal level of noise. On
the other hand, ISI correlations lasting several ISIs have been studied by using
the lagged serial correlation coefficient, which measures linear relations between
sequential interspike intervals [32, 33, 34, 35, 36, 37].
An alternative technique, known as ordinal analysis [38, 39] has also been
used to detect nonlinear ISI correlations.
In general terms, ordinal analysis
transforms a time series into a sequence of symbols, known as ordinal patterns,
considering the temporal order relations among the data points in the time se-
ries. A main advantage of the ordinal symbolic approach is that it provides
a straightforward way to quantify how much information is contained in a se-
quence of spikes (i.e., to apply Information Theory to the study of the neural
code [40, 41]): by counting the number of times each ordinal pattern appears in
the spike sequence, the probabilities of the different patterns can be estimated,
providing a quantification of the information content. Ordinal analysis has been
widely used to investigate biomedical signals, for example, to quantify direction-
ality of coupling in cardio-respiratory data [42], to characterize neuronal spike
3
trains [43, 44, 45], to distinguish healthy subjects from patients suffering from
congestive heart failure [46], to classify neurophysiological data [47, 48, 49, 50],
etc.
In order to understand how a single neuron can encode a weak periodic input
signal (that by itself is unable to generate spikes) in the presence of neural noise,
Aparicio Reinoso et al.
[51] have applied ordinal analysis to spike sequences
simulated with the stochastic FitzHugh-Nagumo (FHN) model [52, 53]. It was
found that the periodic signal induces temporal order in the form of more and
less expressed ordinal patterns. The probabilities of the patterns encode the
signal information, as they depend on both, the amplitude and the period of
the signal.
In a follow up study [54], the role of a second neuron that does
not perceive the weak signal was analyzed.
It was found that the signal is
still encoded in the form of preferred and infrequent ordinal patterns, whose
probabilities again depend on the period and amplitude of the signal.
An open question is whether this encoding mechanism can be employed by a
population of neurons. To answer this question, here we use the stochastic FHN
model to simulate the activity of a group of neurons, when they all perceive a
periodic signal that is weak enough such that by itself (in the absence of noise) it
is unable to generate spikes. Thus, as in previous studies, the neuronal ensemble
encodes the signal in spikes sequences which are generated due to the interplay
of the signal and the noise. Our main findings can be summarized as follows:
(i) the ensemble is able to encode lower amplitude signals, in comparison with
the signal amplitude that can be encoded by a single neuron or by two coupled
neurons; (ii) the noise-induced and period-induced resonances (some ordinal
patterns probabilities are minimum or maximum for particular values of the
noise strength or signal period) observed in one [51] or two coupled neuron
[54] become more pronounced for the neuronal ensemble and (iii) just a few
connections among the neurons can significantly improve the signal encoding.
This paper is organized as follows. Section 2 presents the model equations,
Sec. 3 presents the ordinal analysis method, Sec. 4 presents the results and Sec. 5
presents the conclusions.
4
2. Model
The FitzHugh-Nagumo model is one of the simplest (and yet quite realistic)
models that describe excitable systems [52, 53, 55]. The equations describing
the dynamics of an ensemble of coupled neurons are:
ui = ui − u3
i
3
− vi + a0 cos(2πt/T ) +
σ
ki
aij(uj − ui) +
√
2Dξi(t),
N(cid:88)
j
vi = ui + a.
i (cid:54)= j
(1)
Here N refers to the number of neurons, v is known as the inhibitor variable
and u is known as the activator variable that represents the evolution of the
membrane potential: in the excitable regime, if there is no external perturbation
or it is not strong enough to overcome the threshold, the membrane potential is
held at the resting potential (i.e., stable fixed point) whereas when there is an
external perturbation strong enough to overcome the threshold, the membrane
potential performs a spike (i.e., an action potential). Typical parameters in the
excitable regime are a = 1.05 and = 0.01.
The parameters a0 and T represent the amplitude and period of an external
sinusoidal input, and are chosen such that the signal is sub-threshold: without
noise the neurons do not fire spikes. Dξi(t) represents an stochastic term of
strength D, which is taken as Gaussian distributed, uncorrelated temporally
and across the neuronal ensemble: (cid:104)ξi(t)ξj(t(cid:48))(cid:105) = δijδ(t − t(cid:48)) with (cid:104)ξi(t)(cid:105) = 0
and (cid:104)ξ2
i (t)(cid:105) = 1.
The neurons are mutually coupled with gap-junction connections, charac-
terized by symmetric links (aij = aji = 1 if neurons i and j are connected,
else aij = aji = 0). The coupling strength of each link is σ; to keep the total
coupling strength uniform for all neurons, it is normalized by number of con-
j aij. Regarding the coupling topology, we focus on all-to-all
coupling (in this case ki = N − 1 for all i), but we also consider random connec-
tions. This allows us to analyze the influence of the number of links, as neurons
i and j are connected with probability p that is varied between 0 and 1. It will
5
nections, ki =(cid:80)
be interesting, for future work, to investigate more realistic topologies with, for
example, modular or hierarchical structures.
The model equations are simulated, from random initial conditions, using
the Euler-Maruyama method with an integration step of dt = 10−3. For each
set of parameters, the voltage-like variable of each neuron ui is analyzed and
the sequence of inter-spike-intervals (ISIs) is computed, {Ij i; Ij i = (tj+1 − tj)i}
with tj defined by the condition ui(tj) = 0, considering only the ascensions.
3. Ordinal analysis
The ordinal method [38] is used to analyze each ISI sequence. From the
sequence {I1, . . . Ii, . . . IN} (for clarity, the subindex that labels the neuron is
removed) symbols known as ordinal patterns are obtained by comparing D
consecutive ISIs, based on their temporal relation. For example, if we set D = 2,
the total number of possible ordinal patterns is two: 01, for I1 < I2 and 10, for
I1 > I2, while if we set D = 3, we have 3! = 6 possible ordinal patterns: 012
(I3 > I2 > I1), 021 (I2 > I3 > I1), 102 (I3 > I1 > I2), 120 (I2 > I1 > I3), 201
(I1 > I3 > I2) and 210 (I1 > I2 > I3). The number of possible ordinal patterns
(i.e., the number of possible temporal relations) is determined by the number
of permutations, D!.
Using the function defined in [46] the sequence of ordinal patterns is com-
puted. In order to determine if there are some preferred/infrequent patterns
in the ISI sequences, ordinal patterns probabilities are calculated, taking to-
M =(cid:80)D!
gether all the ISI sequences. The ordinal probabilities are estimated as pi =
Ci/M , where Ci refers to the number of times the i−th pattern appears and
i=1 Ci is the total number of ordinal patterns. If ordinal patterns are
equi-probable it does not exist a preferred order relation among the timing of
spikes. Yet, if there are preferred/infrequent ordinal patterns, a non-uniform
probability distribution is obtained. In order to distinguish between these two
cases (uniform vs. non-uniform ordinal distribution) a binomial test is used: if
all the ordinal patterns are within the interval [p − 3σp, p + 3σp] with p = 1/D!
6
and σp = (cid:112)p(1 − p)/M the ordinal probabilities are consistent with the uni-
form distribution with 99.74% confidence level; else, some patterns are over or
less expressed than others, and there is some degree of temporal order in the
timing of the spikes.
A large number of spikes are needed to precisely estimate the ordinal prob-
abilities (the data requirements for a single neuron were analyzed in [51], see
Fig. 6). As long simulations are computationally demanding, here we limit to
consider ensembles of up to 50 neurons. We have analyzed the role of the num-
ber of neurons, and we expect that our findings will hold for larger ensembles.
The simulations are done for a time long enough to obtain a total number of
105 spikes. As in [51, 54], we use D = 3. This choice is motivated by the fact
that only short ISI correlations are expected since the spikes are noise-induced
(the signal by itself does not induce spikes).
4. Results
The neuronal ensemble displays different dynamical regimes, depending on
the coupling strengh, the signal amplitud and period, the noise strength, and
the coupling topology. Figures 1 and 2 display several examples of the dynamics
of a group of 50 neurons under different conditions: Fig. 1 shows the activity
of an individual neuron (the voltage-like variable of neuron 1), while Fig. 2
displays the raster plot of the ensemble. In panels 1(a) and 2(a) the neurons are
uncoupled and no signal is applied, therefore, random spiking activity occurs
due to the noise. In panels 1(b) and 2(b) the neurons are mutually coupled,
still no signal is applied. Now we see synchronized spiking activity superposed
with random spikes. When the periodic signal is applied, we see in panels 1(c)-
(f) and 2(c)-(f) that the neurons either fire regular and synchronized spikes, or
there is more irregular firing, depending on the period of the signal.
In the following we analyze the influence of the different parameters. To
stress the role of the number of neurons, we compare the results obtained for
50 neurons with those obtained for only two coupled neurons.
7
Fig. 1: Spiking activity of a neuron when no signal is applied (a0 = 0) and the neuron (a)
is uncoupled (σ = 0), (b) is coupled to a group of 50 neurons (all to all coupling, σ = 0.05).
Activity of the neuron when it is coupled and a sinusoidal signal of amplitude a0 = 0.1 and
period (c) T = 10, (d) T = 20, (e) T = 40 is applied. The noise level is D = 2.5 · 10−6.
Fig. 2: Raster plots displaying the spiking activity of the group of 50 neurons for the same
parameters as in Fig. 1
8
We begin by characterizing the role of the signal amplitude, presented in
Fig. 3 that displays probabilities of the six ordinal patterns as a function of a0
for N = 50 [Fig. 3 (a)] and for N = 2 [Fig. 3 (b)]. Here a0 is kept within the
range of values for which, in the absence of noise, the neurons do not fire spikes.
We note that, if the signal amplitude is small enough, as expected, the ordinal
probabilities are within the blue region that indicates values that are consistent
with equal probabilities, with 99.74% confidence level (this region is calculated
as explained in Sec. 3).
As the signal amplitude increases we note that, while for the two coupled
neurons, ordinal probabilities gradually increase (or decrease), for the ensemble
of 50 neurons their variation is more pronounced. Interestingly, the same cod-
ification (i.e., same ordinal patterns probabilities) is obtained for N = 2 and
larger a0. This is shown in Fig. 4, where we analyze the effect of the signal
period (T is kept within the range of values for which, in the absence of noise,
the neurons only display sub-threshold oscillations): comparing Figs. 4 (a) and
4 (d), or Figs. 4 (c) and 4 (f), we see that for two neurons and larger signal
amplitudes we find a very similar set of ordinal probabilities as for 50 neurons
and lower a0. We see that the variation of the ordinal probabilities with the
period is very similar for a0 = 0.025, N = 50 and a0 = 0.05, N = 2 [in Figs. 4
(a) and 4 (d), respectively] and for a0 = 0.05, N = 50 and a0 = 0.1, N = 2 [Figs.
4 (c) and4 (f), respectively]. Therefore, these results suggest that 50 neurons
encode a weak signal in a very similar way as 2 neurons encode a stronger signal.
Regarding how the encoding of the signal depends on its period, in Fig. 4
we verify that the probabilities of the patterns expressed in the spike sequences
depend on the period of the signal (consistent with the observations in [51, 54]).
Comparing the left and right columns of Fig. 4, we note that neuronal coupling
is beneficial for signal encoding because for N = 50 (left column) the ordinal
probabilities take higher or lower values, and the resonances with the period
become more pronounced, as compared to N = 2 (right column).
Interestingly, for N = 50 and a0 = 0.1 the probabilities are nearly constant
in the interval 10 ≤ T ≤ 15 and patterns 012 and 210 have very low or zero
9
Fig. 3: Probabilities of the ordinal patterns as a function of the signal amplitude, a0, for (a)
an ensemble of 50 neurons, all-to-all coupled, and for (b) two mutually coupled neurons. The
parameters are: T = 10, D = 2.5 · 10−6 and σ = 0.05.
probability. The corresponding neuronal activity for T = 10 was displayed in
Figs. 1(c) and 2(c). We see an alternation of long and short intervals between
spikes, while three consecutive increasing or decreasing intervals do not occur
(which would be represented by patterns 012 and 210 respectively).
In Ref. [51] it was shown that the ordinal patterns displayed a noise-induce
resonance, as 012 and 210 reached minimum values when the noise intensity was
such that the mean ISI, (cid:104)ISI(cid:105), was approximately equal to half the signal period.
In Ref. [54] it was demonstrated that this encoding mechanism persisted when
the neuron was coupled to a second neuron that did not perceive the signal.
Here, we show in Fig. 5(a) that the mechanism is robust and the resonance
is more pronounced when the signal is perceived by a group of 50 neurons:
ordinal patterns 012 and 210 are not expressed (have zero probability) when
D = 5·10−6, and for this noise strength, (cid:104)ISI(cid:105) = T /2. For comparison Fig. 5(b)
shows the ordinal probabilities as a function of D for N = 2. Ordinal patterns
012 and 210 are minimum for almost the same noise strength (D = 8 · 10−6)
which gives (cid:104)ISI(cid:105) = T /2. Yet, the minimum is less pronounced, as compared
to the group of 50 neurons.
So far we have seen that for 50 neurons the encoding of the signal is, in
10
Fig. 4: Probabilities of the ordinal patterns as a function of the signal period, T , for (a,b)
a0 = 0.025, (c,d) a0 = 0.05 and (e,f) a0 = 0.1 with N = 50 (a,c,e) and N = 2 (b,d,f). In
panels (e) and (f) we consider T ≥ 8 because for T < 8 the signal by itself triggers spikes.
Other parameters are: D = 2.5 · 10−6 and σ = 0.05.
11
Fig. 5: Probabilities of the ordinal patterns as a function of the noise strength, D, for (a) 50
neurons and for (b) two neurons. Other parameters are: a0 = 0.05, T = 10, and σ = 0.05.
general, improved in comparison with that of only two neurons. Yet, how is the
variation of the ordinal probabilities with the network size? Next, we fix the
period and the amplitude of the signal and we characterize the influence of i)
the number of neurons, N , when they are all-to-all coupled; ii) the number of
links (from zero links to all-to-all coupling, randomly adding links) and iii) the
strength of the coupling, σ, in the all-to-all configuration, from 0 (uncoupled
neurons) to the same coupling strength considered in steps i) and ii). We keep
the coupling level low enough such that, without signal and noise, there are no
spikes. We note that the starting and final points in the three steps are the
same: from the uncoupled neurons to 50 all-to-all coupled neurons.
Figure 6 presents the results: panels (a, b) display the ordinal probabilities
as a function of N ; (c, d) as a function of the percentage of total links; and (e,
f) as a function of the coupling strength. To investigate if these parameters can
play different roles for weak or strong signals, we consider two signal amplitudes:
a0 = 0.05 in panels (a, c, e) and a0 = 0.1 in panels (b, d, f).
In Fig. 6(a) we note that for a0 = 0.05 the probabilities gradually vary,
increasing or decreasing, as N increases up to N = 10. With further increase
of N they remain nearly constant. The signal is encoded (the probabilities are
not in the blue region) but, at least for these parameters, the encoding only
12
slightly improves when increasing N . In contrast, for a0 = 0.1 [Fig. 6(b)] we
observe that the encoding is significantly improved, compared to Fig. 6(a), as
the probabilities of the ordinal patterns 012 and 210 gradually decrease to zero.
An interesting observation is that above a certain number of neurons (which
depends on the parameters) the probabilities saturate and remain stable with
further increase of N .
In Fig. 6(c) and 6(d) we note that for the lower signal amplitude, the
probabilities vary gradually when increasing the number of links, and with just
few links (∼ 10 %) they take the most extreme values, i.e., the encoding is
optimal. In contrast, for the higher signal amplitude the probabilities increase
or decrease fast, and then saturate. Next, in Fig. 6(e) and 6(f) we evaluate the
effect of the coupling strength. We notice that increasing σ tends to improve
the encoding of the signal (the ordinal probabilities tend to higher or lower
values), and the effect is more pronounced if the signal amplitude is high. We
also note a saturation effect, as for the high signal amplitude, patterns 012 and
210 have zero probability for coupling strengths above σ = 0.02. In order to
understand the effect of the coupling strength, Fig. 7 displays the spiking activity
of the neurons for different values of σ. Here we see that when the neurons are
uncopled (σ = 0) their spiking activity is partially synchronized due to the
periodic signal that is perceived by all the neurons. As σ increases, the spikes
gradually become even more synchronized. A similar behavior is found (not
shown) when the number of existing links increases, keeping σ constant. For
future work, it will be interesting to investigate the synchronization transition
using synchronization measures based on the ordinal probabilities [50].
5. Conclusions and discussion
We have analyzed a plausible neuronal mechanism for encoding a weak pe-
riodic signal exploiting neural noise. We have simulated the dynamics of a
neuronal ensemble using the stochastic FitzHugh-Nagumo model with mutual
gap-junction type of coupling, and a sinusoidal signal that is perceived by all
13
the neurons. We applied the ordinal symbolic method to the spike sequences
generated by all the neurons. Considering the variation of the ordinal prob-
abilities with the amplitude of the signal, we have found that a group of 50
neurons encodes a weak amplitude signal in a similar way (similar probabilities)
as two neurons encode a signal of stronger amplitude. We confirmed the results
reported in Refs. [51, 54]: the ordinal probabilities depend on the period and
the amplitude of the signal and thus, they encode the signal information. We
have found that the probabilities have resonances with the period or with the
noise level, which become more pronounced for the neuronal ensemble. Regard-
ing the influence of the number of neurons, N , we have found that increasing
N enhances the signal encoding, but above a certain N (which depends on the
parameters), the ordinal probabilities saturate and remain nearly constant. We
have also investigated the role of the number of links and found that signal
encoding can be enhanced by just a few links. We have also found a gradual
similar effect when increasing the coupling strength.
In sum, our work concludes that the neuronal ensemble improves signal
encoding, in comparison with single or two coupled neurons. We have studied
an homogeneous group of neurons as a first step to understand the ensemble
coding mechanism. Yet, in real biological organisms signal coding is performed
by nonidentical neurons, and for this reason it will be important to understand
the effects of heterogeneous parameters.
The ensemble encoding mechanism proposed here can also allow to encode
aperiodic signals, whose amplitude and/or period vary in time. If the ordinal
probabilities are determined from the spikes of a single neuron, the encoding
mechanim is very slow, because a large number of spikes are needed to estimate
the ordinal probabilities; in contrast, when the signal is perceived by a large
group of neurons and the ordinal probabilities are determined from the spikes
of all the neurons, then signal encoding can be fast, because just a few spikes
per neuron can be sufficient to estimate the probabilities of the different spike
patterns.
As future work, it will also be interesting to study how a weak signal that
14
is perceived by just one neuron (or by a subset of neurons) propagates on the
whole ensemble, which would give information of how the signal is transmitted.
As well, we intent to study how an ensemble of neurons may encode two weak
signals. A recent experimental study [17] of how neurons encode simultaneous
auditory stimuli has found that some neurons fluctuate between firing rates
observed for each individual sound. It would be interesting to compare with our
synthetic model, to contribute to advance the understanding of how neuronal
systems process information of multiple simultaneous stimuli.
6. Acknowledgments
This work was supported in part by Spanish MINECO/FEDER grant FIS2015-
66503-C3-2-P141 and ICREA ACADEMIA, Generalitat de Catalunya.
15
Fig. 6: Probabilities of the ordinal patterns as a function of the number of neurons, N (a,
b), of the percentage of links (c, d), and of the coupling strength (e, f) for a0 = 0.05 and
a0 = 0.1, respectively. In panels (a, b, e, f) the neurons are all-to-all coupled, in panels (c, d)
the coupling topology is random (starting from uncoupled neurons, links are randomly added
until the neurons are all-to-all coupled).
D = 2.5 · 10−6 and T = 10.
In panels (c, d, e, f) N = 50, in all the panels:
16
Fig. 7: Raster plots displaying the spiking activity of the group of 50 neurons all-to-all coupled
over time for (a) σ = 0, (b) σ = 0.01, (c) σ = 0.015 and (d) σ = 0.03. In all the panels:
D = 2.5 · 10−6, T = 10 and a0 = 0.1.
17
References
[1] K. Pearson, J. Gordon, Spinal reflexes, Principles of Neural Science (2000)
713 -- 736.
[2] J. Nolte, The human brain: Introduction to its functional anatomy, Mosby,
St. Louis (2002).
[3] J. J. Knierim, Chapter 19 - information processing in neural networks, in:
J. H. Byrne, R. Heidelberger, M. N. Waxham (Eds.), From Molecules to
Networks, third edition ed., Academic Press, Boston, 2014, pp. 563 -- 589.
doi:https://doi.org/10.1016/B978-0-12-397179-1.00019-1.
[4] C. C. Hunt, Mammalian muscle spindle: peripheral mechanisms, Physio-
logical Rev. 70 (1990) 643 -- 663.
[5] A. W. Monster, H. Chan,
Isometric force production by motor units of
extensor digitorum communis muscle in man, J. of Neurophysiol. 40 (1977)
1432 -- 1443.
[6] B. W. Knight, Dynamics of encoding in a population of neurons, J. Gen.
Physiol. 59 (1972) 734766.
[7] M. Carandini, F. Mechler, C. S. Leonard, J. A. Movshon, Spike train
encoding by regular-spiking cells of the visual cortex, J. Neurophysiol. 76
(1996) 3425 -- 3441.
[8] Y. Sakurai, How do cell assemblies encode information in the brain?, Neu-
roscience & Biobehavioral Reviews 23 (1999) 785 -- 796.
[9] N. Masuda, K. Aihara, Bridging rate coding and temporal spike coding by
effect of noise, Phys. Rev. Lett. 88 (2002) 248101.
[10] E. Arabzadeh, S. Panzeri, M. E. Diamond, Whisker vibration information
carried by rat barrel cortex neurons, J. Neurosci. 24 (2004) 6011 -- 6020.
18
[11] I. Nelken, G. Chechik, T. D. Mrsic-Flogel, A. J. King, J. W. H. Schnupp,
Encoding stimulus information by spike numbers and mean response time
in primary auditory cortex, J. Comput. Neurosci. 19 (2005) 199 -- 221.
[12] B. B. Averbeck, P. E. Latham, A. Pouget, Neural correlations, population
coding and computation, Nat. Rev. Neurosci. 7 (2006) 358366.
[13] R. Quian Quiroga, S. Panzeri, Extracting information from neuronal popu-
lations: information theory and decoding approaches, Nat. Rev. Neurosci.
10 (2001) 173.
[14] X.-J. Wang, Neurophysiological and computational principles of cortical
rhythms in cognition, Physiol. Rev. 90 (2010) 1195 -- 1268.
[15] A. Kumar, S. Rotter, A. Aertsen, Spiking activity propagation in neuronal
networks: reconciling different perspectives on neural coding, Nat. Rev.
Neurosci. 11 (2010) 615 -- 627.
[16] T. Tchumatchenko, A. Malyshev, F. Wolf, M. Volgushev, Ultrafast popu-
lation encoding by cortical neurons, J. Neurosci. 31 (2011) 12171 -- 12179.
[17] V. C. Caruso, J. T. Mohl, C. Glynn, J. Lee, S. M. Willett, A. Zaman, A. F.
Ebihara, R. Estrada, W. A. Freiwald, S. T. Tokdar, J. M. Groh, Single
neurons may encode simultaneous stimuli by switching between activity
patterns, Nat. Comm. 9 (2018) 2715.
[18] E. Lazarov, M. Dannemeyer, B. Feulner, J. Enderlein, M. J. Gutnick,
F. Wolf, A. Neef, An axon initial segment is required for temporal pre-
cision in action potential encoding by neuronal populations, Sci. Adv. 4
(2018).
[19] P. R. Prucnal, B. J. Shastri, T. F. de Lima, M. A. Nahmias, A. N. Tait,
Recent progress in semiconductor excitable lasers for photonic spike pro-
cessing, Adv. Opt. Photon. 8 (2016) 228 -- 299.
19
[20] B. J. Shastri, M. A. Nahmias, A. N. Tait, A. W. Rodriguez, B. Wu, P. R.
Prucnal, Spike processing with a graphene excitable laser, Sci. Rep. 6
(2016) 19126.
[21] Y. Shen, N. C. Harris, S. Skirlo, M. Prabhu, T. Baehr-Jones, M. Hochberg,
X. Sun, S. Zhao, H. Larochelle, D. Englund, M. Soljai, Deep learning with
coherent nanophotonic circuits, Nat. Phot. 11 (2017) 441.
[22] P. R. Prucnal and B. J. Shastri, Neuromorphic Photonics, CRC Press, 2017.
[23] M. D. McDonnell, L. M. Ward, The benefits of noise in neural systems:
Bridging theory and experiment, Nat. Rev. Neurosci. 12 (2011) 415 -- 426.
[24] J. F. Lindner, B. K. Meadows, W. L. Ditto, M. E. Inchiosa, A. R. Bulsara,
Array enhanced stochastic resonance and spatiotemporal synchronization,
Phys. Rev. Lett. 75 (1995) 3 -- 6.
[25] E. Yilmaz, M. Uzuntarla, M. Ozer, M. Perc, Stochastic resonance in hybrid
scale-free neuronal networks, Physica A 392 (2013) 5735 -- 5741.
[26] A. S. Pikovsky, J. Kurths, Coherence resonance in a noise-driven excitable
system, Phys. Rev. Lett. 78 (1997) 775 -- 778.
[27] C. Zhou, J. Kurths, B. Hu, Array-enhanced coherence resonance: Nontriv-
ial effects of heterogeneity and spatial independence of noise, Phys. Rev.
Lett. 87 (2001) 098101.
[28] O. Kwon, H.-T. Moon, Coherence resonance in small-world networks of
excitable cells, Phys. Lett. A 298 (2002) 319 -- 324.
[29] O. Kwon, H.-H. Jo, H.-T. Moon, Effect of spatially correlated noise on
coherence resonance in a network of excitable cells, Phys. Rev. E 72 (2005)
066121.
[30] P. Balenzuela, P. Ru´e, S. Boccaletti, J. Garcia-Ojalvo, Collective stochastic
coherence and synchronizability in weighted scale-free networks, New J.
Phys. 16 (2014) 013036.
20
[31] M. Masoliver, N. Malik, E. Scholl, A. Zakharova, Coherence resonance in a
network of FitzHugh-Nagumo systems: Interplay of noise, time-delay, and
topology, Chaos 27 (2017) 101102.
[32] A. B. Neiman, D. F. Russell, Models of stochastic biperiodic oscillations
and extended serial correlations in electroreceptors of paddlefish, Phys.
Rev. E 71 (2005) 061915.
[33] A. B. Neiman, D. F. Russell, Sensory coding in oscillatory electroreceptors
of paddlefish, Chaos 21 (2011) 047505.
[34] W. H. Nesse, L. Maler, A. Longtin, Biophysical information representation
in temporally correlated spike trains, PNAS 107 (2010) 21973 -- 21978.
[35] O. Avila-Akerberg, M. J. Chacron, Nonrenewal spike train statistics:
causes and functional consequences on neural coding, Exp. Brain Res.
210 (2011) 353 -- 371.
[36] T. Schwalger, B. Lindner, Patterns of interval correlations in neural oscil-
lators with adaptation, Frontiers in Comp. Neurosci. 7 (2013) 164.
[37] W. Braun, A. Longtin, Interspike interval correlations in networks of in-
hibitory integrate-and-fire neurons, Phys. Rev. E 99 (2019) 032402.
[38] C. Bandt, B. Pompe, Permutation entropy: A natural complexity measure
for time series, Phys. Rev. Lett. 88 (2002) 174102.
[39] J. M. Amigo, Permutation Complexity in Dynamical Systems: Ordinal
Patterns, Permutation Entropy and All That. Springer-Verlag Berlin, 2010.
[40] S. P. Strong, R. Koberle, R. R. de Ruyter van Steveninck, W. Bialek,
Entropy and information in neural spike trains, Phys. Rev. Lett. 80 (1998)
197 -- 200.
[41] A. Borst, F. E. Theunissen, Information theory and neural coding, Nat.
Rev. Neurosci. 2 (1999) 947 -- 957.
21
[42] A. Bahraminasab, F. Ghasemi, A. Stefanovska, P. V. E. McClintock,
H. Kantz, Direction of coupling from phases of interacting oscillators: A
permutation information approach, Phys. Rev. Lett. 100 (2008) 084101.
[43] Z. Li, G. Ouyang, D. Li, X. Li, Characterization of the causality between
spike trains with permutation conditional mutual information, Phys. Rev.
E 84 (2011) 021929.
[44] O. A. Rosso and C. Masoller, Detecting and quantifying stochastic and co-
herence resonances via information-theory complexity measurements, Phys.
Rev. E 79, 040106 (2009).
[45] F. Montani, R. Baravalle, L. Montangie, and O. A. Rosso, Causal informa-
tion quantification of prominent dynamical features of biological neurons,
Phil. Trans. Roy. Soc. A. 373, 20150109 (2015).
[46] U. Parlitz, S. Berg, S. Luther, A. Schirdewan, J. Kurths, N. Wessel, Clas-
sifying cardiac biosignals using ordinal pattern statistics and symbolic dy-
namics, Compt. Biol. Med. 42 (2012) 319.
[47] Y. Cao, W.-w. Tung, J. B. Gao, V. A. Protopopescu, L. M. Hively, Detect-
ing dynamical changes in time series using the permutation entropy, Phys.
Rev. E 70 (2004) 046217.
[48] D. Arroyo, P. Chamorro, J. Amig´o, F. Rodr´ıguez, P. Varona, Event detec-
tion, multimodality and non-stationarity: Ordinal patterns, a tool to rule
them all?, Eur. Phys. J. Special Topics 222 (2013) 457 -- 472.
[49] C. Quintero-Quiroz, L. Montesano, A. J. Pons, M. C. Torrent, J. Garca-
Ojalvo, C. Masoller, Differentiating resting brain states using ordinal sym-
bolic analysis, Chaos 28 (2018) 106307.
[50] I. Echegoyen, V. Vera-vila, R. Sevilla-Escoboza, J. Martnez, J. Buld, Or-
dinal synchronization: Using ordinal patterns to capture interdependencies
between time series, Chaos, Solitons & Fractals 119 (2019) 8 -- 18.
22
[51] J. A. Reinoso, M. C. Torrent, C. Masoller, Emergence of spike correlations
in periodically forced excitable systems, Phys. Rev. E 94 (2016) 032218.
[52] R. FitzHugh,
Impulses and physiological states in theoretical models of
nerve membrane, Biophys. J. 1 (1961) 445.
[53] J. Nagumo, S. Arimoto, S. Yoshizawa., An active pulse transmission line
simulating nerve axon, Proc. IRE 50 (1962) 2061 -- 2070.
[54] M. Masoliver, C. Masoller,
Sub-threshold signal encoding in coupled
Fitzhugh-Nagumo neurons, Sci. Rep. 8 (2018) 8276.
[55] J. A. Acebr´on, A. R. Bulsara, W.-J. Rappel, Noisy Fitzhugh-Nagumo
model: From single elements to globally coupled networks, Phys. Rev. E
69 (2004) 026202.
23
|
1705.07614 | 1 | 1705 | 2017-05-22T08:52:31 | Feedback inhibition shapes emergent computational properties of cortical microcircuit motifs | [
"q-bio.NC"
] | Cortical microcircuits are very complex networks, but they are composed of a relatively small number of stereotypical motifs. Hence one strategy for throwing light on the computational function of cortical microcircuits is to analyze emergent computational properties of these stereotypical microcircuit motifs. We are addressing here the question how spike-timing dependent plasticity (STDP) shapes the computational properties of one motif that has frequently been studied experimentally: interconnected populations of pyramidal cells and parvalbumin-positive inhibitory cells in layer 2/3. Experimental studies suggest that these inhibitory neurons exert some form of divisive inhibition on the pyramidal cells. We show that this data-based form of feedback inhibition, which is softer than that of winner-take-all models that are commonly considered in theoretical analyses, contributes to the emergence of an important computational function through STDP: The capability to disentangle superimposed firing patterns in upstream networks, and to represent their information content through a sparse assembly code. | q-bio.NC | q-bio | Feedback inhibition shapes emergent computational
properties of cortical microcircuit motifs
Zeno Jonke∗, Robert Legenstein∗†, Stefan Habenschuss, Wolfgang Maass
Institute for Theoretical Computer Science
Graz University of Technology
Inffeldgasse 16b/I
8010 Graz, Austria
September 14, 2018
Abstract
Cortical microcircuits are very complex networks, but they are composed of a rel-
atively small number of stereotypical motifs. Hence one strategy for throwing light
on the computational function of cortical microcircuits is to analyze emergent compu-
tational properties of these stereotypical microcircuit motifs. We are addressing here
the question how spike-timing dependent plasticity (STDP) shapes the computational
properties of one motif that has frequently been studied experimentally:
intercon-
nected populations of pyramidal cells and parvalbumin-positive inhibitory cells in layer
2/3. Experimental studies suggest that these inhibitory neurons exert some form of
divisive inhibition on the pyramidal cells. We show that this data-based form of feed-
back inhibition, which is softer than that of winner-take-all models that are commonly
considered in theoretical analyses, contributes to the emergence of an important com-
putational function through STDP: The capability to disentangle superimposed firing
patterns in upstream networks, and to represent their information content through a
sparse assembly code.
7
1
0
2
y
a
M
2
2
]
.
C
N
o
i
b
-
q
[
1
v
4
1
6
7
0
.
5
0
7
1
:
v
i
X
r
a
∗ These authors contributed equally to the work.
† Corresponding author (email: [email protected]).
1
1
Introduction
A promising strategy for understanding the computational function of a cortical column
was proposed by [Douglas et al., 1989, Shepherd, 2004, Grillner and Graybiel, 2006],
and others: To probe computational properties of prominent network motifs of a cortical
column, commonly referred to as microcircuit motifs. We are addressing computational
properties of one of the most prominent microcircuit motifs: densely interconnected pop-
ulations of excitatory and inhibitory neurons. We focus on motifs in layer 2/3, where
parvalbumin-positive (PV+) inhibitory neurons (often characterized as fast-spiking in-
terneurons, in particular basket cells) are interconnected with nearby pyramidal cells with
very high connection probability in both directions, see e.g. [Packer and Yuste, 2011, Fino
et al., 2012, Avermann et al., 2012]. One usually refers to this type of inhibition as lateral
or feedback inhibition. The dynamics of this microcircuit motif has frequently been exam-
ined in-vivo [Wilson et al., 2012, Petersen and Crochet, 2013, Pala and Petersen, 2015],
and modelled in [Avermann et al., 2012]. We examine computational properties that
emerge in a model M for this microcircuit motif under spike-timing dependent plasticity
(STDP).
One cannot model this microcircuit motif by the frequently considered winner-take-all
(WTA) model, since this model would require that the firing of a single pyramidal cell (the
"winner") can suppress firing of other pyramidal cells in the motif. But experimental data
show that several pyramidal cells need to fire in order to engage feedback inhibition through
PV+ cells [Isaacson and Scanziani, 2011, Avermann et al., 2012]. Divisive inhibition has
been proposed as a more realistic mathematical model for this softer type of inhibition
[Wilson et al., 2012, Carandini and Heeger, 2012]. Our goal is to understand the impact
of this softer type of inhibition on neural codes and computational properties that emerge
under STDP. There exists a large number of preceding studies of emergent computational
properties of WTA-like microcircuit motifs, from [Rumelhart and Zipser, 1985] to [Nessler
et al., 2013]. But they were based on the assumption of strong WTA-like lateral inhibition.
The functional role of inhibition in this microcircuit motif can be better approximated
by a variation of the k-WTA model [Maass, 2000], where several (k) winners can emerge
simultaneously from a competition of pyramidal cells for firing. We show that this softer
competition leads to the emergence of shared feature selectivity of pyramidal cells, like
in the experimental data of [Lee et al., 2012], where small subsets of pyramidal cells
(assemblies), instead of single neurons, respond to specific input features (see Figure 2).
We also show that an important computational operation, blind source separation
[Foldiak, 1990], also referred to as independent component analysis [Hyvarinen et al.,
2004], emerges in this microcircuit motif through STDP. This operation enables a network
to disentangle and sparsely represent superimposed spike inputs that may result from
separate sources in the environments or upstream neural networks. This modular coding
scheme avoids a combinatorial explosion of the number of neurons that are needed to
encode superimposed sources, since they become encoded by superpositions of neural
codes (assemblies) for each of the sources, rather than by a separate neural code for every
superposition that occurs. An example is given in Figure 4,5 for the case of arbitrarily
superimposed vertical and horizontal bars, a well-known benchmark task for blind source
separation [Foldiak, 1990]. This distributed coding scheme also supports intra-cortical
communication and computation based on spike patterns or spike packets as proposed
in [Luczak et al., 2015], see Figure 3.
2
connection
conn. prob. [%]
symbol w [a.u.]
delay [ms]
Input→E
E→I
I→E
I→I
100
57.5
60
55
w
wEI
wIE
wII
[0.01, 1]
[0, 10]
13.57
1.86
13.57
1
1
1
Table 1: Neuron-type specific synaptic connection parameters in M: Connection prob-
ability (conn. prob.), synaptic weight (w), and synaptic delay. Weights from inputs to
excitatory network neurons are plastic and bounded to the given range. The correspond-
ing delays are uniformly distributed in the given range. See [Legenstein et al., 2017] for a
motivation of synaptic efficacy values from a theoretical perspective.
2 Methods
2.1 Definition of a data-based microcircuit motif model M
We consider in this article a model for interacting populations of pyramidal cells with PV+
inhibitory neurons on layer 2/3 that is based on data from the Petersen Lab [Avermann
et al., 2012] and refer to this specific model as the microcircuit motif model M.
The microcircuit motif model M consists of two reciprocally connected pools of neu-
rons, an excitatory pool and an inhibitory pool. Inhibitory network neurons are recurrently
connected. Excitatory network neurons receive additional excitatory synaptic input from
a pool of N input neurons. Figure 1A summarizes the connectivity structure of the model
together with connection probabilities. Connection probabilities have been chosen accord-
ing to the experimental data described in [Avermann et al., 2012] and listed in Table 1
together with connection-type specific synaptic parameters. For a connection probability
p between two pools, each individual pair of neurons from these two pools is randomly
chosen to be connected by a synapse with probability p.
Input neurons emit Poisson spike trains with time-varying rates. We tested several
temporal profiles of theses rates in different simulations as described below in the corre-
sponding sections. Let t(1)
, . . . denote the spike times of input neuron i. The output
trace yi(t) of input neuron i is given by the temporal sum of unweighted postsynaptic
potentials (PSPs) arising from input neuron i:
, t(2)
i
i
yi(t) = X
f
ǫ(t − t(f )
i
),
(1)
where ǫ is the synaptic response kernel, i.e., the shape of the PSP. It is given by a double-
exponential function
ǫ(s) =
cǫ (cid:0)e−s/τf − e−s/τr(cid:1)
0
,
,
if 0 ≤ s ≤ Tǫ
,
otherwise
(2)
with the rise time constant τr = 1 ms, a fall time constant τf = 10 ms and a cut-off after
Tǫ = 50 ms, see also Figure 1B. The constant cǫ = 1.435 was chosen to assure a peak value
of 1. All synapses in the network have the same response kernel ǫ. For given spike times,
3
output traces of excitatory network neurons and inhibitory network neurons are defined
analogously and denoted by zm(t) and Ij(t) respectively.
The network consists of M = 400 excitatory neurons, modeled as stochastic spike
response model neurons [Jolivet et al., 2006] that we define in the following. The stochas-
ticity of the model stems from its stochastic spike generation, where spikes are generated
according to a Poisson process with a time-varying rate (the instantaneous firing rate of
the neuron). The instantaneous firing rate ρm of a neuron m depends exponentially on its
current membrane potential um,
ρm(t) =
1
τ
exp(γ · um(t)) ,
(3)
where τ = 10 ms and γ = 2 are scaling parameters that control the shape of the response
function. After emitting a spike, the neuron enters an absolute refractory period for 10
ms during which the neuron cannot spike again. These excitatory neurons project to
and receive inputs from a pool of inhibitory neurons. Thus, the membrane potential of
excitatory neuron m is given by the sum of external inputs, inhibition from inhibitory
neurons, and its excitability α
um(t) = X
i
wim yi(t) − X
j∈Im
wIEIj(t) + α,
(4)
where Im denotes the set of indices of inhibitory neurons that project to neuron m, and
wIE denotes the weight of these inhibitory synapses. Ij(t) and yi(t) denote synaptic input
(output traces) from inhibitory neurons and input neurons respectively, see above. We
used α = −5.57 (These parameter values can be motivated from a theoretical perspective,
see [Legenstein et al., 2017]).
Apart from excitatory neurons there are Minh = 100 inhibitory neurons in the network.
While [Jolivet et al., 2006] provides a stochastic model for pyramidal cells, no such model is
available for PV+ inhibitory neurons. Experimental data indicates that in these neurons,
the relationship between the synaptic drive and the firing rate, i.e., the frequency-current
(f-I) curve, is rather linear over a large range of input strengths [Ferguson et al., 2013, Ho
et al., 2012]. We therefore modeled inhibitory neurons as stochastic spike response neurons
with an instantaneous firing rate given by
ρinh
m (t) = σrect(uinh
m (t)),
(5)
where σrect denotes the linear rectifying function σrect(u) = u for u ≥ 0 and 0 otherwise.
The absolute refractory period of inhibitory neurons in the model is 3 ms.
Inhibitory
neurons receive excitatory inputs from excitatory network neurons as well as connections
from other inhibitory neurons. The membrane potentials of inhibitory neurons are thus
given by
m (t) = X
uinh
i∈Em
wEI zi(t) − X
j∈IIm
wIIIj(t) + uopt,
(6)
where zi(t) denotes synaptic input (output trace) from excitatory network neuron i, Em
(IIm) denotes the set of indices of excitatory (inhibitory) neurons that project to inhibitory
neuron m, wEI (wII) denotes the excitatory (inhibitory) weight to inhibitory neurons, and
uopt denotes an external optogenetic activation of inhibitory neurons. uopt was set to 0 in
all simulations except for the simulation shown in Figure 1D, where optogenetic activation
4
was modeled by setting uopt = 50 (arbitrary units). The synaptic weights from excitatory
network neurons to inhibitory neurons imply that a single spike in the excitatory pool
induces a spike in a given post-synaptic inhibitory neuron with a probability of 0.17,
consistent with experimental findings that several excitatory neurons have to be active in
order to induce robust spiking in PV+ interneurons [Avermann et al., 2012].
Synaptic connections from input neurons to excitatory network neurons are subject to
STDP. A standard version of STDP is employed with an exponential weight dependency
for potentiation [Habenschuss et al., 2013b], see Figure 1C. All input weights wij are
updated as follows. For each postsynaptic spike at time tpost, all presynaptic spikes in
the preceding 100 ms are considered. For each such pre-before-post spike pair with time
difference tpost − tpre, the weight is increased by
−
∆wij(tpost − tpre) = ηe−wij +1e
tpost−tpre
τ+
,
(7)
with τ+ = 10 ms. The learning rate η is 0.01 except for Figure 4 where η = 0.02 to speed up
learning. For each presynaptic spike at time t, all postsynaptic spikes in the preceding 100
ms are considered. For each such post-before-pre spike pair with time difference tpre −tpost,
the weight change is given by
∆wij(tpre − tpost) = −ηe
−
tpre−tpost
τ−
,
(8)
with τ− = 25 ms. Synaptic weights are clipped to wmin = 0.01 and wmax = 1.
In all
simulations, initial input weights were drawn from a uniform distribution in the interval
[wmin, wmax]. This concludes the definition of the microcircuit motif model M.
2.2 Details to computer simulations of the model M
Here, we provide details to the computer simulations reported in Results.
It is recom-
mended that the reader skips this section at first reading. References to the individual
subsections are given at the appropriate places in Results.
All simulations were performed in PCSIM, a spiking neural network simulator written
in C++ that provides a Python interface, which was extended in order to support simu-
lation of the model. All simulations were performed with a discretization time step ∆t of
1 ms.
Details to simulations for Figure 1
For Figure 1D, the input to M was given by simulated visual bars stimuli at various ori-
entations (see Details to simulations for Figure 2 below for details). Orientation-tuned
neurons emerged in a learning phase that lasted 400 s of simulated biological time. The
tuning curve of one excitatory neuron was evaluated in the original circuit. Then, opto-
genetic stimulation of inhibitory neurons was mimicked by setting the external activation
uopt in eq. (6) to uopt = 50 (arbitrary units) in all inhibitory neurons, with the effect of
increasing the total rate of inhibition. The tuning curve of the same excitatory neuron was
then evaluated in this modified network. All procedures in the learning and evaluation
phase were the same as described below in Details to simulations for Figure 2.
5
Details to simulations for Figure 2
Here, we tested the behavior of M on an input distribution that mimics visual bar patterns
of various orientations. In this simulation, network inputs were generated from 180 two-
dimensional binary pixel arrays of size 20 × 20. A prototypical horizontal bar of width
2 pixels centered on the array was rotated in steps of 1 degree in order to obtain 180
pixel arrays that span the space of possible bar orientations. These pixel arrays were then
transformed into 400-dimensional rate vectors where each entry had a rate of 75 Hz if the
corresponding pixel was on (i.e., the bar covered that pixel) and 1 Hz otherwise. During a
simulation, a rate vector was chosen randomly (uniformly out of the 180 vectors). The ith
component of this rate vector then defined the firing rate of input neuron i to the network.
One rate vector was presented to the network for 50 ms. During this time, input neurons
produced Poisson spike trains with the rate as defined in the corresponding entry of the
chosen rate vector. Between the presentation of two consecutive bar patterns, all input
neurons spiked with a rate of 2 Hz for a duration drawn from a geometric distribution
with a mean of 50 simulation time steps ∆t, corresponding to 50 ms simulated biological
time.
During the learning phase, the network was presented with such patterns for 400 s.
In a testing phase, STDP in the network was disabled and input patterns were presented
to the network in the same manner as in the training phase for 100 h of simulated time.
Average firing rates of excitatory and inhibitory neurons were computed conditioned on
specific bar orientations for Figure 2E-G.
In the simulation for panel H, the same network input was presented to a WTA network
model proposed in [Nessler et al., 2013]. This model was termed spike-based expectation
maximization (SEM) network. We used a model consisting of 400 neurons that competed
in a WTA-like manner (see [Nessler et al., 2013] for details on the model). The SEM
network was simulated with a time step of 1 ms with rectangular PSPs of length 10 ms, a
total output rate of 100 Hz, initial weights chosen from a uniform distribution in [−0.5, 0.5],
non-adaptive biases of 0, and a learning rate of η = 0.02 (see [Habenschuss et al., 2013b]
for details on the simulated SEM model). The learning phase and the testing phase were
performed in the same manner as for the model M.
Details to simulations for Figure 3
Here we tested our microcircuit motif model M on input that was created by the nonlinear
superposition of 150 ms long spatio-temporal patterns.
Creation of basic rate patterns: Input spike trains to the circuit were created by the
superposition of two basic rate patterns. We first describe the creation of basic patterns,
the superposition of these patterns will be discussed below.
n,s]n=1,...,N ;s=1,...,S with ri
Let Ri denote the ith basic rate pattern. Formally, a rate pattern Ri is a matrix
Ri = [ri
n,s denoting the firing rate of the pattern in channel n at
frame s and S is the number of frames of the pattern. Each frame defines the firing rates
of channels (corresponding to the rates of input neurons) for a discrete time bin of length
∆t = 1 ms.
For Figure 3, we defined a set of two basic rate patterns R1, R2, each consisting of
N = 200 channels with S = 150 frames (i.e., the length of basic patterns was 150 ms). The
firing rate for each channel was obtained by an Ornstein-Uhlenbeck (OU) process drawn
independently for each pattern and each channel. More precisely, it was calculated as ri
j,s =
1.5 exp(xi
j,t is given by a maximum-bounded OU process. The maximum-
j,s∆t), where xi
6
bounded OU process for a variable xt is defined as dxt = ΘOU (µOU − xt)dt + σOU dWt
if xt < log(50) and dxt = 0 otherwise. Here, t denotes continuous time, ΘOU > 0 is the
changing rate (speed), µOU > 0 is the mean, σOU > 0 is the noise variance and Wt is the
standard Wiener process. The parameters were µOU = 0, ΘOU = 5, and σOU = 0.5. The
initial values for xt in the OU process were drawn from a normal distribution with zero
mean and unit variance. The first 50 ms of the OU process were discarded.
Superposition of basic rate patterns: Input spike trains were created by the superpo-
sition of a number of patterns, or more precisely their rates, from the set of basic patterns
P = {R1, R2, . . . }. Since the procedure will below also be used for the superposition of
bar patterns, we describe it here for an arbitrary number of basic patterns.
We first describe the procedure that determines which basic patterns to be superim-
posed at which times (that is, the timing of bars in Figure 3B top). Given is a set of
basic patterns P = {R1, R2, . . . }, each pattern of length S time steps. Let nmax denote
the maximum number of basic patterns that can be superimposed at any time t. We de-
fine nmax registers v1, . . . , vnmax. Each register holds at any time step t either no pattern
(empty register) or one basic pattern, with the constraint that the registers hold different
patterns at any given time step t. The following procedure ensures that at any time,
the probability that a given register holds some pattern is ploaded. At time step t, each
empty register vi is loaded with some pattern independently from other time steps and
. If a register is loaded at time step t,
other registers with probability
the basic pattern to be loaded to this register is chosen uniformly from the set of basic
patterns that are currently not held by any register. This basic pattern is then kept in
the register for the length of its duration S (afterwards the register is empty, but can be
loaded again right away). Note that whether a basic pattern is in register vi or vj at some
time t is irrelevant with respect to the produced superimposed patterns.
1+S(1−ploaded)/ploaded
1
This defines for any time step t, which basic patterns are to be superimposed and also
the frame at which each of these patterns is at that time. Superposition of basic patterns
is then accomplished as described above to obtain the rate for each input neuron. Poisson
spike trains are drawn from the resulting rates.
Patterns were first superimposed linearly, then a nonlinearity was applied. Consider a
time t when a set of patterns should be superimposed (these patterns overlap at this time
point). For the linear superposition, the rate of a particular channel in the superposition
at time t is given by the sum of the rates of this channel in all patterns that overlap at
time t. More formally, let S(t) denote the set of indices of patterns that overlap at time t
and let si(t) denote the frame at which pattern i is at time t (if the pattern presentation
∆t k + 1 at time t, with ∆t being the
started at time t′, the pattern is in frame si(t) = j t−t′
discretization time step). Then the linearly superimposed rate rlinear
(t) for channel j is
given by
j
rlinear
j
(t) = X
i∈S(t)
ri
j,si(t).
In the final nonlinear step, the firing rate rj(t) of input neuron j at time t is squashed by
a sigmoidal nonlinearity
rj(t) =
fH
1 + exp(cid:16)− 2κ
fH
(rlinear
j
(t) − 0.5fH)(cid:17)
,
where fH = 75 Hz is the maximum attainable rate and κ = 5 sets the width of the
sigmoidal function.
In order to avoid completely silent periods in the input between
7
pattern presentations, the rate of each input neuron is set to 2 Hz at times t when no
patterns are superimposed (i.e., S(t) = {}).
For Figure 3, we used 2 basic patterns with a maximum number of superimposed basic
patterns of nmax = 2 and each register had a load probability of ploaded = 0.5 (i.e., each
register was loaded with some basic pattern half of the time).
Pattern-selectivity and activity plots in Figure 3B, C: Network activity was ana-
lyzed after a learning period of 400 s of simulated biological time. Neurons were classified
as preferring pattern 1 (green pattern), as preferring pattern 2 (blue pattern), or as non-
selective based on a procedure similar to the one used in [Harvey et al., 2012]: First, an
activity trace for each neuron was obtained by convolving its spike response with a double
exponential kernel eq. (2) with τr = 1 ms, τf = 20 ms, and cut-off time Tǫ = 200 ms.
A neuron was considered to be active if it had at least 2 spikes during the simulation
time. We classified a neuron as pattern modulated if it was an active neuron and if it had
a twice as high average activity trace during presentations of patterns than during the
times without patterns. From the pattern modulated neurons, a neuron was classified as
pattern selective if it had significantly different activity traces during presentation of blue
and green patterns. This was determined by a two-tailed t-test with significance value set
at p < 0.05. If a neuron was pattern selective, its pattern preference was decided based
on the average activity trace during blue and green pattern presentations: the preferred
pattern was defined as the pattern for which the mean of the activity trace is higher.
Finally, we call a neuron that is not pattern selective a non-selective neuron.
For average activity plots in panel C, the green and blue patterns were presented to the
network in isolation, 200 presentations per pattern. Activity traces of all pattern selective
neurons were averaged over all presentations of the pattern and subsequently normalized
to their peak average activity. Neurons were then sorted by the time of their peak average
activity at presentation of their preferred pattern and average activity was plotted in
the sorted order for both patterns. In panel B, spike trains were also plotted separately
for green pattern preferring neurons, blue pattern preferring neurons and non-selective
neurons. The sorting of the former two groups was the same as in panel C.
Details to simulations for Figure 4
For this simulation, basic rate patterns were superimposed as described above for Figure
3. The experiment however differed in the number and choice of basic patterns. Network
responses, precision measures, and synaptic weight vectors were evaluated and plotted
after a learning period of 400 s simulated biological time.
Creation of basic rate patterns: Basic patterns consisted of 64 channels that were
representing horizontal and vertical bars in a two-dimensional pixel array of size 8 × 8
pixels. The pattern length was 50 ms (50 frames) and in contrast to the basic patterns
for Figure 3, the rate in each individual channel was constant over the period of the
n in a basic pattern Ri
pattern, i.e., ri
corresponded to one pixel in an 8 × 8 pixel array. We defined 16 basic patterns in total,
corresponding to all possible horizontal and vertical bars of width 1 in this pixel array.
For a horizontal (vertical) bar, all pixels of a row (column) in the array attained the value
75 while all other pixels were set to 0. The channel rates ri
n were then defined by the
values of the corresponding pixels in the array.
n for s = 1, . . . , S. Each of the 64 channels, ri
n,s = ri
Superposition of basic rate patterns: Basic rate patterns were superimposed as de-
scribed above for Figure 3. A maximum of nmax = 3 basic patterns were allowed to be
8
superimposed at any time with a load probability of ploaded = 0.9 (see Details to sim-
ulations for Figure 3 above for a definition).
In addition to the rates defined by this
superposition, a noise rate of rnoise(t) = 3(3 − npat(t)) Hz was added to each channel,
where npat(t) denotes the number of patterns that are superimposed at time t.
Precision measure: Our aim was to quantify how well neurons prefer particular basic
patterns (i.e., are tuned to particular basic patterns). To measure tuning properties,
we computed the precision measure [van Rijsbergen, 1974] Precisionij for each pair of
excitatory neuron i and basic pattern (bar) j. To this end, we say that a neuron i indicates
the presence of a pattern whenever the neuron spikes. The precision Precisionij is then the
fraction between the number of times the presence of pattern j is correctly indicated by
neuron i divided by the number of times that neuron i indicates that pattern. Hence, the
precision measures how well one can predict the presence of a pattern j, given that neuron
i spiked. Note that the precision measures whether the pattern is present whenever there
is a spike, and not whether the neuron spikes whenever the pattern is present. Since many
neurons are jointly representing a pattern, the latter question does not make sense on the
individual neuron level (it will be quantified later in Figure 5C on the population level).
Panel C, shows for each pair of excitatory neuron i and basic pattern (bar) j the
precision measure Precisionij. Formally, the precision measure is defined according to [van
Rijsbergen, 1974] as
Precisionij =
TPij
TPij + FPij
,
(9)
where TPij denotes the true positive count and FPij denotes the false positive count for
that pair. The true positive count TPij is given by the number of times that neuron i
spikes while basic pattern j is present in the input. A pattern that starts at time t is
defined to be present in the interval [t, t + S∆t + τ ]. Here, S∆t is the length of the pattern
and τ = 10 ms corrects for PSPs that increase the firing rates of excitatory neurons even
after the pattern disappeared. The false positive count FPij denotes the number of times
that neuron i spikes when pattern j is not present.
We say that a neuron i prefers basic pattern j if the neuron has maximum precision for
pattern j and this precision is larger or equal to 0.8, and if the second largest precision that
neuron i has for any other pattern is lower then 0.7. A neuron is said to be pattern-selective
if it prefers some pattern and non-selective otherwise. In panels B and C, pattern-selective
neurons are shown and sorted according to their preferred basic pattern.
Details to simulations for Figure 5
Figure 5 shows the behavior of M over the course of learning in the overlapping bars task
(same setup as Figure 4). 10 independent simulation runs were performed, each for 1000
s of simulated biological time.
In Figure 5A, a neuron is considered to be recruited if it is pattern selective.
In
Figure 5B a pattern is considered to be represented if at least on neuron prefers that
pattern. Since several excitatory neurons in the circuit can specialize on a given basic
pattern, the network performance shown in Figure 5C was evaluated over ensembles of
neurons, where ensemble Ei is given by the set of neurons that prefer basic pattern i.
To quantify how well basic pattern i is represented by ensemble Ei, we computed the F1
measure [Van Rijsbergen, 2004]. The F1 measure is at its maximum value of 1 if the
following holds true: a neuron in the ensemble Ei is active if and only if basic pattern i
9
is present in the input. False positives (i. e., some neuron in the ensemble is active in the
absence of the basic pattern) and false negatives (i. e., the basic pattern is present in the
input, but no neuron of the ensemble is active) reduce the measure, and the minimum
possible value of the measure is 0. Hence, the F1 measure for basic pattern i measures
how well this basic pattern is represented by the ensemble Ei.
Formally, we computed the F 1 measure [Van Rijsbergen, 2004] for ensemble Ei defined
as
F1i =
2TPi
2TPi + FNi + FPi
,
(10)
where FNi denotes the false negative count. The true positive count TPi is given by the
number of times that basic pattern i is present and detected by ensemble Ei, where the
pattern active during [t, t + S∆t] is detected if some neuron of the ensemble fires at least
one spike within [t, t + S∆t + τ ]. The false negative count FNi denotes the number of
times when the pattern i is active but there is not a single spike from ensemble Ei. To
calculate the false positive count FPi, we split the time between two presentations of the
pattern (time without pattern i) into periods of [t, t + S∆t + τ ] (where the last period
can be shorter). Then the false positive count FPi denotes the number of such periods
during which there is at least one spike from ensemble Ei. In Figure 5C, the mean F1
measure over all 16 basic patterns is plotted, thus indicating how well all the patterns are
represented by the network. For comparison, a SEM network as used for Figure 2H was
trained on the same input for 2000 s simulated time.
Details to simulations for Figure 6
Figure 6 shows analysis regarding the temporal relation between excitation and inhibition
in M in the experiment of Section 3.3 (Figure 3) after learning. Panel A depicts the mean
firing rate of excitatory neurons in the network (blue; mean taken over all excitatory
neurons) and the mean firing rate of inhibitory neurons (red; smoothed through a 100 ms
boxcar filter) as well as the scaled mean firing rate of inhibitory neurons (dashed green).
The mean firing rate of inhibitory neurons is scaled down by the ratio of the average
excitatory and inhibitory firing rates (average taken over the whole simulation time) in
order to facilitate comparison.
The lag between excitation and inhibition was quantified in a similar manner as in
[Okun and Lampl, 2008]. The cross-correlation function between excitatory and scaled
inhibitory firing rate for a duration of 10 seconds was computed (plotted in panel B).
The lag was then given by the offset of the peak in the cross-correlation function (plotted
in panel C) from 0. In order to evaluate the influence of connections between inhibitory
neurons, we performed the same simulations without I-to-I connections and quantified the
lag in the same manner (panel C). In order to facilitate a fair comparison, in addition to
removing I-to-I connections we also scaled down synaptic weights of I-to-E connections
by factor of 0.1155 to obtain the same excitatory firing rate as in the case with I-to-I
connections.
10
Figure 1: A data-based microcircuit motif model M. A) Network anatomy. Circles
denote excitatory (black) and inhibitory (red) pools of neurons. Black arrows indicate
excitatory connections. Red lines with dots indicate inhibitory connections. Numbers
above connections denote corresponding connection probabilities. B) Network physiology.
Same as in (A), but connection delays δ and PSP shapes with decay time constant τf are
indicated for synaptic connections.
Input synapses are subject to STDP. C) Standard
STDP curve that is used in M. Shown is the change of the synaptic efficacy in our
model for 10 pre-post pairings in dependence on the time-difference ∆t = tpost − tpre
between a postsynaptic spike at time tpost and a presynaptic spike at time tpre. D) Divisive
normalization in the model M. The response of an excitatory neuron in the circuit to a
visual bar-stimulus at various orientations (see Methods and below for details) in control
condition (black) and for simulated increased firing of inhibitory neurons (red). Note the
divisive nature of inhibition (stronger responses are more strongly depressed in absolute
terms). Compare to Figs. 2e, 3f in [Wilson et al., 2012].
3 Results
3.1 A data-based model for a microcircuit motif consisting of excitatory
and inhibitory neurons
We analyze computational properties of densely interconnected populations of excitatory
and inhibitory neurons. In particular, we analyze a model for interacting populations of
pyramidal cells with PV+ inhibitory neurons on layer 2/3 that is based on data from the
Petersen Lab [Avermann et al., 2012], see Figure 1A, B. We refer to this specific model as
the microcircuit motif model M.
The excitatory pool in M consists of M stochastic spiking neurons, for which we use
a stochastic version of the spike response model that has been fitted to experimental data
in [Jolivet et al., 2006]. In this model the instantaneous firing rate ρm(t) of neuron m is
approximated by the exponential function applied to the current membrane potential (see
eq. (4) in Methods). These excitatory neurons project to and receive inputs from a pool
11
of inhibitory neurons, that are also interconnected among themselves, with connections
probabilities taken from [Avermann et al., 2012]. Each excitatory neuron m in the network
also receives excitatory synaptic inputs y1(t), .., yN (t) from external input neurons, whose
contribution to its membrane potential at time t depends on the synaptic efficiency wim
between the input neuron i and neuron m. We assume that these afferent connections
are subject to a standard form of STDP, see Figure 1C and Definition of a data-based
microcircuit motif model M in Methods for details.
Negative (inhibitory) contributions Pj∈Im
wIEIj(t) to the membrane potential of pyra-
midal cell m have according to the neuron model a divisive effect on its firing activity,
since its instantaneous firing rate ρm can be written (by substituting eq. (4) in eq. (3)) as:
ρm(t) =
1
τ
exp (γ Pi wim yi(t) + γα)
wIEIj(t)(cid:17)
exp(cid:16)γ Pj∈Im
.
(11)
Here, the numerator includes all excitatory contributions to the firing rate ρm(t), that
is, the synaptic inputs (unweighted sum of EPSPs) yi(t) from input neurons weighted by
the corresponding synaptic weights wim. α denotes the neuronal excitability, and τ, γ are
scaling parameters that control the shape of the response function of the neuron. The
denominator in this term for the firing rate describes inhibitory contributions, thereby
reflecting divisive inhibition [Carandini and Heeger, 2012]. Here, Ij(t) denotes synaptic
input from inhibitory neuron j weighted by some common weight wIE (Im denotes the set
of all inhibitory neurons that connect to neuron m).
Divisive inhibition has been shown to be characteristic for the interaction of pyramidal
cells with PV+ inhibitory neurons [Wilson et al., 2012]. In order to test also on a functional
level the divisive character of inhibition in the model, we artificially increased the firing rate
of inhibitory neurons in the circuit by a constant, corresponding to the in-vivo experiment
described in [Wilson et al., 2012], where activity of the PV+ neurons was increased through
optogenetic stimulation. The response of pyramidal neurons to this increased inhibition
in M resembles the experimental data, see Figure 1D.
3.2 Emergent neural codes: From WTA to k-WTA
In our first test of emergent computational properties of this microcircuit motif model
M we examined the emergence of orientation selectivity. We provided as external spike
inputs pixel-wise representations of bars in numerous random orientations with superim-
posed noise (Figure 2A). Bars were transformed into high-dimensional spike inputs by
representing each black pixel of an oriented bar for 50 ms through a Poisson input neuron
with a Poisson rate of 75 Hz, whereas all other input neurons had a Poisson rate of 1
Hz. See Figure 2B for a typical resulting spike input pattern. The initial network re-
sponse is shown in Figure 2C, and the network response after applying STDP for 400 s
to all synapses from input neurons to excitatory neurons in Figure 2D. One clearly sees
in Figure 2D the emergence of assembly codes for oriented bars. A closer look at the
resulting tuning curves of excitatory neurons in Figure 2E, F shows a dense covering of
orientations by Gaussian-like tuning curves similar as in experimental data from orienta-
tion pinwheels (see Figure 2 d,e in [Ohki et al., 2006]). In contrast, inhibitory neurons did
not become orientation selective (Figure 2G) in accordance with experimental data [Kerlin
et al., 2010, Isaacson and Scanziani, 2011].
The tuning curves of excitatory neurons in Figure 2E, F demonstrate a clear differ-
ence between the impact of divisive inhibition in this data-based model M and previously
12
Figure 2: Emergent neural codes in the microcircuit motif model M. A) Bars at
various orientations serve as network inputs. Shown are network inputs arranged in 2D
for clarity. Gray-level of each pixel indicates the resulting effective network input yi(t)
(see eq. (1)) at some time point t. B) Resulting spike pattern of input neurons (every 4th
neuron shown) for different bar orientations. Gray shading indicates the presence of a bar
in the input with orientation indicated in panel A. C) Example spike pattern of a subset
of excitatory neurons in the circuit to this input before learning. D) Spiking activity of
the same neurons for the same input after applying STDP to all synapses from input
neurons to excitatory neurons for 400 s. Only responses of orientation selective neurons
are shown, sorted by preferred orientation. Spiking activity of a random subset of non-
orientation selective neurons and inhibitory neurons to the same input is shown below. E)
Emergent tuning curves of orientation selective excitatory neurons. F) The same as in E,
but zoomed in on orientations between 90 and 120 degrees. G) Inhibitory neurons are not
orientation selective. H) Emergent tuning curves of neurons in a previously considered
WTA model [Nessler et al., 2013, Habenschuss et al., 2013b].
13
considered idealized strong inhibition in WTA-circuits [Nessler et al., 2013], see Figure 2H
on emergent computational properties. In the data-based model M several (on average
k = 17) neurons respond to each orientation with an increased firing rate. This suggests
that the emergent computational operation of the layer 2/3 microcircuit motif with divi-
sive inhibition is better described as k-WTA computation, where k winners may emerge
simultaneously from the competition.
In contrast, for the WTA model with idealized
strong inhibition [Nessler et al., 2013] at most a single neuron could fire at any moment
of time, and as a result at most two neurons responded after a corresponding learning
protocol with an increased firing rate to a given orientation (see Figure 2H and Figure 5
in [Nessler et al., 2013]).
The k-WTA computation is known to be for k > 1 more powerful than the simple
WTA computation from the perspective of computational complexity theory [Maass, 2000].
However the number k of winners is in this microcircuit motif not fixed: It depends on
synaptic weights and the external input. Hence one can describe its computation best as
an adaptive k-WTA operation.
3.3 Emergent computation on spike patterns
Simultaneous recordings from large numbers of neurons demonstrate the prominence of
large-scale activity patterns in networks of neurons [Luczak et al., 2015]. They are
commonly referred to as assemblies, assembly sequences, or assembly phase sequences.
Since [Hebb, 1949] they have been proposed to reflect tokens of brain computations that
connect the fast time scale of spikes (ms) to the slower time scale of cognition and behaviour
(100's of ms). But their precise role in neural coding and computation has remained un-
known. It is proposed by [Luczak et al., 2015] that they serve as basic information compo-
nents in global cortical communication, where each of these activity patterns is initiated
by a particular cortical region and broadcast to all areas it projects to. We show here
that our microcircuit motif model M is able to carry out a computational operation on
large-scale activity patterns that is fundamental for such a global communication scheme:
It can demix superimposed spike patterns that impinge on a generic cortical area, and
represent the presence of each pattern in their input stream through the firing of sepa-
rate populations of neurons. This suggests that the layer 2/3 microcircuit motif has an
inherent capability to solve the well known cocktail party problem (blind source sepa-
ration) [Cherry, 1953] on the level of larger activity patterns. This capability emerges
automatically through STDP, as demonstrated in Figure 3 for our data-based model M.
The input to the microcircuit motif model M is generated in Figure 3 by 200 spiking
neurons. Two repeating activity patterns (green and blue patterns) are superimposed for
the generation of Poisson spike trains (shown for every 2nd neuron in the top row of Figure
3B ). These two large-scale activity pattern consist of two time varying rate patterns for the
200 input neurons (center of Figure 3A) that are nonlinearly superimposed with random
offsets in the continuous spike input to our model. In spite of these random offsets and
the large trial-to-trial variability of spike times in each of the two patterns (see panels on
the right of Figure 3A), STDP in the synaptic connections from inputs to the excitatory
neurons in the model produced after 400 s two assemblies (green and blue spikes in the
middle row of Figure 3B). Each responded to just one of the two input patterns, and
represented its temporal progress through a stereotypical sequential firing pattern (Figure
3C). This effect occurs even if none of the two input patterns is ever presented in isolation
during learning, as shown for illustration purposes for test inputs after learning at the
right side of Figure 3B. Such emergent demixing of superimposed spike patterns in the
14
Figure 3: (Figure caption on the next page)
15
Figure 3: Emergent computation on large-scale spike patterns. A) Two spatio-
temporal patterns. Each pattern consists of 200 time-varying firing rates over 150 ms
generated by an Ornstein-Uhlenbeck process, see the middle panel (only every 2nd channel
shown for clarity). These rate patterns give rise to highly variable spike patterns, as shown
on the right. Basic rate patterns are superimposed nonlinearly with arbitrary relative
timing. The left panel shows one realization of superimposed patterns for a time segment
in panel B. Spikes are colored according to the basic pattern that most probably caused
the spike (i.e., the one with the higher rate at that time). B) Firing response of the
neurons in our model M for a test input stream after letting STDP be active for synaptic
connections from input neurons to excitatory neurons in M. Two subpopulations emerged
(green and blue), where each neuron specialized on a specific pattern and on a particular
time segment within this pattern. Spiking activity of a subset of non-selective neurons
(black) and inhibitory neurons (red) are shown below. C) Average firing rate of neurons
preferring the green (top) and blue (bottom) input pattern when the green (left) and
blue pattern (right) is shown in isolation. Neurons are ordered according to their peak
firing rate for the preferred pattern as in panel B. Resulting selective firing responses are
qualitatively similar to data from sensory cortices [Luczak et al., 2015] and higher cortical
areas [Harvey et al., 2012].
layer 2/3 microcircuit motif could enable downstream neurons to selectively respond to
just one of the patterns. Furthermore the sequential activation of the two assemblies can
also inform downstream networks through the firing of specific neurons about the current
phase of each of the two input patterns.
3.4 Emergent modular sparse coding
[Foldiak, 1990] suggested that complex objects or scenes are encoded in the brain through
a sparse modular code, where each neuron signals through its firing the presence of a
particular feature in the network input.
In this way a combinatorial explosion of the
number of neurons is avoided, that would be required if each complex external object
or scene is encoded as a whole by separate neurons. [Foldiak, 1990] proposed to use
superpositions of bars (lines), like in the top part of Figure 4A, as benchmark inputs to
test sparse modular coding capabilities of neural network models. A neural network is
able to avoid the combinatorial explosion of the number of neurons that are needed to
encode such complex inputs if it learns to represent them in a modular fashion, where
each neuron encodes the presence of one of the bars (in a particular location) in the
composed input. A key question is how such codes can emerge in a network autonomously
if only composite images (consisting of several superimposed bars) are presented as network
inputs. A WTA circuit is not able to develop a good modular code since it does not allow
that inputs are represented through the firing of more than one neuron. Hence a natural
question is whether biologically more realistic softer lateral inhibition, as implemented in
our model M, supports the emergence of sparse modular codes through STDP. Figure 3
demonstrated already some weak form of modular coding for superpositions of two spatio-
temporal patterns in the input.
Emergent neural codes for Foldiak's superposition-of-bars problem are examined in
Figure 4 and Figure 5 for our model M with 400 excitatory and 100 inhibitory neurons as
before. Superpositions of up to 3 bars were presented through 64 spiking input neurons in
a pixel-wise encoding. Each Poisson input neuron signaled for 50 ms through an increased
16
Figure 4: (Figure caption on the next page)
firing rate if the corresponding pixel was covered by a bar (each bar covered 8 horizontal or
8 vertical pixels in an 8×8 pixel array). Each of the 16 possible bar positions is indicated in
Figure 4A through a different color. Composed network inputs were created by randomly
drawing superposition of bars from the pool of 696 combinations of up to 3 bars. Obviously
our model M would not be able to represent each of these input patterns by a separate
neuron. But nevertheless a complete and noise robust modular code emerged in M.
A typical spike input stream from the 64 input neurons is shown in the middle row
of Figure 4A. The 6 squares at the bottom of Figure 4A show for 6 representative time
points (indicated by grey vertical lines) the resulting pixel-wise code that represents the
17
Figure 4: Test of the emergence of modular sparse codes for a common bench-
mark test. A) A difficult version of Foldiak's superposition-of-bars problem with asyn-
chronously varying numbers of up to 3 superimposed bars. Each of the 16 bar positions
is indicated by a separate color, and its presence in the resulting spike input stream (see
middle row) is indicated by a horizontal colored line above the spike raster. 6 of 696
possible composed input patterns are shown for arbitrarily chosen time points indicated
by grey vertical lines. The pattern at the bottom of each line indicates the effective spike
input that the network receives at that moment in time, see eq. (1). B) Emergent modular
assembly codes in the model M after 400s. A small assembly of neurons emerges for each
of the 16 bar positions (see color code at left axis and background shading). Activity of
non-selective excitatory neurons is shown at the bottom. C) Quantitative analysis of the
precision of the emergent assembly codes measured according to [van Rijsbergen, 1974] for
each of the 310 neurons from the upper part of B) on the x-axis. Dark shading means high
precision for encoding the bar position plotted on the y-axis. D) Typical weight vectors
of neurons from the 16 assemblies that had emerged.
network input (darkness of red color indicates the output trace of that input neuron at
that time, that is, its output spike train convolved with the synaptic response kernel).
After representing such a continuously varying input stream for 400 s to the network,
subpopulations of a few neurons (19.4 on average) emerged that each indicated through
their firing the presence of a bar at a particular position in a noise robust manner through
the firing of several neurons (bar position indicated at the left side of Figure 4B, and
through a corresponding shading in the background of the spike raster).
In this way
each composite input image is represented through an emergent sparse modular neural
code. This holds in spite of the fact that the image presentations were not synchronized,
i.e., individual bars appeared and disappeared at random time points, and the number of
simultaneously present bars varied.
We quantified the learning performance of our model M in extended simulations where
the network was exposed to this input for 1000 s of simulated biological time. The evalua-
tion based on 10 runs with independently drawn initial synaptic weight settings and input
patterns is shown in Figure 5 (see Methods for details). Figure 5A shows the number of
neurons recruited for modular neural coding during learning. Figure 5B shows that the
network rapidly and robustly learns to represent all 16 bar positions. In Figure 5C, network
coding performance is plotted against learning time in terms of the F1-measure [Van Ri-
jsbergen, 2004]. This measure is suitable for analyzing the reliability of assembly codes,
where several neurons in an assembly can become selective for the same feature (here:
bar position) in the network input. The F1-measure was separately computed for each
bar position. An F1 measure of 1 for a bar position indicates that the bar is correctly
reported by those neurons that are selective for a bar at this position, i.e., at least one
of the neurons in the corresponding emergent assembly is active if this bar is present and
all are inactive otherwise. Hence, a high F1 measure indicates a robust encoding of bar
positions by excitatory neurons in M. In Figure 5C, the average F1 measure over all bar
positions is plotted. After 1000 s of learning, an average F1 measure of 0.87 was attained.
For comparison, a WTA network with idealized strong inhibition (as used for Figure 2H)
was trained on the same input. In the WTA circuit, neurons did not develop a modular
code but specialized on combinations of bars. Note that the network already represents
the input very well after about 200 s of learning (see Figure 5C), although only around 200
neurons have become pattern selective at this point (see Figure 5A). Subsequentely, the
18
Figure 5: Quantitative analysis of emergent coding properties in the benchmark
task of Figure 4. A) Evolution of the number of bar selective (blue) and non-selective
(red) neurons during learning. B) The number of bar positions represented by the network
rises rapidly during learning. A bar position is considered to be represented if at least one
excitatory neuron is selective for it. C) Average F1 measure of pattern-selective neural
ensembles during learning. High F1 measure (maximum is 1) indicates emergence of highly
selective assemblies of neurons for all bar positions (see Methods). In all plots, saturated
colors indicate mean and light colored shading indicates STD over 10 runs.
Figure 6: Time course of excitation and inhibition in the model M A) Time course
of the average firing rate of excitatory (blue) and inhibitory neurons (red) during 500 ms
after learning in the experiment shown in Figure 3. The dashed green line shows a scaled
version of the average inhibitory rate for better comparison. B) Cross-correlation func-
tion between the excitatory and inhibitory rate reveals a small lag of about 3 ms between
excitation and inhibition, comparable to in-vivo data. Shown is the cross correlation with
intact connections among inhibitory neurons (black) and without these connections (gray).
Inset shows a zoom into the dotted rectangle. C) Quantification of the lag between excita-
tion and inhibition from panel B. Intact inhibition among inhibitory neurons significantly
reduces the lag between excitation and inhibition.
ensembles that represent basic patterns become larger, but this has only a small impact
on network performance.
3.5 Comparing the resulting temporal dynamics of inhibition with ex-
perimental data
We analyzed the resulting temporal dynamics of inhibition in the model M after the learn-
ing experiment of Section 3.3 (Figure 3). Figure 6A shows the time courses of the average
firing rates of excitatory neurons (blue) and inhibitory neurons (red; scaled inhibitory rate
19
in green for better comparison) during an example time interval of 500 msec. Consistent
with experimental findings [Okun and Lampl, 2008], inhibition tracks excitation quite
precisely, with a small time lag.
We quantified the lag between excitation and inhibition like in [Okun and Lampl, 2008]
as the temporal offset of the peak of the cross-correlation function between the excitatory
and scaled inhibitory firing rates (plotted in Figure 6B, black line). The resulting lag of 3
msec is comparable to the measured mean lag of 3.5 ms in-vivo [Okun and Lampl, 2008].
In accordance with the data in [Avermann et al., 2012] we included inhibitory con-
nections within the pool of inhibitory neurons (I-I connections) in the microcircuit motif
model M. We found that these connections play an important role, because they decrease
the lag between excitation and inhibition. This is quantified in Figure 6B, where the black
line shows the resulting correlation between excitation and inhibition in the model, and
the grey line for a variation of the model where all I-I connections have been deleted. The
average lag between excitation and inhibition increased through this deletion from 3 ms
(black bar in Figure 3C) to 9 ms (gray bar in Figure 3C). With intact I-I connections,
inhibition is sharpened since early inhibitory responses to excitation reduce subsequent
inhibitory spikes with a larger lag. For further details on these experiments see Details to
simulations for Figure 6 in Methods.
4 Discussion
We have investigated the computational properties of interconnected populations of pyra-
midal cells and PV+ interneurons in layer 2/3 (Figure 1), one of the most prominent
motifs in cortical neural networks. Our analysis was based on data from the Petersen
Lab for layer 2/3 of mouse barrel cortex as summarized in [Avermann et al., 2012]. We
have shown that the dynamics of inhibition in a simple model M for this microcircuit
motif is consistent with additional experimental data. Figure 1D shows that the resulting
feedback inhibition is consistent with data from [Wilson et al., 2012]. Furthermore inhi-
bition follows excitation in our model with a lag of around 3 ms (see Figure 6A), a value
that is close to the experimentally measured mean lag of 3.5 ms [Okun and Lampl, 2008].
The model M has produced in addition in Figure 6B,C a hypothesis for the functional
role of synaptic interconnections among PV+ cells in this context: It suggests that these
connections contribute to the small value of this lag.
We found that the role of inhibition in this microcircuit motif cannot be captured
adequately by a WTA model. We are proposing to consider instead a variation of the k-
WTA model, where the k most excited neurons are allowed to fire. The k-WTA model is
well known in computational complexity theory, and tends to produce more computational
power than the simple WTA model [Maass, 2000]. A closer look shows that the dynamics
of the microcircuit motif can even better be captured by an adaptive k-WTA model. In
this model, the actual number of neurons that fire in response to a network input may
vary.
We have investigated the computational properties that emerge in the model M under
STDP for spike input streams that contain superimposed firing patterns. We found the
emergent capability to disentangle these patterns, and represent the occurrence of each
pattern by a separate sparse assembly of neurons (Figure 2-4). Hence we propose that
the ubiquitous microcircuit motif of densely interconnected populations of excitatory and
inhibitory neurons provides an important atomic computational operation to large-scale
distributed brain computations. Through this operation, each network module may pro-
duce one of a small repertoire of stereotypical firing patterns, commonly referred to as
20
assemblies, assembly sequences, or packets of information [Luczak et al., 2015]. If these
assembly activations are fundamental tokens of global cortical computation and communi-
cation, as proposed by [Luczak et al., 2015], then cortical columns have to solve a particular
instance of the well-known cocktail party problem [Cherry, 1953]: They have to recognize
and separately represent spike inputs from different assemblies that are superimposed in
their network input stream.
The existence of blind source separation mechanisms of this type had already been
postulated in [Foldiak, 1990] as a prerequisite for avoiding a combinatorial explosion in
the number of neurons that are needed to represent the information contained in complex
spike input streams. We have shown in Figs. 3 -- 5 that blind source separation for spike
patterns emerges automatically in the microcircuit motif model M through STDP. This
holds even for a more demanding version of the benchmark task that [Foldiak, 1990]
had proposed: Disentangling and representing superpositions of bars not only for a fixed
number, but also for varying numbers of superimposed bars.
Relation to theoretical models for cortical microcircuit motifs
It is natural to ask whether a theoretical analysis can be performed to better understand
the emergence of this fundamental computational capability. Unfortunately, the analysis
from [Nessler et al., 2013] and [Habenschuss et al., 2013b] in terms of mixture distributions
is only applicable to WTA circuits. A new probabilistic model for softer divisive inhibition
is introduced in [Legenstein et al., 2017]. The theoretical analysis in [Legenstein et al.,
2017] shows in particular that one can relate some parameters of the model M -- such as
the neural excitability α and various synaptic efficacies in the network -- to parameters of
this probabilistic model. We used the network parameters that were derived in [Legenstein
et al., 2017] in all our simulations.
Related work
Learning in networks of excitatory and inhibitory neurons was also studied in [Litwin-
Kumar and Doiron, 2014]. They did however not study plasticity of synaptic connections
from inputs to the network. Consequently, their model could not learn to perform any
feature extraction from input patterns, which is the primary emergent computational
property of the model M. Rather, self-organization led in the model of [Litwin-Kumar
and Doiron, 2014] to an associative memory-like network behavior. An interesting feature
of their model was the use of a fast Hebbian STDP rule for synaptic connections from
inhibitory to excitatory neurons (iSTDP), which was in their model essential for maintain-
ing a balance of excitation and inhibition. We did not find a need for such fast inhibitory
plasticity. Instead, we set the strengths of inhibitory connections to fixed values. However,
it would be interesting to study which types of iSTDP would lead to a self-organization
of inhibitory dynamics that also supports blind source separation.
A soft WTA model for cortical circuits with lateral inhibition was previously studied in
[de Almeida et al., 2009]. Consistent with our model, the authors arrived at the conclusion
that lateral inhibition in cortical circuits gives rise to an adaptive k-WTA mechanism,
rather than a strict k-WTA computation. However, since it was essential for their study
that the circuit operates in the limit of no noise, their model is hard to compare to the
stochastic model that we have examined. Further, the authors did not incorporate synaptic
plasticity into their model, which is the focus of this paper.
The model M is also somewhat similar to the models of [Nessler et al., 2013, Haben-
schuss et al., 2013a, Kappel et al., 2014]. However, these studies did not model inhibition
21
through feedback from inhibitory neurons. Instead inhibition was provided in a symbolic
manner as a normalization of network activity, leading to strict WTA behavior.
The emergent computational operation in our model, the extraction of superimposed
components of input patterns, is closely related to independent component analysis (ICA)
[Hyvarinen et al., 2004]. Previous work in this direction includes the classical work
by Foldiak [Foldiak, 1990] and implementations of ICA in artificial neural networks
[Hyvarinen, 1999]. It was shown in [Bell and Sejnowski, 1997] that ICA predicts features
of neural tuning in primary visual cortex. A more recent model for a similar compu-
tational goal was proposed in [Lucke and Eggert, 2010]. This model is more abstract
and only loosely connected to cortical microcircuit motifs. ICA with spiking neurons was
previously considered in [Savin et al., 2010]. The authors derived theoretical rules for in-
trinsic plasticity (i.e., rules for homeostasis of neurons) which, when combined with input
normalization, weight scaling, and STDP, enable each neuron to extract one of a set of
independent components of inputs. While closely related in terms of the computational
function, the data-based form of inhibition in our model M has quite different features.
In [Savin et al., 2010], the main purpose of inhibition is to decorrelate neuronal activity
so that different neurons extract different features. Sparse activity is enforced there by
intrinsic plasticity. Intrinsic plasticity in their model is thus required to work on a fast
time-scale (the time scale of input presentations). In contrast, sparse network activity in
our data-based model M is enforced by inhibition. It is known that feedback inhibition is
very fast and precise [Okun and Lampl, 2008], while it is unclear whether this is also true
for intrinsic plasticity [Turrigiano and Nelson, 2004].
Experimentally testable predictions of our model
A main prediction of our model (see Figure 3) is the emergence of blind source separation
of superimposed spike patterns. In addition, our model predicts that each of the identified
basic patters of the spike inputs becomes represented through some separate assembly of
pyramidal cells. Our model predicts that this effect takes place for any type of network
input, e.g. also for artificially generated stimuli that the organism is never exposed to
in a natural environment. This hypothesis can be tested experimentally, e.g. through
optogenetic control.
In addition our model predicts a specific role of synaptic connections among PV+
inhibitory cells (see Figure 6): They contribute to the experimentally found small time
lag of just a few ms by which inhibition trails excitation. This prediction can be tested
experimentally by silencing synaptic connections among PV+ cells and measuring the
impact on the lag between excitation and inhibition.
Acknowledgments
Written under partial support by the Human Brain Project of the European Union
#604102 and #720270. We would like to thank Carl Petersen for helpful discussions.
22
References
[Avermann et al., 2012] Avermann, M., Tomm, C., Mateo, C., Gerstner, W., and Pe-
tersen, C. (2012). Microcircuits of excitatory and inhibitory neurons in layer 2/3 of
mouse barrel cortex. Journal of neurophysiology, 107(11):3116 -- 3134.
[Bell and Sejnowski, 1997] Bell, A. and Sejnowski, T. (1997). The 'independent compo-
nents' of natural scenes are edge filters. Vision Research, 37:3327 -- 38.
[Carandini and Heeger, 2012] Carandini, M. and Heeger, D. J. (2012). Normalization as
a canonical neural computation. Nature Reviews Neuroscience, 13(1):51 -- 62.
[Cherry, 1953] Cherry, E. C. (1953). Some experiments on the recognition of speech, with
one and with two ears. The Journal of the Acoustical Society of America, 25(5):975 -- 979.
[de Almeida et al., 2009] de Almeida, L., Idiart, M., and Lisman, J. E. (2009). A second
function of gamma frequency oscillations: a E%-max winner-take-all mechanism selects
which cells fire. The Journal of Neuroscience, 29(23):7497 -- 7503.
[Douglas et al., 1989] Douglas, R. J., Martin, K. A., and Whitteridge, D. (1989). A canon-
ical microcircuit for neocortex. Neural Computation, 1(4):480 -- 488.
[Ferguson et al., 2013] Ferguson, K. A., Huh, C. Y., Amilhon, B., Williams, S., and Skin-
ner, F. K. (2013). Experimentally constrained ca1 fast-firing parvalbumin-positive
interneuron network models exhibit sharp transitions into coherent high frequency
rhythms. Frontiers in computational neuroscience, 7(144).
[Fino et al., 2012] Fino, E., Packer, A. M., and Yuste, R. (2012). The logic of inhibitory
connectivity in the neocortex. The Neuroscientist, 19(3):228 -- 237.
[Foldiak, 1990] Foldiak, P. (1990). Forming sparse representations by local anti-hebbian
learning. Biological cybernetics, 64(2):165 -- 170.
[Grillner and Graybiel, 2006] Grillner, S. and Graybiel, A. (2006). Microcircuits: The
Interface between Neurons and Global Brain Function, volume 93. MIT-Press.
[Habenschuss et al., 2013a] Habenschuss, S., Jonke, Z., and Maass, W. (2013a). Stochas-
PLoS Computational Biology,
tic computations in cortical microcircuit models.
9(11):e1003311.
[Habenschuss et al., 2013b] Habenschuss, S., Puhr, H., and Maass, W. (2013b). Emer-
gence of optimal decoding of population codes through STDP. Neural computation,
25(6):1371 -- 1407.
[Harvey et al., 2012] Harvey, C. D., Coen, P., and Tank, D. W. (2012). Choice-specific
sequencis in parietal cortex during a virtual-navigation decision task. Nature, 484:62 -- 68.
[Hebb, 1949] Hebb, D. O. (1949). The Organization of Behavior. Wiley, New York.
[Ho et al., 2012] Ho, E. C., Struber, M., Bartos, M., Zhang, L., and Skinner, F. K.
(2012). Inhibitory networks of fast-spiking interneurons generate slow population activ-
ities due to excitatory fluctuations and network multistability. Journal of Neuroscience,
32(29):9931 -- 9946.
23
[Hyvarinen, 1999] Hyvarinen, A. (1999). Fast and robust fixed-point algorithms for inde-
pendent component analysis. Neural Networks, IEEE Transactions on, 10(3):626 -- 634.
[Hyvarinen et al., 2004] Hyvarinen, A., Karhunen, J., and Oja, E. (2004). Independent
Component Analysis. John Wiley & Sons.
[Isaacson and Scanziani, 2011] Isaacson, J. S. and Scanziani, M. (2011). How inhibition
shapes cortical activity. Neuron, 72(2):231 -- 243.
[Jolivet et al., 2006] Jolivet, R., Rauch, A., Luscher, H., and Gerstner, W. (2006). Predict-
ing spike timing of neocortical pyramidal neurons by simple threshold models. Journal
of Computational Neuroscience, 21:35 -- 49.
[Kappel et al., 2014] Kappel, D., Nessler, B., and Maass, W. (2014). Stdp installs in
winner-take-all circuits an online approximation to hidden markov model learning. PLoS
Comput. Biol, 10:e1003511.
[Kerlin et al., 2010] Kerlin, A. M., Andermann, M. L., Berezovskii, V. K., and Reid, R. C.
(2010). Broadly tuned response properties of diverse inhibitory neuron subtypes in
mouse visual cortex. Neuron, 67(5):858 -- 871.
[Lee et al., 2012] Lee, S. H., Kwan, A. C., Zhang, S., Phoumthipphavong, V., Flannery,
J. G., and Mamanidis, S. C. (2012). Activation of specific interneurons improves V1
feature selectivity and visual perception. Nature, 488(7411):379 -- 383.
[Legenstein et al., 2017] Legenstein, R., Jonke, Z., Habenschuss, S., and Maass, W.
(2017). A probabilistic model for learning in cortical microcircuit motifs with data-
based divisive inhibition.
[Litwin-Kumar and Doiron, 2014] Litwin-Kumar, A. and Doiron, B. (2014). Formation
and maintenance of neuronal assemblies through synaptic plasticity. Nature communi-
cations, 5.
[Lucke and Eggert, 2010] Lucke, J. and Eggert, J. (2010). Expectation truncation and the
benefits of preselection in training generative models. The Journal of Machine Learning
Research, 9999:2855 -- 2900.
[Luczak et al., 2015] Luczak, A., McNaughton, B. L., and Harris, K. D. (2015). Packet-
based communication in the cortex. Nature Review Neuroscience, 16:745 -- 755.
[Maass, 2000] Maass, W. (2000). On the computational power of winner-take-all. Neural
Computation, 12(11):2519 -- 2535.
[Nessler et al., 2013] Nessler, B., Pfeiffer, M., Buesing, L., and Maass, W. (2013). Bayesian
computation emerges in generic cortical microcircuits through spike-timing-dependent
plasticity. PLoS Computational Biology, 9(4):e1003037.
[Ohki et al., 2006] Ohki, K., Chung, S., Kara, P., Hubener, M., Bonhoeffer, T., and Reid,
R. C. (2006). Highly ordered arrangement of single neurons in orientation pinwheels.
Nature, 442(7105):925 -- 928.
[Okun and Lampl, 2008] Okun, M. and Lampl, I. (2008).
Instantaneous correlation of
excitation and inhibition during ongoing and sensory-evoked activities. Nature Neuro-
science, 11(5):535 -- 537.
24
[Packer and Yuste, 2011] Packer, A. M. and Yuste, R. (2011). Dense, unspecific connec-
tivity of neocortical parvalbumin-positive interneurons: a canonical microcircuit for
inhibition? The Journal of Neuroscience, 31(37):13260 -- 13271.
[Pala and Petersen, 2015] Pala, A. and Petersen, C. C. (2015). In vivo measurement of
cell-type-specific synaptic connectivity and synaptic transmission in layer 2/3 mouse
barrel cortex. Neuron, 85(1):68 -- 75.
[Petersen and Crochet, 2013] Petersen, C. C. H. and Crochet, S. (2013). Synaptic Com-
putation and Sensory Processing in Neocortical Layer 2/3. Neuron, 78(1):28 -- 48.
[Rumelhart and Zipser, 1985] Rumelhart, D. E. and Zipser, D. (1985). Feature Discovery
by Competitive Learning. Cognitive Science, 9(1):75 -- 112.
[Savin et al., 2010] Savin, C., Joshi, P., and Triesch, J. (2010). Independent component
analysis in spiking neurons. PLoS computational biology, 6(4):e1000757.
[Shepherd, 2004] Shepherd, G. M. (2004). The Synaptic Organization of the Brain. Oxford
University Press (New York), New York.
[Turrigiano and Nelson, 2004] Turrigiano, G. G. and Nelson, S. B. (2004). Homeostatic
plasticity in the developing nervous system. Nature Reviews Neuroscience, 5(2):97 -- 107.
[van Rijsbergen, 1974] van Rijsbergen, C. J. (1974). Foundation of evaluation. Journal of
Documentation, 30(4):365 -- 373.
[Van Rijsbergen, 2004] Van Rijsbergen, C. J. (2004). The geometry of information re-
trieval, volume 157. Cambridge University Press Cambridge.
[Wilson et al., 2012] Wilson, N. R., Runyan, C. A., Wang, F. L., and Sur, M. (2012).
inhibitory networks in vivo. Nature,
Division and subtraction by distinct cortical
488(7411):343 -- 348.
25
|
1607.03687 | 2 | 1607 | 2016-11-08T13:18:45 | Critical comments on EEG sensor space dynamical connectivity analysis | [
"q-bio.NC",
"stat.AP"
] | Many different analysis techniques have been developed and applied to EEG recordings that allow one to investigate how different brain areas interact. One particular class of methods, based on the linear parametric representation of multiple interacting time series, is widely used to study causal connectivity in the brain. However, the results obtained by these methods should be interpreted with great care. The goal of this paper is to show, both theoretically and using simulations, that results obtained by applying causal connectivity measures on the sensor (scalp) time series do not allow interpretation in terms of interacting brain sources. This is because 1) the channel locations cannot be seen as an approximation of a source's anatomical location and 2) spurious connectivity can occur between sensors. Although many measures of causal connectivity derived from EEG sensor time series are affected by the latter, here we will focus on the well-known time domain index of Granger causality (GC) and on the frequency domain directed transfer function (DTF). Using the state-space framework and designing two simulation studies we show that mixing effects caused by volume conduction can lead to spurious connections, detected either by time domain GC or by DTF. Therefore, GC/DTF causal connectivity measures should be computed at the source level, or derived within analysis frameworks that model the effects of volume conduction. Since mixing effects can also occur in the source space, it is advised to combine source space analysis with connectivity measures that are robust to mixing. | q-bio.NC | q-bio | Critical comments on EEG sensor space dynamical connectivity analysis
Frederik Van de Steen1, Luca Faes2, Esin Karahan3, Jitkomut Songsiri4, Pedro A. Valdes-Sosa3,5,
Daniele Marinazzo1
1Department of Data Analysis, Ghent University, 9000 Ghent
2Healthcare Research and Innovation Program, FBK, Trento and
BIOTech, Dept. of Industrial, Engineering, University of Trento, 38123 Mattarello, Trento, Italy
3Key Laboratory for Neuroinformation of the Ministry of Education, UESTC, 610054 Chengdu, China
4Control Systems Laboratory, Electrical Engineering Department, Chulalongkorn University, 10330
Bangkok, Thailand
5Cuban Neuroscience Center, 15202 La Habana, Cuba
Corresponding Author: Frederik Van de Steen
E-mail:[email protected]
Tel: 003292646482
Acknowledgements
This research was supported by the Fund for Scientific Research-Flanders (FWO-V), grant
FWO14/ASP/255.
2
ABSTRACT
Many different analysis techniques have been developed and applied to EEG recordings that allow one
to investigate how different brain areas interact. One particular class of methods, based on the linear
parametric representation of multiple interacting time series, is widely used to study causal connectivity
in the brain. However, the results obtained by these methods should be interpreted with great care. The
goal of this paper is to show, both theoretically and using simulations, that results obtained by applying
causal connectivity measures on the sensor (scalp) time series do not allow interpretation in terms of
interacting brain sources. This is because 1) the channel locations cannot be seen as an approximation
of a source's anatomical location and 2) spurious connectivity can occur between sensors. Although
many measures of causal connectivity derived from EEG sensor time series are affected by the latter,
here we will focus on the well-known time domain index of Granger causality (GC) and on the
frequency domain directed transfer function (DTF). Using the state-space framework and designing
two simulation studies we show that mixing effects caused by volume conduction can lead to spurious
connections, detected either by time domain GC or by DTF. Therefore, GC/DTF causal connectivity
measures should be computed at the source level, or derived within analysis frameworks that model the
effects of volume conduction. Since mixing effects can also occur in the source space, it is advised to
combine source space analysis with connectivity measures that are robust to mixing.
Keywords:
EEG, Brain connectivity, MVAR, Granger causality, Directed transfer function
3
1. Introduction
With the advent of non-invasive techniques such as, positron emission tomography (PET),
functional
magnetic
resonance
imaging
(fMRI),
electroencephalograph
(EEG),
magnetoencephalography (MEG) and more recently functional near-infrared spectroscopy (fNIRS), the
study of the dynamical behaviour of the human brain has become a major field in science. Along with
these techniques, many data analysis tools have been developed that allow researchers to investigate
brain functioning to a greater extent. For many years, researchers were mostly interested in the
localization of function both in space (e.g. fMRI) and time (e.g. EEG). More recently there has been a
growing interest in functional integration, which refers to the interplay between functionally segregated
brain areas (Friston 2011). This tendency has led to the development, refinement and use of advanced
data analysis techniques such as dynamic causal modelling (DCM; Friston et al. 2003), independent
component analysis (ICA, e.g. Makeig et al. 1996; Van De Ven et al. 2004), and methods based on
multivariate autoregressive (MVAR) modeling. While each of these approaches has its own merits and
pitfalls, here we will focus on the application of MVAR-based connectivity analysis on EEG data.
Mainly due to its close relation with the ubiquitous concept of Granger causality (GC, Granger
1969), the use of MVAR-based measures to investigate human brain connectivity has become
increasingly popular in neuroscience (e.g. Ding et al. 2006; Liao et al. 2011; Porta and Faes, 2016; Seth
et al. 2015). MVAR models are being extensively applied to quantify –in both time and frequency
domains– the concept of GC expressing the direct causal influence from one series to another ( Baccala
et al. 2001; Ding et al. 2006; Seth et al. 2015), as well as the related concept of total causal influence
between two time series –expressed in the frequency domain by the so-called directed transfer function
(DTF; Kamiński and Blinowska 1991). Although the use of MVAR modelling on neural time series has
been debated (e.g. Friston et al. 2013; Friston et al. 2014; Seth et al. 2015), this approach is widely used
in neuroscience due to its high flexibility (no assumptions are made about the underlying neural
mechanisms) and easiness of implementation (the derived measures like GC and DTF are obtained
from the model parameters estimated by standard least squares techniques). Moreover, there is a
continuous refinement of MVAR-based methods in order to deal with issues related to applicative
specific contexts. Recently, for example, linear state space (SS) models have been proposed for GC
analysis, which can overcome the bias due to e.g. sampling and filtering (Barnett and Seth 2015), or
account for the well know effect of volume conduction typical of EEG analysis (Cheung et al. 2010).
In this paper, we do not intend to debate the use of MVAR-based and GC analyses in neuroscience.
Our goal is to show that MVAR-based analysis of causal connectivity performed at the level of EEG
sensors does not allow interpretation in terms of anatomically interacting sources. There are two
4
reasons why this interpretation is problematic. First, no inferences can be made about the anatomical
location of the (possibly) interacting sources based on the location of the sensors (Nunez and
Srinivasan 2006; Haufe et al. 2013; Papadopoulou et al. 2015). Secondly, since MVAR-based measures
are influenced by volume conduction, their straight application to sensor EEG data can lead to the
detection of spurious connections. The former problem is generally accepted in the literature.
Regarding the latter, there is less consensus. This issue will be the main focus of the present work
which mainly disagrees with what said in Kaminski and Blinowska (2014). In that paper the authors
claim that MVAR-based measures such as the DTF are not influenced by volume conduction. Their
main argument is that volume conduction is an instantaneous effect and hence cannot induce a phase
shift at the sensors. Along that line on the other hand, several papers were published where the DTF
was applied to the EEG sensor data to make claims regarding interacting brain sources (e.g. Wyczesany
et al. 2014; Ligeza et al. 2015; Wyczesany et al. 2015). On the other hand, several authors have pointed
to the adverse effects of volume conduction on sensor space connectivity analysis (Gómez-Herrero et
al. 2008; Schoffelen and Gross 2009; Haufe et al. 2010; Haufe et al. 2013; Bastos and Schoffelen
2016). Vinck et al. (2015) for example showed that linearly mixed additive noise (as a consequence of
e.g. volume conduction) can lead to spurious connectivity. Here we will investigate the effect of
volume conduction on two of the most popular MVAR measures of causal connectivity: the time
domain GC index (Geweke 1982) and the DTF (Kamiński and Blinowska 1991).
The paper is organized as follows. First, we show how brain dynamics can be modelled at the source
level using MVAR models and how source activity is projected to the EEG channels (section 2.1).
Then, we discuss how these dynamics can be quantified both in the time and frequency domain using
GC and DTF measures (section 2.2). In section 2.3, we outline three strategies for causal connectivity
analysis that may be applied to EEG data: 1) the analysis performed directly on sensor data; 2) a two
stage approach where the source time series are first estimated after which the analysis is performed on
these reconstructed time series; 3) the analysis performed within the state space (SS) framework. Then,
we show how the first strategy can lead to spurious connectivity results (section 3). This is shown first
by making some theoretical considerations regarding the effects of volume conduction on causal
connectivity analysis of EEG sensor time series (section 3.1), and then illustrating the issue with two
simulations studies (section 3.2). In the first simulation, the adverse effect of mixing are demonstrated
with a simple 'toy' model. In the second simulation study, more realistic EEG data are generated.
Finally, we discuss our findings in light of the existing literature.
5
2. Causal Connectivity Analysis of EEG Time Series
2.1 Linear Modelling of EEG Source Interaction and Volume Conduction
The general aim of the application to neural data of methods for connectivity analysis like GC and
DTF is to investigate how different functionally segregated brain areas interact with each other. In the
case of the EEG, the data consists of electrical potentials which are measured using multiple electrodes
that are placed at different locations of the scalp. These potentials are generally believed to originate
from the summation of post-synaptic current flows that are generated by local synchronized activity of
many neurons. These locally synchronized neurons are conveniently represented as a cortical dipole
which we call a source (Baillet et al. 2001). Due to the propagation effect known as volume
conduction, these source activations are instantly mixed resulting in electrical potentials that are
measured at the scalp. In this study we represent the time series of M cortical sources as realizations of
the stochastic process (cid:1)((cid:3))=[(cid:1)(cid:7)((cid:3))⋯(cid:1)(cid:9)((cid:3))](cid:11)s, and the corresponding measured time series of L
scalp sensors as realizations of the process (cid:12)((cid:3))=[(cid:12)(cid:7)((cid:3))⋯(cid:12)(cid:13)((cid:3))](cid:11) (where t = 1,...,T and T is the
length of the realization). The dynamical interactions between the sources can be modelled by means of
a MVAR model (Lütkepohl 2005)
(cid:1)((cid:3))=(cid:15)(cid:16)((cid:17))(cid:1)((cid:3)−(cid:17))+(cid:20)((cid:3))
(cid:21)
(cid:22)(cid:23)(cid:7)
(1)
where A(d), d=1,...,p, are M x M coefficient matrices, p is the maximum time lag that is used for
quantifying the influences of past states on the current state and U(t) is a white noise innovation
process with diagonal covariance matrix (cid:24)=(cid:25)[(cid:20)((cid:3))(cid:20)((cid:3))(cid:11)]. The mapping from source space to sensor
space can be expressed by (Haufe et al. 2013)
(cid:12)((cid:3))=(cid:26)(cid:1)((cid:3))+(cid:27)((cid:3))
(2)
where (cid:27)((cid:3))=[(cid:27)(cid:7)((cid:3))⋯(cid:27)(cid:13)((cid:3))](cid:11) is the L-dimensional vector of the measurement noise superimposed to
each channel and L is a L x M lead field matrix. The columns of the lead field matrix summarize the
so-called forward model for the different sources. Due to volume conduction, the sources are mixed
(more specifically, linearly superimposed) and hence the lead field matrix is non-sparse so that each
source contributes to some extent to the measured scalp potentials. Assuming that different sources are
activated, the EEG channels therefore contain information from a mixture of sources.
2.2 Causal Connectivity Measures
6
Granger causality (GC) is a useful tool to quantify directed neural interactions in neuroscience (Porta
and Faes 2016). The general principle of GC is based on predictability and temporal precedence: given
two time series S1 and S2, respectively considered as the target and the driver, S2 is said to Granger
cause S1 if the prediction of its current value is improved by considering the past values of S2 compared
to the case in which only the past value of S1 are exploited for prediction (Granger 1969). When M>2
time series are considered, a distinction must be made between "direct" and "indirect" GC, where the
former is obtained by conditioning on all the M-2 times-series other than the driver and the target, i.e.,
including all these other time series in the prediction models. The classical definition of "direct" GC
from a driver series Sj to a target series Si (1≤i,j≤M) is based on linear MVAR models in the form of (1),
where two models are formulated and predictability is assessed in terms of prediction error variance:
the first model (i.e the "full" model) contains all time series exactly as in (1), and the predictability of
"reduced" model) contains all time series except the driver Sj, so that the predictability of Si is now
the target series Si is assessed through the variance of Ui, (cid:24)(cid:28)(cid:28)(cid:29)=(cid:25)[(cid:20)((cid:3))(cid:28)(cid:30)]; the second model (i.e., the
assessed from the reduced innovations (cid:20)((cid:3))(cid:28)(cid:31) as (cid:24)(cid:28)(cid:28)(cid:31)=(cid:25)[(cid:20)((cid:3))(cid:28)(cid:31)(cid:30)]. Then, time domain GC (TGC) from
Sj to Si is quantified in its logarithmic formulation (Geweke 1982) as
The superscripts F and R refer to the full and reduced MVAR model respectively. From (3), it is clear
that if the inclusion of the driver does not reduce the variance of the error (i.e. the denominator) of the
target compared to the model without the driver (i.e. the numerator), then the ratio of the error
variances will be 1 and hence GC will be zero. If the variance in the error is reduced, the ratio will be
bigger than one and hence the GC values will be larger than 0. This implies that at least one of the
coefficients in the A matrices that relates the past values of the driver to the current value of the target
is different from zero (Haufe et al. 2013).
Frequency domain measures of causal connectivity can be obtained by taking the Fourier transform
of (1), yielding
&
(cid:22)(cid:23)(cid:7)
(cid:1)(%)=(cid:15)(cid:16)((cid:17))
(cid:1)(%)'((cid:28)(cid:30))(cid:22)(cid:11)*+(cid:20)(%)
(cid:1)(%)=+(ω)(cid:20)(ω),+(ω)=(.(cid:9)−(cid:15)(cid:16)((cid:17))
&
(cid:22)(cid:23)(cid:7)
'((cid:28)(cid:30))(cid:22)(cid:11)*)((cid:7)
(4a)
(4b)
(cid:27)!(cid:28)"=ln(cid:24)(cid:28)(cid:28)(cid:31)(cid:24)(cid:28)(cid:28)(cid:29)
(3)
/(cid:28)"(%)(cid:30)=
∣∣1(cid:28)"(%)∣∣(cid:30)
(cid:30)
2 ∣∣1(cid:28)3(%)∣∣
43(cid:23)(cid:7)
(5)
7
where IM is the M×M identity matrix, i = √-1, and H(ω) is the transfer function acting as a linear filter
which relates the zero mean white noise processes U(ω) to the observed processes S(ω). More
specifically, the element Hij(ω) is the frequency response of Si to a zero phase unit amplitude sinusoidal
chock with frequency ω in Sj. In the case of DTF, the elements of the transfer function are used to
quantify in normalized terms the influence of the j-th innovation process on the i-th observed process,
resulting in the definition (Kamiński et al. 2001)
The DTF is classically adopted to infer in the frequency domain the total directed influence from Sj to
Si, i.e. the directed influence arising from both the direct and indirect causal connections between the
two time series (Eichler 2006; Eichler 2012; Faes 2014). Importantly, since the DTF is based on the
elements of the transfer function, one should keep in mind that the DTF measures the frequency
response of one time series to a random sinusoidal chock (i.e. noise input) at another time series and
not the frequency response to the time series itself. The interpretation of the DTF deviates from the GC
notion and should be treated as a complementary measure of causal connectivity. If there are no
interactions among the time-series and when each time-series only depends on its own past, the transfer
function is a diagonal matrix and the DTF will be uniformly zero over frequencies for each time-series
pair. If, on the contrary, there is at least one interaction among a pair of the time-series, then at least one
of the off-diagonal elements in one of the A(d) matrices is non-zero and hence the transfer function
H(ω) will be a non-diagonal matrix, leading to a DTF which is not uniformly zero over the frequency
range for at least one pair of channels.
2.3 Assessment of Causal Connectivity for EEG time series
In the assessment of directional coupling among different neural sources by means of GC and DTF
analyses, it is clear that the goal is to make inferences on the interactions between the source time
series St. A major issue in this case is that the source time series are not directly observable, and
therefore the methods presented above cannot be applied in a straightforward way. Three possible
approaches can be followed in this case, as discussed in the following.
The first approach is to perform causal connectivity analysis as described in Sect. 2.2 directly on the
available sensor time series Y(t) (Kaminski and Blinowska 2014; Wyczesany et al. 2014; Wyczesany et
al. 2015). This implies considering the MVAR model
8
(cid:12)((cid:3))=2 5 6((cid:17))(cid:12)((cid:3)−(cid:17))+(cid:20)67((cid:3))
(cid:21)(cid:22)(cid:23)(cid:7)
,
(6)
which, being estimated from different time series than the sources S(t), will in general yield MVAR
parameters (56((cid:17)), (cid:24)8=(cid:25)[(cid:20)6((cid:3))(cid:20)6((cid:3))(cid:11)]) that differ from those relevant to the interacting ( (cid:16)((cid:17)), (cid:24)),
sources and thus will lead to different values of GC and DTF.
The second way to proceed is to employ a two-stage approach, by (i) reconstructing the time course
of the neural sources by means of some source reconstruction method (e.g. Michel et al. 2004) to get
the time series (cid:1)9((cid:3))=[(cid:1)9(cid:7)((cid:3))⋯(cid:1)9(cid:9)((cid:3))](cid:11) which represent an estimate of the unmeasured sources S(t);
and (ii) performing causal connectivity analysis on the reconstructed sources after identification of the
MVAR model:
(cid:21)(cid:22)(cid:23)(cid:7)
In this case it is expected that, if the reconstruction method and the MVAR estimation approach are
(cid:1)9((cid:3))=2 5:((cid:17))(cid:1)9((cid:3))+(cid:20):((cid:3))
accurate enough, the MVAR parameters (5:((cid:17)), (cid:24);) will be a good estimate of the true parameters
( (cid:16)((cid:17)), (cid:24)), and so will be the corresponding GC and DTF values.
A third possibility is to exploit the framework of state-space (SS) models to combine source
(7)
.
interaction and volume conduction models and estimate the measures of causal connectivity within this
framework (Cheung et al. 2010; Barnett and Seth 2015). The general linear time-invariant SS model
can be formulated as
<((cid:3)+1)=><((cid:3))+?((cid:3))
(cid:12)((cid:3))=@<((cid:3))+A((cid:3))
(8a)
(8b)
where X(t) is an K-dimensional unobserved state process, Y(t) the L-dimensional observed process, the
K×K state transition matrix B describes the update of the hidden states at each discrete time increment,
and the L x K observation matrix C describes the instantaneous mapping of the state process X to the
observed process Y; W(t) (state-equation error) and V(t) ( measurement equation error) are zero-mean
white noise processes with covariances B=(cid:25)[?((cid:3))?((cid:3))(cid:11)] and C=(cid:25)[A((cid:3))A((cid:3))(cid:11)], and cross-
covariance D=(cid:25)[?((cid:3))A((cid:3))(cid:11)]. The source interaction and volume conduction models of Eqs. (1,2)
can be re-written in SS form by defining the state process <7=[(cid:1)((cid:3))(cid:11)⋯(cid:1)((cid:3)−E)(cid:11)](cid:11) of dimension
K=Mp, defining the state and observation errors ?((cid:3))=[(cid:20)((cid:3)+1)(cid:11)0⋯0](cid:11) and A((cid:3))=(cid:27)((cid:3)), letting
the state and measurement errors be uncorrelated (D=0), and setting the parameters of the state and
observation equations (8a,b) as:
9
>=G(cid:16)(1) ⋯ (cid:16)(E−1) (cid:16)(E)
.(cid:9) ⋯
H(cid:9)
H(cid:9)
H(cid:9) K,
⋮
⋮
⋱
⋮
.(cid:9)
H(cid:9) ⋯
@=[(cid:26) H(cid:13)×(4((cid:9))] .
(9)
An important property of the SS representation is that it allows to compute the time domain measure of
GC directly from the model parameters, without the need of identifying the reduced model and estimate
its prediction error variance (Barnett and Seth 2015). Indeed, time domain GC among the sources is
assessed in a straightforward way deriving the parameters ( (cid:16)(cid:22), (cid:24)) from B and C; GC among the
sensors is assessed in the time domain from the "innovations form" representation of the SS model. In
the frequency domain, the DTF is computed from the moving-average representation of the SS model,
which can be easily derived from the original MVAR parameters (see, e.g., Barnett and Seth 2015, for a
detailed description of the procedure).
3. Effect of Volume Conduction on Sensor-Space Causal Connectivity
3.1 Theoretical Considerations
The easiest way to accomplish causal connectivity analysis on EEG datasets is to follow the first
approach described in Sect. 2.3 (Eq. (6)), since it only requires the application of MVAR modelling
routines on the observed sensor time series. However, in this case simplicity comes at a price. First of
all, if one finds significant GC or DTF among pairs of sensors, one cannot make inferences about the
spatial location of the underlying sources except for the (rare) case when one knows (or assumes) that
the activity of one single source is expressed at only one EEG channel and that one knows where these
source are a priori (Papadopoulou et al. 2015). If one cannot make these assumptions, the channel time
series are a mixture of activated sources and therefore it is impossible to infer which source's past
values that are expressed in the assigned driver sensor contribute to the prediction of another source's
future which is expressed in the target sensor. Even when there are no interacting sources, a significant
interaction between two sensors can be found applying standard connectivity analysis either in the time
domain (e.g., the TGC index (3)) or in the frequency domain (e.g., the DTF (5)). This latter point has
been recently debated (Kaminski and Blinowska 2014). In that paper, the authors claim the DTF is not
influenced by volume conduction. If this would be true, then non-interacting VAR sources should not
lead to a significant DTF at some frequencies between any pair of sensors. However, if the sources can
be expressed by (1), combined with (2) to obtain the scalp potentials, we obtain the following relations
10
for the time and frequency domain representations of the sensor series:
(cid:21)
(cid:12)((cid:3))=(cid:26)((cid:15)(cid:16)((cid:17))S(t−d)+(cid:20)((cid:3))
(cid:22)(cid:23)(cid:7)
(cid:12)(%)=(cid:26)+(ω)(cid:20)(%)+G(ω)
)+(cid:27)((cid:3))
(10a)
(10b)
In case of non-interacting MVAR sources, the coefficient matrix A and the transfer function matrix
H(ω) are both diagonal matrices. However, given that the lead field matrix L is non-diagonal, the
products LA(d) in (10a) and LH(ω) in (10b) result in a non-diagonal matrix. This implies that the
sensors can be expressed as a weighted sum of past values of all the sources (though in practice the
contribution of some specific sources to a specific EEG channel can be negligible). In addition, the
mixing of the sources is different for each sensor. This implies that the past values of the sensors will
improve the prediction of another sensor current state (even though the mixing is instantaneous, past
values are also instantaneous mixtures of sources) and hence, spurious causal connectivity will occur
both when measured by the TGC (non-diagonal coefficient matrix) and when measured by the DTF
(non-diagonal transfer function).
The above considerations show theoretically that non-interacting MVAR sources may, in
general, yield causal connectivity among sensors. In the following we identify some particular
conditions for which the mixing of non-interacting sources yields zero causal connectivity between
sensors. The first case is that of mixing of pure random statistically independent noise time series, for
situation is that of non-interacting MVAR sources with proportional power spectra projected without
exhibiting zero-lag correlation but no temporal structure, and thus no time-lagged correlations. Another
which the sensor series are (cid:1)((cid:3))=(cid:20)((cid:3)) and the sensor series become (cid:12)((cid:3))=Q(cid:20)((cid:3))+(cid:27)((cid:3)), thus
measurement noise. In this case the sources obey to Eq. (1) with (cid:16)((cid:17))=R(cid:22).(cid:9), where ad are constant
scalars and .(cid:9) is the M×M identity matrix, and the mapping occurs according to Eq. (2) with G(t)=0.
Under these assumptions, Eq. (10a) can be equated to Eq. (6) yielding (cid:20)6((cid:3))=(cid:26)(cid:20)((cid:3)) and (cid:16)6((cid:17))=
(cid:26)(cid:16)((cid:17))(cid:26)S, where (cid:26)S is the Moore-Penrose pseudoinverse of L. Then, we distinguish three cases: if L is
square (there are as many sources as sensors, L=M), we have (cid:16)6((cid:17))=(cid:26)(cid:16)((cid:17))(cid:26)(T=R(cid:22).(cid:13); if L is fat
(there are more sources than sensors, L<M), we have (cid:16)6((cid:17))=(cid:26)(cid:16)((cid:17))(cid:26)U((cid:26)(cid:26)U)((cid:7)=(cid:26)R(cid:22).(cid:9)(cid:26)U((cid:26)(cid:26)U)((cid:7)=
R(cid:22).(cid:13); if L is skinny (there are more sensors than sources, L>M), we have (cid:16)6((cid:17))=(cid:26)(cid:16)((cid:17))((cid:26)U(cid:26))((cid:7)(cid:26)U=
R(cid:22)(cid:26)((cid:26)U(cid:26))((cid:7)(cid:26)U. In the first two cases, (cid:16)6((cid:17)) is diagonal for any full rank leadfield matrix L, while in the
third case it is still possible that (cid:16)6((cid:17)) is non-diagonal depending on L (see appendix A for more
details). Thus, non-interacting sources with proportional power spectra lead to non-interacting sensors,
11
provided that the sensors are not more than the sources and are probed without measurement noise.
Note that, in the particular case in which there is only one source (M=1) and the mixing still occurs
without measurement noise, the sensor time series will be scaled versions of the source time series
(with different scales determined by the entries of L). In such a case, Granger causality among the
sensor time series is still absent because the inclusion of the other channels' past values would be
completely redundant compared to the use of the target past; however, from a practical perspective, due
to numerical inaccuracies and differences in algorithms, estimates of (cid:16)6((cid:17)) will be highly unstable,
yielding very different solutions or no-solution depending on the algorithm.
3.2 Simulation I: Theoretical Connectivity Analysis of Uncoupled Sources
In this section we report the analysis of GC in both time and frequency domains for a simple
theoretical example demonstrating how spurious connectivity can be measured in the sensor space as a
result of volume conduction. Specifically, we consider the case of two non-interacting sources
described by the following MVAR process of order 2:
(cid:1)(cid:7)((cid:3))=0.95(cid:1)(cid:7)((cid:3)−1)−0.70(cid:1)(cid:7)((cid:3)−2)+?(cid:7)((cid:3))
(cid:1)(cid:30)((cid:3))=0.50(cid:1)(cid:30)((cid:3)−1)−0.90(cid:1)(cid:30)((cid:3)−2)+?(cid:30)((cid:3))
(11)
where the innovations W1(t) and W2(t) are independently drawn from the Gaussian distribution with
zero mean and unit variance. Then, mixed time series representative of the sensor activity are obtained
by pre-multiplying S(t) = [S1(t) S2(t)]T with the following mixing matrix
(cid:26)=[0.8 0.3
0.4 0.7_
and considering zero measurement noise, i.e., Y(t) = L S(t).
Then, the derivations in Sect 2.3 are exploited to define the SS model descriptive of this parameter
setting, and connectivity analysis is performed in both time and frequency domains computing the TGC
and DTF measures from the SS representation. The TGC is zero along both directions between the two
sources S1 and S2 (TGC12=TGC21=0), while the results are substantially larger than zero between the
two sensors Y1 and Y2 (TGC12=0.1064, TGC21=0.1417). The profiles of the DTF obtained between the
two original and mixed sources are shown in Fig. 1 together with the corresponding spectral density
functions. These results of both time and frequency domain analysis show no interactions among the
12
Fig. 1 DTF analysis for simulation I, showing the profiles of DTF (off diagonal plots) and of the power
spectrum (diagonal plots) of two non-interacting VAR sources before (green) and after (red) linear
mixing. The simulated sampling frequency is 128 Hz
original source time series. On the other hand, the mixed sources reveal a clear bidirectional interaction
detected by significant TGC and DTF values along the two directions of interaction.
3.2 Simulation II: Connectivity Analysis of a Realistic Pseudo-EEG dataset
This second simulation was aimed at showing the effects of volume conduction on realistic pseudo-
EEG data generated using an adapted version of the simulation framework proposed by Haufe and
Ewald (2016). Here, we generated three sources with only one unidirectional coupling between two
sources. The MVAR model of the sources was:
13
(cid:1)(cid:30)((cid:3))=0.7(cid:1)(cid:30)((cid:3)−1)−0.5(cid:1)(cid:30)((cid:3)−2)+0.60(cid:1)(cid:7)((cid:3)−1)+?(cid:30)((cid:3))
(cid:1)(cid:7)((cid:3))=0.55(cid:1)(cid:7)((cid:3)−1)−0.80(cid:1)(cid:7)((cid:3)−2)+?(cid:7)((cid:3))
(cid:1)a((cid:3))=1.1(cid:1)a((cid:3)−1)−0.50(cid:1)a((cid:3)−2)+?a((cid:3))
,
(12)
with noise terms drawn again from the Gaussian distribution with zero mean and unit variance. We see
that the first source (S1) Granger-causes the second source (S2). The third source (S3) is not connected
to the first two sources.
First, we generated a realization of the three sources lasting T=10000 points, and computed the TGC
and DTF over this realization from the MVAR parameters estimated according to the standard least
squares approach (Faes et al. 2012). While the DTF was obtained straightforwardly as described in
Sect. 2.2 from the MVAR parameters, the TGC was computed from the SS representation of the
estimated MVAR parameters, thus avoiding the fit of full and reduced MVAR models (Barnett and Seth
2015). The TGC values were: TGC12=TGC13= TGC23 = TGC31 = TGC32 = 0 and TGC21=0.499. The
DTF profiles are shown in Fig. 2.
Next, each of the three simulated sources was assigned to a specific anatomical location: Source one
was assigned to the right occipital cortex (vertex MNI coordinates: [29 -89 4]), source two was
assigned to the right middle orbital frontal cortex (vertex MNI coordinates: [41 52 -8], and the third
source was assigned to the right superior temporal cortex (vertex MNI coordinates: [45 -15 3]), see fig.
3.
In addition to these three sources, 500 mutually statistically independent pink noise sources (i.e.
brain noise) were generated and each brain noise source was randomly allocated to 1 of the vertices of
the source model. All the generated sources were assumed to be dipoles perpendicularly oriented
toward the surface. Here the ICBM-NY head combined with the boundary element method (BEM) was
used to solve the forward problem (i.e. projecting source space time series to channel space time series
using the lead field matrix). The sources were projected to 108 EEG channels defined in 10-5 electrode
placement system (Oostenveld and Praamstra 2001). We used a similar procedure as in Haufe and
Ewald (2016) for the projection of the brain sources (i.e. the MVAR sources and the 500 pink noise
sources). More specifically, the MVAR sources and the brain noise sources were projected to the
sensors, normalized by their Frobenius norm and summed:
(cid:12)bcd(cid:28)e((cid:3))=a (cid:1)((cid:3))ghijklmh
‖(cid:1)((cid:3))ghijklmh‖(cid:29)+(1−a) o((cid:3))gbep(cid:28)qr
‖o((cid:3))gbep(cid:28)qr‖(cid:29)
(13)
14
Fig. 2 The DTF for the generated sources are show. The diagonal panels show the power spectrum of
the three sources
where Ybrain(t) are the channel time series without measurement noise. S(t) and Z(t) are the three MVAR
sources and the 500 brain noise sources respectively.Q(t)F denotes the Frobenius norm of Q(t). The
matrices Lsources and Lbnoise are the leadfield matrices for the MVAR sources and the brain noise
respectively. The signal-to-noise (SNR) parameter ,a, was set to 0.8. The overall channel data, Y(t), was
generated according to
(cid:12)((cid:3))= 0.9
(cid:12)bcd(cid:28)e(cid:2)(cid:3)(cid:4)
‖(cid:12)bcd(cid:28)e(cid:2)(cid:3)(cid:4)‖(cid:29)
(cid:19) 0.1
(cid:27)(cid:2)(cid:3)(cid:4)
‖(cid:27)(cid:2)(cid:3)(cid:4)‖(cid:29)
(14)
where G(t) is independent measurement noise drawn from the standard normal distribution.
The following channels were selected for connectivity analysis in the sensor space: AF8,O2 and T8.
15
These channels were selected as those being the closest to the three source locations. We applied
eLORETA to the channel time series to reconstruct the three MVAR sources at the a priori defined
anatomical locations (Pascual-Marqui 2007). An MVAR model of order 2 was fitted to these
Fig. 3 Anatomical locations of the three generated sources (red dots). The arrow indicates the
unidirectional connection from source one to source two
reconstructed time series using ordinary least squares method. The DTF and TGC were derived from
the estimated MVAR parameters. In addition to the two stage approach, we applied the expectation
maximization algorithm developed by (Cheung et al. 2010) to estimate the SS parameters. This
algorithm was developed specifically for EEG (MEG) applications and requires that the source
locations to be specified a priori (see Cheung et al. 2010 for more details).The DTF and TGC were
obtained from the estimated state equation parameters. The TGC values and DTF profile for the sensor-
space analysis can be found in table 1 and fig. 4 respectively. The TGC values and DTF profiles for
source reconstructed analysis and SS analysis can be found in table 2 and fig. 5. The sensor-space DTF
analysis reveals clear non-zero DTF's from O2 to AF8, from O2 to T8 and from AF8 to T8. The DTF
Table 1 Values of sensor-space TGC
Target channel
O2
AF8
T8
Driver channel
AF8
0.0199
0.0123
T8
0.0042
0.0807
O2
0.1120
0.0763
16
for the other connections are much smaller. The trends of DTF for the source-reconstructed time series
and SS analysis were almost identical to those of the originally generated time-series.
Fig. 4 Results of causal connectivity analysis showing DTF and power spectra (on the diagonal)
computed for sensor space analysis
Table 2 Values of source-reconstructed and SS-based TGC
SS analysis
Driver source
Target source
S1
S2
S3
S1
S2
S3
0.0011
0.0002
0.4938
0.0009
0.0006
0.0004
Source reconstructed analysis
Driver source
Target source
S1
S2
S3
S1
S2
S3
0.0028
0.0000
0.4132
0.0008
0.0030
0.0037
17
Fig. 5 Results of causal connectivity analysis showing DTF and power spectra (on the diagonal)
computed for source-reconstructed and SS analysis
4. Discussion
The main objective of this study was to show that the use of traditional MVAR-based connectivity
analysis, yielding classic measures such as the GC in the time domain and the DTF in the frequency
domain, does not allow interpretation of anatomically interacting brain areas when performed on EEG
sensor-space. First of all, due to volume conduction there is no one-to-one relation between the EEG
channel time series and source activity. More specifically, multiple spatially distinct neural sources
contribute to the measured EEG scalp potentials. In addition, the depth and orientation of the sources
has a major impact on how a source is projected to the channels. A single deep source for example will
have a very diffuse effect across many channels. A tangential source that is relatively close to the
18
surface will result in larger activity in channels that are not closest to that source location. A
perpendicular source close to the surface will project most strongly to the nearest channel. See fig 6 for
an instructive example of the effect of the orientation of a source that is close to the surface.
Fig. 6 An instructive example of the effect of source orientation on the scalp electrodes. The sources
are of unit strength and projected to the scalp. Top row and bottom row shows the locations (dot),
orientation (arrow) of the sources (left: tangential source; right: perpendicular source) and
corresponding topography in axial and sagittal view respectively
Therefore the spatial location of the EEG channels cannot be seen as an approximation of a source's
anatomical location (such as the anatomical location underneath the electrode). Secondly, we show that,
since the EEG time series are a mixture of sources, spurious connectivity can occur quite easily
between pairs of channels even in the absence of any causal interaction at the source level. Since the
latter has been debated in the past by Kaminski and Blinowska (2014), this issue is the main focus of
this paper. More specifically, the authors claim that the DTF is not influenced by volume conduction.
Volume conduction in EEG is mathematically equivalent to a linear superposition (mixing) of brain
sources. If the time evolution of one or more sources can be characterized by a VAR model, then the
mixed time series will also contain an autoregressive component. However, since the mixing of sources
is not identical for the different channels, the inclusion of past values of another (driver) channel can
improve the prediction of another (target) channels' current state. The same argument was given in
(Haufe et al. 2013). In that paper, the authors state that the even in the case only one MVAR source is
19
active, the inclusion of some channels' past values will improve the prediction of the current state in
another channel because both channels measure the same source but with different noise realizations
(i.e. measurement noise). Crucial in their argument is the measurement noise. In case of no
measurement noise, the two channels will be differently scaled versions of that source. as we also
discuss in Sect. 3.1, no causal connectivity will be present according to the Granger's principle. In
addition, the adverse effect of additive noise (and more specifically linearly mixed additive noise,
which can be caused by volume conduction) on GC-based connectivity analysis was also shown by
Vinck et al. (2015)
In case of multiple non-interacting MVAR sources, differentially mixing of these sources will also
result in spurious connections even in the absence of additive measurement noise. In this case the
inclusion of another channel past values in the model will not introduce redundancy. This was shown in
the first simulation where two non-interacting MVAR sources were specified and then differentially
mixed to obtain two mixed time-series. No measurement noise was added to the mixed time series. The
DTF was calculated on both the original and the mixed time series. In fig. 1, one can clearly see that,
although the DTF for the original sources is uniformly zero over frequencies in both directions, the
DTF of the mixed times-series clearly show a bidirectional interaction. Kaminski and Blinowska
(2014), however, claim that the DTF is insensitive to volume conduction. They argue that the linear
superposition of the sources is instantaneous and hence cannot produce a phase difference at the
sensors. However, linear superposition of sources can change the phase difference at the sensors under
certain conditions (see appendix B for an illustration).In the case of non-interacting sources, however,
the phase of the cross spectrum will be zero. Nevertheless, it is not true that signals without a phase
difference should result in vanishing DTF. An example of this is given by our first simulation (Eq. (11)
and Fig. 1), for which the cross-spectrum computed either between two sources or between the two
sensors has vanishing phase, while the DTF is zero if computed between the sources and nonzero if
computed between the sensors. Here, the motivation behind the non-vanishing sensor DTF is not the
presence or absence of phase difference, but rather the mixing of the same source activity in the sensor
space, which lets the source act as a 'common driver' of the activity measured at the sensors; the result
is that, given that these common effects are not (and cannot be) conditioned out in causal connectivity
analyses, the time lagged dependencies within the source are misinterpreted as time-lagged cross-
dependencies between the sensors into which these effects are mixed, even though the mixing occurs
instantaneously.
The bottom line is that volume conduction can result in a change in phase difference under certain
conditions and that a vanishing phase difference does not imply a vanishing DTF. Of course, the same
20
holds for many other measure of causal connectivity computed in the time domain (e.g., the TGC
studied here) or in the frequency domain (e.g., the partial directed coherence (Baccala et al. 2001)). On
the contrary, in this study we have shown that the measures of causal connectivity are in general
sensitive to volume conduction, while they are robust to it –in the sense that they do not indicate false
positive sensor connectivity in the absence of source connectivity– only in some particular conditions
that include the lack of autocorrelations for each individual source, the presence of the same
autocorrelation structure for all sources and in the case of a single MVAR source that is projected
without measurement noise. These conditions are arguably unrealistic in case of real EEG data. First of
all, the absence of uncorrelated measurement noise is highly unlikely (if not impossible). Secondly,
even if the (few) sources of interest show the same autocorrelation structure, many other brain sources
that are not of interest (i.e. brain noise sources) will be superimposed on the sensor level. It is however
unreasonable that all of these brain noise sources are either none interacting ,show no autocorrelation or
have the same autocorrelation structure as the signals of interest. Therefore, sensor space analysis will
in general show significant connectivity even if the sources of interest are non-interacting.
Furthermore, Kaminski and Blinowska (2014) suggest that no preprocessing of the channels should
be done to reduce the volume conduction effect because these procedures would destroy the causal
information between sources. In our second simulation study, we show that (in principle) applying the
DTF to source reconstructed time-series or by means of the SS approach can result in the correct
recovery of the original (anatomical) causal structure of the dynamical system and is thus not a priori
wrong. It should be noted that in our simulations, we used the same head model for source
reconstruction and the time series were obtained at the a priori spatial locations. In practice, however,
the head model can only be an approximation of the 'true' head model. It has been shown that errors in
the head model can result in errors in source-space GC based connectivity analysis (Cho et al. 2015). In
addition, the true locations of activated sources are not known a priori and hence need to be chosen
based on the data itself or prior knowledge (e.g. from fMRI studies using a similar paradigm; see e.g.
Schoffelen and Gross (2009) for selection strategies of sources in the context of source connectivity
analysis). Since source-space causal connectivity analysis requires estimates of the source time series, a
source reconstruction algorithm needs to be applied to the data. Since the inverse problem is ill-posed,
the solution depends on (biologically) informed constraints. Different inverse solutions pose different
constraints and the accuracy of the output of the algorithm depends on the correctness of these
constraints. Furthermore, inverse solution does not undo the mixing completely and hence spatial
leakage will still be present. This is especially the case for sources that are close to each other. In our
simulations, the sources were placed far apart from each other so that mixing effects at these source
21
locations was minimal. Nevertheless, if we would have taken a source estimate close to one of the two
interacting sources, it is likely that we would have found spurious connectivity. We emphasize that
when applying source reconstruction on real EEG data, mixing effects in the source space remain.
Nevertheless, we show that use of the DTF on source reconstructed time series is not a priori wrong
and may be used under certain circumstances.
In order to overcome the leakage effects in the source space, other methods that are more robust to
mixing are required to assess brain interaction between the estimated source time series. One such
method that is specifically designed for measures such as TGC and DTF is time reversal (Haufe et al.
2013). Time reversed data are used as surrogates for statistical testing, effectively alleviating spurious
causality due to volume conduction Importantly, (under certain conditions) reversal testing correctly
identifies causality in the presence of true interaction (Winkler et al. 2016). Although the methods that
are robust to mixing can be applied to the sensor space, their anatomical interpretability is still
problematic given that the sensor location cannot be seen as proxy to anatomical location. Moreover,
(Mahjoory et al. 2016) have shown that the choice of reference has a major impact on the sensor
connectivity profile. Using the phase slope index (i.e. a robust measure that assess the directionality of
information flow, see Nolte et al. 2008), they showed that changing the reference can reverse the
directionality of the connectivity pattern. In sum, even though source-space causal connectivity
analysis is a promising avenue, many issues still needs to be addressed in future research.
Another possibility is to use the SS framework to assess causal interactions among neural sources.
When using SS models for EEG, both volume conduction and the temporal evolution of the neural state
are modeled within the same framework. The SS framework has, for example, been used to solve the
dynamic inverse solution in EEG, which results in better spatiotemporal recovery of the hidden neural
sources (Galka et al. 2004; Yamashita et al. 2004). In the context of causal connectivity analysis,
Cheung et al. (2010), recently developed an expectation maximization algorithm to obtain model
parameter of (8a) and (8b). In that algorithm, the anatomical locations of the sources needs to be
specified a priori and hence prior knowledge or EEG source imaging needs to be applied first.
Nevertheless, the authors showed that GC analysis using SS models was less sensitive to noise
compared to the two-stage procedure. In addition, other issues (such as downsampling and other
preprocessing procedures) with respect to the MVAR modeling in causal connectivity analysis can be
overcome by using the SS framework (see Barnett and Seth 2015). In the second simulation, we
showed that the connectivity analysis using the SS framework resulted also in the correct recovery of
the DTF profiles.
To conclude, sensor-space causal connectivity analysis, performed either in time or frequency
22
domain, does not allow interpretation in terms of interacting brain sources. We also showed that the use
of measures such as the DTF and TGC in the source space is not a priori wrong as claimed by
Kaminski and Blinowska (2014). However, in order to mitigate mixing effect in the source space, it is
advised to used measures which are robust to volume conduction.
References
Baccala LA, Sameshima K, Baccalá LA, Sameshima K (2001) Partial directed coherence: a new concept
in neural structure determination. Biol Cybern 84:463–474. doi: 10.1007/PL00007990
Baillet S, Mosher JC, Leahy RM (2001) Electromagnetic brain mapping. IEEE Signal Process Mag 18:14–
30. doi: 10.1109/79.962275
Barnett L, Seth AK (2015) Granger causality for state space models. arXiv Prepr arXiv150106502v2 1–
13. doi: 10.1103/PhysRevE.91.040101
Bastos AM, Schoffelen J-M (2016) A Tutorial Review of Functional Connectivity Analysis Methods and
Their Interpretational Pitfalls. Front Syst Neurosci 9:1–23. doi: 10.3389/fnsys.2015.00175
Cheung BLP, Riedner BA, Tononi G, Van Veen BD (2010) Estimation of cortical connectivity from EEG
using state-space models. IEEE Trans Biomed Eng 57:2122–34. doi: 10.1109/TBME.2010.2050319
Cho J-H, Vorwerk J, Wolters CH, Knösche TR (2015) Influence of the head model on EEG and MEG
source connectivity analyses. Neuroimage 110:60–77. doi: 10.1016/j.neuroimage.2015.01.043
Ding M, Chen Y, Bressler SSL (2006) Granger Causality : Basic Theory and Application to Neuroscience.
Handb time Ser Anal 451–474. doi: 10.1002/9783527609970.ch17
Eichler M (2006) On the evaluation of information flow in multivariate systems by the directed
transfer function. Biol Cybern 94:469–82. doi: 10.1007/s00422-006-0062-z
Eichler M (2012) Causal inference in time series analysis. Causality Stat Perspect Appl 327–354. doi:
10.1002/9781119945710.ch22
Faes L (2014) Assessing Connectivity in the Presence of Instantaneous Causality. In: Methods in Brain
Connectivity Inference through Multivariate Time Series Analysis. pp 87–112
Faes L, Erla S, Nollo G, et al (2012) Measuring Connectivity in Linear Multivariate Processes:
Definitions, Interpretation, and Practical Analysis. Comput Math Methods Med 2012:1–18. doi:
10.1155/2012/140513
Friston K, Moran R, Seth AK (2013) Analysing connectivity with Granger causality and dynamic causal
modelling. Curr Opin Neurobiol 23:172–8. doi: 10.1016/j.conb.2012.11.010
Friston KJ (2011) Functional and effective connectivity: a review. Brain Connect 1:13–36. doi:
10.1089/brain.2011.0008
23
Friston KJ, Bastos AM, Oswal A, et al (2014) Granger causality revisited. Neuroimage 101:796–808.
doi: 10.1016/j.neuroimage.2014.06.062
Friston KJ, Harrison L, Penny W (2003) Dynamic causal modelling. Neuroimage 19:1273–1302. doi:
10.1016/S1053-8119(03)00202-7
Galka A, Yamashita O, Ozaki T, et al (2004) A solution to the dynamical inverse problem of EEG
generation using spatiotemporal Kalman filtering. Neuroimage 23:435–453. doi:
10.1016/j.neuroimage.2004.02.022
Geweke J (1982) Measurement of Linear Dependence and Feedback between Multiple Time Series. J
Am Stat Assoc 77:304–313. doi: 10.1080/01621459.1982.10477803
Gómez-Herrero G, Atienza M, Egiazarian K, Cantero JL (2008) Measuring directional coupling between
EEG sources. Neuroimage 43:497–508. doi: 10.1016/j.neuroimage.2008.07.032
Granger CWJ (1969) Investigating Causal Relations by Econometric Models and Cross-spectral
Methods. Econometrica 37:424–438. doi: 10.2307/1912791
Haufe S, Ewald A (2016) A Simulation Framework for Benchmarking EEG-Based Brain Connectivity
Estimation Methodologies. Brain Topogr 1–18. doi: 10.1007/s10548-016-0498-y
Haufe S, Nikulin V V., Müller KR, Nolte G (2013) A critical assessment of connectivity measures for EEG
data: A simulation study. Neuroimage 64:120–133. doi: 10.1016/j.neuroimage.2012.09.036
Haufe S, Tomioka R, Nolte G, et al (2010) Modeling Sparse Connectivity Between Underlying Brain
Sources for EEG/MEG. IEEE Trans Biomed Eng 57:1954–1963. doi: 10.1109/TBME.2010.2046325
Kaminski M, Blinowska KJ (2014) Directed Transfer Function is not influenced by volume conduction-
inexpedient pre-processing should be avoided. Front Comput Neurosci 8:61. doi:
10.3389/fncom.2014.00061
Kamiński M, Ding M, Truccolo WA, Bressler SL (2001) Evaluating causal relations in neural systems:
Granger causality, directed transfer function and statistical assessment of significance. Biol
Cybern 85:145–157. doi: 10.1007/s004220000235
Kamiński MJ, Blinowska KJ (1991) A new method of the description of the information flow in the
brain structures. Biol Cybern 65:203–210. doi: 10.1007/BF00198091
Liao W, Ding J, Marinazzo D, et al (2011) Small-world directed networks in the human brain:
Multivariate Granger causality analysis of resting-state fMRI. Neuroimage 54:2683–2694. doi:
10.1016/j.neuroimage.2010.11.007
Ligeza TS, Wyczesany M, Tymorek AD, Kamiński M (2015) Interactions Between the Prefrontal Cortex
and Attentional Systems During Volitional Affective Regulation: An Effective Connectivity
Reappraisal Study. Brain Topogr 253–261. doi: 10.1007/s10548-015-0454-2
Lütkepohl H (2005) New Introduction to Multiple Time Series Analysis. Springer Berlin Heidelberg,
Berlin, Heidelberg
Mahjoory K, Nikulin V V, Botrel L, et al (2016) Consistency of EEG source localization and connectivity
24
estimates. Cold Spring Harbor Labs Journals
Makeig S, J. Bell. A, Jung T-P, Sejnowski TJ (1996) Independent Component Analysis of
Electroencephalographic Data. Adv Neural Inf Process Syst 8:145–151. doi:
10.1109/ICOSP.2002.1180091
Michel CM, Murray MM, Lantz G, et al (2004) EEG source imaging. Clin Neurophysiol 115:2195–2222.
doi: 10.1016/j.clinph.2004.06.001
Nolte G, Ziehe A, Nikulin V V., et al (2008) Robustly estimating the flow direction of information in
complex physical systems. Phys Rev Lett 100:1–4. doi: 10.1103/PhysRevLett.100.234101
Nunez PL, Srinivasan R (2006) Electric fields of the brain: the neurophysics of EEG. Oxford University
Press,USA
Oostenveld R, Praamstra P (2001) The five percent electrode system for high-resolution EEG and ERP
measurements. Clin Neurophysiol 112:713–719. doi: 10.1016/S1388-2457(00)00527-7
Papadopoulou M, Friston K, Marinazzo D (2015) Estimating Directed Connectivity from Cortical
Recordings and Reconstructed Sources. Brain Topogr. doi: 10.1007/s10548-015-0450-6
Pascual-Marqui RD (2007) Discrete, 3D distributed, linear imaging methods of electric neuronal
activity. Part 1: exact, zero error localization. 1–16.
Porta A, Faes L (2016) Wiener–Granger Causality in Network Physiology With Applications to
Cardiovascular Control and Neuroscience. Proc IEEE 104:282–309. doi:
10.1109/JPROC.2015.2476824
Schoffelen J-M, Gross J (2009) Source connectivity analysis with MEG and EEG. Hum Brain Mapp
30:1857–65. doi: 10.1002/hbm.20745
Seth AK, Barrett AB, Barnett L (2015) Granger Causality Analysis in Neuroscience and Neuroimaging. J
Neurosci 35:3293–3297. doi: 10.1523/JNEUROSCI.4399-14.2015
Van De Ven VG, Formisano E, Prvulovic D, et al (2004) Functional connectivity as revealed by spatial
independent component analysis of fMRI measurements during rest. Hum Brain Mapp 22:165–
178. doi: 10.1002/hbm.20022
Vinck M, Huurdeman L, Bosman CA, et al (2015) How to detect the Granger-causal flow direction in
the presence of additive noise? Neuroimage 108:301–318. doi:
10.1016/j.neuroimage.2014.12.017
Winkler I, Panknin D, Bartz D, et al (2016) Validity of Time Reversal for Testing Granger Causality. IEEE
Trans Signal Process 64:2746–2760. doi: 10.1109/TSP.2016.2531628
Wyczesany M, Ferdek MA, Grzybowski SJ (2014) Cortical functional connectivity is associated with the
valence of affective states. Brain Cogn 90:109–115. doi: 10.1016/j.bandc.2014.06.001
Wyczesany M, Ligeza TS, Grzybowski SJ (2015) Effective connectivity during visual processing is
affected by emotional state. Brain Imaging Behav 9:717–728. doi: 10.1007/s11682-014-9326-8
25
Yamashita O, Galka A, Ozaki T, et al (2004) Recursive penalized least squares solution for dynamical
inverse problems of EEG generation. Hum Brain Mapp 21:221–235. doi: 10.1002/hbm.20000
1
The sources can be written as
Appendix A
(cid:1)(cid:2)(cid:3)(cid:4) (cid:5) (cid:6)(cid:7)(cid:8)(cid:9)(cid:2)(cid:3)(cid:4) (cid:10) (cid:11)(cid:2)(cid:3)(cid:4)(cid:12) (cid:13) (cid:14)(cid:2)(cid:3)(cid:4).
(A1)
Assuming that L is full rank and Y(t) is in the range of L, (A1) has at least one solution where
Z(t) is in the nullspace of L. L† denotes the pseudo inverse of L. Substituting (A1) into (10a)
gives
(cid:9)(cid:2)(cid:3)(cid:4) (cid:5) (cid:27) (cid:28)(cid:29)(cid:2)(cid:17)(cid:4)(cid:30)(cid:6)(cid:7)(cid:8)(cid:9)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4) (cid:10) (cid:11)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4)(cid:12) (cid:13) (cid:14)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4)(cid:31)
(cid:13) (cid:11)(cid:2)(cid:3)(cid:4) (cid:13) (cid:6)(cid:18)(cid:2)(cid:3)(cid:4)
(A2a)
(cid:9)(cid:2)(cid:3)(cid:4) (cid:5) (cid:27) (cid:28)(cid:29)(cid:2)(cid:17)(cid:4)(cid:6)(cid:7)(cid:9)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4) (cid:13)
(cid:27) (cid:28)(cid:29)(cid:2)(cid:17)(cid:4)(cid:14)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4)
(cid:21)!(cid:23)
(cid:10) (cid:27) (cid:28)(cid:29)(cid:6)(cid:7)(cid:2)(cid:17)(cid:4)(cid:11)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4) (cid:13)
(cid:21)!(cid:23)
(cid:11)(cid:2)(cid:3)(cid:4)
(A2b)
(cid:21)!(cid:23)
(cid:21)!(cid:23)
(cid:13) (cid:6)(cid:18)(cid:2)(cid:3)(cid:4)
We can define
(cid:15)(cid:16)(cid:2)(cid:17)(cid:4) (cid:5) (cid:6)(cid:15)(cid:2)(cid:17)(cid:4)(cid:6)(cid:7) and (cid:18)(cid:16)(cid:2)(cid:3)(cid:4) (cid:5) (cid:6)(cid:19)(cid:2)(cid:3)(cid:4).
Assuming no measurement noise, (A2b) becomes
(cid:9)(cid:2)(cid:3)(cid:4) (cid:5) (cid:27) (cid:15)(cid:16)(cid:2)(cid:17)(cid:4)(cid:9)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4) (cid:13)
(cid:21)!(cid:23)
(cid:27) (cid:28)(cid:29)(cid:2)(cid:17)(cid:4)(cid:14)(cid:2)(cid:3) (cid:10) (cid:17)(cid:4)
(cid:21)!(cid:23)
(cid:13) (cid:18)(cid:16)(cid:2)(cid:3)(cid:4).
Suppose Ad = adI for all d. We can describe three cases: L is square, fat or skinny.
When L is square, we have that Z(t) =0 and L†=L-1. Hence
(cid:15)(cid:16)(cid:2)(cid:17)(cid:4) (cid:5) (cid:6)(cid:15)(cid:2)(cid:17)(cid:4)(cid:6)(cid:7) (cid:5) (cid:20)(cid:21)(cid:6)(cid:6)(cid:22)(cid:23) (cid:5) (cid:20)(cid:21)(cid:24).
When L is fat, LA(d)Z(t-d)= adLZ(t-d)=0 and (cid:6)(cid:7) (cid:5) (cid:6)(cid:25)(cid:2)(cid:6)(cid:6)(cid:25)(cid:4)(cid:22)(cid:26), hence
(cid:15)(cid:16)(cid:2)(cid:17)(cid:4) (cid:5) (cid:6)(cid:15)(cid:2)(cid:17)(cid:4)(cid:6)(cid:7) (cid:5) (cid:20)(cid:21)(cid:6)(cid:6)(cid:25)(cid:2)(cid:6)(cid:6)(cid:25)(cid:4)(cid:22)(cid:26) (cid:5) (cid:20)(cid:21)(cid:24).
When L is skinny we have that Z(t) =0 and (cid:6)(cid:7) (cid:5) (cid:2)(cid:6)(cid:25)(cid:6)(cid:4)(cid:22)(cid:26)(cid:6)(cid:25). Hence
(cid:15)(cid:16)(cid:2)(cid:17)(cid:4) (cid:5) (cid:6)(cid:15)(cid:2)(cid:17)(cid:4)(cid:6)(cid:7) (cid:5) (cid:20)(cid:21)(cid:6)(cid:2)(cid:6)(cid:25)(cid:6)(cid:4)(cid:22)(cid:26)(cid:6)(cid:25).
(A3)
(A4)
(A5)
(A6)
(A7)
2
From the derivations above we can conclude that in the case of no measurement noise and
when Ad = adI for all d, non-interacting sources leads to non-interacting sensors when L is
either fat or square. When L is skinny, spurious interaction at the sensors can occur.
1
Appendix B
Here phase difference is defined as the phase of the cross spectrum. The cross-spectrum of
two time series is (the frequency index is drop for notational simplicity):
Where the phase of the cross spectrum can be obtained by:
Suppose the time series S1 and S2 are linearly superimposed into time series Y1 and Y2 by:
(B1)
(B2)
(B3)
(B4a)
(B4b)
(B5)
(cid:1)(cid:2)(cid:3)(cid:2)(cid:4) (cid:6) (cid:7)(cid:8)(cid:9)(cid:10)(cid:9)(cid:11)∗(cid:13)
(cid:14)(cid:2)(cid:3)(cid:2)(cid:4) (cid:6) arctan (cid:20)(cid:21)(cid:22)(cid:1)(cid:2)(cid:3)(cid:2)(cid:4) (cid:23)
(cid:24)(cid:22)(cid:1)(cid:2)(cid:3)(cid:2)(cid:4) (cid:23)(cid:25)
(cid:26)(cid:27)(cid:10)(cid:10)
(cid:27)(cid:11)(cid:10)
(cid:27)(cid:10)(cid:11)
(cid:27)(cid:11)(cid:11)(cid:28)(cid:26)(cid:9)(cid:10)(cid:9)(cid:11)(cid:28) (cid:6) (cid:26)(cid:29)(cid:10)(cid:29)(cid:11)(cid:28)
Then the cross-spectrum of Y1 and Y2 equals
(cid:1)(cid:30)(cid:3)(cid:30)(cid:4) (cid:6) (cid:7)(cid:8)(cid:22)(cid:27)(cid:10)(cid:10)(cid:9)(cid:10) !(cid:22)(cid:27)(cid:10)(cid:11)(cid:9)(cid:11)(cid:23)(cid:22)(cid:27)(cid:11)(cid:10)(cid:9)(cid:10)!(cid:22)(cid:27)(cid:11)(cid:11)(cid:9)(cid:11)(cid:23)∗(cid:13)
(cid:1)(cid:30)(cid:3)(cid:30)(cid:4) (cid:6) (cid:27)(cid:10)(cid:10)(cid:27)(cid:11)(cid:10)(cid:7)(cid:8)(cid:9)(cid:10)(cid:9)(cid:10)∗(cid:13)!(cid:27)(cid:10)(cid:11)(cid:27)(cid:11)(cid:11)(cid:7)(cid:8)(cid:9)(cid:11)(cid:9)(cid:11)∗(cid:13)!(cid:22)(cid:27)(cid:10)(cid:10)(cid:27)(cid:11)(cid:11) (cid:31)(cid:27)(cid:10)(cid:11)(cid:27)(cid:11)(cid:10)(cid:23)(cid:7)(cid:8)(cid:9)(cid:10)(cid:9)(cid:11)∗(cid:13)
The phase difference between Y1 and Y2 can then be formulated as
(cid:27)(cid:10)(cid:10)(cid:27)(cid:11)(cid:10)(cid:7)(cid:8)(cid:9)(cid:10)(cid:11)(cid:13)!(cid:27)(cid:10)(cid:11)(cid:27)(cid:11)(cid:11)(cid:7)(cid:8)(cid:9)(cid:11)(cid:11)(cid:13)!(cid:22)"(cid:10)(cid:10)"(cid:11)(cid:11) (cid:31)"(cid:10)(cid:11)"(cid:11)(cid:10)(cid:23)(cid:7)(cid:8)(cid:24)(cid:22)(cid:9)(cid:10)(cid:9)(cid:11)∗(cid:23)(cid:13)(cid:25)
(cid:22)(cid:27)(cid:10)(cid:10)(cid:27)(cid:11)(cid:11)(cid:31)(cid:27)(cid:10)(cid:11)(cid:27)(cid:11)(cid:10)(cid:23)(cid:7)(cid:8)(cid:21)(cid:22)(cid:9)(cid:10)(cid:9)(cid:11)∗(cid:23)(cid:13)
(cid:14)(cid:30)(cid:3)(cid:30)(cid:4) (cid:6) arctan (cid:20)
From (B5), we can derive two conditions in which the linear superposition does not change
the phase difference (i.e. when (B5) = (B2)): 1) when L is either diagonal or anti diagonal and
2) when (16) is zero or π (i.e when the sources have a phase difference of 0 or π and hence the
imaginary part of their cross spectrum is zero). To illustrate that volume conduction can lead
to a phase shift. Two signals were generated with a phase difference of π/3. The off diagonal
of the leadfield matrix was varied according to L12= L11k and L21= L22k.The phase of the cross
spectrum was then calculated for k =0,..2 See fig. B1 for the results.
2
Fig. B1 The phase difference of the mixed signals is plotted against varying values of k.
To conclude under certain conditions, volume conduction can lead to a change in phase
difference
|
1902.03121 | 3 | 1902 | 2019-05-27T19:07:56 | Can biological quantum networks solve NP-hard problems? | [
"q-bio.NC"
] | There is a widespread view that the human brain is so complex that it cannot be efficiently simulated by universal Turing machines. During the last decades the question has therefore been raised whether we need to consider quantum effects to explain the imagined cognitive power of a conscious mind.
This paper presents a personal view of several fields of philosophy and computational neurobiology in an attempt to suggest a realistic picture of how the brain might work as a basis for perception, consciousness and cognition. The purpose is to be able to identify and evaluate instances where quantum effects might play a significant role in cognitive processes.
Not surprisingly, the conclusion is that quantum-enhanced cognition and intelligence are very unlikely to be found in biological brains. Quantum effects may certainly influence the functionality of various components and signalling pathways at the molecular level in the brain network, like ion ports, synapses, sensors, and enzymes. This might evidently influence the functionality of some nodes and perhaps even the overall intelligence of the brain network, but hardly give it any dramatically enhanced functionality. So, the conclusion is that biological quantum networks can only approximately solve small instances of NP-hard problems.
On the other hand, artificial intelligence and machine learning implemented in complex dynamical systems based on genuine quantum networks can certainly be expected to show enhanced performance and quantum advantage compared with classical networks. Nevertheless, even quantum networks can only be expected to efficiently solve NP-hard problems approximately. In the end it is a question of precision - Nature is approximate. | q-bio.NC | q-bio |
Can biological quantum networks solve NP-hard problems?
Goran Wendin
Microtechnology and Nanoscience - MC2
Chalmers University of Technology
SE-41296 Gothenburg, Sweden
Email: [email protected]
Abstract.
There is a widespread view that the human brain is so complex that it cannot
be efficiently simulated by universal Turing machines. During the last decades the
question has therefore been raised whether we need to consider quantum effects to
explain the imagined cognitive power of a conscious mind.
This paper presents a personal view of
several fields of philosophy and
computational neurobiology in an attempt to suggest a realistic picture of how the
brain might work as a basis for perception, consciousness and cognition. The purpose
is to be able to identify and evaluate instances where quantum effects might play a
significant role in cognitive processes.
Not
surprisingly,
the conclusion is
that quantum-enhanced cognition and
intelligence are very unlikely to be found in biological brains. Quantum effects may
certainly influence the functionality of various components and signalling pathways
at the molecular level in the brain network, like ion ports, synapses, sensors, and
enzymes. This might evidently influence the functionality of some nodes and perhaps
even the overall intelligence of the brain network, but hardly give it any dramatically
enhanced functionality. So, the conclusion is that biological quantum networks can
only approximately solve small instances of NP-hard problems.
On the other hand, artificial intelligence and machine learning implemented in
complex dynamical systems based on genuine quantum networks can certainly be
expected to show enhanced performance and quantum advantage compared with
classical networks. Nevertheless, even quantum networks can only be expected to
efficiently solve NP-hard problems approximately.
In the end it is a question of
precision - Nature is approximate.
1
Contents
1 Introduction
2 What is the power of computing?
3 What do philosophers say about mind and matter?
3
5
7
4 The brain as a dynamic physical system
9
9
4.1 Large-scale brain networks . . . . . . . . . . . . . . . . . . . . . . . . . .
. . . . . . . . . . . 10
4.2 Brain networks show self-organised critical dynamics
4.3 The brain performs reservoir computing
. . . . . . . . . . . . . . . . . . 10
4.4 Brain network percolation model of sensory transmission . . . . . . . . . 11
5 Consciousness and cognition
13
5.1 Current paradigms: GNW and IIT . . . . . . . . . . . . . . . . . . . . . 13
5.2 Global dynamics and EEG patterns in models of brain networks . . . . . 14
5.3 Default Mode Network - the key to the self? . . . . . . . . . . . . . . . . 14
. . . . . . . . . . . . . . . . . . . . 15
5.4 Levels and disorders of consciousness
5.5 From coma to completely-locked-in syndrome
. . . . . . . . . . . . . . . 16
5.6 Reading the brain with image processing and machine learning . . . . . . 16
6 Quantum cognition
18
6.1 Quantum effects in the brain - would we even notice? . . . . . . . . . . . 18
6.2 Biochemical quantum effects on cognition . . . . . . . . . . . . . . . . . . 19
6.3 Fisher's idea of flying spin qubits
. . . . . . . . . . . . . . . . . . . . . . 19
6.4 Stapp's view of a quantum brain . . . . . . . . . . . . . . . . . . . . . . 20
6.5 Penrose/Hameroff and orchestrated objective reduction . . . . . . . . . . 20
6.6 Back to reality - the classical brain . . . . . . . . . . . . . . . . . . . . . 21
7 Quantum neural networks - quantum brains ?
22
7.1 Quantum reservoir computing (QRC) . . . . . . . . . . . . . . . . . . . . 22
7.2 Machine learning in feed-forward and recurrent QNNs . . . . . . . . . . . 23
8 Perspectives
25
2
1. Introduction
Human consciousness is often considered to be a hard problem that cannot be treated
efficiently by classical computers. To explain the assumed superior capacity of the brain,
it is sometimes speculated that quantum coherence and entanglement may be part of
the solution.
Despite the present lack of understanding of the physical mechanisms behind the
information processing of the brain, there is however one thing to be noted - quantum
computing does not solve NP-hard problems efficiently (see e.g. [1 -- 4]). A quantum
processor in the brain could possibly accelerate certain cognitive abilities, but hardly
provide any major contribution to explaining the "Hard problem of consciousness" [5].
A well-suppported current view is that the brain is a classical complex dynamical
system showing self-organised critical behaviour (SOC) [6 -- 14].
In SOC models, the
physical basis for consciousness may be emergent long-range functional correlations.
Obviously,
local quantum coherence and entanglement in specific brain and sensor
components might influence the details of the complex dynamics. However, this will
hardly enhance the computational power of the brain. And how that power is related
to human perception, consciousness and thinking is probably a very different question.
The core of the problem is that quantum physics is often conjectured or postulated
to add power to cognition without knowledge of what cognition really is. It certainly
seems reasonable to assume that cognition is connected with physical processes in the
brain, and quantum effects could possibly affect such processes. Whether that would
influence consciousness, cognition and subjective experience is a different matter.
In
order to discuss quantum enhanced cognition, we need to relate to concrete classical
computational models of the brain that effectively have not yet been formulated.
However, even if they were known, the answer to the question of "Can biological
quantum networks solve NP-hard problems?" would be a clear personal "no". Not
even quantum computers can efficiently solve NP-hard problems.
It is extremely interesting to see how the topic of consciousness and cognition
in neuroscience and philosophy, with or without quantum effects, has evolved during
the last three decades through the opinions expressed in the work of professionals in
neuroscience, psychiatry, psychology, physics and philosophy that forcefully defend their
established lines of thought and structures of knowledge [15 -- 33]. On the whole, there
is little convergence. It is difficult to avoid the feeling(!) that the good old dividing line
is more visible than ever: materialists versus dualists.
Only recently becoming conscious of those paradigms and "schools", I tend to agree
with Dennett [30] that the canonical "Hard problem of consciousness" of Chalmers [5]
is a chimera, leading us astray. This is also the view of Dehaene [18]. To me, the
origins of consciousness and cognition are very concrete things that are already basically
understood from a systems point of view of the brain. How that connects to subjective
experience and personal perceptions - qualia - is a question of internal representation
and reporting for human individuals. "Intentionality, meaning, and value" are certainly
3
The mission of this paper is threefold:
valid abstract concepts, but are necessarily coded and processed as physical information.
(i) to put quantum physics in its
proper place biologically, (ii) to make probable that there are relevant top-down
computational classical models of the brain meeting established bottom up experimental
neurophysiological models, and (iii) to clarify where quantum effects just might play a
role for cognition. No need to invent a universal mind, just hard persistent work to
convince Nagel [32] that the modern materialist approach to life will not fail to explain
the physical/biological foundations for consciousness, intentionality, meaning, and value.
What is then the meaning and value of existence is a very different matter.
First we outline some views on the power of computation (Sect. 2), present some
philosophical background (Sect. 3), and describe coupled brain networks as dynamical
complex systems (Sect. 4). Section 5 presents a review and analysis of the status of
experimental knowledge of the structure and dynamics of global brain networks. The
quantum aspects of cognition are discussed in Sect. 6, followed by some comments on
quantum networks in Sect. 7.
4
2. What is the power of computing?
Unconventional computing (UCOMP) is a vast field covering various kinds of
information processing different from von Neumann type of classical digital computing
[3, 34]. The field deals with various forms of analogue computing, and key questions
are [3]: (i) Can UCOMP provide solutions to classically undecidable problems? (ii) Can
UCOMP provide more effective solutions to NP-complete and NP-hard problems? (iii)
Are classical complexity classes and measures appropriate to any forms of UCOMP? (iv)
Which forms of UCOMP are clearly and easily amenable to characterisation and analysis
by these? And why? (v) Are there forms of UCOMP where traditional complexity
classes and measures are not appropriate, and what alternatives are then available?
There is a long-held notion that some forms of unconventional classical computing
can provide tractable solutions to NP-hard problems that take exponential time or
memory resources for classical digital machines to solve [35 -- 38], and the idea of solving
NP-hard problems in polynomial time with finite resources is still actively explored
[39 -- 43]. The question is then, what is the real computational power of UCOMP
machines: are there UCOMP solutions that can at least provide significant polynomial
speed-up and energy savings, or more cost-effective solutions?
The brain can be modelled as a heterogenous adaptive recurrent neural network
(RNN). RNNs with rational weights are Turing-complete, meaning that they can
simulate a universal Turing machine and vice versa.
If the weights represented real
numbers, the RNN would be super-Turing. However, this requires infinite precision and
is not a realistic situation. The best we can achieve is most probably the computational
power of adaptive RNNs [44, 45].
Since analogue machines like RNNs can be efficiently simulated by digital computers
[46] with "only" polynomial overhead, the conclusions of several decades of discussion
and debate [1, 2, 47, 48] can be illustrated in terms of a number of "Laws" [3]: 1st Law
of Computing: You cannot solve uncomputable or NP-hard problems efficiently unless
you have a physical infinity or an efficient oracle. 2nd Law of Computing: There are
no physical infinities or efficient oracles. 3rd Law of Computing: Nature is physical
and does not solve uncomputable or NP-hard problems efficiently. Corollary: Nature
necessarily solves uncomputable or NP-hard problems only approximately.
A natural consequence of the Corollary is the following: Natural problems solved by
the Evolution do not represent really hard instances. Protein folding is in principle NP-
hard, scaling exponentially with time. The protein chains chosen by Nature to build
biological systems must therefore involve fairly "simple" instances of NP-problems -
otherwise we would not exist, folding would take too long. However, protein folding
in biological systems does not take place from scratch - testing all
instances and
configurations. Nature is adaptive and coevolution has created efficient processes for
assembling complicated proteins on catalysing enzymatic substrates (chaperones) in
finite short time.
It follows that if we simulate a such a process digitally we will need adaptive
5
machine-learning algorithms that can find the optimal routes through the energy
landscape. In principle it also implies that it would be more efficient to use physical
implementations of RNNs. The problem is that we cannot build physical RNNs with
sufficient complexity - so far it has been necessary to simulate them in software using
powerful digital computers.
This is now beginning to change by the recent technological development (IBM,
TrueNorth [49]), making it possible to emulate spiking neural networks directly in state-
of-the-art semiconductor hardware [49 -- 51], building exremely fast event-based systems
for image processing [50, 52, 53] and deep learning [54].
These conclusions also apply to quantum information processing (QIP). Certainly
QIP is able to efficiently compute many problems that will take exponential resources
for classical computers. But in the end, not even QIP can solve NP-hard problems
efficiently. It follows that quantum networks cannot solve NP-hard problems efficiently
- no good reason for making the brain a quantum processor in order to explain its
computational capacity.
6
3. What do philosophers say about mind and matter?
The question of the nature of consciousness and self-awareness has bothered philosophers
from little Lucy until today. The litmus-test is supposedly how you react when you
look into a mirror - always a troublesome experience. Over the years, the issue of
consciousness - especially self-consciousness - was mostly addressed by philosophers and
theologists. During the 1960s a philosophical materialist view - Functionalism [55] -
was developed, assuming that the brain is a neuronal machine, and that everything can
be explained on that basis. The quality of self-consciousness and the human brain has
then come into new light with the advent of powerful computers and "deep-learning"
adaptive algorithms, as well as through the quite dramatic progress in dynamic brain
imaging.
The basic problem is the lack of operational definition and measure of consciousness
and qualia. Philosophers have tended to group into opposing categories of "materialists"
and "dualists". Materialists assume that consciousness is governed by physical law
and can be understood in physical and chemical terms. They start from generic brain
network models, investigate the dynamics, and compare with experimental results of
brain imaging. Dualists argue that mind and brain are radically different kinds of
thing - the mind cannot be treated as simply part of the physical world [56]. The
dualist perspective is expressed by David Chalmers in terms of the "Hard problem of
consciousness" [5]. Chalmers claims that the problem of subjective experience - qualia
- is distinct from the set of physically computable brain processes and that the problem
of subjective experience will "persist even when the performance of all the relevant
functions is explained".
Dualists tend to invoke concepts that look supernatural from a materialist point
of view. And, so far, materialists have not been able to explain the physical nature
of consciousness and perception based on the workings of neural circuits and systems.
This has led to a "battle between philosophers", illustrated by a series of papers and
exchanges over the last two decades [5, 28, 29, 32, 33, 57 -- 59]. The essence is illustrated
by Nagel's conclusion [32] that the modern materialist approach to life has failed to
explain consciousness, intentionality, meaning, and value. Nagel states that "this failure
threatens the entire naturalistic world picture, extending to biology, evolutionary theory,
and cosmology", underpinning the dualist view that it is impossible to reduce the mental
level to the physical level.
An opposing view is that it rather represents a failure to recognise a basic
fact of physics, formulated by the late Rolf Landauer:
information is physical [60].
"Intentionality, meaning, and value" are certainly abstract concepts, but they are
also consequences of human cognition and thinking - they don't exist in free space
independent of brains thinking about them. To have an opinion about the meaning or
value of something takes some thinking and costs energy. Subjective experience - qualia
- necessarily involves information, and information is not abstract - it has substance and
it involves energy and entropy. Transfer of information - communication - takes power.
7
This was succinctly expressed in 2014 by Dehaene [18]: Chalmers' hard problem
of representing subjective experience is an ill-posed problem - the really hard problem
is the complex of computationally hard problems describing the intricate dynamics of
complex brain networks. Chalmers' hard problem will dissolve when computational
modelling and brain imaging advances.
We might actually be there already: what is the neural representation of subjective
experience when a dose of the psychedelic psilocybin dissolves a person's entire
self, traceable to changes in specific synaptic couplings of specific brain networks,
disconnecting the individual from previous "programming" [61 -- 64]?
Arguably these changes of consciousness and self-awareness must correlate with
changes of the intrinsic dynamic states of the network. These can be very sensitive
to functional network connectivity. What is considered as "the self" in the absence
of the drug, might be regarded as illusions (or the true self!) in the presence of the
drug, or schizophrenic delusions in the case of pathological synaptic imbalance. A
radical approach is to regard consciousness and self-awareness as illusions by default,
like Dennett does [29, 30], but this feels like a philosophical impasse.
8
4. The brain as a dynamic physical system
4.1. Large-scale brain networks
The brain forms a complex dynamic system [65] that can be modelled as a highly
clustered small-world type of network [66] of highly inter-connected modules built from
networks of spiking neurons. These modules form nodes that are sparsely connected to
each other [67 -- 71], allowing both local and global information processing on a variety
of timescales [14, 19, 67].
Multiscale modelling of the structure and non-linear dynamics of the brain networks
is a vast and active research field [7 -- 9, 14 -- 19, 67, 72, 78 -- 102], providing a computational
foundation for understanding information processing in the brain and the emergence of
consciousness and cognition.
To be able to critically assess the potential role of quantum effects on consciousness
and cognition, we must first outline an up-to-date picture of the brain's network
structure, couplings and dynamics. Spontaneous brain activity is organised into resting-
state networks (RSNs) involved in internally guided, higher-order mental functions
(default mode (DMN), central executive (CEN), and salience (SN) networks [69]),
and externally-driven, specialised motor and sensory processing (sensorimotor (SMN),
auditory (AN), and visual (VN) networks) [72]. Cerebellar (CN) networks lie in between,
involving both mental and motor processing.
The Default (Mode) Network (DMN) [70, 71] describes the resting state function
of the brain and is responsible for intrinsic awareness and self-generated thought.
The Central Executive Network (CEN) (also called Executive Control Network, ECN
[73]) is responsible for extrinsic awareness (focused attention) that regulates executive
functions that control and mediate cognitive processes,
including working memory,
reasoning, flexibility, problem solving, and planning. The Salience Network (SN)
[73] is implicated in the detection and integration of emotional and sensory stimuli,
and in modulating the switch between the DMN and CEN, i.e.
switching between
introspection and attention. The connectivity within the SN distinguishes between
unconscious and minimally conscious states [73]. The Frontal Mesocircuit Network
(FMN) [76] serves as the basic underlying dynamic driver of the forebrain cortical
networks and is necessary for supporting consciousness and governing the quality of
consciousness. The Cerebellar Network (CN) [74] is of central importance for motor
control, and is involved in some cognitive functions such as attention and language
In addition, there is a number
as well as in regulating fear and pleasure responses.
of more local specialised attention/action networks for input/output functions:
the
Sensorimotor Network (SMN) [77], the Auditory network (AN), the Visual Networks
(VN), and the Olfactory Network (ON). Disruptions in activity in various networks have
been implicated in neuropsychiatric disorders such as depression, Alzheimer's, autism
spectrum disorder, schizophrenia and bipolar disorder (see e.g. [69, 75 -- 77]).
9
4.2. Brain networks show self-organised critical dynamics
There is a host of evidence that much of the brain dynamics is poised at the edge
of chaos [14, 84 -- 89]. Three decades ago, Kauffman [103] described criticality and
coevolution to the edge of chaos and avalanches in biological systems, and Bak [6]
introduced the concept of self-organised criticality (SOC) and illustrated the fractal
structure of avalanches at all scales with the formation and behaviour of a pile of
sand. SOC represents a "global" systems description in terms of universality and critical
exponents, and applications to real systems for interpretation of the dynamics must be
done with great care, as is evident from a history of successes, controversies and open
questions [83, 85, 86, 104, 105].
4.3. The brain performs reservoir computing
As already discussed, there is now plenty of experimental support for modelling the
conscious brain as a complex dynamic system poised at the edge of critical behaviour,
separating chaotic and ordered states [9 -- 12].
In SOC models, the physical basis for
consciousness may be emergent long-range functional correlations. In this situation the
system is very sensitive to external perturbations, like input from different kinds of
sensors, and it can produce fast response.
system,
that employs a signal-driven dynamical
Reservoir computing (RC) [14, 106, 108 -- 113] is a brain-inspired machine learning
framework [14]
in particular
harnessing common-signal-induced synchronisation which is a widely observed nonlinear
phenomenon. RC has been applied to a variety of physical systems [114 -- 118], and in
particular to information processing in brain-like networks [92,93,110,113]. In particular,
Enel et al. [92] compared reservoir model activity to neural activity of the dorsal anterior
cingulate cortex of monkeys, training a reservoir to perform a complex cognitive task
initially developed for monkeys. The reservoir model inherently displayed a dynamic
form of mixed selectivity [119], key to the representation of the behavioral context over
time, and revealed similar network dynamics as the cortex [92].
A comprehensive and very readable discussion of the present status of brain
reservoir computation is provided by Maass [14]. Maass emphasises that neural networks
in the brain are not the kind of homogeneous artificial neural networks that are
suitable for generic computation of Boolean circuits, or searches in energy landscapes
for optimisation of cost functions. Neural networks in the brain consist of feed-forward
and feed-back coupled heterogeneous nodes containing a variety of spiking neurons that
are connected by different types of synapses with short-term and long-term dynamics.
The short-time dynamics of excitatory and inhibitory neurons/synapses depend on the
structure of the pulse trains, making it history-dependent. All of this puts important
constraints on computational models.
One might think that such a heterogeneous recurrent neural network might not
be suitable for computations. However, Nature has successfully evolved such brains,
and they are sometimes regarded as superior to digital computers. Naturally they may
10
therefore be difficult to simulate. Nevertheless, there is a computational model for which
a diversity of computational units and time constants is a clear advantage, namely the
liquid computing paradigm [14], which together with the "echo-state" model [107] is
referred to as reservoir computing (RC). The liquid computing model is designed for
biological neural networks governed by noise, diversity of units, and temporal aspects
related to spikes.
Following Maass [14], RC models conceptually divide neural network computations
into two stages, a fixed generic nonlinear pre-processing stage and a subsequent stage of
linear readouts that are trained for specific computational tasks. In the pre-processing
stage, different types of neurons and subcircuits within the recurrent network produce a
large number of potentially useful features and nonlinear combinations of such features,
out if which projection/readout neurons can select useful information for its target
networks.
In this way even seemingly chaotic dynamics of a recurrent local network
can make a useful computational contribution. Furthermore, there is a connection
to machine learning using Support Vector Machines (SVMs). An SVM also consists
of two stages: a generic nonlinear preprocessing stage (kernel) and linear readouts.
The kernel projects external input vectors nonlinearly onto vectors in a much higher
dimensional space. One can view a large nonlinear recurrent neural network as an
implementation of such a kernel, where the network response to an input corresponds to
the kernel output [14]. An important advantage of a two-stage computational model is
multiplexing: the same first stage (the kernel) can be accessed by an unlimited number
of linear readout neurons that learn to extract different computational results for their
specific target networks.
Experiments show that the networks in the brain are changing continuously:
connectivity, neurotransmitters and neural codes. Nevertheless, those spiking networks
provide stable computational
function in spite of ongoing rewiring and network
perturbations [14]. Probabilistic computations are now playing a prominent role in
many models in neuroscience. Stochasticity of spiking neurons enables the network to
solve problems in a heuristic manner - each network state represents a possible solution
to a problem, and the frequency of being in this network state encodes the fitness of
the solution. This suggests a possible relation between flickering internal states of brain
networks, and the response in terms of perception and behaviour.
4.4. Brain network percolation model of sensory transmission
Current neurocomputational efforts often state that the goal
is to understand
consciousness and cognition manifested in sufficiently complex biological networks. In
order to do precisely that, in a recent theoretical study of a model for thalamo-cortical
structure, Zhou et al. [87] described a systems-level network approach to explaining brain
network processing under deep anesthesia, to account for EEG patterns, disruption of
information flow, and neurobiological function together. The work also compared to
clinical observations during loss of consciousness under general anesthesia.
11
Zhou et al.
[87] used a network model with a single percolation-probability
parameter without considering many biological details. The model [87] consists of
a layered fractal expansion of nodes from a single source (input) node to multiple
destination (output) nodes. The nodes represent anatomical or functional simple or
complex brain units. The directional edges represent structural, functional, or causal
communications between nodes. Nodes are arranged into layers in 1-to-4 and 1-to-
9 fractalisation schemes. Within each layer, nodes are connected by the small-world
topology according to the Watts-Strogatz algorithm [66].
The node amplitudes Ai(t) denote neural activity and are given by Ai(t) =
Pj wijPj(t), where the input function Pj(t) = Pm
τ =1 e−τ Aj(t − τ ) represents the history
of activity during m previous time steps. The weights wij of the directional edges are
stochastically defined by a single parameter p representing the percolation probability
of information transmission. The system is driven by white noise at the first input node
i = 1, and the output noise is detected at randomly selected single nodes in the topmost
layer. The time-dependent output is analysed in terms Fourier spectra, spectral power
and relative spectral entropy [87]. The network percolation algorithm produces patterns
that are similar to human EEG patterns, and broadly describes how the frequency and
power distributions of the α, β, γ, δ signals change going from conscious to unconscious
states during anesthesia. In particular, the model reveals emergence of phase coherence
and synchronisation at a critical threshold.
In summary: Zhou et al. [87] propose that the behaviour of the model supports the
notion that consciousness may arise from the same basic statistical processes as those
governing the self-organised criticality, regulated by a single connectivity parameter.
They state that the results agree with clinical observations of (i) a spectral shift of power
toward lower frequency during sedation and (ii) a sharp transition between conscious
and unconscious states within an extremely narrow anesthetic concentration range.
One may now wonder what is the relevance of the thalamo-cortical SMN
communication network for identifying consciousness? A way (my way) to see this
is to note that the model is generic, including no biological detail. The fractal structure
is generic, and should describe the path from any specific "input" node to any other
distant "output" node in any of the major networks discussed before: DMN, CEN, SN,
CN, SMN.
Since the model [87] connects the computed EEG time pattern for the model SMN
with actual clinical results for the global brain, the result should be relevant for the
appearance of phase coherence and synchronisation in generic large-scale brain networks.
Which specific network, if any, is the home of consciousness is a different matter, to be
discussed in the next Section.
12
5. Consciousness and cognition
5.1. Current paradigms: GNW and IIT
For the last two decades, neurobiological paradigms have been largely dominated by
Global Neural Workspace (GNW) theory [15 -- 17, 19, 78, 79] and Integrated Information
Theory (IIT) [23 -- 25, 27, 121]. Both GNW [19] and IIT [27] are founded in neurobiology
and the functionality of neural circuits, but the way they make use of the common
experimental knowledge is fundamentally different. Broadly speaking, GNW [19]
describes consciousness as a computable global property of a neural network, while
IIT [27] posits that consciousness is connected with high values of irreducible global
network information.
IIT predicts that a sophisticated simulation of a human brain
running on a digital computer cannot be conscious - even if it can speak in a manner
indistinguishable from a human being [27]. This is in stark contrast to the predictions of
GNW which posits that machines (AIs) will be able to show consciousness at a number
of different high levels [19] independently of the substrate.
Both schools are addressing the same basic experimental facts, and the differences
seem rooted in concepts and beliefs developed three decades ago. During the last
decades, brain imaging technologies like EEG, MEG, PET and fMRI have made it
possible to study the behaviour of the brain in time and space, discovering a wide
range of brain patterns connected with all sorts of conditions from vegetative states, via
various forms of unconsciousness, via mental diseases, to normal brains and behaviours.
Moreover, by stimulating the brain with external magnetic pulses - transcranial magnetic
stimulation (TMS) - and recording EEG spectra and/or fMRI images, one can get
information about correlations in time and space [20, 122 -- 127].
Tononi's information integration theory (IIT) [22, 23] treats consciousness as a
global irreducible state of a complex system characterised by a single parameter Φ
describing the information content. The Φ parameter is effectively obtained by averaging
the EEG output from repeated TMS pulses and zip-compressing the output data. Simple
regular patterns characterising unconscious states will have low Φ, while conscious states
with high activity will show high Φ. Koch [27] refers to it as a "consciousness meter".
Maquire et al. have discussed understanding consciousness as data compression [128],
and argued that consciousness may not be computable [129].
IIT is axiomatic and puts itself at the side of Dualism, and there is no bottom-
up computational basis for predictions. A critical analysis of IIT is given by
Aaronson [130], with response by Tononi [131]. Further critical analyses of IIT versus
Functionalism/GNW have been provided by Cerullo [132], Fallon [133, 134] and Bayne
[135].
In contrast, computational neurobiology is based on concrete models of brain
networks, and GNW belongs to this class. Real brain networks have a hierarchical
network structure [19, 20, 67, 136]. GNW posits that local structures are able to access
collective properties of the entire network, accessing the global workspace, and that this
possibly represents consciousness. Like seeing an object (a local visual response) may
13
trigger the recall of a scene, an episode, out of memory.
5.2. Global dynamics and EEG patterns in models of brain networks
Arguably, the only systematic, albeit time consuming, way to make progress is to develop
and test top-down and bottom-up models against each other, in order to converge to a
consistent picture. The top-down models must correctly describe behavioural patterns.
The bottom-up models should describe the corresponding properties of the physical
network. Together they must provide verifiable neural correlates of consciousness
(NCC) [120]. The percolation model [87] is an example of a top-down global model
that captures some essential global properties.
It seems that the most rational approach is to define basic consciousness in terms
of the real-time dynamics of the default mode DMN resting state network, the highest-
level global network, when identified as consciousness in brain imaging. In this state,
the subject is "day-dreaming", thinking about things with all couplings to attention
networks effectively turned off. This is at the core of self-awareness, at the stage where
psychedelics will dissolve the "self". At the extreme, this could perhaps be a person with
completely-locked-in syndrome (CLIS), immersed in thought but unable to communicate
(except via the EEG pattern [137 -- 141]). Since the DMN is a global network, there is no
reason to associate NCC with any particular module or hub - the proper NCC should
be the dynamic global pattern. Nevertheless, features of the DMN dynamics should be
able to read out locally, e.g. at nodes by other networks.
Cognition is quite a different matter, involving attention and action. Cognition is
defined as (i) "the mental action or process of acquiring knowledge and understanding
thought, experience, and the senses" as well as (ii) "a perception, sensation, idea, or
intuition resulting from the process of cognition". The first definition (i) is operational
and tangible, and can be characterised by network activities. The second definition (ii) is
abstract: the concrete manifestations of subjective experience - qualia - must necessarily
be identified via the results of mental action and subsequent physical information
processing.
5.3. Default Mode Network - the key to the self ?
The concept of a top-level Default Mode Network (DMN; see Sect. 4) was introduced
by Reichle et al. in 2001 [143] in order to describe the "introspective mind-wandering"
resting state of the brain as evidenced by fMRI and PET [144].
As discussed before, there are good reasons for regarding the global DMN as the
home of self-consciousness. Ideally it represents the brain and mind of a person totally
disconnected from input from the environment. The effects of psychedelics induce a
deeper introspection and dissolution of the "self" and can be traced to changes in the
connectivity of the DMN [61 -- 64].
The DMN is often characterised as a "task-negative network" [143, 144] because
is requesting attention (like good old
it "shuts down" when an interrupt signal
14
microprocessors). The generic term task-positive network (TPN) is then sometimes
used to refer to a collection of attention and action networks (CEN, SN, CN, SMN)
that are anti-correlating with the DMN: when they "turn" on because of some external
input, they make the DMN "turn off" ("stop thinking"). The command can of course
also come from inside, as a self-conscious decision to pay attention, and even to perform
a task.
Cognition implies interaction with the external environment through sensors
involving touch (SMN), hearing, seeing and smelling. Cognition also involves interaction
with the internal environment, e.g. with memory, through pure contemplation. An
external sound or an internal bright idea will excite the CEN to demand attention and
action via the DMN. The internal environment involves the salience network (SN) and
the cerebellum (CN) expressing perception, emotions, fear, anger. Evolution must have
used it to reward us and to save us from danger [99, 145 -- 149].
To be able to build a network that displays some elementary form of consciousness
that can be identified in operational terms (e.g. EEG pattern), it then seems reasonable
to expect that one "only" needs a minimal model of the brain containing the essential
collective (global) and local network properties. This may have nothing to do with
human consciousness, but that is a different matter. To go on, one can include
elementary versions of major components of the TPN to give the system memory and
some ability of self-monitoring and self-reflection. It may be a formidable task, but it
seems doable.
5.4. Levels and disorders of consciousness
"Level of consciousness (LoC)" is an exceedingly complex concept. It is at the core of
much of the philosophical confusion, but also, and more importantly, is at the focus
of current clinical medical attention [150, 151]. "Disorders of Consciousness (DoC)"
range from coma and vegetative state (VS)/unresponsive wakefulness (UWS) through
minimally conscious states (MCS) to full awareness and cognition. This is then to
be compared with normal states of sleep and states induced by sedation and drugs
(including psychedelics). The MCS-label is problematic, because MCS is not a well-
defined state:
it covers a broad range of responses and needs to be "expanded" in a
multi-dimensional space of brain patterns and behaviours [150 -- 153].
In a recent review in Science: "What is consciousness, and could machines have
it?", Dehaene, Lau, and Kouider [19] discuss how to define different LoCs, and how
artificial systems might perceive things.
In particular Dehaene et al. [19] propose
that the word "consciousness" combines two different types of information-processing
computations in the brain:
information for global broadcasting,
thus making it flexibly available for computation and reporting (C1), and the self-
monitoring of those computations, leading to a subjective sense of certainty or error
(C2). Moreover, Dehaene et al. argue, not surprisingly, that current machines are
still mostly implementing computations that reflect unconscious processing (C0) in the
the selection of
15
human brain.
A recent example could be AlphaGo Zero, the AI that is the new master of the
game of Go [154]. A way to characterise the ability of machines might be to apply the
emerging multi-dimensional knowledge and taxonomy of "Disorders of Consciousness
(DoC)" [150 -- 153] to descibe their behaviour. Like "What can a human do that a
machine can't"? Actually, one might argue that most challenging question might be
"What are intentions, emotions, and common sense, and could machines have them?"
For AlphaGo Zero to reach the C1 level, it must be able to explain how it arrived
at various decisions during a game of Go and report to the world - this needs episodic
memory. And to reach level C2 it must be able to remember its reasoning and critically
analyse its behaviour - this takes a lot of memory and higher network functions. This
memory and higher functions is the crux of the matter, still to be developed and evolved.
5.5. From coma to completely-locked-in syndrome
Seen from the outside, coma/VS/UWS and completely-locked-in syndrome (CLIS) look
similar - no signs of motor response from the patient to any external input. On the
inside, however, VS/UWS [142] looks very different from CLIS [137 -- 141].
A completely-locked-in state (CLIS) is typical for the final state of ALS, but can
also result from stroke. CLIS is characterised by broken connections with the output
circuits, specifically the motor circuits - the person is completely paralysed. ALS is
a complex case of neurodegeneracy [155], with motor neurons and axons dying in the
motor cortex and/or in the spine. After stroke, the muscles and nerves are still there,
but important motor regions of the brain are damaged. Nevertheless, CLIS patients
can hear, which means they can show attention and respond internally "in the mind"
to commands, but not activate the SMN.
If we take the experimentally established small-world-network view, the brain
consists of a number of tightly connected local nodes coupled together to a global
network. Moreover, many of these local nodes are connected to internal and external
sensors for monitoring the environment. This means that they together could produce
a number of collective (global) signals, e.g. a variety if EEG patterns, including neural
correlates for perceptual and attentional states [156]. It also means that perception and
awareness may critically depend on the sensor environment - multisensory perception
[157 -- 160].
5.6. Reading the brain with image processing and machine learning
Neural Correlates of Consciousness (NCC) necessarily involve time dependent brain
patterns, resulting from intrinsic or extrinsic actions and detected with EEG, MEG,
PET, fMRI, fNIRS [138], etc. Structural patterns will underlie it all, but dynamic
patterns are essential for NCC. This is particularly important because it opens the way
for direct communication with persons with severe disabilities [138 -- 140, 142, 164].
16
To make use of the time-dependent EEG spectrum one needs to develop useful
measures to extract the most relevant information. A commonly used measure is spectral
entropy in various forms [87, 140, 161 -- 166]. However, one can anticipate that advanced
methods to extract time-dependent correlations will eventually be developed.
Important progress has been made possible by the recent development in real-
time spectral and image processing using machine learning (ML) based on deep
neural networks [167 -- 169]. Recent examples involve cognitive functioning in fragile-X
syndrome [170], Alzheimer's disease [171], detection of depression [172], and predicting
sex from brain rhythms [173].
Recently Shen et al. [174, 175] performed deep image reconstruction by recording
human brain activity. A picture of an animal was shown to a person and to a camera.
A deep ANN was trained to identify the pixelated camera picture and compare with the
fMRI signal pattern. In the end, the trained ANN was able to display, pixel by pixel,
unknown pictures shown to the human. The accuracy was low, but there was some clear
resemblance.
In related work, Fong et al. [176] used human brain activity to guide a machine
learning algorithm, taking fMRI measurements of human brain activity from subjects
viewing images, and infusing the data into the training process of an object recognition
learning algorithm. After training, the neurally-weighted classifiers were able to classify
images without requiring any additional neural data [176].
It follows that EEG/fMRI readout plus real-time data analysis and image processing
plus deep ML with ANNs may provide amazing opportunities for feedback and control
in both directions: the real brain can train the ANN, and the ANN can train the brain.
Training a human brain could take the form of interaction via the normal senses, as well
as through artificial sensors, including implanted ones. An example is an implanted
electrode array picking up signals from the motor cortex of an ALS patient [177],
replacing part of the brain's sensorimotor system by an artificial one.
17
6. Quantum cognition
Cognition is the (human) mental process of acquiring knowledge and understanding via
thought, experience, and senses.
If consciousness and cognition are global emergent
phenomena in a sufficiently complex biological brain, how can cognition be influenced
by quantum physics? And, would it make any difference, would we even notice?
If one searches for "Quantum consciousness" one gets right into a world of quantum
mysticism, quantum mind and quantum healing. If one instead searches for "Quantum
cognition" one meets an active research field in psychology [178 -- 182], applying quantum
probability theory to human behaviour to model the phenomenology of human decision
making:
incompatibility, and
dependence on order. An alternative view [182] suggests that contextual effects may
be better modelled by allowing non-observable probabilities to take negative values to
describe certain contextual incompatible stochastic processes, rather than describing
them by quantum probabilities.
contextuality (experimental conditions), uncertainty,
Further search along the lines of "Quantum games", "Quantum decisions" and
"Quantum finance" opens up a world of quantum applications for rational decision
making [183 -- 189]. All these fields are variants of optimisation problems, usually
involving Ising-type quantum spin Hamiltonians generalising classical binary problems.
To provide quantum advantage (i.e. exponential speedup) the algorithms must run on
quantum hardware.
6.1. Quantum effects in the brain - would we even notice?
But what about quantum enhanced cognition in biological brains? Arguably, from a
materialist/functionalist point of the potential power of quantum cognition must be
related to a corresponding classical biological model of cognition. One needs to start
from a reasonable model for consciousness and cognition and argue for how physical
quantum effects will improve performance.
There is a variety of
ideas how quantum effects might influence or control
perception, consciousness and cognition (see e.g. [190 -- 193] for discussion and references).
In some scenarios, quantum effects in the brain tend to be linked to local, or even global,
quantum coherent processing and entanglement, but few of them are anchored in realistic
brain models of biology and cognition.
As discussed in Sect. 5, experimentally based realistic network models of the
brain have been developed during the last two-three decades. However,
issues of
neural correlates of consciousness (NCC) have tended to depreciate those models as
reductionist, unable to explain the essential hard questions.
This has now arguably changed for good: the development during the last three to
five years has confirmed basic network models, demonstrated global dynamic brain-like
behaviour and identified reasonable NCCs. We therefore now have concepts and tools
- levels of consciousness (LoC) and disorders of consciousness (DoC) (Sect. 5) - for
18
discussing and assessing in a verifiable manner how quantum effects might influence -
but not create! - consciousness and cognition.
6.2. Biochemical quantum effects on cognition
At the local biochemical molecular level,
it is all quantum. Quantum tunnelling,
interference and entanglement may play decisive roles in biochemical reactions via
intermediate transition-state complexes. Quantum effects underlie protein folding, and
the functioning of enzymes and ribosomes. Malfunction and destruction of neurons and
nerves in ALS [155] might be influenced by biochemical quantum effects. Nevertheless,
these are "trivial quantum effects" in the context of consciousness and cognition.
On the contrary, the balance between two types of serotonine receptors in the
cortex [63]
is perhaps the best example of how biochemical quantum effects can
directly influence consciousness and cognition ("dissolution of the self") by modifying
the functional connectivity in the default mode network (DMN). This can then be
generalised to all sorts of chemical agents and external probes that modify the functional
connectivity of the DMN, globally or locally, perhaps also via coupling to and feedback
from the task-positive attention networks (TPN).
The nature and range of coherence in time and space in receptor complexes is still
an open question (see Jedlicka [190] for a recent review of quantum biology). However,
ideas of macroscopic quantum coherence in brain networks remain pure speculation and
are completely unwarranted.
6.3. Fisher's idea of flying spin qubits
Nevertheless, a recent paper by Fisher [194] presents an interesting proposal for how it
might be possible to influence chemistry in the brain by entangled quantum processes
involving well-protected flying spin qubits. Fisher's approach can be viewed as a
bottom-up approach, developing a concrete mechanism for coherent quantum processing
based on physical long-coherence quantum components, with no need for quantum
error correction. The model explicitly deals with quantum computing in the brain,
encapsulating phosphorous spin qubits in phosphate ions, paired into "Posner molecules"
to provide potentially long-lived spin qubits, entangled pairs, and quantum memory
(many seconds).
Fisher argues that the enzyme-catalysed chemical reactions can entangle spin qubit
pairs, and that the pairs can be transported to synapses and released in such ways that
resulting neurotransmitter cascades are entangled and give rise to non-local correlations.
Fisher takes the capability of modulating the excitability and signalling of neurons as a
working definition of "quantum cognition".
The obvious objection is the same as before: the proposed coherent processes can
possibly locally entangle some components or nodes, but hardly globally entangle the
brain network. It might influence rate coefficients in the classical brain network, but
19
hardly provide anything like quantum computing, or genuine "quantum cognition" that
would not exist in the classical brain.
Even if Fisher's quantum processes do not exist in the brain, it seems technically
possible to build artificial chemical systems that might display such effects - that the
outcome of the chemical reactions depend on entanglement created at some point. This
is similar to the proposed mechanism for magnetoreception [195 -- 197].
6.4. Stapp's view of a quantum brain
Stapp takes a top-down dualist view, arguing for a global quantum brain where
coherence is maintained by mental action [198 -- 200]. Stapp's core argument is that basic
brain components are best understood in quantum terms, and the brain itself must be
treated as a quantum system. This evidently leads to the need for a quantum coherent
brain, and Stapp developed a theory describing the effects of mental action on brain
activity as achieved by a Quantum Zeno Effect that is not weakened by decoherence (see
[192] for an opposing view). Thus, terms having intrinsic mentalistic and/or experiential
content (qualia), like "feeling", "knowing" and "effort", are included as primary causal
factors. Nevertheless, like in all other theories, there is really no functional relation
between the quantum brain network and qualia, modelling the emergence of feelings in
a quantum network.
6.5. Penrose/Hameroff and orchestrated objective reduction
During the 1980s, Stuart Hameroff proposed a classical version of protein-connected
microtubules as classical molecular automata performing basic computational logic of
the living state [201] . In that work, there was no mention of quantum oscillators, and it
was in line with other ideas within the fields of molecular computing and artificial life.
However, the collaboration with Penrose took microtubules to a hypothetical quantum
level. In his 1994 book "Shadows of the Mind", Penrose [202] hypothesised that human
consciousness is non-algorithmic (non-computable) and therefore cannot be simulated
by a universal Turing machine, let alone a classical digital computer.
Penrose chose to connect his view of non-computable consciousness with his views
on the foundation of quantum mechanics: objective reduction (OR) postulates that
quantum gravitational interaction at the Planck scale will inevitably make quantum
wavefunctions collapse physically, and that the collapse is non-computable (but still
physical!). Penrose apparently regarded quantum mechanics to be fundamentally non-
computable, thus providing a tool for understanding a likewise non-computable brain.
In a subsequent 1996 paper [203] Hameroff and Penrose introduced microtubule
networks as quantum coherent networks performing quantum computing. They
the quantum network wavefunction
postulated that
(orchestrated objective
creates an
instantaneous "now" event, and that sequences of such events create a flow of time,
and consciousness [203]. It is interesting to note that the Orch-OR picture does not
reduction, Orch-OR),
self-collapse of
resultant
irreversible
in time,
20
require very long coherence times - "only" in the 100 millisecond range corresponding
to the time frame for classical nervous processes and human cognition. However, in
2000 Tegmark [204] calculated that the coherence time in microtubules would fall in the
picosecond range or shorter (requiring a very fast-thinking brain).
Even if a "warm, wet and noisy" environment can allow longer coherence times,
milliseconds is still a very long time. Moreover, it does not really matter: it all hinges
on Penrose postulating that non-computational quantum mechanical ingredients are
present in nature, and that the wavefunction collapse - not the quantum computation
itself - causes the quality of consciousness. If this were true, which is not likely, there
would necessarily exist some physical mechanism to be found.
Most other models are based on efficient quantum computation in a computable
brain, and then very long coherence times become essential.
6.6. Back to reality - the classical brain
At this point, the still engaged reader may ask why bother arguing about theories
and hypotheses that one anyway rejects? Views that can be regarded as outdated
and/or manifestly unrealistic? The answer is that many of these ideas about quantum
physics, human mind and free will are still highly visible, and often promoted by highly
respected, famous and influential scientists and philosophers. They present topics for
deep and interesting discussions, and the popular books describing the ideas are often
highly rated.
The problem is that they also contribute to creating an artificial conflict between
up-to-date materialism/functionalism and quantum physics.
It is time to develop
narratives that are driven by a genuine interest in presenting modern views on how to
integrate functionalism and quantum physics with consciousness and cognition, rather
than promoting far-out views based on often speculative ways of how to interpret
quantum mechanics. The discussion in the first part of this section hopefully provides
a realistic view of how local quantum effects on the functional connectivity of classical
brain networks may influence consciousness and cognition.
21
7. Quantum neural networks - quantum brains ?
The brain is certainly not a quantum computer, and significant quantum-enhanced
functionality influencing consciousness and cognition in biological brains is very unlikely.
However, small quantum computers and simulators are emerging:
it is now possible
to implement quantum algorithms and protocols on systems with 10-20 qubits [4], and
operational systems with 50-100 qubits will be available in the immediate future [4,205].
Such systems form real-space physical qubit networks, and the gate operations creates
networks in state (Hilbert) space.
The notion of "Quantum brain" refers to quantum neural networks (QNN) and
recurrent QNNs (RQNN). It may be anything from a simple network of connected qubits
to networks of important complexity, e.g. a hierachial distributed network of quantum
nodes that are networks in themselves. The obvious question is: can such quantum
networks mimic a biological brain, combining the strengths of neural processing with the
exponentially larger state space and potential algorithmic speedup of quantum systems
to achieve some form of quantum AI - QAI? A generic answer is "yes, in principle" -
and QNNs are obvious platforms for hybrid analogue-digital QC (ADQ).
Actually, the present and near-future NISQ processors [205] face similar problems
as biological brains: in practice a "noisy, warm and wet" environment. A major problem
for quantum computing is decoherence and loss of entanglement. This means that the
present applications of digital algorithms necessarily can only have low depth and a
rather small number of gates.
In classical CMOS hardware is now possible to physically configure chips to function
as a network of neurons and synapses [49, 51]. One can also design and simulate the
neural network and all of its functions in software (see e.g. Human Brain project -
HBP) [206].
In principle one could imagine a similar quantum software approach if
one has access to a large coherent QC, e.g. simulating quantum perceptrons [207, 208],
and synaptic weights and neurons [209, 210]. However, this may be neither realistic nor
efficient: such large QCs will need error correction, and are for the distant future, if at all.
A digital QC approach may not be particularly useful for brain-like applications. There
are however interesting proposals for building QNN components in hardware [211 -- 214].
7.1. Quantum reservoir computing (QRC)
4 we described reservoir computing (RC) with recurrent classical neural
In Sect.
networks as a very promising approach to modelling the dynamics and functionality
of components of a brain. As discussed by Enel et al. [92], mixed selectivity in cortical
neurons [119], representing a variety of situations defined by current sensor inputs, is
readily obtained in randomly connected RNNs. In this context, these reservoir networks
reproduce the highly recurrent nature of local cortical connectivity [92].
It is then natural to investigate the computational power of complex dynamical
quantum systems in a similar way using recurrent quantum networks - RQNN. Quantum
reservoir computing (QRC) might be particularly useful for quantum information
22
processing because loss of memory - fading memory - is a key RC resource, and there
is no need for global control of the reservoir network. Present and emerging quantum
hardware with limited coherence should be able to serve as a fading-memory reservoir
for QRC.
The field is just opening up, and there are so far only a few papers on the topic.
Fujii and Nakajima [215] have performed numerical QRC experiments that show that
quantum systems consisting of 5-7 qubits possess computational capabilities comparable
to conventional RNNs with 100-500 nodes. Nakajima et al. [216] introduce a spatial
multiplexing scheme to boost the computational power via multiple different quantum
systems driven by common input streams in parallel. Kutvonen et al. [217] study
memory capacity and accuracy of a QRC based on the fully connected transverse field
Ising model. They show that variation in inter-spin interactions leads to a better memory
capacity in general, by engineering the type of interactions the capacity can be greatly
enhanced and there exists an optimal timescale at which the capacity is maximised.
7.2. Machine learning in feed-forward and recurrent QNNs
There is a huge number of recent papers using machine learning (ML) to train artificial
neural networks to implement algorithms and protocols, e.g. for classifying data. There
is an excellent review article by Biamonte et al. [218] which, together with a number
of recent research papers on Hopfield recurrent QNNs [219, 220], quantum Boltzmann
machines [221,222], quantum circuit QNNs [223,224], and quantum agents and artificial
intelligence [225 -- 230], covers most of the field.
Here we will only briefly discuss two papers [223, 224] that promise applications for
"Noisy Intermediate-Scale Quantum" (NISQ) processors [205] by generating QNNs as
sequences of quantum circuit gates implemented in a "standard" quantum computer.
Farhi and Neven [224] construct the quantum circuit as a sequence of parameter
dependent unitary transformations which acts on an input quantum state.
For
classification of binary data, the input quantum state is an n-bit computational basis
state corresponding to a sample string. Farhi and Neven [224] show how to design a
circuit made from two qubit unitaries that can correctly represent the label (effectively
parity) of any Boolean function of n bits. Through classical simulation the parameters
can be found that allow the QNN to learn to correctly distinguish the two data sets.
Farhi and Neven [224] anticipate that it will be possible to run this QNN on a near
term gate model quantum computer where its power can be explored beyond what can
be explored with simulation.
Mitarai et al. [223] propose a hybrid classical-quantum algorithm for machine
learning - quantum circuit learning (QCL) - that they suggest can provide an alternative
to the Variational Quantum Eigensolver (VQE) and the Quantum Approximate
Optimization Algorithm (QAOA). QCL and QRC both pass the central optimisation
procedure to a classical computer. In QRC, the output is defined by a set of observables
taken from quantum many-body dynamics driven with a fixed Hamiltonian, and the
23
weight vector is tuned on a classical device to minimise the cost function. In contrast,
QCL tunes the whole network and can be regarded as a QNN with tunable couplings,
the quantum counterpart of a basic neural network.
24
8. Perspectives
Thomas Nagel's book "Why the Materialist Neo-Darwinian Conception of Nature is
Almost Certainly False" [32] played a pivotal role in forming my view of how philosophers
underestimate the power of science, especially classical physics. The view in that book
is that even the modern materialist approach to life fails to explain consciousness,
intentionality, meaning, and value. Therefore, according to Nagel, this threatens
the entire naturalistic world picture, extending to biology, evolutionary theory, and
cosmology. The simple fact is, of course, that we have no idea of whether modern
materialist approaches to life fail before they provably fail. And that may take a long
time and lots of patience to find out.
Or maybe not:
the recent experimental and modelling progress to understand
the brain's global and local network structure and dynamics is amazing: the "hard
problem of consciousness" has dissolved, and the classical complexity of the brain
is huge but not overwhelming. There is no reason to consider quantum complexity:
quantum effects are basically biochemical, able to influence both the local and global
functional connectivity of brain networks. This can evidently influence consciousness,
cognition and motor responses, but has nothing to do with a "quantum mind". An
individual's subjective experience of "value, intention and meaning" is the result of
internal classical information processing of interacting networks, comparing the various
physical experiences of that individual. Functional network disorder can then obviously
create highly personal views of "value, intention and meaning".
It is interesting to note that the most visible actors have often been people in
quantum physics of particles and fields who have a strong interest in the "physics" of
the mind, or people in biology, medicine and philosophy with a strong interest in esoteric
concepts in quantum physics, like superposition, coherence, entanglement, measurement
and wavefunction collapse. As a result, views of quantum physics have often bordered
to quantum mysticism, and been used to support panpsychism and a universal mind
based on tentative quantum processes in biological brain components.
The optimistic conclusion of this paper is that we may soon be facing a paradigm
shift where some basic mechanisms of consciousness and cognition can be efficiently
described by classical analogue and digital systems. The problem will then be to
understand the spectrum of consciousness and self-awareness, from bacteria and C.
elegans via Lucy and modern man, to future AIs that report experiences and emotions.
At the present level of knowledge in neurobiology, quantum processes probably
play no significant role for cognition. However, this can change in future artificial
cognitive systems involving quantum circuits, with quantum agents living in quantum
networks and demonstrating free will [225]. Nevertheless, even these powerful quantum
agents [225] will still have limited power - Nature will only solve really hard problems
approximately, and quantum computers may not always have any advantage.
There is an old Swedish proverb: "Nar fan blir gammal blir han religios" ("When
the devil grows old he becomes religious"). Closing in, we tend to challenge established
25
views, sometimes moving to the extreme edges, searching for the meaning of it all. The
radical new insight is that we might in fact be able to understand who we are, given a
bit of time.
For some people, the year is 2023. In a recent article in New Scientist a Swedish
science writer describes an interesting encounter in Europe 1998. Quoting Per Snaprud
[234]: "Twenty years ago this week, two young men sat in a smoky bar in Bremen,
northern Germany. Neuroscientist Christof Koch and philosopher David Chalmers had
spent the day lecturing at a conference about consciousness, and they still had more to
say. After a few drinks, Koch suggested a wager. He bet a case of fine wine that within
the next 25 years someone would discover a specific signature of consciousness in the
brain. Chalmers said it wouldn't happen, and bet against."
The core of the bet is to demonstrate a convincing neural correlate of consciousness
(NCC) by 2023. For Koch it would be an experimental "consciousness meter",
providing convincing response seen in EEG brain waves after zapping the brain with an
electromagnetic pulse [26, 27].
The recent development, as described in Sect. 5, suggests that David Chalmers will
have to buy the wine. However, they should drink the wine together, and invite Stanislas
Dehaene to join, because Christof Koch will not really win. The delocalised global
default mode network (DMN) is verifiably the best model for the home of consciousness,
and the NCC is given by the dynamics of the DMN resting state as manifested in the
information content (entropy) of the EEG spectrum. In practice, the NCC is already
used clinically to map disorders of consciousness and to create methods for two-way
communication, training and control. This will probably lead to a quite dramatic
development during the next decades, assisted by a parallel development of machine
learning and AI. However, one can expect exciting progress already in the near future -
there will be good reasons for Koch and Chalmers to share a case of wine in 2023.
26
Acknowledgement
This work has been partially supported by the Knut & Alice Wallenberg Foundation
(KAW) (WACQT project), and by the EU Quantum Technologies Flagship (820363 -
QpenSuperQ).
27
References
[1] S. Aaronson, NP-complete problems and physical reality?, ACM SIGACT News 2005, 36, 30.
[2] S. Aaronson, Why philosophers should care about computational complexity, in Computability:
Godel, Turing, Church, and Beyond (Eds. B. Copeland, C. J. Posy, O. Shagrir), MIT Press,
Cambridge, MA 2013, pp. 261-328.
[3] H. Broersma, S. Stepney, G. Wendin, Computability and Complexity of Unconventional
Computing Devices, in Computational Matter, (Eds. S. Stepney, S. Rasmussen, M. Amos),
Springer 2018; arXiv:1702.02980v2.
[4] G. Wendin, Quantum information processing with Superconducting circuits: a review, Rep. Prog.
Phys. 2017, 80, 106001; arXiv: 1610.02208v2.
[5] D. J. Chalmers, Facing up to the problem of consciousness, J. Consciousness Studies 2017, 2, 200.
[6] P. Bak, How Nature Works: The Science of Self-Organised Criticality, New York, NY: Copernicus
Press 1996.
[7] E. Bullmore, A. Barnes, D. S. Bassett, A. Fornito, M. Kitzbichler, D. Meunier, J. Suckling, Generic
aspects of complexity in brain imaging data and other biological systems, NeuroImage 2009,
47, 1125.
[8] D. S. Bassett, M. S. Gazzaniga, Understanding complexity in the human brain, Trends in Cognitive
Sciences 2009, 15, 200.
[9] D. R. Chialvo, Emergent complex neural dynamics, Nature Phys 2010, 6, 744.
[10] E. Tagliazucchi, D. R. Chialvo, M. Siniatchkin, E. Amico, J.-F. Brichant, V. Bonhomme, Q.
Noirhomme, H. Laufs, S. Laureys, Large-scale signatures of unconsciousness are consistent with
a departure from critical dynamics, J. R. Soc. Interface 2016, 13, 20151027.
[11] E. Tagliazucchi, The signatures of conscious access and its phenomenology are consistent with
large-scale brain communication at criticality, Consciousness and Cognition 2017, 55, 136.
[12] B. M. Lake, T. D. Ullman, J. B. Tenenbaum, S. J. Gershman, Building machines that learn and
think like people, Behavioral and Brain Sciences 2017, 40, e253.
[13] L. Cocchi, L. L. Gollo, A. Zalesky, M. Breakspear, Criticality in the brain: A synthesis of
neurobiology, models and cognition, Progress in Neurobiology 2017, 158, 132.
[14] W. Maass, Searching for principles of brain computation, Current Opinion in Behavioral
Sciences 2016, 11, 81.
[15] J.-P. Changeux, S. Dehaene, Neuronal models of cognitive functions, Cognition 1989, 33, 63.
[16] S. Dehaene, L. Naccache, Toward a cognitive neuroscience of consciousness: basic evidence and a
workspace framework, Cognition 2001, 79, 1.
[17] S. Dehaene, J.-P. Changeux, Experimental and Theoretical Approaches to Conscious Processing,
Neuron, 2011, 70, 200.
[18] S. Dehaene, Consciousness and the Brain: Deciphering How the Brain Codes Our Thoughts, 2014
Penguin Publishing Group.
[19] S. Dehaene, H. Lau, S. Kouider, What is consciousness, and could machines have it?, Science
2017, 358, 486.
[20] P. Barttfeld, L. Uhrig, J. D. Sitt, M. Sigman, B. Jarraya, S. Dehaene, Signature of consciousness
in the dynamics of resting-state brain activity, PNAS, 2015, 112, 887.
[21] C. Koch, K. Hepp, Quantum mechanics in the brain, Nature 2006, 440, 611.
[22] G. Tononi, An information integration theory of consciousness, BMC Neurosci. 2004, 5, 42.
[23] G. Tononi, Consciousness as integrated information: a provisional manifesto, Biol. Bull. 2008,
215, 216.
[24] A. G. Casali, O. Gosseries, M. Rosanova, M. Boly, S. Sarasso, K. R. Casali, S. Casarotto, M.-
A. Bruno, S. Laureys, G. Tononi, M. Massimini,, A theoretically based index of consciousness
independent of sensory processing and behavior, Sci. Transl. Med. 2013, 5, 198ra105.
[25] G. Tononi, M. Boly, M. Massimini, C. Koch, Integrated information theory: from consciousness
to its physical substrate, Nature Reviews: Neuroscience 2016, 17, 450.
28
[26] C. Koch, How to make a consciousness meter, Sci. Am., November 2017.
[27] C. Koch, What Is Consciousness?, Nature 2018, 557, S9.
[28] D. C. Dennett, Consciousness Explained, Little, Brown and Co. 1991.
[29] D. C. Dennett, From Bacteria to Bach and Back: The Evolution of Minds, W. W. Norton &
Company, New York/London 2017.
[30] D. C. Dennett, Facing up to the hard question of consciousness, Phil. Trans. R. Soc. B 2018, 37,
20170342.
[31] T. Nagel, What is it like to be a bat?, Philos. Rev. 1974, 83, 435.
[32] T. Nagel, Mind and Cosmos: Why the Materialist Neo-Darwinian Conception of Nature is Almost
Certainly False, Oxford University Press 2012.
[33] T. Nagel, Is Consciousness an Illusion?, The New York Review of Books, March 9, 2017.
[34] Computational Matter, (Eds. S. Stepney, S. Rasmussen, M. Amos), Springer 2018.
[35] L. M. Adleman, Molecular computation of solutions to combinatorial problems, Science 1994,
266, 1021.
[36] H. Siegelmann, Computation beyond the Turing limit, Science 1995, 268, 545.
[37] H. T. Siegelmann, E. Sonntag, On the Computational Power of Neural Nets, J. Computer and
System Sciences 1995, 50, 132.
[38] J. Cabessa and H. T. Siegelmann, Evolving recurrent neural networks are super-Turing?, Proc.
IJCNN 2011, IEEE 2011, 3200.
[39] M. Di Ventra, F. L. Traversa, Perspective: Memcomputing: Leveraging memory and physics to
compute efficiently, J. Appl. Phys. 2018, 123, 180901.
[40] Y. R. Pei, F. L. Traversa, M. Di Ventra, On the Universality of Memcomputing
IEEE Transactions on Neural Networks and Learning Systems 2018; DOI:
Machines,
10.1109/TNNLS.2018.2872676.
[41] H. Manukian, F. L. Traversa, M. Di Ventra, Accelerating Deep Learning with Memcomputing,
Neural Networks 2019 110, 1.
[42] F. Sheldon, P. Cicotti, F. L. Traversa, M. Di Ventra, Stress-testing memcomputing on hard
combinatorial optimization problems (2018); arXiv:1807.00107v1
[43] F. L. Traversa, P. Cicotti, F. Sheldon, M. Di Ventra, Evidence of Exponential Speed-Up in the
Solution of Hard Optimization Problems, Complexity 2018, 2018, 7982851
[44] D. Sussillo, Neural circuits as computational dynamical systems, Current Opinion in Neurobiology
2014, 25, 156.
[45] G. S. Carmantini, P. beim Graben, M. Desroches and S. Rodrigues, Turing Computation with
Recurrent Artificial Neural Networks, CoCoNIPS, 2015; arXiv:1511.01427.
[46] A. Vergis, K. Steiglitz, B. Dickinson, The complexity of analog computation, Mathematics and
Computers in Simulation, Mathematics and Computers in Simulation 1986, 28, 91.
[47] F. Denef, M. R. Douglas, Computational complexity of the landscape: Part I, Annals of Physics
2007, 322, 1096.
[48] K. Douglas, Learning to Hypercompute? An Analysis of Siegelmann Networks, in Computing
Nature: Turing Centenary Perspective (Eds. G. Dogic-Crnkovic, R. Giovagnoli), Springer 2013,
pp. 201-211.
[49] P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada, F. Akopyan, B. L. Jackson,
N. Imam, C. Guo, Y. Nakamura, B. Brezzo, I. Vo, S. K. Esser, R. Appuswamy, B. Taba, A.
Amir, M. D. Flickner, W. P. Risk, R. Manohar, D. S. Modha, A million spiking-neuron integrated
circuit with a scalable communication network and interface, Science 2014, 345, 668.
[50] S. K. Esser, P. A. Merolla, J. V. Arthur, A. S. Cassidy, R. Appuswamy, A. Andreopoulos, D.
J. Berg, J. L. McKinstry, T. Melano, D. R. Barch, C. di Nolfo, P. Datta, A. Amir, B. Taba,
M. D. Flickner, D. S. Modha, Convolutional networks for fast, energy-efficient neuromorphic
computing, PNAS 2016, 113, 11441.
[51] W. Maass, Energy-efficient neural network chips approach human recognition capabilities, PNAS
2016, 113, 11387.
29
[52] A. Amir, B. Taba, D. Berg, T. Melano, J. McKinstry, C. Di Nolfo, T. Nayak, A. Andreopoulos,
G. Garreau, M. Mendoza, J. Kusnitz, M. Debole, S. Esser, T. Delbruck, M. Flickner, D.
Modha, A Low Power, Fully Event-Based Gesture Recognition System, 2017 IEEE Conference
on Computer Vision and Pattern Recognition 2017.
[53] A. Andreopoulos, H. J. Kashyap, T. K. Nayak, A. Amir, M. D. Flickner, A Low Power, High
Throughput, Fully Event-Based Stereo System, The IEEE Conference on Computer Vision and
Pattern Recognition (CVPR) 2018, pp. 7532-7542.
[54] S. Ambrogio, P. Narayanan, H. Tsai, R. M. Shelby, I. Boybat, C. di Nolfo, S. Sidler, M. Giordano,
M. Bodini, N. C. P. Farinha, B. Killeen, C. Cheng, Y. Jaoudi, G. W. Burr, Equivalent-accuracy
accelerated neural-network training using analogue memory, Nature 2018, 558, 60.
[55] P. M. Churchland, Functionalism at Forty: A Critical Retrospective, J. Philosophy 2005, 102, 33.
[56] Stanford Encyclopedia of Philosophy; https://plato.stanford.edu/entries/dualism/
[57] D. C. Dennett, Quantum Incoherence, Nature 1996, 381, 486. Review of A. G. Cairns-Smith,
Evolving the Mind: on the nature of matter and the origin of consciousness, Cambridge Univ.
Press, 1996.
[58] M. Chorost, Where Thomas Nagel Went Wrong, The Chronicle of Higher Education 2013, May
13.
[59] D. Papineau, Competence without comprehension: The peculiar philosophical assumptions of
Daniel Dennett, The Times Literary Supplement 2017, June 28.
[60] R. Landauer, Information is Physical, Physics Today 1991, 44, 23.
[61] R. L. Carhart-Harris, R. Leech, P. J. Hellyer, M. Shanahan, A. Feilding, E. Tagliazucchi, D.
R.Chialvo, D. Nutt, The entropic brain: a theory of conscious states informed by neuroimaging
research with psychedelic drugs, Frontiers in Human Neuroscience 2014, 8, 1.
[62] R. L. Carhart-Harris, L. Roseman, M. Bolstridge, L. Demetriou, J. N. Pannekoek, M. B. Wall, M.
Tanner, M. Kaelen, J. McGonigle, K. Murphy, R. Leech, H. V. Curran, D. J. Nutt, Psilocybin
for treatment-resistant depression: fMRI-measured brain mechanisms, Sci. Rep. 2017, 7, 13187.
[63] R. L. Carhart-Harris, D. J. Nutt, Serotonin and brain function: a tale of two receptors, Journal
of Psychopharmacology 2017, 731,1091.
[64] F. Palhano-Fontes, K. C. Andrade, L. F. Tofoli, A. C. Santos, J. A. S. Crippa, J. E. C. Hallak, S.
Ribeiro, D. B. de Araujo, The psychedelic state induced by ayahuasca modulates the activity
and connectivity of the default mode network, PLoS ONE 2015, 10, e0118143.
[65] T. M. McKenna, T. A. McMullen, M. F. Shlesinger, The brain as a dynamic physical system,
Neuroscience Neuroscience 1994, 60, 587.
[66] D. J. Watts, S. H. Strogatz, Collective dynamics of "small-world" networks, Nature 1998, 393,
440.
[67] D. Meunier, R. Lambiotte, E. T. Bullmore, Modular hierarchically modular organization of brain
networks, Front. Neurosci. 2010, 4, 200.
[68] S. L. Bressler, V. Menon, Large-scale brain networks in cognition:
emerging methods and
principles, Trends Cognit. Sci. 2010, 14, 277.
[69] V. Menon, Large-scale brain networks and psychopathology: a unifying triple network model,
Trends Cognit. Sci. 2011, 15, 483.
[70] J. R. Andrews-Hanna, The Brain's Default Network and Its Adaptive Role in Internal Mentation,
The Neuroscientist 2012, 18, 251.
[71] J. R. Andrews-Hanna, J. Smallwood, R. N. Spreng, The default network and self-generated
thought: component processes, dynamic control, and clinical relevance, Ann. N.Y. Acad. Sci.
2014, 1316, 29.
[72] W. H. Lee and S. Frangou, Linking functional connectivity and dynamic properties of resting-state
networks, Sci. Rep. 2017, 7, 16610
[73] P. Qin, X. Wu, Z. Huang, N. W. Duncan, W. Tang, A. Wolff, J. Hu, L. Gao, Y. Jin, X. Wu, J.
Zhang, L. Lu, C. Wu, X. Qu, Y. Mao, X. Weng, J. Zhang, G. Northoff, How are different neural
networks related to consciousness? Ann Neurol. 2015, 78, 594.
30
[74] S. Marek, J. S.Siegel, Evan M.Gordon, R. V. Raut, C. Gratton, D. J. Newbold, M. Ortega, T.
O. Laumann, B. Adeyemo, D. B.Miller, A. Zheng, K. C.Lopez, J. J. Berg, R. S.Coalson, A.
L. Nguyen, D. Dierker, A. N.Van, C. R. Hoyt, N. U. F. Dosenbach, Spatial and Temporal
Organization of the Individual Human Cerebellum, Neuron 2018, 100, 977.e7.
[75] R. A. Berman, S. J. Gotts, H. M. McAdams, D. Greenstein, F. Lalonde, L. Clasen, R. E.
Watsky, L. Shora, A. E. Ordonez, A. Raznahan, A. Martin, N. Gogtay, J. Rapoport, Disrupted
sensorimotor and social-cognitive networks underlie symptoms in childhood-onset schizophrenia,
BRAIN 2016, 139, 276
[76] N. D. Lant, L. E. Gonzalez-Lara, A. M. Owen, Davinia Fern´andez-Espejo, Relationship between the
anterior forebrain mesocircuit and the default mode network in the structural bases of disorders
of consciousness, NeuroImage: Clinical 2016, 10, 27
[77] M. Martino, P. Magioncalda, Z. Huang, B. Conio, N. Piaggio, N. W. Duncan, G. Rocchi, A.
Escelsior, V. Marozzi, A. Wolff, M. Inglese, M. Amore, G. Northoff, Contrasting variability
patterns in the default mode and sensorimotor networks balance in bipolar depression and
mania, PNAS 2016, 113, 4824
[78] B. J. Baars, A cognitive theory of consciousness, Cambridge University Press, New York 1988.
[79] B. J. Baars, The conscious access hypothesis: Origins and recent evidence. Trends in Cognitive
Science 2002, 6, 47.
[80] B. Baars, Global workspace theory of consciousness: Towards a cognitive neuroscience of human
experience? Progress in Brain Research 2005, 150, 45.
[81] M. Breakspear, Dynamic models of large-scale brain activity, Nature Neuroscience 2017, 20, 340.
[82] E. Olbrich, P. Achermann, T. Wennekers, The sleeping brain as a complex system, Phil. Trans.
R. Soc. A 2011, 369, 3697.
[83] G. Werner, Metastability, Criticality and Phase Transitions in brain and its models, Biosystems
2007, 90, 496.
[84] M. G. Kitzbichler, M. L. Smith, S. R. Christensen, E. Bullmore, Broadband Criticality of Human
Brain Network Synchronization. PLoS Comput. Biol. 2009, 5, e1000314.
[85] M. G. Kitzbichler, E. T. Bullmore, Power Law Scaling in Human and Empty Room MEG
Recordings, PLoS Comput. Biol. 2015, 11, e1004175.
[86] J. M.Beggs, N. Timme, Being critical of criticality in the brain, Frontiers in Physiology 2012, 3,
163.
[87] D. W. Zhou, D. D. Mowrey, P. Tang, Y. Xu, Percolation Model of Sensory Transmission and Loss
of Consciousness Under General Anesthesia, Phys. Rev. Lett. 2015, 115, 108103.
[88] D. Papo, J. Goni, J. M. Buldu, Editorial: On the relation of dynamics and structure in brain
networks, CHAOS 2017, 27, 047201.
[89] F. S. Racz, Peter Mukli, Zoltan Nagy, Andras Eke, Multifractal dynamics of resting-state functional
connectivity in the prefrontal cortex Physiol. Meas. 2018, 39, 024003.
[90] G. Werner, Fractals in the nervous system: conceptual implications for theoretical neuroscience,
Frontiers in Physiology 2010, 1, 15.
[91] M. E. J. Newman, Modularity and community structure in networks, PNAS 2006, 103, 8577.
[92] P. Enel, E. Procyk, R. Quilodran, P. F. Dominey, Reservoir Computing Properties of Neural
Dynamics in Prefrontal Cortex, PLoS Comput. Biol. 2016, 12, e1004967.
[93] N. Rodriguez, E.
Izquierdo, Y.-Y. Ahn, Optimal modularity and memory capacity of
neural networks, Network Neuroscience (MIT) 2019; https://doi.org/10.1162/netn a 00082;
arXiv:1706.06511v2.
[94] P. N. Taylor, Y. Wang, M. Kaiser, Within brain area tractography suggests local modularity using
high resolution connectomics, Sci. Rep. 2017, 7, 39859.
[95] S. Grossberg, Towards solving the hard problem of consciousness: The varieties of brain resonances
and the conscious experiences that they support, Neural Networks 2017, 87, 38.
[96] C. Seguin, Martijn P. van den Heuvel, A. Zalesky, Navigation of brain networks, PNAS 2018, 115,
6297.
31
[97] M. P. van den Heuvel, A. Fornito, Brain Networks in Schizophrenia, Neuropsychol Rev. 2014, 24
32.
[98] L. Witter, C. I. De Zeeuw, Regional
functionality of the cerebellum, Current Opinion in
Neurobiology 2015, 33, 150.
[99] M. J. Wagner, T. H. Kim, J. Savall, M. J. Schnitzer, L. Luo, Cerebellar granule cells encode the
expectation of reward, Nature 2017, 544, 96.
[100] E. Tognoli, J. A. Scott Kelso, The Metastable Brain, Neuron 2014, 81, 35.
[101] C. Rossert, P. Dean, J. Porrill, At the Edge of Chaos: How Cerebellar Granular Layer Network
Dynamics Can Provide the Basis for Temporal Filters, PLoS PLoS Comput. Biol, 2015, 11,
e1004515.
[102] N. Stepp, D. Plenz, N. Srinivasa, Synaptic Plasticity Enables Adaptive Self-Tuning Critical
Networks, PLoS Comput. Biol. 2015, 11, e1004043.
[103] J. S. A. Kauffman, S. Johnsen, Coevolution to the edge of chaos: Coupled fitness landscapes,
poised states, and coevolutionary avalanches, J. Theor. Biology 1991, 149, 467.
[104] N. W. Watkins, G. Pruessner, S. C. Chapman, N. B. Crosby, H. J. Jensen, 25 Years of Self-
organized Criticality: Concepts and Controversies, Space Sci. Rev. 2016, 198, 3.
[105] A. Roli, M. Villani, A. Filisetti, R. Serra, Dynamical Criticality: Overview and Open Questions,
J. Syst. Sci. Complex. 2018, 31, 647.
[106] B. Schrauwen, L. Busing, R. Legenstein, On Computational Power and the Order-Chaos Phase
Transition in Reservoir Computing, Advances in Neural Information Processing Systems t21
(NIPS 2008) 2008.
[107] H. Jaeger, The "echo state" approach to analysing and training recurrent neural networks. GMD
Report 148, German National Research Institute for Computer Science 2001.
[108] M. Lukosevicius, H. Jaeger, Reservoir computing approaches to recurrent neural network
trainings, Computer Science Review 2009, 3, 127.
[109] W. Maass, W. Liquid state machines: motivation, theory, and applications, Computability in
context: computation and logic in the real world. in Computability in context: computation and
logic in the real world, (Eds: B. Cooper, A. Sorbi), World Scientific 2009, pp. 275-296.
[110] T. Yamazaki, S. Tanaka, The cerebellum as a liquid state machine, Neural Networks 2007, 20
290.
[111] Z. Konkoli, On developing theory of reservoir computing for sensing applications: the state
weaving environment echo tracker (SWEET) algorithm, Int. J. Parallel, Emergent Distrib. 2018,
33, 121.
[112] Z. Konkoli, S. Nichele, M. Dale, S. Stepney, Reservoir Computing with Computational Matter, in
Computational Matter, (Eds: S. Stepney, S. Rasmussen, M. Amos), Springer 2018, pp. 269-293.
[113] M. Inubushi, K. Yoshimura, Reservoir Computing Beyond Memory-Nonlinearity Trade-off, Sci.
Rep. 2017, 7, 1019.
[114] H. O. Sillin, R. Aguilera, H.-H. Shieh, A. V. Avizienis, M. Aono, A. Z. Stieg, J. K. Gimzewski,
A theoretical and experimental study of neuromorphic atomic switch networks for reservoir
computing, Nanotechnology 2013, 24, 3840.
[115] E. C. Demis, R. Aguilera, H. O. Sillin, K. Scharnhorst, E. J. Sandouk, M. Aono, A. Z. Stieg,
J. K. Gimzewski, Atomic switch networks - nanoarchitectonic design of a complex system for
natural computing, Nanotechnology 2015, 26, 204003.
[116] G. Van der Sande, D. Brunner, M. C. Soriano Advances in photonic reservoir computing,
Nanophotonics 2017, 6, 561.
[117] C. Du, F. Cai, M. A. Zidan, W. Ma, S. Hwan Lee, W. D. Lu, Reservoir computing using dynamic
memristors for temporal information processing, Nature Commun. 2017, 8, 2204.
[118] S. Pecqueur, M. Mastropasqua Talamo, David Gu´erin, P. Blanchard, J. Roncali, D. Vuillaume,
F. Alibart, Neuromorphic Time-Dependent Pattern Classification with Organic Electrochemical
Transistor Arrays, Adv. Electron. Mater. 2018, 4, 1800166.
[119] S. Fusi, E. K. Miller, M. Rigotti, Why neurons mix: high dimensionality for higher cognition,
32
Current Opinion in Neurobiology 2016, 37, 66.
[120] S. B.Fink, A Deeper Look at the "Neural Correlate of Consciousness", Front. Psychol. 2016, 7,
1044.
[121] M. Tegmark, Consciousness as a state of matter, Chaos, Solitons & Fractals 2015, 76, 238..
[122] S. Bestmann, E. Feredoes, Combined neurostimulation and neuroimaging in cognitive
neuroscience: past, present, and future, Ann. N.Y. Acad. Sci 2013, 1296, 11.
[123] S. Sarasso, M. Rosanova, A. G. Casali, S. Casarotto, M. Fecchio, M. Boly, O. Gosseries, G. Tononi,
S. Laureys, M. Massimini, Quantifying Cortical EEG Responses to TMS in (Un)consciousness,
Clin. EEG Neurosci. 2014, 5, 40.
[124] G. P. Hallam, C. Whitney, M. Hymers, A. D. Gouws, E. Jefferies, Charting the effects of TMS with
fMRI: Modulation of cortical recruitment within the distributed network supporting semantic
control, Neuropsychologia 2016, 93, 40.
[125] C. Hawco, J. L. Armony, Z. J. Daskalakis, M. T. Berlim, M. M. Chakravarty, G. B. Pike, M.
Lepage, Differing Time of Onset of Concurrent TMS-fMRI during Associative Memory Encoding:
A Measure of Dynamic Connectivity, Front. Hum. Neurosci. 2017, 11, 404.
[126] J. Saari, E. Kallioniemi, M. Tarvainen, P. Julkunen, Oscillatory TMS-EEG-Responses as a
Measure of the Cortical Excitability Threshold, IEEE Trans. Neural Syst. Rehab. Eng. 2018,
26, 383.
[127] J. Leitao, A. Thielscher, J. Tuennerhoff, U. Noppeney, Comparing TMS perturbations to occipital
and parietal cortices in concurrent TMS-fMRI studies: Methodological considerations, PLoS
ONE 2017, 12, e0181438.
[128] P. Maguire, P. Moser, R. Maguire, Understanding Consciousness as Data Compression, Journal
of Cognitive Science 2016 17, 63.
[129] P. Maguire, P. Moser, R. Maguire, V. Griffith, Is consciousness computable? Quantifying
integrated information using algorithmic information theory, in Proceedings of the 36th Annual
Conference of the Cognitive Science Society, 2615-2620. Austin, TX, (Eds. P. Bello, M. Guarini,
M. McShane, B. Scassellati), Cognitive Science Society, 2014.
[130] S. Aaronson, Why I Am Not An Integrated Information Theorist (or, The Unconscious Expander),
Shtetl-Optimized 2014; https://www.scottaaronson.com/blog/?p=1799.
[131] G. Tononi 2014; https://www.scottaaronson.com/blog/?p=1823
[132] M. A. Cerullo, The Problem with Phi: A Critique of Integrated Information Theory, PloS Comput.
Biol. 2015, 11, e1004286.
[133] F. Fallon, Integrated Information Theory of Consciousness, Internet Encyclopedia of Philosophy
2016; http://www.iep.utm.edu/int-info/.
[134] F. Fallon, Integrated Information, The Routledge Handbook of Consciousness (Ed. R. J. Gennaro),
Taylor and Francis 2018.
[135] T. Bayne, On the axiomatic foundations of the integrated information theory of consciousness,
Neuroscience of Consciousness 2018, 4, niy007.
[136] K. Hilger, M. Ekman, C. J. Fiebach, U. Basten, Intelligence is associated with the modular
structure of intrinsic brain networks, Sci. Rep. 2017, 7, 16088.
[137] A. Ramos Murguialday, J. Hill, M. Bensch, S. Martens, S. Halder, F. Nijboer, B. Schoelkopf, N.
Birbaumer, A. Gharabaghi, Transition from the locked in to the completely locked-in state: A
physiological analysis, Clinical Neurophysiology 2011, 122, 925.
[138] U. Chaudhary, B. Xia, S. Silvoni, L. G. Cohen, N. Birbaumer, Brain-Computer Interface-Based
Communication in the Completely Locked-In State, PLoS Biol 2017, 15, e1002593.
[139] C. Guger, R. Spataro, B. Z. Allison, A. Heilinger, R. Ortner, W. Cho, V. La Bella, Complete
Locked-in and Locked-in Patients: Command Following Assessment and Communication with
Vibro-Tactile P300 and Motor Imagery Brain-Computer Interface Tools, Front. Neurosci. 2017
11, 251.
[140] D. Lesenfants, D. Habbal, C. Chatelle, A. Soddu, S. Laureys, Q. Noirhomme, Toward an
Attention-Based Diagnostic Tool for Patients With Locked-in Syndrome, Clinical EEG and
33
Neuroscience 2018, 49, 122.
[141] B. Rohaut, F. Raimondo, D. Galanaud, M. Valente, J. D. Sitt, L. Naccache, Probing consciousness
in a sensory-disconnected paralyzed patient, Brain Injury 2017, 31, 1398.
[142] D. Espejo, S. Rossit, A. M. Owen, A Thalamocortical Mechanism for the Absence of Overt Motor
Behavior in Covertly Aware Patients, JAMA Neurol. 2015, 72, 1442.
[143] M. E. Raichle, A. M. MacLeod, A. Z. Snyder, W. J. Powers, D. A. Gusnard, G. L. Shulman, A
default mode of brain function, PNAS, 2001, 98, 676.
[144] M. E. Raichle, A. Z. Snyder, A default mode of brain function: A brief history of an evolving
idea, NeuroImage, 2007, 37, 1083.
[145] M. Hausknecht, W.-K. Li, M. Mauk, P. Stone, Machine Learning Capabilities of a Simulated
Cerebellum, IEEE Trans. Neural Networks and Leaning Systems 2017, 28, 510.
[146] S. Clausi, C. Iacobacci, M. Lupo, Giusy Olivito, M. Molinari, M. Leggio, The Role of the
Cerebellum in Unconscious and Conscious Processing of Emotions: A Review, Appl. Sci., 2017,
7, 521.
[147] J. J. Fahrenfort, J. van Leeuwen, C. N. L. Olivers, H. Hogendoorn, Perceptual integration without
conscious access, PNAS 2017, 114, 3744.
[148] D. S. Roy,T. Kitamura, T. Okuyama, S. K. Ogawa, C. Sun, Y. Obata, A. Yoshiki, S. Tonegawa,
Distinct Neural Circuits for the Formation and Retrieval of Episodic Memories, Cell 2017, 170,
1000.
[149] E. H. Thompson, K. Kinden Lensjø, M. Braenne Wigestrand, A. Malthe-Sørenssen, T. Hafting,
M. Fyhn, Removal of perineuronal nets disrupts recall of a remote fear memory, PNAS 2018,
115, 607.
[150] C. Sergent, F. Faugeras, B. Rohaut, F. Perrin, M. Valente, C. Tallon-Baudry, L. Cohen, L.
Naccache, Multidimensional cognitive evaluation of patients with disorders of consciousness
using EEG: A proof of concept study, NeuroImage: Clinical 2017, 13, 455.
[151] L. Naccache, Minimally conscious state or cortically mediated state? BRAIN 2018, 141, 949.
[152] T. Bayne, J. Hohwy, A. M. Owen, Are There Levels of Consciousness? Ann. Neurol. 2017, 82,
866.
[153] T. Bayne, J. Hohwy, and A. M. Owen. Reforming the Taxonomy in Disorders of Consciousness,
Ann. Neurol. 2017, 82, 866.
[154] D. Silver, J. Schrittwieser, K. Simonyan, I. Antonoglou, A. Huang, A. Guez, T. Hubert, L. Baker,
M. Lai, A. Bolton, Y. Chen, T. Lillicrap, F. Hui, L. Sifre, G. van den Driessche, T. Graepel, D.
Hassabis, Mastering the game of Go without human knowledge, Nature 2017, 550, 354.
[155] P. Van Damme, W. Robberecht, L. Van Den Bosch, Modelling amyotrophic lateral sclerosis:
progress and possibilities, Disease Models & Mechanisms 2017, 10, 537.
[156] A. List, M. D. Rosenberg, A. Sherman, M. Esterman, Pattern classification of EEG signals reveals
perceptual and attentional states, PLoS ONE 2017, 12, e0176349.
[157] O. Deroy, C. Spence, U. Noppeney, Metacognition in Multisensory Perception, Trends in
Cognitive Sciences 2016, 20, 736..
[158] O. Deroy, Y.-C. Chen, C. Spence, Multisensory constraints on awareness, Phil. Trans. R. Soc. B
2014, 369, 20130207.
[159] C. O'Callaghan, Grades of Multisensory Awareness, Mind & Language 2017, 32, 155.
[160] R. E. Briscoe, Multisensory processing and perceptual consciousness: Part II, Philosophy Compass
2017, 12, e12423.
[161] Z. Liang, Y. Wang, X. Sun, D. Li, L. J. Voss, J. W. Sleigh, S. Hagihira, X. Li, EEG entropy
measures in anesthesia, Front. Comput. Neurosci. 2015, 9,16.
[162] J. Wirsich, A.-L. Giraud, S. Sadaghiani, Concurrent EEG- and fMRI-derived func-
doi:
linked dynamics, 2018, bioRxiv (Nov. 7,
exhibit
connectomes
tional
http://dx.doi.org/10.1101/464438
2018);
[163] D. Vidaurre, R. Abeysuriya, R. Becker, A. J. Quinn, F. Alfaro-Almagro, S. M. Smith, M. W.
Woolrich, Discovering dynamic brain networks from big data in rest and task, NeuroImage 2018,
34
180, 646.
[164] B. S. Oken, U. Orhan, B. Roark, D. Erdogmus, A. Fowler, A. Mooney, B. Peters, M. Miller, M. B.
Fried-Oken, Brain-computer interface with language model-EEG fusion for locked-in syndrome,
Neurorehabil. Neural Repair. 2014, 28, 387.
[165] Y. Tian, H. Zhang, W. Xu, H. Zhang, L. Yang, S. Zheng, Y. Shi, Spectral Entropy Can Predict
Changes of Working Memory Performance Reduced by Short-Time Training in the Delayed-
Match-to-Sample Task, Front. Hum. Neurosci. 2017, 11, 437.
[166] D. M. Mateos, R. Wennberg, R. Guevara, J. L. Perez Velazquez, Consciousness as a global
property of brain dynamic activity, Phys. Rev. E 2017, 96, 062410.
[167] R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger,
M. Tangermann, F. Hutter, W. Burgard, T. Ball, Deep Learning With Convolutional Neural
Networks for EEG Decoding and Visualization, Human Brain Mapping 2017, 38, 5391.
[168] D. Wen, Z. Wei, Y. Zhou, G. Li, X. Zhang and W. Han, Deep Learning Methods to Process fMRI
Data and Their Application in the Diagnosis of Cognitive Impairment: A Brief Overview and
Our Opinion, Front. Neuroinform. 2018, 12, 23.
[169] K.-S. Hong, M. J. Khan, M. J. Hong, Feature Extraction and Classification Methods for Hybrid
fNIRS-EEG Brain-Computer Interfaces. Front. Hum. Neurosci. 2018, 12, 46.
[170] I. S. Knoth, T. Lajnef, S. Rigoulot, K. Lacourse, P. Vannasing, J. L. Michaud, S. Jacquemont,
P. Major, K. Jerbi, S. Lipp´e, Auditory repetition suppression alterations in relation to cognitive
functioning in fragile X syndrome: a combined EEG and machine learning approach, J.
Neurodevelopmental Disorders 2018, 10, 4.
[171] S. Simpraga, R. Alvarez-Jimenez, H. D. Mansvelder, J. M. A. van Gerven, G. Jan Groeneveld,
S.-S. Poil, K. Linkenkaer-Hansen, EEG machine learning for accurate detection of cholinergic
intervention and Alzheimer's disease, Sci. Rep. 2017, 7, 5775.
[172] M. Cuki´c, D. Pokrajac, M. Stoki´c, S. Simi´c, V. Radivojevi´c, M. Ljubisavljevi´c, EEG machine
learning with Higuchi's fractal dimension and Sample Entropy as features for successful detection
of depression, 2018; arXiv:1803.05985.
[173] M. J. A. M. van Putten, S. Olbrich, M. Arns, Predicting sex from brain rhythms with deep
learning, Sci. Rep. 2018, 8, 3069.
[174] G. Shen, T. Horikawa, K. Majima, Y. Kamitani, Deep image reconstruction from human brain
activity, PLoS Comput. Biol. 2019, 15, e100663.
[175] G. Shen, K. Dwivedi, K. Majima, T. Horikawa, Y. Kamitani, End-to-end deep image
reconstruction from human brain activity, 2018, bioRxiv (2018).
[176] R. C. Fong, W. J. Scheirer, D. D. Cox, Using human brain activity to guide machine learning,
Sci. Rep. 2018, 8, 5397.
[177] M. J. Vansteensel, E. G. M. Pels, M. G. Bleichner, M. P. Branco, T. Denison, Z. V. Freudenburg,
P. Gosselaar, S. Leinders, T. H. Ottens, M. A. Van Den Boom, P. C. Van Rijen, E. J. Aarnoutse,
N.F. Ramsey, Fully Implanted Brain-Computer Interface in a Locked-In Patient with ALS, N.
Engl. J. Med. 2016, 375, 2060.
[178] Z. Wang, J. R. Busemeyer, H. Atmanspacher, E. M. Pothos, The Potential of Using Quantum
Theory to Build Models of Cognition, Topics in Cognitive Sciences 2013, 5, 672.
[179] A. Lukasik, Quantum models of cognition and decision, Int. J. Parallel, Emergent and Distributed
Systems 2018, 33, 3, 336.
[180] J. Broekaert, I. Basieva, P. Blasiak, E. M. Pothos, Quantum-like dynamics applied to cognition:
a consideration of available options, Phil. Trans. R. Soc. A 2017, 375, 20160387.
[181] E. M. Pothos, J. R. Busemeyer, R. M. Shiffrin, J. M. Yearsley, J. Experimental Psychology:
General 2017, 146, 968.
[182] J. Acacio de Barros, G. Oas, Quantum Cognition, Neural Oscillators, and Negative Probabilities,
(Eds. E. Haven, A.
in The Palgrave Handbook of Quantum Models in Social Science,
Khrennikov), 2017.
[183] A. P. Flitney, D. Abbott, An introduction to quantum game theory, Fluctuation and Noise Letters
35
2002, 2(4), R175.
[184] N. Brunner, N. Linden, Connection between Bell nonlocality and Bayesian game theory, Nature
Communications 2013 4, 2057.
[185] J. Bang, J. Ryu, M. Paw lowski, B. S. Ham, J. Lee, Quantum-mechanical machinery for rational
decision-making in classical guessing game, Sci. Rep. 2016, 6, 21424.
[186] J. R. Busemeyer, P. Bruza, Quantum models of cognition and decision making, Cambridge, UK,
Cambridge University Press, 2012.
[187] L. C. White, E. M. Pothos, J. R. Busemeyer, Insights from quantum cognitive models for
organizational decision making, J. Applied Research in Memory and Cognition 2015, 4, 229.
[188] I. Mart´ınez-Mart´ınez, E. S´anchez-Burillo, Quantum stochastic walks on networks for decision-
making, Sci. Rep. 2016, 6, 23812.
[189] B. E. Baaquie, Financial modeling and quantum mathematics, Computers and Mathematics with
Applications 2013, 65 1665.
[190] P. Jedlicka Revisiting the Quantum Brain Hypothesis: Toward Quantum (Neuro)biology? Front.
Mol. Neurosci. 2017, 10, 366.
[191] D. Bourget, Quantum Leaps in Philosophy of Mind, J. Consciousness Studies 2004, 11, 17.
[192] D. Georgiev, Mind Efforts, Quantum Zeno Effect and Environmental Decoherence,
NeuroQuantology 2015, 13, 179.
[193] D. Georgiev, Quantum Information and Consciousness a gentle introduction 2018.
[194] M. P. A. Fisher, Quantum cognition: The possibility of processing with nuclear spins in the brain,
Annals of Physics 2015, 362, 593.
[195] J. Cai, M. Plenio, Chemical Compass Model for Avian Magnetoreception as a Quantum Coherent
Device, Phys. Rev. Lett. 2013, 111, 230503.
[196] M. Tiersch, G. G. Guerreschi, J. Clausen, H. J. Briegel, Approaches to Measuring Entanglement
in Chemical Magnetometers, J. Phys. Chem. A 2014, 118, 13.
[197] M. Mohseni, Y. Omar, G. S. Engel, M. B. Plenio, Quantum Effects in Biology, Cambridge
University Press, 2014.
[198] J. M. Schwartz, H. P. Stapp, M. Beauregard, Quantum physics in neuroscience and psychology:
a neurophysical model of mind-brain interaction, Phil. Trans. R. Soc. B 2005, 60, 1309.
[199] H. P. Stapp, Mindful Universe: Quantum Mechanics and the Participating Observer, 2nd edition,
Springer, 2011.
[200] H. P. Stapp, Retrocausation in quantum mechanics and the effects of minds on the creation of
physical reality, AIP Conf. Proc 2017, 1841, 040001.
[201] S. R. Hameroff, S. Rasmussen, B. Mansson, Molecular automata in microtubules: basic
computational logic of the living state?, in Artificial Life, SF/studies in the sciences of complexity
(Ed. C. Langton). AddisonWesley, New York, 1988..
[202] R. Penrose, Shadows of the Mind, A Search for the Missing Science of Consciousness, Oxford
University Press, 1994.
[203] S. Hameroff, R. Penrose, Orchestrated reduction of quantum coherence in brain microtubules: A
model for consciousness, Mathematics and Computers in Simulation 1996, 40, 453.
[204] M. Tegmark, The Importance of Quantum Decoherence in Brain Processes, Phys. Rev. E 2000,
1, 4194.
[205] J. Preskill, Quantum Computing in the NISQ era and beyond, Quantum 2018, 2, 79.
[206] The Human Brain Project
https://www.humanbrainproject.eu/en/brain-
(HBP 2018;
simulation/brain-simulation-platform/
[207] M. Schuld, I. Sinayskiy, F. Petruccione, Simulating a perceptron on a quantum computer, Phys.
Lett. A 2015, 379, 660.
[208] N. Wiebe, A. Kapoor, K. Svore, Quantum Perceptron Models, 30th Conference on Neural
Information Processing Systems (NIPS 2016), Barcelona, Spain.
[209] Y. Cao, G. G. Guerreschi, A. Aspuru-Guzik, Quantum Neuron: An elementary building block
for machine learning on a quantum computer 2017; arXiv:1711.11240.
36
[210] K. H. Wan, O. Dahlsten, H. Kristj´ansson, R. Gardner, M. S. Kim, Quantum generalisation of
feedforward neural networks, npj Quantum Information, 2017, 3, 36.
[211] P. Pfeiffer, I. L. Egusquiza, M. Di Ventra, M. Sanz, E. Solano, Quantum Memristor, Sci. Rep.
2016, 6, 29507.
[212] J. Salmilehto, F. Deppe, M. Di Ventra, M. Sanz, E. Solano, Quantum Memristors with
Superconducting Circuits, Sci. Rep. 2017, 7, 42044.
[213] X.-H. Cheng, T. Gonzalez-Raya, X. Chen, M. Sanz, E. Solano, Quantized Hodgkin-Huxley Model
for Quantum Neurons 2018; arXiv:1807.10698v1.
[214] F. Silva, M. Sanz, J. Seixas, E. Solano, Y. Omar, Perceptrons from Memristors 2018;
arXiv:1807.04912v1.
[215] K. Fujii, K. Nakajima, Harnessing Disordered-Ensemble Quantum Dynamics for Machine
Learning, Phys. Rev. Applied 2017, 8, 024030.
[216] K. Nakajima, K. Fuji, M. Negoro, K, Mitarai, M. Kitagawa, Boosting computational power
through spatial multiplexing in quantum reservoir computing, Phys. Rev. Applied 2019, 11,
034021.
[217] A. Kutvonen, K. Fujii, T. Sagawa, Recurrent neural networks running on quantum spins: memory
accuracy and capacity 2018; arXiv:1807.03947v1.
[218] J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, S. Lloyd, Quantum machine
learning, Nature 2017, 549, 195.
[219] P. Rebentrost, T. R. Bromley, C. Weedbrook, S. Lloyd, A Quantum Hopfield Neural Network,
Phys. Rev. A 2018, 98, 042308.
[220] P. Rotondo M. Marcuzzi, J. P. Garrahan, I. Lesanovsky, M Muller, Open quantum generalisation
of Hopfield neural networks, J. Phys. A: Math. Theor. 2018, 51, 115301.
[221] M. H. Amin, E. Andriyash, J. Rolfe, B. Kulchytskyy, B. Melko, Quantum Boltzmann machine,
2016; arXiv:1601.02036.
[222] D.-L. Deng, X. Li, S. Das Sarma, Quantum Entanglement in Neural Network States, Phys. Rev.
X 2017, 7, 021021.
[223] K. Mitarai, M. Negoro, M. Kitagawa, K. Fujii, Quantum circuit learning, Phys. Rev. A 2018,
98, 032309.
[224] E. Farhi, H. Neven, Classification with Quantum Neural Networks on Near Term Processors,
2018; arXiv:1802.06002v1.
[225] H. J. Briegel, On creative machines and the physical origins of free behavior, Sci. Rep. 2012, 2,
522.
[226] J. Mautner, A. Makmal, D. Manzano, M. Tiersch, H. J. Briegel, Projective simulation for classical
learning agents: a comprehensive investigation, New Gener. Comput. 2015, 3, 69.
[227] V. Dunjko, N. Friis, H. J. Briegel, Quantum-enhanced deliberation of learning agents using
trapped ions, New J. Phys. 2015, 17, 023006.
[228] V. Dunjko, J. M. Taylor, H. J. Briegel, Quantum-Enhanced Machine Learning, Phys. Rev. Lett.
2016, 117, 130501.
[229] A. A. Melnikov, A. Makmal, V. Dunjko, H. J. Briegel, Projective simulation with generalization,
Sci. Rep. 2017 7, 14430.
[230] A. A. Melnikov, H. Poulsen Nautrup, M. Krenn, V. Dunjko, M. Tiersch, A. Zeilinger, H. J.
Briegel, Active learning machine learns to create new quantum experiments, PNAS 2018, 115,
1221.
[231] Y. Yamamoto, K. Aihara , T. Leleu, K.-i. Kawarabayashi, S. Kako, M. Fejer, K. Inoue, H.
Takesue, Coherent Ising machines - optical neural networks operating at the quantum limit, npj
Quantum Information 2017, 49.
[232] R. Hamerly, T. Inagaki, P. L. McMahon, D. Venturelli, A. Marandi, T. Onodera, E. Ng, C.
Langrock, K. Inaba, T. Honjo, K. Enbutsu, T. Umeki, R. Kasahara, S. Utsunomiya, S. Kako,
K.-i. Kawarabayashi, R. L. Byer, M. M. Fejer, H. Mabuchi, E. Rieffel, H. Takesue, Y. Yamamoto,
Scaling advantages of all-to-all connectivity in physical annealers: the Coherent Ising Machine
37
vs. D-Wave 2000Q, 2018, arXiv:1805.05217v1.
[233] A. D. King, W. Bernoudy, J. King, A. J. Berkley, T. Lanting, Emulating the coherent Ising
machine with a mean-field algorithm, 2018; arXiv:1806.08422v1.
[234] P. Snaprud, The consciousness wager, New Scientist, 2018, 23 June, p.23.
38
|
1708.02423 | 4 | 1708 | 2018-04-09T09:02:57 | From Correlation to Causation: Estimation of Effective Connectivity from Continuous Brain Signals based on Zero-Lag Covariance | [
"q-bio.NC"
] | Knowing brain connectivity is of great importance both in basic research and for clinical applications. We are proposing a method to infer directed connectivity from zero-lag covariances of neuronal activity recorded at multiple sites. This allows us to identify causal relations that are reflected in neuronal population activity. To derive our strategy, we assume a generic linear model of interacting continuous variables, the components of which represent the activity of local neuronal populations. The suggested method for inferring connectivity from recorded signals exploits the fact that the covariance matrix derived from the observed activity contains information about the existence, the direction and the sign of connections. Assuming a sparsely coupled network, we disambiguate the underlying causal structure via $L^1$-minimization. In general, this method is suited to infer effective connectivity from resting state data of various types. We show that our method is applicable over a broad range of structural parameters regarding network size and connection probability of the network. We also explored parameters affecting its activity dynamics, like the eigenvalue spectrum. Also, based on the simulation of suitable Ornstein-Uhlenbeck processes to model BOLD dynamics, we show that with our method it is possible to estimate directed connectivity from zero-lag covariances derived from such signals. In this study, we consider measurement noise and unobserved nodes as additional confounding factors. Furthermore, we investigate the amount of data required for a reliable estimate. Additionally, we apply the proposed method on a fMRI dataset. The resulting network exhibits a tendency for close-by areas being connected as well as inter-hemispheric connections between corresponding areas. Also, we found that a large fraction of identified connections were inhibitory. | q-bio.NC | q-bio |
From Correlation to Causation: Estimation of Effective Connectivity
from Continuous Brain Signals based on Zero-Lag Covariance
Jonathan Schiefer, Alexander Niederbuhl, Volker Pernice, Carolin Lennartz, Pierre LeVan,
Jurgen Henning and Stefan Rotter
Abstract
Knowing brain connectivity is of great importance both in basic research and for clinical applications.
We are proposing a method to infer directed connectivity from zero-lag covariances of neuronal activity
recorded at multiple sites. This allows us to identify causal relations that are reflected in neuronal
population activity. To derive our strategy, we assume a generic linear model of interacting continuous
variables, the components of which represent the activity of local neuronal populations. The suggested
method for inferring connectivity from recorded signals exploits the fact that the covariance matrix
derived from the observed activity contains information about the existence, the direction and the
sign of connections. Assuming a sparsely coupled network, we disambiguate the underlying causal
structure via L1-minimization, which is known to prefer sparse solutions.
In general, this method
is suited to infer effective connectivity from resting state data of various types. We show that our
method is applicable over a broad range of structural parameters regarding network size and connection
probability of the network. We also explored parameters affecting its activity dynamics, like the
eigenvalue spectrum. Also, based on the simulation of suitable Ornstein-Uhlenbeck processes to model
BOLD dynamics, we show that with our method it is possible to estimate directed connectivity from
zero-lag covariances derived from such signals.
In this study, we consider measurement noise and
unobserved nodes as additional confounding factors. Furthermore, we investigate the amount of data
required for a reliable estimate. Additionally, we apply the proposed method on full-brain resting-
state fast fMRI datasets. The resulting network exhibits a tendency for close-by areas being connected
as well as inter-hemispheric connections between corresponding areas. In addition, we found that a
surprisingly large fraction of more than one third of all identified connections were of inhibitory nature.
Introduction
The networks of the brain are key to understanding its function and dysfunction [1]. Depending on
the methods employed to assess structure and to record activity, networks may be defined at different
levels of resolution. Their nodes may be individual neurons, linked by chemical or electrical synapses.
Alternatively, nodes may also be conceived as populations of neurons, with links represented by the
net effect of all synaptic connections that exist between two populations. In any case, this defines
the structural substrate of brain connectivity, representing the physical (causal) basis of neuronal
interactions. Nodes in a brain network influence each other by sending signals. For example, the
activities of nodes in a network are generally not independent, and neuronal dynamics are characterized
by correlations among the nodes involved in the network. This suggests an alternative perspective
on active brain networks: Functional connectivity assigns a link to a pair of nodes to the degree to
which their activities are correlated.
It has been argued that this concept emphasizes connections
that "matter", including the possibility that the same substrate may give rise to different networks,
depending on how they are used. As a consequence, functional connectivity and structural connectivity
are not equivalent. A well-known phenomenon is that two nodes may be correlated, even if there
is no direct anatomical link between them. For example, a shared source of input to both nodes
may generate such a correlation, which does not correspond to a direct interaction between the two
nodes. Apart from that, correlation is a symmetric relation between two nodes, whereas a physical
1
connection implies a cause-effect relation that is directed. There have, in fact, been multiple attempts
to overcome the shortcomings of functional connectivity, especially the lack of directed interaction.
The term effective connectivity has been suggested for this [2]. The idea is to bring the networks,
inferred from activity measurements, closer to structural connectivity, which can only be inferred with
anatomical methods. The dichotomy between structural and functional aspects of connectivity raises
the general question whether it is possible to infer brain networks from recorded activity. We are only
beginning to understand the forward link between structural connectivity and functional connectivity.
As a consequence, it is possible to compute correlations from connectivity in certain simplified network
scenarios [3]. The correspondence between connectivity and correlation, however, is not one-to-one.
Networks with different connectivity may lead to exactly the same correlations between nodes. As a
consequence, the inverse problem of inferring connectivity from correlation is generally ill-defined. As
we will demonstrate in this paper, additional assumptions about the connectivity can help to resolve
the ambiguity. Specifically, we search for the network with the lowest number of nonzero edges (via
L1-minimization) to disambiguate the problem. Structural, functional and effective connectivity are
not equally well accessible. Some aspects of the anatomical structure can be assessed post mortem by
invasive tracing methods, or non-invasively by Diffusion Tensor Imaging, DTI. In contrast, functional
connectivity is based on statistical relationships between the activity of neuronal populations and can
be easily estimated from recorded signals. For estimating effective connectivity there are methods
like Dynamic Causal Modelling, DCM [4, 5], Granger causality [6] and others [7, 8, 9, 10, 11, 12, 13].
Only few methods to infer effective connectivity, however, can deal with large numbers of nodes (40
or more) based on zero-lag correlation only. However, they are either limited to small networks [14],
or to directed acyclic graphs [15]. Here, we are proposing a new method for the estimation of effective
connectivity from population activity in the brain, especially BOLD-related signals. The new method
is a variant of the procedure described in [16], based on a L1-minimization. For the method proposed
here it is sufficient to use zero-lag covariances to estimate directed effective connectivity.
Materials and methods
Estimation method
The main idea of our estimation method is inspired by the finding, "that the key to determining the
direction of the causal relationship between X and Y lies in 'the presence of a third variable Z that
correlates with Y but not with X,' as in the collider X → Y ← Z . . . " [17, 18].
Similarly, assuming a linear interaction model, the presence of a collider structure in a network
(see Fig 1) produces specific entries in the corresponding inverse covariance (precision) matrix. Fig 1
shows a disconnected network in the left column, and a network which induces the same covariance
matrix if all links have opposite direction in the middle column. In the latter case an estimation of
the direction is impossible, because there is simply no information about it in the covariance matrix.
Whenever a collider structure is present, however, the entry in the inverse covariance matrix for the
two source nodes (here, 2 and 3) is non-zero. This is due to the fact that in a linear model the entry
in the inverse covariance matrix depends not only on the connections of the nodes 2 and 3, but also
whether these nodes have a common target. This means the presence of a collider structure allows us
to disambiguate the direction of this particular connection.
We consider here a scenario, where the interaction between nodes is described by a generic linear
model. Assuming stationarity, let the neural activity x(t) be implicitly defined by the consistency
equation
x(t) = (G ∗ x)(t) + v(t)
(1)
where G(t) is a matrix of causal interaction kernels and v(t) denotes fluctuating external inputs
("driving noise"). All variables are also listed in table 1. Fourier transformation of Eq (1) yields
x(f ) = G(f )x(f ) + v(f )
and simple rearrangement leads to
x(f ) =
(cid:17)−1
v(f )
(cid:16)
1 − G(f )
2
Figure 1: Collider structures are encoded in the inverse covariance matrix. Upper row: Three simple
network architectures. Lower row: The corresponding inverse covariance matrices, red color represents
positive entries, blue color stands for negative ones. In the left and middle column, the entries (2, 3)
and (3, 2) are 0. The only difference between the right column and the middle column is that the
connection between node 1 and 2 is flipped, such that nodes 1, 2 and 3 form a collider structure.
Although there is still only an indirect connection between node 2 and 3, the entry in the corresponding
inverse covariance matrix is now non-zero.
where x denotes the Fourier transform of x. The cross spectral density of the signals is then given by
C(f ) =
1 − G(f )
1 − G∗(f )
where Z(f ) is the cross-spectral density of the external inputs. It follows
(cid:16)
(cid:16)
(cid:16)
(cid:17)−1 Z(f )
(cid:17) Z−1(f )
(cid:16)
(cid:17)−1
(cid:17)
1 − G(f )
C−1(f ) =
1 − G∗(f )
= B∗(f )B(f )
(cid:17)
(cid:113)
(cid:16)
Z−1(f )
1 − G(f )
(2)
with B(f ) =
. In our model, we assume that the components of the external
fluctuating input are pairwise stochastically independent for all nodes. Then, Z is a diagonal matrix,
and we make the additional assumption that Z = 1. For the linear model considered here, there is a
relation between covariance and connectivity, which can be exploited for the estimation of connectivity
from correlation. In the case Z = 1 it is given by
1 − G(f )
= 1 − G(f ) − G∗(f ) + G∗(f ) G(f )
1 − G∗(f )
C−1(f ) =
(cid:17)(cid:16)
(cid:16)
(cid:17)
If the matrix product
where the last term contributes the information of the collider structures.
G∗(f ) G(f ) has a non-zero off-diagonal entry the corresponding nodes have outgoing connections ter-
minating at the same node, which means these nodes form a collider.
It is clear that for any unitary matrix U ∈ U(n) the product U B is still a solution of Eq (2), as
U∗U = 1. We will resolve this ambiguity with an L1 minimization which is known to prefer sparse
solutions under certain conditions [19]. In order to find G from a given C we first fix an initial matrix
B, and then search for a unitary matrix U ∈ U(n) such that (cid:107)U B(cid:107)1 is minimal, so we are minimizing
the function
Γ : U(n) −→ R
U (cid:55)−→ (cid:107)U B(cid:107)1.
(3)
Gradient descent
To estimate the connectivity matrix from the covariance matrix we use a conjugate gradient descent
algorithm similar to [20, 21] for minimizing the function Γ(U ) given in Eq (3), implemented in Python.
3
Table 1: Variables used for the estimation method and simulation
Variable name
node activity
network connectivity
external inputs
cross-spectral density
covariance of external input
unitary matrix
L1-norm cost function
gradient
initial matrix
Wiener process
stationary covariance matrix
simulation step
time constant of activity
regularisation-controlling parameter for regularized ICOV
Symbol
x(t)
G
v(t)
C(f )
Z(f )
U
Γ
d
B0
w(t)
σ
∆t
τ
λ
For details please see supporting information, algorithm 1. For the gradient
dij =
∂Γ(U )
∂Ui,j
2 (d − d∗) is skew-hermitian, and the matrix exponential of a skew-
of the cost function Γ(U ), a = 1
hermitian matrix is unitary. This means, starting in a point Uact and choosing an appropriate step
size δ, we obtain a point Unew = exp(−δa)Uact with Γ(Unew) < Γ(Uact).
In other words, the new
point has a smaller L1-norm than the old one and still satisfies the condition C−1 = B∗U∗newUnewB.
Iterating this procedure until convergence leads to a point with locally minimal L1-norm.
The two conditions for convergence are inspired by [21]. The first one is a condition on the norm
of the gradient. In each step, it is checked if
(cid:107)d − U d∗U(cid:107)F =
dij(U d∗U )ij2 < gtol
(cid:115)(cid:88)
is fulfilled, where (cid:107) . . .(cid:107)F is the Frobenius norm and gtol > 0 is the convergence tolerance. As a second
(alternative) condition, it is checked whether simultaneously
i,j
(cid:107)U − Uold(cid:107)F
√
N
< xtol
and
Γ(Uold) − Γ(U )
Γ(Uold + 1) < ftol
are fulfilled. The values used are listed in table 2. Before convergence the cost function typically
oscillates around a certain value. To avoid stopping at a random phase of this oscillation, as a final
step we apply a line-search, for details see supporting information, algorithm 2. The described gradient
descent algorithm provides an efficient way for minimizing Eq (3). When calculating the gradient, we
neglect the diagonal. Consequently, we also neglect the diagonal of the resulting estimated matrix, so
we are not able to study self connections of the nodes.
Initial condition
As starting condition for the gradient descent we use a matrix B0(f ) such that
B∗0(f )B0(f ) = C−1(f ).
There are many ways to choose a B0 with this property, we found the following choice efficient: As C
is the cross-spectral density it is positive definite, and so is C−1. Thus, there is exactly one positive
definite square root of C−1 [22] which can be calculated by
√
B0 =
EW ∗
(4)
(cid:112) C−1 = W
4
Table 2: Parameter used for estimation
Parameter
xtol
ftol
gtol
κ
λ
Value
0.7· 10−2
0.7· 10−4
0.7· 10−2
500
5
where the columns of W are the eigenvectors of C−1, and E is the matrix with the corresponding
eigenvalues of C−1 on the diagonal. Thus we initialize the gradient descent with U0 = 1 and B0 given
by Eq (4).
Step size selection
A critical part of the optimization is the selection of an appropriate step size. If the step size is too
large, one might miss the minimum. If the step size is too small, the optimization converges very
slowly. For the gradient descent, we use an adaptive scheme inspired by [20], where the step-size
depends on the largest eigenvalue of the actual gradient: Let λmax be the largest eigenvalue of dact,
the step size is given by
δact =
2π
λmax · κ
where κ is constant. The intuition behind that, is that smooth cost functions along a geodesic on
the unitary manifold are almost periodic. So the step size should be a fraction of the period of
this function. This is achieved by the scaling with the largest eigenvalue, which allows us to take a
scale-invariant fraction of this period.
Validation methods
Noise-free covariance matrices
We assume that the interactions among neuronal populations can be described by a linear model, see
Eq (2) with Z = 1. This model allows us to derive a relation between the connectivity matrix of the
network G and the inverse cross-spectral density matrix C−1 of the measured activity
C−1 =
1 − G∗
1 − G
= 1 − G∗ − G + G∗ G.
(5)
(cid:16)
(cid:17)(cid:16)
(cid:17)
Given a sampled connectivity matrix G we can calculate the inverse covariance matrix directly using
Eq (5). For all simulations, half of the connections were negative (inhibitory) connections, the absolute
strength was the same for all connections and 20 repetitions were simulated. As connectivity profiles
we used random Erdos-R´enyi networks.
Ornstein-Uhlenbeck processes
To validate our inference procedure before applying it to the network inference from measurements
of neuronal activity we simulated stationary signals. Since there is no gold-standard for simulations
of fMRI data[23], we based our simulations on the Ornstein-Uhlenbeck process [24], which provides a
simple linear model for neural activity.
dx(t) = Ax(t)dt + dW (t)
(6)
where A is a matrix and W a Wiener process. In our applications, we parametrize this matrix as
A = 1
possible to calculate the stationary covariance matrix Σ from the continuous Lyapunov equation
τ (G − 1) with real-valued connectivity matrix G and time constant τ . For this process, it is
1 = AΣ + ΣAT .
5
In fact, we simulated the process in discrete time. In analogy with [25] we use a multivariate version
of the exact update formula
x(t + ∆t) = eA∆tx(t) + n(t),
(7)
where n(t) ∼ N (0, Σ) is normally distributed, with Σ being the stationary covariance matrix described
above. As a final step, we filter the time series x(t) with the canonical hemodynamic response function
(HRF) [26, 27]. To match the data obtained in brain scans sampled at a temporal resolution of
∆t = 0.1 s, we used random connectivity profiles G with a connection probability p = 0.1 (Erdos-
R´enyi model), 50% negative entries, and a spectral radius of ρ = 0.3. All parameters used are listed,
once more, in table 3.
Table 3: Parameter used for simulations
Parameter
repetitions
network type
N
p
T
dt
τ
ρ
Value
20
Erdos-R´enyi
100
0.1
350 000 s
0.1 s
0.1 s
0.3
Before calculating the covariance C, the data is standardized such that the mean is 0 and the
variance is 1 for all components of the time series. We add normally distributed observation noise
uobs with a N (0, σobs) distribution to the simulated signal. After the simulation we calculated the
signal-to-noise ratio according to
SNR =
σ2
X
σ2
obs
where σ2
X denotes the variance of the signal.
Performance measures
When estimating connectivity from simulations with known underlying network structure (ground
truth), one can quantify the performance of the estimation. For measuring the accuracy of our
estimation we employ three different methods.
First, we use the area under the ROC-curve (AUC). The ROC (receiver operating characteristic)
curve is obtained as following: For each possible parameter value (in our case the threshold for the
existence of a connection), the number of true-positives (TP) and false-positives (FP) is used to
calculate the true-positive rate (or recall) TP/(TP + FN) and the false-positive rate FP/(FP + TN).
The ROC curve is then obtained by plotting the true-positive-rate against the false-positive rate.
Secondly, we use the average precision score (PRS) which is the area under the precision-recall
curve. This also includes the false-negatives (FN) (precision:
TP+FP ). If both AUC and PRS are
equal to 1, the connections in the network are perfectly estimated. Sample curves are shown in Fig 2
D.
TP
Thirdly, we calculate the Pearson Correlation Coefficient (PCC) which in contrast to the measures
defined before also take the strength and the sign of the interactions into account. This also means
that this measure is less suited to assess whether a connection exists or not.
It rather measures
whether the estimated connections have the same strength as the original ones. We consider all three
performance measures simultaneously to establish the quality of our estimates.
Experimental fMRI data
Seven healthy subjects underwent a 20-minute resting-state fMRI experiment on a 3 T Siemens Prisma
scanner. The data was acquired using the MREG sequence [28], yielding a high temporal resolution
6
Figure 2: Networks inferred from a simulated Ornstein-Uhlenbeck process.A shows the original net-
work. B shows the network inferred with our new method from the zero-lag covariances. White and
black entries indicate true negative (TN) and true positive (TP) connections, blue and red entries
indicate false negative (FN) and false positive (FP) connections, respectively. In this example, the
performance measures are AUC = 0.98, PRS = 0.97 and PCC = 0.95. C depicts the sample co-
variance (functional connectivity) matrix directly estimated from the data. In C, as a consequence
of symmetry, the number of wrongly estimated connections is quite high, the performance measures
are AUC = 0.93, PRS = 0.54, and PCC = 0.29. D shows the Receiver Operating Characteristic
Curve and the Precision Recall Curve for the networks estimated from zero-lag covariance Gest in
blue/orange and of the functional connectivity C in green/red. The areas under these curves are the
AUC and PRS, respectively.
7
020406080Nodes020406080NodesATrue Connectivity Matrix020406080NodesBEstimated Connectivity MatrixNodes020406080NodesCCovariance Matrix0.00.20.40.60.81.0False Positive Rate / Recall0.50.60.70.80.91.0True Positive Rate / PrecisionDPerformance MeasuresROC Curve GestPR Curve GestROC Curve CPR Curve CFNTNTPFP (TR = 0.1 s, 12000 time points) that facilitates functional connectivity analyses [29]. The other se-
quence parameters were TE = 36 ms, FA = 25◦, 64 × 64 × 50 matrix and 3 mm isotropic voxel size.
Additionally, cardiac and respiratory signals were recorded with the ECG and abdominal breathing
band from the scanner's physiological monitoring unit. Motion correction was done with FSL and
physiological noise correction was performed with RETROICOR [30]. Average CSF and white matter
signals were regressed out, but no global signal regression was performed. Following image normal-
ization to MNI space, voxels were parcellated according to the AAL atlas (excluding the cerebellum),
and the mean activity within each atlas region was calculated. The connectivity was then estimated
using zero-lag covariances of the standardized signals.
Ethics statement
The experiments have been approved by the Ethics Committee of the University Medical Center
Freiburg.
Results
Noise-free covariance matrices
Intrinsic properties of our new estimation procedure can be identified by studying the performance of
the method for perfectly estimated (noise-free) covariance matrices. This way we address properties
that do not depend on any particular feature of the underlying data, and that are not due to the
success of the measurement process. In particular, we show for which types of networks our estimation
procedure gives good results on technical grounds, with a wide range of networks hopefully including
those arising in applications. We used random Erdos-R´enyi connectivity profiles for all simulations.
The macro-connectivity between neuronal populations has to satisfy certain conditions in order to
be tractable by our methods. Two of these conditions concern the dynamic stability of the network
and the strength of the interactions. There is a trade-off between the number of physical links and
the resulting strength of macro-connections, and the dynamic stability of the network. To study the
performance of our method in these various regimes, we separately varied the network size N , the
connection probability p, and the absolute strength of connections J in the connectivity matrix G,
while the fraction of inhibitory couplings was kept at 50%. The spectral radius ρ of the bulk eigenvalue
spectrum is approximately given by
ρ2 = J 2p(1 − p)N.
(8)
The default values of the parameters used in our study were N = 100, p = 0.1 and ρ = 0.7, where
only one of them at a time was systematically varied. Low values of the spectral radius ρ correspond
to networks with weak recurrent interaction and high values to networks with strong interaction,
respectively. According to the model of network interaction assumed here, the networks need to have
a spectral radius ρ > 0 for network interaction to be present and ρ < 1 for the dynamics to be stable.
First, our results in Fig 3 A indicate that a certain minimal level of interaction is necessary to be
able to estimate the connections reliably. Above a value of ρmin = 0.2, the influence of the spectral
radius on the performance of the estimation is weak, but the larger the spectral radius is the better
the estimation gets.
Secondly, the connection probability of the network influences the quality of the estimation. For
all connection probabilities tested here the network size was kept constant at N = 100 nodes. The
networks were constructed such that the strength J of all connections was the same and such that the
spectral radius ρ was constant according to Eq (8). Fig 3 B shows that the estimation works very well
for sparse matrices with a connection probability in the range between 5% and 15%. For networks with
higher connection probability and equally strong connections, the performance decreases as expected,
due to the bias associated with L1-minimization. But even for a connection probability of p = 0.21,
a fraction of 14.2% of the estimated connections are false negative, and 3.3% are false positive. More
than 90% of the correctly estimated connections have the correct sign. In applications, the focus of
the estimation often lies on the strongest connections in the network. In networks with a background
of weak connections and a sparse skeleton of stronger connections, it is possible to selectively estimate
8
Figure 3: Effects of spectral radius ρ, connection probability p and network size N . Here we consider
the case of noise-free covariance matrices, which were created based on the theory of the underlying
model. The quantities considered are the area under the ROC curve (AUC; green), the precision
recall score (PRS; orange) and the Pearson correlation coefficient (PCC; purple). The shaded areas
indicate the mean ± standard deviation computed over 20 realizations. A If the network interaction
is larger than ρmin, it has relatively little effect on the performance of the estimation. Even in the
extreme case, where ρ is close to 1, the estimation works well. B Performance of the estimation
for different sparsity levels, encoded by the respective connection probabilities p. As expected, for
non-sparse networks the performance of the algorithm degrades dramatically. C Performance of the
estimation for increasing network size. Our results indicate clearly that bigger networks can be better
reconstructed. Applicability may be limited by the numerical effort associated with the optimization.
D Performance of the estitmation in presence of weak background connections.
It is nevertheless
possible to infer the skeleton of strong connections with high fidelity.
these strong links although, strictly, the assumption of a sparse network is violated. Fig 3 shows the
performance of our method for such networks: the networks consist of a skeleton of strong connections
with connection probability p = 10% and a connection strength derived from Eq (8) for ρ = 0.7.
Additionally, we created a second network with weaker connections for various connection probabilities
q. The two networks were combined by adding the connectivity matrices. The connection strength of
this weaker connections is also derived from Eq (8), with a spectral radius of the background network
being 20% of the spectral radius of the skeleton network. Then the performance of the estimation is
calculated with respect to the skeleton of strong connections.
Thirdly, to be applicable to a broad range of data types, a method of connectivity estimation should
perform stable for different network sizes N . For most common types of non-invasive recordings of
population activity the number of nodes considered is in the range between 30 and 150.
It is, of
course, possible to consider larger networks, although the estimation becomes computationally more
expensive. The runtime of the algorithm for networks with 200 nodes still in the range of seconds on
a state-of-the-art desktop computer, but even networks with 1 000 nodes or more are tractable. The
strength of the connections J are set such that the spectral radius ρ of G is constant; the connection
probability is constant at p = 0.1. Fig 3 C shows that our method performs better for bigger networks.
We have observed that the L1 cost landscape becomes smoother for larger networks.
9
0.20.40.60.8Spectral radius 0.70.80.91.0ScoreA0.10.2Connection probability pB50100150Network size N0.70.80.91.0ScoreC0.20.30.40.5Connection probability qof background connectivityDArea under the ROC CurvePrecision-Recall ScoreCorrelation Coefficient Ornstein-Uhlenbeck processes as model for BOLD signals
In order to create surrogate data which fit fast fMRI data [28], we simulated interacting stochastic
processes known as Ornstein-Uhlenbeck processes.
In this case, the performance of the network
inference depends on how well the inverse covariance matrix, which is the basis of the estimation, can
be derived from the data. In addition to finite size effects, we studied the impact of observation noise
on the performance, see Fig 4. We used N = 100, p = 0.1, dt = 0.1 s, ρ = 0.74 and τ = 0.1 s as default
values of the parameters. Generally, it seems natural to use Welch's method to calculate cross-spectral
densities directly, and then to estimate the connectivity for each frequency band separately. For the
data described here, however, we can estimate the connectivity from zero-lag sample covariances in
the time domain. This is possible when the mass of the covariance function is concentrated very close
around lag 0. Then lag 0 is the only one contributing to the integral of the covariance function, which
corresponds to the cross-spectral density C(0).
As shown in Fig 4 A, with noisy data the AUC is still good, but the PRS is lower than in the
case, where the covariance is known without error. However, for a signal-to-noise ratio above 1 the
performance improves very quickly.
Figure 4: Performance of network inference based on simulated Ornstein-Uhlenbeck processes. Same
colors as in Fig 3. A Performance of the estimation when measurement noise is added. B Performance
of the estimation if only parts of the network are observable. The fraction of observed nodes in a
network are indicated on the x-axis. The total number of nodes in the network was N = 180.
In the case of fMRI usually the whole brain is scanned, and there are no unobserved nodes in
the network. However, for other data types (e.g. fNIRS) only parts of the brain can be observed.
The question then is, whether this sub-network can nevertheless be reconstructed from the recorded
signals. To model this scenario, we took simulated data and removed randomly a certain subset of
components from the dataset. The interaction of the removed nodes is then not part of the covariance
matrix of the reduced dataset, although the unobserved nodes of course still exert their influence on
the observed ones. The performance of the estimation of the sub-network based on the reduced dataset
is shown in Fig 4 B. Our analysis shows very clearly that the estimation still leads to reasonable results
under these conditions. In fact, we can demonstrate that we are inferring causal connections only:
For unconnected observed nodes X, Y and a latent node L connected to both X and Y , our method
10
0246810Signal-to-noise ratio0.50.60.70.80.91.0ScoreA16253341505866758391100Percentage of observed nodes0.50.60.70.80.91.0ScoreBArea under the ROC curvePrecision-Recall scoreCorrelation Coefficient does not erroneously indicate a link between X and Y .
One key factor for a reliable estimation of the covariance matrix is the amount of data available.
This depends on the length of the measurement or simulation, and on the sampling rate. Since fast
fMRI time series are obtained by measuring the BOLD response as a proxy of neuronal activity,
the time scale of the measured data is relatively slow compared to the time scale of the underlying
neuronal activity. Fig 5 shows the performance of network inference depending on the amount of data
available, and on the time-scale of the neuronal activity. Not surprisingly, the more voluminous the
dataset is, the better the estimation gets. On the other hand, it shows that the estimation generally
leads to better results for slower temporal dynamics. Also, for data of sufficient length with a fairly
good signal-to-noise ratio, the estimation of the connectivity is possible even when only a part of the
network is observed. To allow comparison of our new method with other known methods for network
inference [31, 32, 33], we applied it to the NetSim dataset provided by [34]. For details on the result
of this, please see Fig 10 in the supporting information.
Figure 5: Performance (color coded) of the estimation depending on data length (y-axis) and time
scale of the activity (x-axis). Both scales are logarithmic. For interpolation a bilinear method is used.
fMRI data
We estimated connectivity from seven fast fMRI datasets, for details see the methods section. The
resulting networks, after a threshold of 10% was applied, consist of 810 connections for each dataset.
The threshold of 10% was chosen arbitrarily.
In the supporting information (Fig 9) we show the
histogram of estimated connection strengths for all seven reconstructed networks before thresholding.
The threshold is derived from the 10% strongest connections, disregarding their signs. As there is
generally no full ground truth for the connectivity inferred from human fMRI recordings available
[35, 31, 32], we cannot definitely assess the degree to which the result of our inference are correct. We
can, however, establish whether they are plausible. One representative connectivity matrix is shown in
Fig 6. On average, 34% of the connections were inhibitory, with negligible variability across subjects.
Of all connections found, 301(37%) were found in four subjects or more, and 4 872 out of 8 100 possible
connections were absent in all subjects. On average, 245 of the connections were bi-directional and
565 connections were identified only for one direction. In general, close-by areas are more likely to
be connected than more distant ones. This fact is (approximately) represented by a concentration
of connections along secondary diagonals in the within-hemisphere blocks. Also, there are frequent
inter-hemispheric connections between corresponding areas. This fact is represented by the diagonal
entries in the across-hemisphere blocks.
11
0.010.020.050.10.20.50.751.0τ[s]1248163264128#Samples(×17500)0.010.020.050.10.20.50.751.0τ[s]PrecisionRecallScore0.010.020.050.10.20.50.751.0τ[s]0.20.40.60.81.0 Figure 6: Left panel: Directed connectivity estimated with our new method from one sample MREG
data set. Voxels were parceled using the AAL90-atlas. In the top-left block of the connectivity matrix
connections within the left hemisphere are shown, in the lower-right block connections within the
right hemisphere. The off-diagonal blocks represent the inter-hemispheric connections from the left to
the right hemisphere (lower left) and from the right to the left hemisphere (top right). The strength
of all connections is color coded, with red representing positive (excitatory) connections and blue
representing negative (inhibitory) connections. Only the strongest 10% of connections are shown.
Right panel: Functional connectivity matrix derived from the same data.
Comparison with the Regularized Inverse Covariance (RIC) method
As mentioned above, different heuristics have been suggested to reconstruct networks from neuronal
signals.
In Fig 7 we compare the performance of the new method we propose here and the es-
tablished method of Regularized Inverse Covariance [34], based on the implementation provided at
https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSLNets. Our comparison clearly shows that our new
method performs significantly better than the Regularized Inverse Covariance method, mainly, be-
cause the latter cannot establish the direction of connections. The superior performance of the new
method is reflected in higher values for all three performance measures, in particular PRS and PCC.
As regularization parameter required by the software toolbox, we used λ = 5.
Furthermore, we applied the RIC method on all seven MREG datasets described before. A thresh-
old was applied, such that only the 10% strongest connections are retained. To compare the outcome of
both methods, we only condidered the existence of connections (binary and symmetric connectivity)
and disregarded weights and directions (weighted nonsymmetric connectivity). One representative
example of the comparison of both methods is shown in Fig 8. For RIC, 376 out of 810 possible
connections where identified in four subjects or more out of seven, the corresponding number for
our method is 392 out of 810 possible connections. If any method produced directed networks with
10% connection probability at random, this would yield an average count of less than 25 connections
(3% of 810 connections) that agree for least four out of seven independently generated networks. On
average, 290.5 out of 4 050 possible connections (undirected) are identified by both methods, 3 530.5
connections were found by neither of the methods. This means that both methods agree on 3 821 out
of 4 050 connections on average. The two methods disagreed on the remaining 229 connections.
12
01020304050607080Nodes01020304050607080NodesEstimated Effective Connectvity01020304050607080NodesleftrightleftrightleftrightFunctional Connectvity1.00.50.00.51.0 Figure 7: Comparison of performance with the Regularized Inverse Covariance (RIC) method based on
numerical simulations of Ornstein-Uhlenbeck processes. Shown are the results from the reconstruction
of 20 different networks with Erdos-R´enyi connectivity profiles as described before (cf. Fig 2). AUC,
PRS and PCC of our new method and of the RIC method, respectively, are shown side-by-side.
Figure 8: Estimated networks for one representative MREG dataset. The left panel shows the sym-
metrized network reconstructed with our estimation method, the middle panel shows the network
found with the RIC method. The right panel shows the connections which are identified by both
methods (EB, black), by none of the methods (EN, white), the connections found only by the RIC
(ERIC, blue) and the connections found only by our method, but not by the RIC method (NERIC,
red).
Discussion
With the described method we can estimate directed and signed effective connectivity between neural
populations from measured brain signals, based on zero-lag covariances only. To investigate the
reliability of our estimated connections we used simulations of Ornstein-Uhlenbeck processes mimicking
BOLD-related signals generated by interacting neuronal populations. Our method shows very good
performance, if enough data is available and the observation noise is not too strong. Also, even
in cases with relatively poor performance (e.g. if the network is too dense) more than 90% of the
estimated connections have the correct sign. Applying the method on measured fast fMRI data,
we found that about 34% of all identified connections have an inhibitory effect on their respective
target population. In general, inhibitory synapses are mainly formed within local populations, and
13
AUCRIC AUCPRSRIC PRSPCCRIC PCC0.60.70.80.91.0Score01020304050607080Nodes01020304050607080NodesA01020304050607080NodesB01020304050607080Nodes01020304050607080GG RICCERICENEBNERIC typically do not project to distant targets. An inhibitory connection between populations, however,
can also be achieved by excitatory neurons preferentially terminating on the inhibitory neurons of the
target region. The comparison with the Regularized Inverse Covariance (RIC) method shows good
agreement with regard to the existence of connections. Directions cannot be disambiguated with the
RIC method. Our results based on simulated surrogate data reflect what one would expect from the
design of an estimation procedure. For large, sparse networks with sufficiently strong interaction,
our network estimation procedure works reliably. However, as expected if the network is not sparse,
or the time series is too short, the quality of the estimate drops. Nevertheless, in most cases the
main interest lies on the strongest connections, which can be reliably estimated with our method
even when the network is not sparse. For the experimental data shown, individual connections may
be unreliable because of the limited size of the dataset. Also, it is unclear whether the biological
network to be analyzed is really sparse, and if the assumption of pairwise independent external input
is really justified. On the other hand, due to the higher likelihood of a coupling between close-by areas
and between inter-hemispheric counterparts, the resulting network looks plausible. For interpreting
individual connections longer recordings would certainly be beneficial. Also, one could then use
temporal information from additional frequency bands. Of high interest is also the comparison with
structural measures as the ones obtained by diffusion tensor imaging. To the best knowledge of the
authors, this is the first time that effective whole-brain connectivity has been estimated from zero-
lag covariances. Other methods [8] rely on lagged covariances, where the correct lag parameter is
critical, and needs to be inferred from the exponential decay of the observed auto-covariances. Also,
our proposed method is the only one that can detect directed inhibitory connections on the whole-
brain scale. The estimation procedure is fast and easy to apply. As it uses no temporal information,
our method can also be applied on other data types that rely on the BOLD effect, e.g. fNIRS, but
also data types measuring electrical population activity directly. This makes it a good candidate for,
among other things, studying changing connectivity in neurodegenerative diseases, like Parkinson's or
Alzheimer's.
Conclusion
With the presented method we can estimate directed effective connectivity on a whole-brain scale.
Also we are able to detect whether connections are excitatory or inhibitory. The estimation is possible
based on zero-lag covariances, but can also be applied to frequency-resolved cross spectral densities.
Supporting information
Algorithm 1 Conjugate gradient descent
1: initialize U, B0
2: calculate gradient d and a ←− 1
3: calculate step size δ
4: U ←− exp(−δa)
5: while not converged do
6:
2 (d − d∗)
7:
a(cid:48) ←− a
calculate gradient a
β ←− (cid:104)a,a+a(cid:48)
(cid:105)
(cid:104)a(cid:48),a(cid:48)
(cid:105)
g ←− −a − βa(cid:48)
calculate step size δ of g
U ←− exp(δg)
11:
12: Linesearch(U, B0)
10:
8:
9:
14
Algorithm 2 Line Search with Armijo step size rule
1: function LINESEARCH(Uact, d, α)
2:
(cid:46) current estimate Uact, gradient d, initial-step length α
3:
4:
5:
6:
7:
8:
9:
10:
U ←− exp(−αd)
Q ←− U U
while Γ(UactB0) − Γ(QB0) (cid:62) α(cid:104)d, d(cid:105) do
U ←− Q
Q ←− U U
α ←− 2α
U ←− exp(−αd)
α ←− 0.5α
while Γ(UactB0) − Γ(U B0) (cid:62) 0.5α(cid:104)d, d(cid:105) do
return U, α
Figure 9: Histogram of estimated connection strengths taken from the reconstructed networks of all
seven subjects. The vertical lines show the thresholds for the excitatory and inhibitory connections,
respectively. Only a part of the histogram is shown, the actual range of values is between −0.99 and
2.54.
Figure 10: Performance of our new inference method on the NetSim dataset published by [34]. Other
methods have also been tested on these simulated data sets [31, 32, 33]. The x-axis represent the
indices of simulated data sets, as in the original publication. The y-axis shows the AUC and PRS of
our estimations. We estimated the connectivity for every individual subject and applied a threshold
of 50%, the resulting networks were then averaged over all available subjects/trials. Although the
networks considered in this paper cover a range of parameters, where we found that our method
performs sub-optimally (the networks are generally too small), it still performs reasonably well on
these synthetic data. We obtained average values for AUC and PRS of 0.94 and 0.79, respectively.
15
0.60.40.20.00.20.40.6Value050001000015000200002500030000Count0510152025Simulation number0.30.40.50.60.70.80.91.0ScoreArea under the ROC curvePrecision Recall Score Acknowledgments
Supported by the DFG (grant EXC 1086). The HPC facilities are funded by the state of Baden-
Wurttemberg through bwHPC and DFG grant INST 39/963-1 FUGG. We thank Uwe Grauer from
the Bernstein Center Freiburg as well as Bernd Wiebelt and Michael Janczyk from the Freiburg
University Computing Center for their assistance with HPC issues.
References
[1] Li B, Razi A, Friston KJ. Editorial: Mapping Psychopathology with fMRI and Effective Con-
nectivity Analysis. Front. Hum. Neurosci. 2017; 11:151. doi:10.3389/fnhum.2017.00151.
[2] Friston KJ. Functional and effective connectivity: a review. Brain connectivity. 2011;1(1):13–36.
doi:10.1089/brain.2011.0008.
[3] Pernice, V, Staude, B, Cardanobile, S, Rotter, S. How structure determines correlations in
neuronal networks. PLoS Comput. Biol. 2011;7,. doi:10.1371/journal.pcbi.1002059,
[4] Havlicek M, Roebroeck A, Friston KJ, Gardumi A, Ivanov D, Uludag K.
On the im-
portance of modeling fMRI transients when estimating e ff ective connectivity : A dy-
namic causal modeling study using ASL data. NeuroImage. 2017;155(July 2016):217–233.
doi:10.1016/j.neuroimage.2017.03.017.
[5] Stephan KE, Friston KJ. Analyzing effective connectivity with fMRI. Wiley interdisciplinary
reviews Cognitive science. 2010;1(3):446–459. doi:10.1002/wcs.58.
[6] Smith JF, Pillai A, Chen K, Horwitz B.
Identification and validation of effective connectiv-
ity networks in functional magnetic resonance imaging using switching linear dynamic systems.
NeuroImage. 2010;52(3):1027–1040. doi:10.1016/j.neuroimage.2009.11.081.
[7] Freestone DR, Karoly PJ, Nesi´c D, Aram P, Cook MJ, Grayden DB.
effective connectivity via data-driven neural modeling.
doi:10.3389/fnins.2014.00383
Front. Neurosci. 2014;
Estimation of
8:383.
[8] Gilson M, Moreno-Bote R, Ponce-Alvarez A, Ritter P, Deco G. Estimation of Directed Effective
Connectivity from fMRI Functional Connectivity Hints at Asymmetries of Cortical Connectome.
PLoS computational biology. 2016. doi:10.1371/journal.pcbi.1004762.
[9] Ting CM, Seghouane AK, Member S, Salleh SH, Noor AM. Estimating Effective Connectivity
from fMRI Data Using Factor-based Subspace Autoregressive Models. IEEE Signal Processing
Letters. 2015;22(6):757–761.
[10] Roebroeck A, Formisano E, Goebel R. The identification of interacting networks in the brain
using fMRI : Model selection , causality and deconvolution. NeuroImage. 2011;58(2):296–302.
doi:10.1016/j.neuroimage.2009.09.036.
[11] Mehta-Pandejee G, Robinson PA, Henderson JA, Aquino KM, Sarkar S.
Inference of di-
rect and multistep effective connectivities from functional connectivity of the brain and
of relationships to cortical geometry.
Journal of Neuroscience Methods. 2017;283:42–54.
doi:10.1016/j.jneumeth.2017.03.014.
[12] Marrelec G, Krainik A, Duffau H, Doyon J, Benali H.
Partial correlation for func-
tional brain interactivity investigation in functional MRI. NeuroImage. 2006;32:228–237.
doi:10.1016/j.neuroimage.2005.12.057.
16
[13] Timme M, Casadiego J. Revealing networks from dynamics: an introduction. Journal of Physics
A: Mathematical and Theoretical. 2014;47(34):343001.
[14] Gates KM, Molenaar PCM. Group search algorithm recovers effective connectivity maps for
individuals in homogeneous and heterogeneous samples. NeuroImage. 2012; 63(1):310 – 319.
doi:10.1016/j.neuroimage.2012.06.026.
[15] Ramsey J, Glymour M, Sanchez-Romero, Glymour C. A million variables and more: the Fast
Greedy Equivalence Search algorithm for learning high-dimensional graphical causal models, with
an application to functional magnetic resonance images. International Journal of Data Science
and Analytics. 2016. doi:10.1007/s41060-016-0032-z.
[16] Pernice V, Rotter S. Reconstruction of sparse connectivity in neural networks from spike train
covariances. Journal of Statistical Mechanics: Theory and Experiment. 2013;2013(03):P03008.
doi:10.1088/1742-5468/2013/03/P03008.
[17] Pearl J. Causality: Models, Reasoning, and Inference. New York, NY, USA: Cambridge University
Press; 2000.
[18] Rebane G, Pearl J. The recovery of causal poly-trees from statistical data. Proceedings of the
Third Workshop on Uncertainty in AI; 1987; 222–228
[19] Candes EJ, Tao T. Decoding by Linear Programming. IEEE Trans Inf Theor. 2005;51(12):4203–
4215. doi:10.1109/TIT.2005.858979.
[20] Abrudan T, Eriksson J, Koivunen V.
mization Under Unitary Matrix Constraint.
doi:doi:10.1016/j.sigpro.2009.03.015.
Conjugate Gradient Algorithm for Opti-
2009;89:1704–1714.
Signal Processing.
[21] Wen Z, Yin W. A feasible method for optimization with orthogonality constraints. Mathematical
Programming. 2013;142(1):397–434. doi:10.1007/s10107-012-0584-1.
[22] Horn RA, Johnson CR. Matrix Analysis. Cambridge University Press, Cambridge, MA, 1985
[23] Welvaert M, Rosseel Y. A review of fMRI simulation studies. PLoS ONE. 2014;9(7):e101953.
doi:10.1371/journal.pone.0101953.
[24] Gardiner CW. Handbook of stochastic methods for physics, chemistry and the natural sciences.
vol. 13 of Springer Series in Synergetics. 3rd ed. Berlin: Springer-Verlag; 2004.
[25] Gillespie DT. Exact numerical simulation of the Ornstein-Uhlenbeck process and its integral.
Phys Rev E. 1996;54:2084–2091. doi:10.1103/PhysRevE.54.2084.
[26] Friston KJ, Fletcher P, Josephs O, Holmes A, Rugg MD, Turner R. Event-Related fMRI: Char-
acterizing Differential Responses. NeuroImage. 1998;7(1):30–40. doi:10.1006/nimg.1997.0306.
[27] Glover GH. Deconvolution of Impulse Response in Event-Related BOLD fMRI1. NeuroImage.
1999;9(4):416–429. doi:10.1006/nimg.1998.0419.
[28] Asslander J, Zahneisen B, Hugger T, Reisert M, Lee HL, LeVan P, Hennig J. Single shot
whole brain imaging using spherical stack of spirals trajectories. NeuroImage. 2013;73:59–70.
doi:10.1016/j.neuroimage.2013.01.065.
[29] LeVan P, Akin B, Hennig J. Fast imaging for mapping dynamic networks. NeuroImage. 2017
doi:10.1016/j.neuroimage.2017.08.029.
[30] Glover GH, Li TQ, Ress D.
Image-based method for retrospective correction of physiological
motion effects in fMRI: RETROICOR. Magnetic Resonance in Medicine. 2000;44(1):162–167.
doi:10.1006/nimg.1998.0419.
17
[31] Nie L, Yang X, Matthews PM, Xu ZW, and Guo YK. Inferring functional connectivity in fMRI
using minimum partial correlation. International Journal of Automation and Computing, 14:371–
385, 2017. doi:10.1007/s11633-017-1084-9
[32] Ryali S, Chen T, Supekar K, Tu T, Kochalka J, Cai W, and Menon V. Multivariate dynami-
cal systems-based estimation of causal brain interactions in fMRI: Group-level validation using
benchmark data, neurophysiological models and human connectome project data. Journal of
Neuroscience Methods, 268:142–153, 2016. doi:10.1016/j.jneumeth.2016.03.010
[33] Hyvarinen A and Smith SM. Pairwise likelihood ratios for estimation of non-gaussian structural
equation models. Journal of Machine Learning Research, 14(1):111–152, 2013.
[34] Smith SM, Miller KL, Salimi-Khorshidi G, Webster M, Beckmann CF, Nichols TE, Ramsey
JD, Woolrich MW. Network modelling methods for fMRI. NeuroImage 2011;54;2:875–891.
doi:10.1016/j.neuroimage.2010.08.063.
[35] Zaghlool SB and Wyatt CL. Missing data estimation in fMRI dynamic causal modeling. Frontiers
in Neuroscience, 8:191, 2014. doi:10.3389/fnins.2014.00191
18
|
1604.00087 | 2 | 1604 | 2017-03-30T16:50:50 | Automated point-neuron simplification of data-driven microcircuit models | [
"q-bio.NC"
] | A method is presented for the reduction of morphologically detailed microcircuit models to a point-neuron representation without human intervention. The simplification occurs in a modular workflow, in the neighborhood of a user specified network activity state for the reference model, the "operating point". First, synapses are moved to the soma, correcting for dendritic filtering by low-pass filtering the delivered synaptic current. Filter parameters are computed numerically and independently for inhibitory and excitatory input using a Green's function approach. Next, point-neuron models for each neuron in the microcircuit are fit to their respective morphologically detailed counterparts. Here, generalized integrate-and-fire point neuron models are used, leveraging a recently published fitting toolbox. The fits are constrained by currents and voltages computed in the morphologically detailed partner neurons with soma corrected synapses at three depolarizations about the user specified operating point. The result is a simplified circuit which is well constrained by the reference circuit, and can be continuously updated as the latter iteratively integrates new data. The modularity of the approach makes it applicable also for other point-neuron and synapse models. The approach is demonstrated on a recently reported reconstruction of a neocortical microcircuit around an in vivo-like working point. The resulting simplified network model is benchmarked to the reference morphologically detailed microcircuit model for a range of simulated network protocols. The simplified network is found to be slightly more sub-critical than the reference, with otherwise good agreement for both quantitative and qualitative validations. | q-bio.NC | q-bio |
Automated point-neuron simplification of data-driven
microcircuit models
Christian Rossert 1,∗, Christian Pozzorini 2, Giuseppe Chindemi 1, Andrew P.
Davison 3, Csaba Eroe 1, James King 1, Taylor H. Newton 1, Max Nolte 1, Srikanth
Ramaswamy 1, Michael W. Reimann 1, Willem Wybo1,2, Marc-Oliver Gewaltig 1,
Wulfram Gerstner2, Henry Markram 1,4, Idan Segev 5,6, Eilif Muller 1,∗
1Blue Brain Project, ´Ecole Polytechnique F´ed´erale de Lausanne (EPFL) Biotech
Campus, Geneva, Switzerland
2Laboratory of Computational Neuroscience (LCN), Brain Mind Institute, ´Ecole
Polytechnique F´ed´erale de Lausanne, Lausanne, Switzerland
3Neuroinformatics group, Unit´e de Neurosciences, Information et Complexit´e,
Centre National de la Recherche Scientifique, Gif sur Yvette, France
4Laboratory of Neural Microcircuitry, Brain Mind Institute, ´Ecole Polytechnique
F´ed´erale de Lausanne, Lausanne, Switzerland
5Department of Neurobiology, Alexander Silberman Institute of Life Sciences, The
Hebrew University of Jerusalem, Jerusalem, Israel
6The Edmond and Lily Safra Centre for Brain Sciences, The Hebrew University of
Jerusalem, Jerusalem, Israel
March 31, 2017
Abstract
A method is presented for the reduction of morphologically detailed microcircuit models
to a point-neuron representation without human intervention. The simplification occurs in
a modular workflow, in the neighborhood of a user specified network activity state for the
reference model, the "operating point". First, synapses are moved to the soma, correcting for
dendritic filtering by low-pass filtering the delivered synaptic current. Filter parameters are
computed numerically and independently for inhibitory and excitatory input using a Green's
function approach. Next, point-neuron models for each neuron in the microcircuit are fit to
their respective morphologically detailed counterparts. Here, generalized integrate-and-fire
point neuron models are used, leveraging a recently published fitting toolbox. The fits are
constrained by currents and voltages computed in the morphologically detailed partner neurons
with soma corrected synapses at three depolarizations about the user specified operating point.
The result is a simplified circuit which is well constrained by the reference circuit, and can
be continuously updated as the latter iteratively integrates new data. The modularity of the
approach makes it applicable also for other point-neuron and synapse models.
The approach is demonstrated on a recently reported reconstruction of a neocortical
microcircuit around an in vivo-like working point. The resulting simplified network model
1
is benchmarked to the reference morphologically detailed microcircuit model for a range of
simulated network protocols. The simplified network is found to be slightly more sub-critical
than the reference, with otherwise good agreement for both quantitative and qualitative
validations.
Comments on version 2
The following changes have been made to the document since version 1: The soma-synaptic filter
fitting approach using PRAXIS has been replaced by a new method to directly extract the filters
for each dendritic compartment using a Green's functions approach. Methods Section 2.1, Results
Section 3.1 and Figures 1, 2, 4, 5 have been updated to reflect these changes. Furthermore the
Discussion has been updated to incorporate the new findings on reduction of post-synaptic filter
variability.
1 Introduction
To understand the brain, it could be said that we must simultaneously appreciate its daunting
complexity, and grasp its essential mechanisms. While modern experimental neuroscience offers us
a perpetually expanding view of the former, there remain major barriers to achieving the latter,
an integrated and synthesized view of pertinent experimental facts for the functional principles of
brain systems. A first-draft integrated view of a piece of the neocortex as recently been reported by
an international collaboration (Markram et al., 2015), resulting in precisely defined mathematical
models accounting for much of the known cellular and synaptic anatomy and physiology of this
part of the brain. A method for synthesizing the resulting mathematical models to a minimal form
is, thus formulated, an ill-posed problem. Minimal for what purpose?
On the other hand, (spiking) point-neuron network models are widely used in theoretical
neuroscience to describe various functions of the brain (Thorpe, 2002; Maass et al., 2004; Eliasmith
et al., 2012; Zenke et al., 2015). They are easy to analyze and numerically light-weight, making
them suitable for real-time and accelarated execution on GPU-based and neuromorphic platforms
(Nageswaran et al., 2009; Fidjeland and Shanahan, 2010; Bruderle et al., 2009; Galluppi et al.,
2010; Bruderle et al., 2011; Yavuz et al., 2016). However, these models for the most part have been
generated using adhoc assumptions rather than constraining them to biological data making them
disjunct from modern experimental neuroscience (Eliasmith and Trujillo, 2014). To bridge the
gap between these two areas of neuroscience, the goal of our simplification approach was to derive
a point-neuron network from the experimentally constrained morphologically detailed network
model in an automated, repeatable and quantitatively verifiable way. The result is a data-driven
point-neuron network model which can be continuously updated as the data-driven reference
continues to integrate the latest experimental data, without the need for human intervention.
While approaches exist to simplify the morphological complexity of biophysical neuron models
by reducing dendritic arbors (Marasco et al., 2012, 2013), the phenomenological approach pursued
here focuses on point neuron models as the specific target. The primary difficulty hereby is that
the synapses which excite or inhibit separate dendritic branches of the neuron have to be moved to
the somatic compartment, in a way that accounts for the transformations of the synaptic responses
due to the intermingling dendritic cable (Rall, 1967).
2
For our simplification procedure, we further employ the idea of the "operating point" a construct
widely used in e.g. nonlinear control theory where the complex system is linearized at the point
of interest (Slotine et al., 1991). In our case, the point of interest is the detailed network during
in vivo-like conditions and background activity. We use this in vivo network state to constrain
synaptic correction factors and to extract parameters for the point neuron models using synaptic
current and membrane potential data of individual neurons.
To benchmark the approach, we examine the simplified network by repeating a barrage of in
silico experiments used for the validation of the reference morphologically detailed network model
(Markram et al., 2015). Besides providing unprecedented validation of the point neuron network
model, this approach is fundamentally important, as it tells us if and when a certain property is
lost during the simplification pipeline.
The detailed network model further serves as a reference model for point neuron models and the
automated simplification pipeline makes it easy to iteratively update the simplified model as new
data is integrated into the detailed circuit. Our approach will further allow us bridge from point
neurons to population density and mean-field models in a quantitative manner, allowing dynamical
systems and phase-plane analysis of the dynamics of data-driven networks.
To begin, we present a phenomenological method for soma-synaptic replacement and correction.
We show examples of this procedure for individual neurons and report on the quality of our
approach. Next, we show how detailed soma-synaptic corrected networks can be used to constrain
point neuron models for each individual neuron in the network and we report on the results of
constraining Generalized Integrate-and-Fire (GIF) models (Pozzorini et al., 2015). Finally we
elaborate in detail on the validation of the simplified GIF point neuron network by comparing it to
the detailed network based on synaptic physiology, synchronous-asynchronous spectrum of network
states, sensory-evoked responses, temporal structure of in vivo spontaneous activity and spatial
correlations.
2 Methods
The automated point-neuron simplification procedure is modularized into two main steps: soma-
synaptic correction and neuron simplification (Figure 1). In the "soma-synatic correction" step, all
synapses on dendritic locations are moved to the soma while accounting for dendritic attenuation and
delay (dendritic filtering) by applying a numerically computed corrective filter that approximately
preserves the effective post-synaptic response (EPSP) of each synapse. In the "neuron simplification"
step, the total somatic stimulation- and synaptic currents and voltage responses for soma-synaptic
corrected network simulations are used to constrain Generalized Integrate-and-Fire (GIF) models
for each individual neuron in the network (Pozzorini et al., 2015).
2.1 Soma-synaptic correction
To account for the dendritic filtering of synaptic responses which will be lost when moving them to
the soma, all excitatory and inhibitory synapse models were extended by a simple low-pass filter of
the synaptic current, isyn, of the form
isyn soma = −isyn soma + w · isyn
τ
d
dt
(1)
3
Figure 1: Automated point-neuron simplification procedure. The procedure of automated point-neuron
simplification can be separated into two main parts: 1. soma-synaptic correction and 2. neuron simplification.
where w/τ and τ are the amplitude and time constant of the filter respectively. This filter
approximates the effective attenuation and delay that a synaptic current would experience when
flowing from its dendritic location to the soma (Rall, 1967; Berger et al., 2001; Williams and
Stuart, 2002; Nevian et al., 2007). The filter parameter estimation was however not conducted
in the passive neuron, but during the replay of post synaptic activation coming from a network
simulation mimicking physiological depolarization levels (100% rheobase) and extracellular calcium
([Ca2+] = 1.25 mM) (Figure 2A) (Markram et al., 2015). Parametrizing the dendritic filtering
during this bombardment of synaptic input, resulting in the so-called "high-conductance state"
(Destexhe et al., 2003), approximates the neuron's operating point in vivo, and therefore accounts
also for first order effects of the interaction of dendritic non-linearities with on-going synaptic
activity, and their role in altering dendritic properties such as input resistance, and shunting
inhibition (London and Hausser, 2005).
In the detailed network model synaptic connections between different neuron types have
been grouped into different synaptic types being comprised of different synaptic properties, e.g.
excitatory/inhibitory, ratio of NMDA/GABAB, synaptic decay time constants etc. to allow for
variation in reported synaptic properties (Markram et al., 2015). The maximum number of different
synaptic types is 12 with a distribution of 7.22 ± 2.6.
Ideally the parameters w and τ have to be supplied for each individual synapse. However,
to limit the computational cost of filter estimation, filter parameters for the same synaptic type
and electrical compartment were only extracted once: for each synaptic type a "probe" synapse
S (t), Figure 2A, blue trace) or at the soma (V S
having respective mean synaptic parameters was either activated at each dendritic compartment
(V D
S (t), red trace). In addition also the background
activity without probe synapse stimulation was recorded (VS(t), gray trace). After subtraction of
the background activity
V D(t) = V D
V S(t) = V S
S (t) − VS(t)
S (t) − VS(t).
the resulting traces (Figure 2B) were centered around the time of spike arrival at the synapse of
interest, so that the post-synaptic potential (PSP) starts at t = 0. Our goal was to extract the
parameters w and τ (1) that modify the synaptic current so that the PSP when the probe synapse
is at the soma optimally reproduces the PSP when the probe synapse is on the dendrites. To do so,
4
1. Soma-synaptic correction1b. Network simulationsynapses at soma+ correctionsynapses at somacontrolτw / τisynisyn soma1a. Compare cell responses 4. Compare network behavior 2. Neuron simplification3. GIF network simulationisyn somaGIFvsomacorrected synapses + GIF modelscreate GIF models for all cells in the network we first derived a full kernel κ in the Fourier domain that, when convolved with the synaptic input
current when the probe synapse is at the soma, yields an exact reproduction of the PSP when
the probe synapse is on the dendrites. Then we approximated this kernel as a single exponential
function
κ1(t) =
e− t
τ ≈ κ(t).
w
τ
(2)
Convolving the synaptic current of the probe synapse with this exponential function is equivalent
to (1).
The kernel in the Fourier domain. Following the Green's Function formalism (Koch, 1998;
Wybo et al., 2013, 2015), V D(t) is found by convolving the synaptic input current isyn with a
transfer kernel from dendrite to soma:
(cid:90) ∞
V D(t) =
ds KD(s) isyn(t − s),
for which we will use the shorthand:
0
V D(t) = KD(t) ∗ isyn(t).
Analogously, V S(t) is found by convolving the synaptic input current with a somatic input kernel:
V S(t) = KS(t) ∗ isyn(t).
(3)
Our goal was to find a kernel κ(t) that allows computing a modified somatic voltage when the
synapse of interest is at the soma:
V S(cid:48)(t) = KS(t) ∗ (κ(t) ∗ isyn(t)) ,
so that this voltage is as close as possible to the original voltage:
V S(cid:48)(t) ≈ V D(t).
(4)
(5)
To do so, we work in the Fourier domain, where the convolutions become simple multiplications
(Bracewell, 1978). Hence (3) becomes:
V S(ω) = KS(ω) isyn(ω),
(6)
while for (4) we get:
V S(cid:48)(ω) = KS(ω) κ(ω) isyn(ω).
Substituting (6) in this equation yields:
V S(cid:48)(ω) = κ(ω) V S(ω).
Converting (5) to the Fourier domain and combining it with the previous equation then allows the
kernel κ(ω) to be computed from the Fourier transforms of the two known voltage traces:
κ(ω) =
V D(ω)
V S(ω)
5
(7)
Figure 2: Soma-synaptic correction method. A: Resulting change in postsynaptic membrane potential
recorded at the soma when moving a single synapse i from its dendritic location on the apical dendrite (V D
S (t),
blue) at a L5 pyramidal cell (with path-distance of 230 µm from the soma) to the soma (V S
S (t), red). Baseline
activity of the latter cell without activation of synapse i (VS (t), gray). B: Postsynaptic potentials of dendritically
(V D(t), blue) and somatically (V S (t), red) activated synapse responses after subtraction of baseline activity. C:
Frequency-domain response (real (green) and imaginary (black) parts) of full dendritic filter (κ(ω), dashed lines)
extracted from baseline-subtracted dendritically and somatically activated postsynaptic responses and response of
fitted single exponential filter (κ1(ω), solid lines). D: Time-domain impulse response of single exponential filter
(κ1(ω), as extracted in C) with resulting parameters of τ = 4.65 ms and w=0.77.
The kernel as a single exponential. The Fourier transform of a single exponential function
ceat (a < 0) is given by the fraction
iω+a . Such a fraction was fitted to κ(ω) (Figure 2C, dashed
lines) as obtained from (7) in the Fourier domain by means of the vector fitting algorithm (Gustavsen
and Semlyen, 1999; Wybo et al., 2015) resulting in κ1(ω) (solid lines). In equation (1), τ is then
given by −1
a , resulting in the time domain representation κ1(t) (Figure 2D)
a and w is given by −c
c
Quality of filter correction To test synaptic filter correction, 100 excitatory neurons and
inhibitory interneurons were randomly chosen from the microcircuit and were simulated for 40 s
with replay of identical network activity under three configurations:
1. control, with synapses at default locations;
2. synapses moved to the soma without correction and
6
01020304050t (ms)58.859.259.660.0mVbaseline (VS)control, 230m (VDS)synapse at soma (VSS)01020304050t (ms)0.00.20.40.6mVcontrol - baseline (VD)synapse at soma - baseline (VS)100075050025002505007501000 (Hz)0.40.00.40.8real(())imag(())real(1())imag(1())01020304050t (ms)0.00.10.21(t), =4.65 ms, w=0.77ABCD 3. synapses moved to the soma while applying soma-synaptic filter correction as described above.
The quality has been estimated by comparing the coincidence factor Γ (Jolivet et al., 2004), and
the root mean square (RMS) voltage error of the non-spiking membrane potential response without
sodium channels, mimicking bath application of tetrodotoxin (in silico TTX), between configuration
1 and configuration 2 or 3.
Reduction of filter variability While the filter weight can be easily incorporated into the
synaptic weight of each synapse, the synaptic filter delay τ increases the complexity and variability
of the synapses which counteracts the goal to increase simulation efficiency. We therefore also tested
reduction of the variability of parameter τ by applying k-means clustering into sets of k = 1, 2, 3, 5
or 9 clusters using Lloyd's algorithm as implemented in the python function sklearn.cluster.KMeans
2.2 Neuron simplification
Soma-synaptic corrected neurons were simplified to Generalized Integrate-and-Fire (GIF) point-
neuron models about an in vivo-like operating point using a previously published automated
approach (Pozzorini et al., 2015). GIF model fitting was performed individually for each of 31346
neurons in the microcircuit, and constrained to individual cell responses during network simulations
mimicking physiological concentrations of extracellular calcium in vivo (1.25 mM) (Markram et al.,
2015). To broaden the generalization power of the GIF models, simulations for three different
depolarization levels (expressed in % rheobase of individual neurons) of 93%, 100% and 120% were
used as constraints for the fitting process (Figure 3). For each of these three depolarization levels,
four repetitions of 20 s were simulated leading to a total length of 240 s, and divided into 180 s
used as a training set for GIF model parameter extraction, and 60 s used as a validation set. To
generate current and voltage traces, network simulations were executed with all synapses moved to
the soma and applying the synaptic correction (configuration 3). This consequently allowed the
recording of the total input current (due to depolarization and and synaptic inputs) at the soma
for all depolarization levels (Figure 3A), together with the individual membrane potential response
of the cell (Figure 3B, green traces), which served as constraints for the GIF model fitting process.
To allow for fast simulations, the spike-triggered current η(t) [nA] and the spike-triggered firing
threshold γ(t) [mV] have been described by three exponential functions each. For the spike-triggered
current, the kernel η(cid:48)(t) extracted using rectangular basis functions Pozzorini et al. (2015) was
approximated by three exponentials. Timescales of γ(t) were fixed to 10, 50 and 250 ms and only
their amplitudes were optimized. The refractory period was fixed to Tref = 3 ms and λ0 set to 1 Hz
leading to a total of 14 extracted parameters per GIF model. After the GIF parameter extraction
procedure using the method described in (Pozzorini et al., 2015), the performance of each GIF
model was evaluated by estimating the log-likelihood L i.e. probability that the validation spike
train was produced by the model (Pozzorini et al. (2015), equation 20), and by comparing variance
explained V of membrane potential fluctuations between validation set and model (Pozzorini et al.
(2015), equation 27). Cells matching any of the following criteria
1. less than 5 spikes in the training set
2. Likelihood < 2 bits/spike
3. mean variance explained of membrane potential fluctuations of V < 40%
7
Figure 3: GIF Model data generation and fitting procedure. A: Recorded sum of synaptic currents during
network simulation with depolarization levels (% rheobase) of 93%, 100% and 120%. B: Comparison of responses
from detailed neuron (green) and GIF model (black) to corresponding depolarizing and synaptic currents (as shown
in A) for 93% (B1), 100% (B2) and 120% (B3).
4. ∆V > 2 mV
were excluded from the results. For the 6412 cells which were excluded according to these criteria,
they were replaced with a random GIF model drawn from successful optimizations from the same
morpho-electrical class.
2.3 Validations
2.3.1 Network simulations
Simulations of a single cortical column containing a total of 31346 neurons, as described in Markram
et al. (2015) were conducted for four configurations as sketched above: 1. control, 2. synapses at
soma, 3. synapses at soma + correction and 4. simplified GIF circuit (configuration 3. with all
detailed neurons replaced by GIF neurons).
In brief, all simulations were executed on the EPFL Blue Brain IV BlueGene/Q supercomputer
hosted at the Swiss National Supercomputing Centre in Lugano using NEURON (Hines and
Carnevale, 1997) with a time-step of 0.025 ms unless otherwise specified. Custom built software
assisted in the model setup, definition of experiment, and online recording of state-variables of
interest to disk, as described in Markram et al. (2015). Simulations results were analyzed in Python
using standard scientific python tools and also custom software.
2.3.2 Validation of synaptic physiology
In the reference circuit, the kinetics and dynamics for individual synaptic contacts and connections
were modeled by prescribing experimentally measured parameters, validated by comparing the
amplitude, rise time, latency, decay time constant, transmission failure rate, and the coefficient of
variation of resulting postsynaptic potentials (PSPs) against relevant in vitro data, and corrected
8
−0.10.1nADepol. 120%Depol. 100%Depol. 93%−63−62−61mV−60−40−20020mV0246810s−60−40−20020mVAB1B2B3synapses at soma + correctionisyn somavsomaGIF for extracellular [Ca2+] by downscaling unitary release probability, as described in (Markram et al.,
2015). In brief, excitatory synaptic transmission was modeled using both AMPA and NMDA
receptor kinetics (Jahr and Stevens, 1990; Tsodyks and Markram, 1997; Hausser and Roth, 1997;
Markram et al., 1998; Fuhrmann et al., 2002; Ramaswamy et al., 2012).
Inhibitory synaptic
transmission was modeled with a combination of GABAA and GABAB receptor kinetics (Khazipov
et al., 1995; De Koninck and Mody, 1997; Mott et al., 1999; Gupta et al., 2000). Stochastic
synaptic transmission was implemented as a two-state Markov model of dynamic synaptic release
by extending the Tsodyks-Markram dynamic synapse model to incorporate trial-to-trial variability
(Tsodyks and Markram, 1997; Fuhrmann et al., 2002).
Apart from the inclusion of the additional low-pass filter given in (1) to account for dendritic
filtering as described above, synapse models used in the simplified microcircuit were otherwise
unmodified from the reference.
To assess the correspondence of post-synaptic responses between the reference and simplified
circuit, on average, about 100 pairs of pre-postsynaptic neurons at inter somatic distances of
100 m were chosen for each of the 1941 biologically viable pathways. For individual pairs of
pre-postsynaptic neurons at resting membrane potential, the presynaptic neuron was stimulated
with a brief somatic current injection to elicit a unitary action potential, resulting in a postsynaptic
response. The postsynaptic response was determined as the average of 30 individual presynaptic
stimulation trials.
2.3.3 Sensory-evoked spike sequences
The analysis of simulated single-whisker deflection (Reyes-Puerta et al., 2015) in the GIF neuron
network (configuration 4) was performed on exactly the same cells as in the detailed reconstruction
Markram et al. (2015), with one exception as follows. One L5 Martinotti cell was added as an
exemplary OFF cell, since the previously chosen L5 Double Bouquet cell did not pass the significance
test (p > 0.05, as used for classification by Reyes-Puerta et al. (2015)) and is classified as not
responding (NR) in the simplified circuit (it was OFF in the original circuit, the p value changed
slightly from below 0.05 to above 0.05). All parameters are otherwise the same. The firing rate is
the average firing rate over 200 trials from −500 to 500 ms, relative to the stimulus.
All cells in the circuit were included in the analysis with the exception of 21 cells: There are
18 cells with a FR of more than 150 Hz in the simplified circuit, and another 3 cells with more
than 100 Hz in the original circuit. The former cells indicate an issue in the fallback solution for
failed GIF model fits, which will be addressed in future versions of the automated simplification
procedure.
2.3.4 Temporal structure of in vivo spontaneous activity
Luczak et al. (2007) report on the temporal structure of in vivo spontaneous activity in the
somatosensory cortex of anesthetized and awake rats. They show that during periods of global
activity (UP states), trios of neurons generate spike motifs with fixed temporal relationships that
occur more frequently than predicted by chance. Applying the same analysis techniques to the
reference circuit yielded qualitatively similar results Markram et al. (2015). We sought to investigate
whether neuron trios in our simplified circuit would also exhibit these properties. Simulations
were performed exactly as described in Markram et al. (2015) on the simplified circuit. Briefly,
to simulate UP state onsets, in the in vivo-like state (100% depolarization, [Ca2+] = 2 mM) we
9
stimulated the center-most 20 thalamic fibers of our circuit with single synchronous spikes after an
initial 1500 ms of relaxation time. This experiment was repeated 25 times, with each trial differing
only in terms of random seeds. From layer 5 of the circuit, we selected 50 cells (corresponding to
19600 unique cell trios) at random whose firing rate was greater than 3 Hz over the 500 ms active
period. Next, we concatenated the active period of each experiment, thus obtaining spike trains
for each neuron with a duration of 12.5 s. From the spike trains of the three cells in each trio, we
calculated all possible spike triplets (see Figure 12A), and extracted the mode (see Figure 12B)
for the count histogram of a representative neuron trio). Count histograms were smoothed with a
spatial Gaussian filter with a kernel of 10 ms. Additionally, we created a scatter plot of individual
neural latencies (defined as the average center of mass of the PSTH of a cell over a given UP state)
against the inter-spike intervals associated with the triplet mode of the trios in which that cell
participates. Finally, we investigated two null hypotheses, namely, that spikes occur at random (the
independent Poisson hypothesis) and that triplet structure is merely a result of increased overall
activity shortly following UP state onset (the common excitability hypothesis). To evaluate both
hypotheses, the analysis was repeated with shuffled spike train data. For the independent Poisson
hypothesis, we preserved the number of spikes per train, but regenerated the individual spike times
from a Poisson process. For the common excitability model, we preserved the individual spike times
and number of spikes per cell, but exchanged spikes at random between cells across time bins. The
procedures for both null hypotheses were repeated to obtain the mean and standard deviation over
multiple trials.
2.3.5 Spatial correlation
The spatial correlation between local groups of neurons as function of distance between groups was
investigated for the simplified circuit, and compared to the reference microcircuit model as described
in Markram et al. (2015). In brief, the minicolumns of the microcircuit were grouped into local
spatial clusters via a k-means clustering algorithm with a mean cluster size of 10 minicolumns. Mean
PSTHs for each cluster were then computed using a time bin of 5 ms. Pairwise cross-correlation
coefficients were calculated between all clusters. Distances between clusters refer to k-means centroid
distances. Exponential fits were obtained using the python scipy.curve fit routine.
2.3.6 Response reliability
We compared the reliability of neuron responses between the reference and simplified circuits when
stimulated by a single pulse stimulus delivered to 60 thalamocortical fibers, as assessed by delay
until the first spike after stimulation as described in Markram et al. (2015).
3 Results
3.1 Synaptic replacement and correction
Synaptic correction filters properties τ and w exhibited a systematic dependence on the distance
between soma and the original dendritic location, as shown exemplarily for Layer 5 thick-tufted
pyramidal cells, type 1 (L5 TTPC1) (Figure 4). As expected the delay τ was lowest at the soma
and reached values as high as 35 ms (A1,B1) in the apical dendrites. Weight w was high for filters
close to the soma but synaptic signals had to be reduced by more than 90% to account for signal
10
Figure 4: Example of individual filters for Layer 5 thick-tufted pyramidal cells, type 1 (L5 TTPC1).
A,B: Parameters of 10 L5 TTPC1 neurons plotted as color coded points onto the xy plane of individual compartment
locations. Location dependence for filter parameters τ (A1,B1) and w (A2,B2) shown separately for excitatory (A)
and inhibitory (B) synapses. C: Difference of filter parameters τ (C1) and w (C2) between excitatory and inhibitory
synapses for filters on the same compartment.
attenuation in the apical dendrites (A2,B2). While the difference for filter delay between excitatory
and inhibitory synapses on the same compartments was insignificant (∆τ = −0.088 ± 1.068 ms),
the attenuation was higher in filters for inhibitory synapses on the same compartment than for
excitatory synapses with ∆w = 0.062 ± 0.054. This justifies the need for separate filter properties
per synapse type.
To validate our procedure of soma-synaptic correction, the different configurations of replaced
uncorrected synapses (2) and replaced synapses + correction (3) were compared to the control case
(1) during replay of synaptic activity in single neurons. Direct examination of membrane potentials
revealed that, especially for pyramidal cells, the inhibition was often too high and consequently
the activity of these cells was reduced since inhibition was directly affecting the soma (Figure 5A,
compare blue control and red uncorrected traces). Application of the synaptic correction filter
qualitatively restored the membrane and spiking dynamics (Figure 5A), compare blue to green
traces).
Direct comparison of simulations without sodium currents (in silico TTX) (Figure 5B1) showed
that the error without soma-synaptic correction ("no correction") was higher for excitatory (red)
than for inhibitory interneurons (blue). This error consequently lead to a mean spike coincidence
smaller than 50% in excitatory cells (Figure 5B2). Applying the soma-synaptic correction for each
11
distributionExcitatorysynapses200mInhibitorysynapsesw distributionEx-Inhsynapses05101520253035 (ms)0.00.20.40.60.81.0weight w25201510505difference (ms)0.20.10.00.10.20.30.40.5difference weight wA1B1A2B2C1C2 compartment ("individual filters") as estimated above greatly reduced the RMS error and increased
the spike coincidence factor (Γ) especially for excitatory neurons (Figure 5B).
To further simplify the model and reduce synaptic variability we analyzed the grouping of filter
delays τ using k-means clustering (Figure 5B, "k-means(τ )"). This analysis revealed that only a
small number of different filter delays (k = 3 − 5) per synapse type were effective to obtain small
RMS voltage errors and high coincidence factors. Surprisingly just using one cluster ("k-means(τ )
k=1") which is equivalent to applying the mean delay τ for all synapses was less effective than not
using any filter delay correction at all ("no delay τ ") which suggest that fast inputs close to the
soma have been overcompensated in the former case.
To be able to significantly increase simulation speed also the variability of the synapses themselves
have to be reduced. Here the main variability in the detailed model (Markram et al., 2015) lies in
a normal distributed decay time constant (dtc) of the post-synaptic process of the fast synaptic
transmitters (AMPA or GABAA). Reducing this variability and using mean decay time constant for
each of the synapse types ("mean(dtc)") in combination with individual filters for each compartment
increased RMS and decreased Γ. Combining this with three soma-synaptic filters ("mean(dtc),
k=3") slightly increases the error again while having the potential to greatly increase simulation
speed by decreasing the number of post-synaptic processes to be simulated by two orders of
magnitude (from thousands of synapses per cell to a maximum of 36 = 3 · 12 (maximum synaptic
types)).
3.2 Neuron simplification and validation
GIF model fitting was performed individually for each of the 31346 neurons in the reference
microcircuit, and constrained to individual cell responses during network simulations mimicking
physiological concentrations of extracellular calcium in vivo (1.25 mM) (Markram et al., 2015). In
general GIF parameter extraction (Figure 6A) showed large likelihood (A1) and good explained
variance of the membrane potential fluctuations (A2) for the majority of cells. Since most of
the optimized cells were deterministic, this can also be observed in the low extracted level of
stochasticity ∆V (A3). Only for the stochastic stuttering electrical types (cSTUT, bSTUD and
dSTUT) was a higher ∆V observed (Figure 6B1).
In general the electrical cell types (e-types) showed a large variability in extracted parameters
(Figure 6B1). Consistent with our expectations, the extracted spike-triggered current η(t) (Figure
6B2, left column) was most pronounced in adapting e-types (cADPyr, cACint, bAC) while it
was fast in irregular (cIR, bIR) and most non-adapting types (bNAC, dNAC). Interestingly some
non-adapting e-types (cNAC) showed a large variability in η(t). Furthermore, stuttering cells
of type bSTUT and dSTUT showed longest time constants of spike-triggered currents. For the
spike-triggered firing threshold, all e-types showed a large variability in γ(t), with bIR, dNAC and
dSTUT types having the largest mean thresholds (Figure 6B2, right column).
3.3 Synaptic physiology validation
To validate the post-synaptic potential (PSP) responses of the simplified circuit conform to the
reference model, we compared the average PSPs between pairs of neurons of specific pre- and
post-synaptic m-types. Compared to the reference model, shown in Figure 7A, the simplified model
PSP amplitudes, shown in Figure 7B, are qualitatively similar. Figure 7C reveals similar PSP
amplitude distributions between the simplified and reference models, but with simplified circuit
12
Figure 5: Validation of synaptic filter correction. A: Comparison of membrane potential responses for a
L5 TTPC1 cell during post-synaptic replay in the three configurations of control (green), synapses moved to the
soma (red) and applied soma-synaptic correction (green). Action potentials in are cut. B: Comparison of mean
and standard deviation of RMS membrane potential error (B1) and Γ coincidence factor (B2) for 100 excitatory
neurons (red) and 100 inhibitory interneurons (blue) for different soma-synaptic correction configurations. Results
shown using no correction, using individual filters for each synapse, k-means binning of filter delays τ , using no delay
correction, using mean decay time constant for each synapse and in combination with k-means clustering (k=3).
13
010002000ms6050mV01mVmean(dtc), k=3mean(dtc)no delay k-means() k=1k-means() k=2k-means() k=3k-means() k=5k-means() k=9individual filtersno correctionRMS "TTX" voltage error0.20.40.60.81mean coincidence factorAB1B2controlall synapses at somaall synapses at soma + correction Figure 6: Results of GIF Model fitting procedure. A: Distribution of likelihood L (bits/spike; see Methods)
(A1), explained variance of the membrane potential fluctuations V (%) (A2), and extracted level of stochasticity
parameter ∆V (mV) (A3). B1: Mean and standard deviation for membrane capacity C (nF ), reversal potential EL
(mV), membrane conductance gL (µS), threshold baseline V ∗
T (mV), voltage reset Vreset (mV), and ∆V , L and V
grouped per electrical type of target neuron. B2: Mean and standard deviation of estimated filters for spike-triggered
current η(t) (left) and spike-triggered firing threshold γ(t) (right) grouped per electrical type as indicated in A1.
14
024681012Likelihood (bits/spike)0300600020406080100²V (%)050010000.00.20.40.60.81.01.21.41.6¢V050010000.10.3C (nF)dSTUTdNACbIRbSTUTbNACbACcIRcSTUTcNACcACintcADpyr−80−70EL (mV)0.000.02gL (uS)−64−58V¤T (mV)−75−55Vreset (mV)0.41.2¢V (mV)48L (bits/spike)70100²V (%)0.00.2´(t) (nA)-0.50.5°(t) (mV)0.00.2-110.00.2-1001000.00.2-110.00.2-10100.00.2-2e42e40.00.2-2002000.00.2-220.00.2-1e101e100.00.2-1e151e15050ms0.00.20100200ms-1e151e15A1A2A3B1B2 exhibiting a longer tail. Figure 7D shows the numerical correspondence of average PSPs for the
reference and simplified circuit. While there is as a good degree of correspondence, some pathways
are overly strong in the simplified circuit.
3.4 Validation of network properties
3.4.1 Synchronous-asynchronous spectrum of network states
To evaluate the effect of the various stages of simplification on the basic network behavior, the
microcircuit was simulated under various depolarization and calcium conditions (Figure 8). In
addition to the three previous configurations 1. control (blue), 2. synapses moved to the soma
without correction (red) and synapses moved to the soma + correction (green) the fourth condition
examined was the network with relocated synapses and GIF replacement of all cells (black).
The most prominent change seen from the raster plots with 100% threshold depolarization is
that simply replacing the synapses to the soma leads to a shift in critical calcium concentration for
network oscillations from 1.3 to 1.35 mM (compare first row (blue) to second row (red)). Furthermore
the initial frequency of oscillations was reduced but increases back to 1 Hz like in the control case
for 1.4 mM calcium. This change in emergent synchronous behavior can however be recovered
completely when accounting for dendritic filtering by applying soma-synaptic correction (third row,
green). Even exchanging all neurons by the GIF equivalents retains the basic synchronous behavior
(last row, black) as seen in the control.
Further investigating the cumulated spiking activity (PSTH) for different depolarization condi-
tions (93, 100 and 107%) (Figure 8B) reveals that without any synaptic corrections (red traces)
the network shows prominent fast oscillations especially during low depolarization (first row) and
lower calcium ([Ca2+]o = 1.2 to 1.35 mM). These fast oscillations are however dominated by slow
oscillations during high calcium (1.4 mM). Using soma-synaptic correction (green traces), a close
match was achieved to the control configuration (blue traces) for all depolarization conditions
leading to an almost exact match of bursting activity. While this match in bursting activity was
retained in the GIF neuron network (black traces), it did not generalize well to low levels of
depolarization (93%) leading to a higher baseline activity than control. This difference can also be
observed during oscillatory network states in all depolarization levels where the activity of the GIF
network does not decay to zero between bursts. These discrepancies highlight the need to pick an
operating point for the fitting of point-neuron models to a morphologically detailed reference. For
the targeted operating point of the simplification at in vivo-like conditions ([Ca2+]o = 1.25 mM
and 100% depolarization), the GIF neuron network shows a good agreement of baseline network
activity to the control case, however without an initial equilibratory transient as seen in the detailed
network.
3.4.2 Sensory-evoked responses
To assess the stimulus response properties, several simulations from a study on the reference model
where reproduced in the simplified model Markram et al. (2015). The PSTHs in response to
simulated single-whisker deflection (see Methods) in the GIF neuron network, depicted in Figure
9A, are qualitatively very similar to the results of the detailed circuit. The PSTHs of the same
eight cells as shown in Markram et al. (2015) are depicted, however, one L5 Martinotti cell was
added as an exemplary OFF cell, since the chosen L5 Double Bouquet cell for the reference model
no longer not passed the significance test (p > 0.05, as used for classification by Reyes-Puerta et al.
15
Figure 7: Validating synaptic physiology. A: Matrix representation of the average PSP amplitude for synaptic
connections (pathways) formed between the 55 m-types (1,941 biological viable connections) in the reference
microcircuit. B: Same as in A, but for the simplified microcircuit. C1: Distribution of PSP amplitudes. Histogram
of the average PSP amplitude for 1,941 biologically viable connections in the reference microcircuit. C2: Same as in
C1, but for the simplified microcircuit. D: Comparison of PSP amplitudes. Comparison of average PSP amplitudes
between the reference and simplified microcircuits. Error bars show the standard deviation (SD). Dashed line shows
the identity line.
16
L1DACL1NGC-DAL1NGC-SAL1HACL1LACL1SACL23PCL23MCL23BTCL23DBCL23BPL23NGCL23LBCL23NBCL23SBCL23ChCL4PCL4SPL4SSL4MCL4BTCL4DBCL4BPL4NGCL4LBCL4NBCL4SBCL4ChCL5TTPC1L5TTPC2L5UTPCL5STPCL5MCL5BTCL5DBCL5BPL5NGCL5LBCL5NBCL5SBCL5ChCL6TPCL1L6TPCL4L6UTPCL6IPCL6BPCL6MCL6BTCL6DBCL6BPL6NGCL6LBCL6NBCL6SBCL6ChCL1DACL1NGC-DAL1NGC-SAL1HACL1LACL1SACL23PCL23MCL23BTCL23DBCL23BPL23NGCL23LBCL23NBCL23SBCL23ChCL4PCL4SPL4SSL4MCL4BTCL4DBCL4BPL4NGCL4LBCL4NBCL4SBCL4ChCL5TTPC1L5TTPC2L5UTPCL5STPCL5MCL5BTCL5DBCL5BPL5NGCL5LBCL5NBCL5SBCL5ChCL6TPCL1L6TPCL4L6UTPCL6IPCL6BPCL6MCL6BTCL6DBCL6BPL6NGCL6LBCL6NBCL6SBCL6ChCPostsynaptic m-typePresynaptic m-typeL1DACL1NGC-DAL1NGC-SAL1HACL1LACL1SACL23PCL23MCL23BTCL23DBCL23BPL23NGCL23LBCL23NBCL23SBCL23ChCL4PCL4SPL4SSL4MCL4BTCL4DBCL4BPL4NGCL4LBCL4NBCL4SBCL4ChCL5TTPC1L5TTPC2L5UTPCL5STPCL5MCL5BTCL5DBCL5BPL5NGCL5LBCL5NBCL5SBCL5ChCL6TPCL1L6TPCL4L6UTPCL6IPCL6BPCL6MCL6BTCL6DBCL6BPL6NGCL6LBCL6NBCL6SBCL6ChCL1DACL1NGC-DAL1NGC-SAL1HACL1LACL1SACL23PCL23MCL23BTCL23DBCL23BPL23NGCL23LBCL23NBCL23SBCL23ChCL4PCL4SPL4SSL4MCL4BTCL4DBCL4BPL4NGCL4LBCL4NBCL4SBCL4ChCL5TTPC1L5TTPC2L5UTPCL5STPCL5MCL5BTCL5DBCL5BPL5NGCL5LBCL5NBCL5SBCL5ChCL6TPCL1L6TPCL4L6UTPCL6IPCL6BPCL6MCL6BTCL6DBCL6BPL6NGCL6LBCL6NBCL6SBCL6ChCPostsynaptic m-typePresynaptic m-typeAB PSP amplitude [mV]605040302010 PSP amplitude [mV]01020304050607010410310210110010-110-2No. of pathways PSP amplitude [mv]No. of pathways PSP amplitude [mv]C1C2PSP amplitude [mV] simplified microcircuitPSP amplitude [mV] reference microcircuitD010203040506070Simplified microcircuit10-210-1100101102Reference microcircuitReference microcircuitSimplified microcircuit10210110010-110-260504030201010410310210110010-110-2 Figure 8: Comparison of synchronous-asynchronous spectrum of network states. A: Spike raster plots
of the detailed control network (blue, first row), network with synapses moved to the soma (red, second row), network
with synapses moved to the soma and soma-synaptic correction applied (green, third row), and network with all
cells replaced by GIF models (black, fourth row). Exploration of emergent network states for a range of [Ca2+]o
from 1.2 to 1.4 mM at one depolarization level of 100% threshold. B: Exploration of three different depolarization
levels: 93% (first row), 100% (second row) and 107% (third row)) and range of different [Ca2+]o concentrations
(horizontal axis) as in A, assessed by peristimulus time histogram (PSTH). As before, comparison of four different
configurations: control network (blue), all synapses replaced to soma (red), soma-synaptic correction applied (green)
and GIF network (black).
17
VIVIVIIIIIIcontrol[Ca2+]O 1.2 mM1.25 mM1.3 mM1.35 mM1.4 mMVIVIVIIIIIIsomaVIVIVIIIIIIsoma+filter0123t [s]VIVIVIIIIIIGIF0123t [s]0123t [s]0123t [s]0123t [s]024Depol. 93%[Ca2+]O 1.2 mM1.25 mM1.3 mM1.35 mM1.4 mM048Depol. 100%0123t [s]0816Depol. 107%0123t [s]0123t [s]0123t [s]0123t [s]AB (2015)) in the simplified network. A quantitative PSTH correlation analysis confirmed that there
is a high positive linear correlation between the PSTHs in the detailed circuit (as computed in
Markram et al. (2015) and the PSTHs in the GIF neuron network (see Figure 9E). The distribution
of cell-type firing rates, in terms of firing rate before versus after the stimulus, is qualitatively
similar too (see Figure 9B), as are the first spike response latencies (Figure 9C). While overall
mean firing rates are linearly correlated, some inhibitory cells have a sharply increased firing rate
compared to the detailed microcircuit (see Figure 9D), which indicates a fit-failure for certain cell
types that will need to be fixed in future versions of the GIF model fitting module of the workflow.
Response variability as assessed by the SD of the time-to-first-spike distribution across stimulation
trials reveal good agreement between the reference and simplified models, as shown in Figure 10.
Next we assessed and compared stimulus response curves under in vivo-like between the reference
and simplified circuits. Stimulating an increasing number of central thalamocortical fibers with
a synchronous volley of spikes revealed initial central responses (max of PSTH within the 30 ms
after stimulus) that were qualitatively similar, but 25% reduced in the simplified circuit compared
to the reference, over the range of stimuli and for [Ca2+] in the range 1.1 − 1.3 mM (see Figures
11A and 11C). Propagation of evoked responses to the periphery was markedly reduced in the
simplified circuit. As shown in Figures 11B and 11A, the peak response at the outermost neurons
in the reference circuit showed a gradual increase with number of fibers stimulated (minicols) and
increasing [Ca2+]. In contrast, the simplified circuit showed almost no dependence of the peripheral
response on number of fibers stimulated (minicols) for [Ca2+] < 1.25 mM.
3.4.3 Temporal structure of in vivo spontaneous activity
As described in Markram et al., 2015, simulations of the reference model revealed temporal structure
comparable to recent findings in vivo Luczak et al. (2007). To assess if such temporal structure
was preserved in the simplified network, the same analysis were performed for the simplified circuit.
As shown in Figure 12, the observed precisely repeating triplets are also found to be present in the
simplified circuit, albeit with significantly reduced frequency. Increasing the extracellular calcium
level to within 0.02 mM of the critical point for the simplified network markedly increased the
frequency to exceed those levels found in the reference circuit, indicating disagreement between
the simplified and reference circuit on this validation are due to the minor differences in level of
criticality between the two. This highlights the need to develop a calibration technique to make
network quantitatively compatible under validations which are criticality-level dependent.
3.4.4 Spatial correlation
The previously described spatial dependence of PSTH correlations between clusters of neurons in
the reference circuit (Markram et al., 2015), is qualitatively replicated in the GIF neuron network
(see Figure 13, left), but the simplified circuit has a significantly faster fall-off of spatial correlations
(τ = 176 ± 4 µm versus 284 ± 4 µm for the reference), and a significantly reduced correlation overall
(see Figure 13, middle) for all calcium conditions except for in the immediate neighborhood of the
critical point. Plotting τ as a function of extracellular calcium concentration reveals (Figure 13,
right) that only for calcium concentrations within 0.02 mM of the critical point, could the τ of
the reference circuit at 1.25 mM be approached. Taken together, these simulations indicate that
the simplified circuit appears to have genuinely reduced firing rate correlation between clusters of
neurons compared to the reference circuit.
18
Figure 9: Sensory-evoked spike sequences A: Different cell response-types to simulated single-whisker deflection.
The subplots depict the activity of the same cells as in the original detailed microcircuit experiment, showing PSTH
and the aligned raster plot. After the stimulus, neurons either increased their firing rate (ON cells), showed no
significant change in firing rate (NR cells), or decreased their firing rate (OFF cells). An additional OFF cell was
added, since the depicted L5DBC is now classified as NR (p > 0.05). B: Comparison of mean firing rates before
and after whisker deflection plotted in logarithmic scale (2630 excitatory and 550 inhibitory neurons, same cells as
in original microcircuit experiment). Empty symbols show NR cells, and filled symbols represent neurons showing
significantly different (p < 0.05) activity (ON and OFF cells). C: Mean first-spike latencies of inhibitory (Inh.) and
excitatory (Exc.) neurons to simulated whisker deflection (first spike occurrence within 30 ms after stimulation,
mean over 200 trials, for all 31346 neurons in the stimulated column) Each box plot represents median, interquartile,
and range of latencies; crosses represent outliers (2.5 times interquartile range). D: Scatter plot showing the mean
firing rate over 200 trials of 1 ms each (500 ms before stimulus to 500 ms after stimulus), for almost all cells in the
microcircuit. 3 cells with a firing rate of over 100 Hz in the original simulation are not included, and 18 cells with a
firing rate of over 140 Hz in the simplified simulation are not included, since they are most likely due to failed GIF
parameter extraction. Note that this is likely true for all neurons close to the simplified axis. Colors are according
to Gaussian kernel density estimation (blue-low density, red-high density). The red line depicts a slope of 1, and
the dashed lines 0.5 and 2 respectively. E: Left: Ratio of cells that do not fire in either the original or simplified
circuit (40 ms before, or 60 ms after stimulus), do not fire in one of the simulation sets, or fire in both. Right: Linear
correlation (Pearson's r) between PSTHs before and after the stimulus (as depicted in panel A).
19
1200TrialsL5NBCON1200L4MCON1200L5STPCON1200L23LBCNR1200L6TPCL4ON1200L5TTPC1ON1200L4PCON1200L5DBCNR−40−200204060Time relative to stimulus (ms)1200L5MCOFF070Spikes/s04000500.04.0080015007006014Cell-type responses<0.020.1110100FR before stimulus (spikes/s)<0.020.1110100FR after stimulus (spikes/s)Cell-type (cid:31)ring ratesInh., signi(cid:30)cant.Exc., non-signi(cid:30)cantInh., signi(cid:30)cantInh., non-signi(cid:30)cantSelected NRSelected ONSelected OFFABCCell-type response latenciesInh.Exc.020406080100FR (Hz) - Original020406080100120140FR (Hz) - Simpli(cid:30)edResponse-latency (ms)DFiring rate change between original and simpli(cid:31)ed circuitEPSTH correlation between original and simpli(cid:31)ed circuit0.00.20.40.60.81.0Ratio of cells−1.0−0.50.00.51.0Correlation r0500100015002000250030003500Cell count0.00.20.40.60.81.0Ratio of cells−1.0−0.50.00.51.0Correlation r0500100015002000250030003500Cell countFR > 0,bothFR = 0,bothFR = 0,oneFR > 0,bothFR = 0,bothFR = 0,onePre stimulus(-40 to 0 ms)Post stimulus(0 to 60 ms)L4 and L5,centerAll otherL4 and L5,centerAll other Figure 10: Response variability. A: Distributions of response variability, calculated as the SD over trials of the
delay, for excitatory (top) and inhibitory (bottom) m-types. Red lines show the median, boxes the 25th and 75th
percentiles and whiskers the full data spread (excluding outliers) across neurons. Left: In the simplified circuit, right:
in the reference circuit. B: Distributions of response variability in layer IV (top) and layer V (bottom) compared
to the variability between the reference and simplified circuit considering the same neuron in the same trial (grey
shaded background).
Figure 11: Stimulus response curves for various levels of Ca2+ and somatic depolarization to 100%
threshold. A: Response amplitudes defined as peak response of the 1000 most central (A, C) or most peripheral
(B, D) neurons in the 30 ms after injection of thalamocortical stimuli of various strengths. (A, B) in the simplified
circuit; (C, D) in the reference circuit.
20
050100150200250III/IIIIVVVIII/IIIIVVVI050100150200250ExcitatoryInhibitorySimplifiedReferenceResponse variability(SD of delay to first spike)050100IVacross Neurons(smpl)across Neurons(ref)Circuits(smpl vs ref)050100VVariability across trialsAB0102030405060minicols0102030405060Hzca=1.10ca=1.15ca=1.20ca=1.25ca=1.30ca=1.350102030405060minicols0102030405060minicols0102030405060Hz0102030405060minicolsSimplifiedReferenceCenterPeripheryAABCD Figure 12: Temporal structure of in vivo spontaneous activity. A: Schematic depiction of the structure
of a spike triplet for a trio of neurons. From a pool of 50 randomly selected L5 neurons, every possible triplet for
every possible cell trio was computed to produce the data plotted in figures B through D. B: Count histogram for a
representative neuron trio. Black box indicates region containing "precisely repeating triplets" (those that occur
sufficiently close to mode). White square signifies mode. C: Plot of correlation between neural latency differences
and inter spike intervals for corresponding triplet modes. Positive slope indicates that individual neural latencies
are predictive of triplet structure for trios in which those neurons participate. Data points whose triplet inter spike
intervals exceed m ± 150 ms are not shown. D: Probability of occurrence of precisely repeating triplets peaks shortly
after onset of activated state (black, solid line). The significance of this peak is compared with two alternative null
hypotheses (independent Poisson model, blue curve; common excitability model, red curve). The independent Poisson
model preserves overall number spikes per train, but draws spike times at random from a Poisson distribution. The
common excitability model preserves individual spike times, but randomly exchanges these spikes between neurons
in the trio. Dashed lines indicate standard deviation.
21
t2−t1t3−t1neuron123-150.0150.0t2- t1(ms)-150.0150.0t3- t1(ms)012.0triplet probability - Hz20100200300Time from UP state onset (ms)00.18Triplet probability - Hzoriginalshuffling b/n neuronsshuffling in timeA1A2A3A4 Figure 13: Spatial correlation. A: Correlation Scatter. The pairwise activity correlation coefficient between
clusters of mini columns decays with the distance of the clusters. Overall, the correlation is lower in the simplified
circuit. Results obtained at 1.25 mM extra cellular calcium. Exponential fits showed by dashed lines. B: Neighbors
Average Correlation. The average correlation between neighboring clusters with a maximum distance of 60 µm
increases with extra cellular calcium concentration, with a sharp jump at the transition point from the spontaneous
activity regime to the synchronous one. Error bars show error of fit. C: Space Constant. Fitted space constant τ (see
panel A) as a function of extra cellular calcium concentration. Higher calcium concentration is needed in the case of
the simplified circuit to reach the same tau of the detailed circuit. For higher calcium concentrations, correlation
data are no longer well approximated by an exponential curve (data not shown). Error bars show error of fit.
4 Discussion
Here we present a pipeline for automated point-neuron simplification of data-driven microcircuit
models, and demonstrate the technique on a recently reported reconstruction of a neocortical
microcircuit around an in vivo-like working point. Our modular method first uses a phenomenological
approach that corrects for the effective PSP response when relocating synapses from the dendrites
to the soma. It then uses a high-throughput parameter extraction of Generalized Integrate-and-Fire
(GIF) point-neuron models (Pozzorini et al., 2015) to replace each detailed neuron. We consider first
the findings and implications raised by our applied synaptic correction and cellular simplification
and then discuss the results of network simulations and validations.
4.1 Soma-synaptic correction and cellular simplification
While many different approaches exist to simplify dendritic complexity by e.g. reducing the number
of dendritic arbors, this approach is to our knowledge the first to target point-neuron models and
directly and systematically correct for dendritic attenuation by modifying the synaptic dynamics
directly. Our approach is not intended and cannot account for non-linear local dendritic computation
but is nevertheless very powerful in correcting for the change in synaptic delay and effectiveness
when neurons have to be reduced to a point model.
It can reduce main errors in membrane
activity of neurons especially with long dendritic arbors i.e. apical dendrites in pyramidal cells.
Soma-synaptic correction is most effective on the membrane potential error especially in simulations
without non-linear activation of sodium channels (in silico TTX) but also spiking precision can be
recaptured to some extent (Figure 5B).
Using the soma-synaptic correction approach, it is further straight-forward to generate data
that can be used to constrain simplified neuron models. Since all synapses are located at the soma
the sum of synaptic currents recorded during a simulation of the detailed network in combination
with its somatic membrane response can be used to simplify the neuron around its operating point.
While this data can be used to constrain any point-neuron model, we chose the GIF neuron fitting
22
050100150200250300350400450Distance d [¹m]−0.10.00.10.20.30.40.50.6Correlation coefficient r¿=284§4¿=176§4referencesimplified1.201.221.241.261.281.301.321.34Calcium [mM]0.00.20.40.60.81.0Average correlation neighborsreferencesimplified1.201.211.221.231.241.251.261.271.28Calcium (mM)120140160180200220240260280300¿referencesimplifiedABC procedure for its fast convergence properties.
Our extracted GIF models from these data showed overall good performance on the validation
traces. Nevertheless 20% of the cells had to be replaced by other GIF models of the same respective
morpho-electrical type due to either non converging optimization or low performance on the
validation set.
4.2 Effect of simplification on network behavior
Our soma-synaptic correction approach was able to recapture the transition from asynchronous to
synchronous network states with increasing calcium concentrations as seen in the detailed control
network simulations, with only minor shifts in the critical calcium concentration. Furthermore it
was able to remove fast oscillations especially appearing in layer 6 when synapses were displaced to
the soma without compensation. The basic asynchronous and synchronous behavior was retained
when all detailed models were replaced by simplified GIF models (Figure 8). However, GIF models
could not recapture the low baseline activity as observed during low depolarizations or in between
synchronous burst, where the cumulated spike activity did not decay down close to zero.
Stimulus response properties were largely preserved, though the simplified circuit was slightly
less responsive. Interestingly, the spiking identity/uniqueness of each neuron is largely preserved by
this approach as shown qualitatively in Figure 9A compared to Figure 17A1 in Markram et al.,
2015 (gids preserved), and quantitatively in Figure 9E. Appearance of previously reported precisely
repeating triplet structures is markedly reduced, but could be compensated by moving the network
closer to the critical point. Correlation structure spatially and temporally is found to be significantly
reduced in the simplified circuit, even if the reduced criticality of the circuit is compensated. This
could indicate a shortcoming of point-neurons in general, but might be a short-coming of the
specific simplification approach and neuron model taken here. The existence of these benchmarks
to systematically address quality of simplification is an important step towards addressing such
questions systematically.
4.3 Outlook
One objective of simplification is to reduce required compute resources, or target real-time and
accelarated execution on GPU-based and neuromorphic platforms (Nageswaran et al., 2009; Fidjeland
and Shanahan, 2010; Bruderle et al., 2009; Galluppi et al., 2010; Bruderle et al., 2011; Yavuz et al.,
2016). Targeting a simulator such as BRIAN (Goodman and Brette, 2008), NEST (Gewaltig and
Diesmann, 2007) or Nengo (Stewart et al., 2009) , which are optimized for point-neuron simulations
is desirable, either directly or through PyNN (Davison et al., 2009), as these together account for
the majority of the point-neuron simulation user community, according to a recent survey (Hanke
and Halchenko, 2011). For these simulators, a common optimization technique is to employ linear
synapse models, and lump them all into one state variable. This approach is not immediately
possible with the method proposed here, because the synapses have a variety of time constants
determined by their dendritic location, and variability in parameters inherited from the reference
model. Nevertheless we could already show for the single neuron case (Figure 5B) that it is possible
to reduce the large number of thousands of post-synaptic processes to 3-36 processes per cell with
only moderate increase of error by using k-means clustering of the synaptic delay τ . Future work
will further analyse and validate this technique to simplify synapses for running on these popular
simulators.
23
An important opportunity which arises with the availability of data-driven point neuron network
is to utilize them in turn as a reference model for the subsequent simplification to population density
approaches, for which ample tools and analytical strategies already exist(Renart et al., 2004; Muller
et al., 2007). Such models are highly suitable for mathematical analysis of population dynamics.
The reference model used here is itself a moving target, and will be refined over time to
incorporate new data on gap junctions, interactions with extracellular space, emphatic effects,
plasticity, glial cells, and neural-glial-vasculature interactions. Further simplification techniques
may need to be developed to simplify these new aspects to point-neuron representations. Herein
lies the fundamental important of an automated simplification pipeline presented here. Coupled
with the data-driven reference model which is continuously evolving to integrate experimental data,
a continuous bridge for exchange is achieved between experimental neuroscience and data-driven
"bottom-up" models on the one side, and the predominantly "top-down" point-neuron modeling
community on the other.
Disclosure/Conflict-of-Interest Statement
The authors declare that the research was conducted in the absence of any commercial or financial
relationships that could be construed as a potential conflict of interest.
Acknowledgments
Funding: The work was supported by funding from the EPFL to the Laboratory of Neural Microcir-
cuitry (LNMC) and funding from the ETH Domain for the Blue Brain Project (BBP). Additional
support was provided by funding for the Human Brain Project from the European Union Seventh
Framework Program (FP7/2007- 2013) under grant agreement no. 604102 (HBP). The BlueBrain
IV BlueGene/Q and Linux cluster used as a development system for this work is financed by ETH
Board Funding to the Blue Brain Project as a National Research Infrastructure and hosted at the
Swiss National Supercomputing Center (CSCS).
References
Berger, T., M. E. Larkum, and H.-R. Luscher (2001). High i h channel density in the distal
apical dendrite of layer v pyramidal cells increases bidirectional attenuation of epsps. Journal of
neurophysiology 85 (2), 855–868.
Bracewell, R. (1978). The Fourier Transform and its Applications (Second ed.). Tokyo: McGraw-Hill
Kogakusha, Ltd.
Brent, R. P. (2013). Algorithms for Minimization Without Derivatives. Mineola, N.Y: Dover
Publications Inc.
Bruderle, D., E. Muller, A. P. Davison, E. Muller, J. Schemmel, and K. Meier (2009). Establishing
a novel modeling tool: a python-based interface for a neuromorphic hardware system. Frontiers
in Neuroinformatics 3, 17.
24
Bruderle, D., M. A. Petrovici, B. Vogginger, M. Ehrlich, T. Pfeil, S. Millner, A. Grubl, K. Wendt,
E. Muller, M.-O. Schwartz, D. H. Oliveira, S. Jeltsch, J. Fieres, M. Schilling, P. Muller, O. Bre-
itwieser, V. Petkov, L. Muller, A. P. Davison, P. Krishnamurthy, J. Kremkow, M. Lundqvist,
E. Muller, J. Partzsch, S. Scholze, L. Zuhl, C. Mayr, A. Destexhe, M. Diesmann, T. C. Pot-
jans, A. Lansner, R. Schuffny, J. Schemmel, and K. Meier (2011). A comprehensive workflow
for general-purpose neural modeling with highly configurable neuromorphic hardware systems.
Biological Cybernetics 104 (4), 263–296.
Davison, A. P., D. Bruderle, J. M. Eppler, J. Kremkow, E. Muller, D. Pecevski, L. Perrinet,
and P. Yger (2009). Pynn: a common interface for neuronal network simulators. Frontiers in
Neuroinformatics 2, 11.
De Koninck, Y. and I. Mody (1997, April). Endogenous GABA Activates Small-Conductance K+
Channels Underlying Slow IPSCs in Rat Hippocampal Neurons. pp. 2202–2208.
Destexhe, A., M. Rudolph, and D. Par´e (2003). The high-conductance state of neocortical neurons
in vivo. Nature reviews neuroscience 4 (9), 739–751.
Eliasmith, C., T. C. Stewart, X. Choo, T. Bekolay, T. DeWolf, Y. Tang, and D. Rasmussen (2012).
A large-scale model of the functioning brain. science 338 (6111), 1202–1205.
Eliasmith, C. and O. Trujillo (2014). The use and abuse of large-scale brain models. Current
Opinion in Neurobiology 25, 1 – 6. Theoretical and computational neuroscience.
Fidjeland, A. K. and M. P. Shanahan (2010). Accelerated simulation of spiking neural networks
using gpus. In Neural Networks (IJCNN), The 2010 International Joint Conference on, pp. 1–8.
IEEE.
Fuhrmann, G., I. Segev, H. Markram, and M. Tsodyks (2002, January). Coding of temporal
information by activity-dependent synapses. J Neurophysiol. 87 (1), 140–148.
Galluppi, F., A. Rast, S. Davies, and S. Furber (2010). Neural Information Processing. Theory and
Algorithms: 17th International Conference, ICONIP 2010, Sydney, Australia, November 22-25,
2010, Proceedings, Part I, Chapter A General-Purpose Model Translation System for a Universal
Neural Chip, pp. 58–65. Berlin, Heidelberg: Springer Berlin Heidelberg.
Gewaltig, M.-O. and M. Diesmann (2007). Nest (neural simulation tool). Scholarpedia 2 (4), 1430.
Goodman, D. F. and R. Brette (2008). Brian: a simulator for spiking neural networks in python.
Frontiers in Neuroinformatics 2, 5.
Gupta, A., Y. Wang, and H. Markram (2000, January). Organizing Principles for a Diversity of
GABAergic Interneurons and Synapses in the Neocortex. Science 287 (5451), 273–278.
Gustavsen, B. and A. Semlyen (1999). Rational approximation of frequency domain responses by
vector fitting. IEEE Transaction on Power Delivery 14 (3), 1052–1061.
Hanke, M. and Y. O. Halchenko (2011). Neuroscience runs on gnu/linux. Frontiers in neuroinfor-
matics 5, 8.
Hausser, M. and A. Roth (1997, October). Estimating the time course of the excitatory synaptic
conductance in neocortical pyramidal cells using a novel voltage jump method. The Journal of
neuroscience: the official journal of the Society for Neuroscience 17 (20), 7606–7625.
25
Hines, M. L. and N. T. Carnevale (1997). The neuron simulation environment. Neural computa-
tion 9 (6), 1179–1209.
Jahr, C. E. and C. F. Stevens (1990, June). A quantitative description of NMDA receptor-
channel kinetic behavior. The Journal of neuroscience: the official journal of the Society for
Neuroscience 10 (6), 1830–1837.
Jolivet, R., T. J. Lewis, and W. Gerstner (2004). Generalized Integrate-and-Fire Models of Neuronal
Activity Approximate Spike Trains of a Detailed Model to a High Degree of Accuracy. Journal
of Neurophysiology 92 (2), 959–976.
Khazipov, R., P. Congar, and Y. Ben-Ari (1995, November). Hippocampal CA1 lacunosum-
moleculare interneurons: modulation of monosynaptic GABAergic IPSCs by presynaptic GABAB
receptors. Journal of neurophysiology 74 (5), 2126–2137.
Koch, C. (1998). Biophysics of Computation: Information Processing in Single Neurons (Computa-
tional Neuroscience) (1 ed.). Oxford University Press.
London, M. and M. Hausser (2005). Dendritic Computation. Annual Review of Neuroscience 28 (1),
503–532.
Luczak, A., P. Barth´o, S. L. Marguet, G. Buzs´aki, and K. D. Harris (2007). Sequential structure of
neocortical spontaneous activity in vivo. Proceedings of the National Academy of Sciences 104 (1),
347–352.
Maass, W., T. Natschlager, and H. Markram (2004). Fading memory and kernel properties of
generic cortical microcircuit models. Journal of Physiology-Paris 98 (4), 315–330.
Marasco, A., A. Limongiello, and M. Migliore (2012). Fast and accurate low-dimensional reduction
of biophysically detailed neuron models. Scientific Reports 2, 928.
Marasco, A., A. Limongiello, and M. Migliore (2013). Using strahler's analysis to reduce up to
200-fold the run time of realistic neuron models. Scientific Reports 3, 2934.
Markram, H., E. Muller, S. Ramaswamy, M. W. Reimann, M. Abdellah, C. A. Sanchez, A. Ailamaki,
L. Alonso-Nanclares, N. Antille, S. Arsever, G. A. A. Kahou, T. K. Berger, A. Bilgili, N. Bun-
cic, A. Chalimourda, G. Chindemi, J.-D. Courcol, F. Delalondre, V. Delattre, S. Druckmann,
R. Dumusc, J. Dynes, S. Eilemann, E. Gal, M. E. Gevaert, J.-P. Ghobril, A. Gidon, J. W.
Graham, A. Gupta, V. Haenel, E. Hay, T. Heinis, J. B. Hernando, M. Hines, L. Kanari, D. Keller,
J. Kenyon, G. Khazen, Y. Kim, J. G. King, Z. Kisvarday, P. Kumbhar, S. Lasserre, J.-V. Le B´e,
B. R. C. Magalhaes, A. Merch´an-P´erez, J. Meystre, B. R. Morrice, J. Muller, A. Munoz-C´espedes,
S. Muralidhar, K. Muthurasa, D. Nachbaur, T. H. Newton, M. Nolte, A. Ovcharenko, J. Palacios,
L. Pastor, R. Perin, R. Ranjan, I. Riachi, J.-R. Rodr´ıguez, J. L. Riquelme, C. Rossert, K. Sfyrakis,
Y. Shi, J. C. Shillcock, G. Silberberg, R. Silva, F. Tauheed, M. Telefont, M. Toledo-Rodriguez,
T. Trankler, W. Van Geit, J. V. D´ıaz, R. Walker, Y. Wang, S. M. Zaninetta, J. DeFelipe, S. L. Hill,
I. Segev, and F. Schurmann (2015). Reconstruction and Simulation of Neocortical Microcircuitry.
Cell 163 (2), 456–492.
Markram, H., Y. Wang, and M. Tsodyks (1998, April). Differential Signaling Via the Same Axon
of Neocortical Pyramidal Neurons. pp. 5323–5328.
26
Mott, D. D., Q. Li, M. M. Okazaki, D. A. Turner, and D. V. Lewis (1999, September). GABAB-
Receptor–Mediated Currents in Interneurons of the Dentate-Hilus Border. pp. 1438–1450.
Muller, E., L. Buesing, J. Schemmel, and K. Meier (2007). Spike-frequency adapting neural
ensembles: beyond mean adaptation and renewal theories. Neural Computation 19 (11), 2958–
3010.
Nageswaran, J. M., N. Dutt, J. L. Krichmar, A. Nicolau, and A. V. Veidenbaum (2009). A
configurable simulation environment for the efficient simulation of large-scale spiking neural
networks on graphics processors. Neural networks 22 (5), 791–800.
Nevian, T., M. E. Larkum, A. Polsky, and J. Schiller (2007). Properties of basal dendrites of layer
5 pyramidal neurons: a direct patch-clamp recording study. Nature Neuroscience 10 (2), 206–214.
Pozzorini, C., S. Mensi, O. Hagens, R. Naud, C. Koch, and W. Gerstner (2015). Automated
High-Throughput Characterization of Single Neurons by Means of Simplified Spiking Models.
PLoS Comput Biol 11 (6), e1004275.
Rall, W. (1967). Distinguishing theoretical synaptic potentials computed for different soma-dendritic
distributions of synaptic input. Journal of Neurophysiology 30 (5), 1138–1168.
Ramaswamy, S., S. L. Hill, J. G. King, F. Schurmann, Y. Wang, and H. Markram (2012). Intrinsic
morphological diversity of thick-tufted layer 5 pyramidal neurons ensures robust and invariant
properties of in silico synaptic connections. The Journal of Physiology 590 (4), 737–752.
Renart, A., N. Brunel, and X.-J. Wang (2004). Mean-field theory of irregularly spiking neuronal
populations and working memory in recurrent cortical networks. Computational neuroscience: A
comprehensive approach, 431–490.
Reyes-Puerta, V., J.-J. Sun, S. Kim, W. Kilb, and H. J. Luhmann (2015). Laminar and Columnar
Structure of Sensory-Evoked Multineuronal Spike Sequences in Adult Rat Barrel Cortex In Vivo.
Cerebral Cortex 25 (8), 2001–2021.
Slotine, J.-J. E., W. Li, and others (1991). Applied nonlinear control, Volume 199. Prentice-Hall
Englewood Cliffs, NJ.
Stewart, T. C., B. Tripp, and C. Eliasmith (2009). Python scripting in the nengo simulator.
Frontiers in Neuroinformatics 3, 7.
Thorpe, S. (2002). Biologically Motivated Computer Vision: Second International Workshop,
BMCV 2002 Tubingen, Germany, November 22–24, 2002 Proceedings, Chapter Ultra-Rapid
Scene Categorization with a Wave of Spikes, pp. 1–15. Berlin, Heidelberg: Springer Berlin
Heidelberg.
Tsodyks, M. and H. Markram (1997, May). The neural code between neocortical pyramidal neurons
depends on neurotransmitter release probability. Proc Natl Acad Sci USA 94 (10), 5495.
Williams, S. R. and G. J. Stuart (2002). Dependence of epsp efficacy on synapse location in
neocortical pyramidal neurons. Science 295 (5561), 1907–1910.
Wybo, W. A. M., D. Boccalini, B. Torben-Nielsen, and M.-O. Gewaltig (2015, December). A Sparse
Reformulation of the Green's Function Formalism Allows Efficient Simulations of Morphological
Neuron Models. Neural computation 27 (12), 2587–622.
27
Wybo, W. A. M., K. M. Stiefel, and B. Torben-Nielsen (2013, September). The Green's function
formalism as a bridge between single- and multi-compartmental modeling. Biological cybernet-
ics 107 (6), 685–694.
Yavuz, E., J. Turner, and T. Nowotny (2016, January). GeNN: a code generation framework for
accelerated brain simulations. Scientific Reports 6, 18854.
Zenke, F., E. J. Agnes, and W. Gerstner (2015). Diverse synaptic plasticity mechanisms orchestrated
to form and retrieve memories in spiking neural networks. Nature communications 6.
28
|
1810.10498 | 5 | 1810 | 2019-11-18T13:01:01 | Sleep-like slow oscillations improve visual classification through synaptic homeostasis and memory association in a thalamo-cortical model | [
"q-bio.NC",
"cs.AI",
"cs.DC"
] | The occurrence of sleep passed through the evolutionary sieve and is widespread in animal species. Sleep is known to be beneficial to cognitive and mnemonic tasks, while chronic sleep deprivation is detrimental. Despite the importance of the phenomenon, a complete understanding of its functions and underlying mechanisms is still lacking. In this paper, we show interesting effects of deep-sleep-like slow oscillation activity on a simplified thalamo-cortical model which is trained to encode, retrieve and classify images of handwritten digits. During slow oscillations, spike-timing-dependent-plasticity (STDP) produces a differential homeostatic process. It is characterized by both a specific unsupervised enhancement of connections among groups of neurons associated to instances of the same class (digit) and a simultaneous down-regulation of stronger synapses created by the training. This hierarchical organization of post-sleep internal representations favours higher performances in retrieval and classification tasks. The mechanism is based on the interaction between top-down cortico-thalamic predictions and bottom-up thalamo-cortical projections during deep-sleep-like slow oscillations. Indeed, when learned patterns are replayed during sleep, cortico-thalamo-cortical connections favour the activation of other neurons coding for similar thalamic inputs, promoting their association. Such mechanism hints at possible applications to artificial learning systems. | q-bio.NC | q-bio | opeN
Received: 24 January 2019
Accepted: 3 June 2019
Published: xx xx xxxx
sleep-like slow oscillations improve
visual classification through
synaptic homeostasis and memory
association in a thalamo-cortical
model
Cristiano Capone1, elena pastorelli1,2, Bruno Golosio
3,4 & pier stanislao paolucci
1
The occurrence of sleep passed through the evolutionary sieve and is widespread in animal species.
Sleep is known to be beneficial to cognitive and mnemonic tasks, while chronic sleep deprivation is
detrimental. Despite the importance of the phenomenon, a complete understanding of its functions
and underlying mechanisms is still lacking. In this paper, we show interesting effects of deep-sleep-like
slow oscillation activity on a simplified thalamo-cortical model which is trained to encode, retrieve
and classify images of handwritten digits. During slow oscillations, spike-timing-dependent-plasticity
(STDP) produces a differential homeostatic process. It is characterized by both a specific unsupervised
enhancement of connections among groups of neurons associated to instances of the same class (digit)
and a simultaneous down-regulation of stronger synapses created by the training. This hierarchical
organization of post-sleep internal representations favours higher performances in retrieval and
classification tasks. The mechanism is based on the interaction between top-down cortico-thalamic
predictions and bottom-up thalamo-cortical projections during deep-sleep-like slow oscillations.
Indeed, when learned patterns are replayed during sleep, cortico-thalamo-cortical connections favour
the activation of other neurons coding for similar thalamic inputs, promoting their association. Such
mechanism hints at possible applications to artificial learning systems.
Human brains spend about one-third of their life-time sleeping. Sleep is present in every animal species that has
been studied1. This happens notwithstanding two negative facts: the danger caused by sleep, that diminishes the
capability to defend from predators and other threats, and the reduction of time available for activities targeting
immediate rewards (e.g. hunting or gathering food). Having survived the evolutionary selection in all species,
sleep must therefore provide strong advantages. Another notable fact is that newborns' human brains occupy the
majority of their time asleep, nevertheless they learn at a very fast rate. For this and other motivations for studying
sleep, also in relation to consciousness, see e.g.2. Moreover, even if occasional awakening does not seriously impair
brain biology and cognitive functions, in the long term chronic deprivation of sleep produces measurable effects
on cognition, mood and health3. Experimental studies like1,4 investigated the effects of sleep on firing rates and
synaptic efficacies. In1 the authors formulated hypotheses about homeostatic processes occurring during sleep.
A possible motivation for the brain entering in the sleep activity would be to set a better energy consumption
regime during next wakefulness cycle. This might be obtained by reducing firing rates and the amplitude of
evoked post-synaptic potentials. It could be reached by pruning non necessary synapses and by reducing synaptic
weights within the limits imposed by the conservation of adequate coding for memories. In4 Watson et al. propose
a novel intriguing experimental evidence. They used large-scale recordings to examine the activity of neurons in
the frontal cortex of rats and observed that neurons with different pre-sleep firing rate are differentially modu-
lated by different sleep substates (REM, non-REM and micro arousal). Sleep activity such as slow waves activity
and sharp-waves ripples have been shown to be beneficial for memory consolidation5,6 and task performances
1infn Sezione di Roma, Rome, italy. 2PhD Program in Behavioural neuroscience, "Sapienza" University of Rome,
Rome, italy. 3Dipartimento di fisica, Università di cagliari, cagliari, italy. 4infn Sezione di cagliari, cagliari, italy.
correspondence and requests for materials should be addressed to c.c. (email: [email protected])
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
1
www.nature.com/scientificreports .
Hz
++Ca
[0 5, 4]
optimization7. A few computational models have been developed to investigate the interaction of sleep-like activ-
ity and plasticity. In8, the authors showed that Up states specifically mediate synaptic down-scaling with beneficial
effect on signal to noise ratio received by post-synaptic neurons. The effect of thalamo-cortical sleep on pre-stored
time sequences is explored in9,10.
Here, we focus on sleep mediated memory association and its implications on cognitive tasks performances.
We present a minimal thalamo-cortical model which, after being trained on handwritten characters in unsuper-
vised mode, is induced to express sleep-like dynamics. We measure its effects on the classification accuracy, the
structure of the synaptic matrix and firing rate distributions with findings that are consistent with experimental
observations4,8,11,12.
Slow oscillations (SO) are considered the default emergent activity of the cortical network13 and are observed
during the deepest of physiological non-REM sleep stages as an alternation between Down states (characterized
by nearly silent neurons) and Up states (in which a subset of neurons goes in a high firing rate regime) occurring
(delta band). We set the model SO at a comparable frequency. SO activity
at a frequency in the range
is expected to play two complementary roles, which are separately mediated by Up states and Down states. Down
states would play a purely biological function, with a lower immediate impact on cognitive performance. The role
of Down states, with a majority of neurons put in a silent state for a long fraction of deep sleep time, would serve
a restoration purpose, enabling periodic biological maintenance and recovery, as it happens in the whole body
when at rest14. Our modeling and investigation is focused only on the effects mediated by the Up states dynamics.
During sleep external perceptions are, at least, strongly attenuated, and the majority of the motor system is
blocked15. For this reason in our model the local interaction between cortex and thalamus is crucial during sleep
rather than contextual signal coming from other cortical modules and sensory input coming from thalamic
pathways.
In order to make a biologically realistic learning protocol and to implement the role of the context in the learn-
ing phase, we took inspiration from the "organizing principle" in16 for the Cerebral Cortex, which describes spe-
cific computational strategies implemented by the neuronal structure of Layer 5. The architecture is grounded on
the separation of intra-areal and inter-areal contextual information (reaching the apical dendrites of pyramidal
neurons) from the feed-forward flow of area specific information (targeting its basal synapses). Cellular mecha-
nisms, like
spikes, promote the detection of coincidence between contextual and specific activity.
High-frequency bursts of spikes are emitted when the coincidence is detected. Relying on these observation we
introduced in our model external stimuli mimiking contextual information which changes the effective firing
threshold of specific subsets of neurons during the presentation of examples in the training phase. For each exam-
ple, a vector of features is projected toward a cortical network by a thalamic neural network. Due to the change in
the perceptual effective firing threshold, spike-timing-dependent-plasticity (STDP) creates stronger bottom-up
(thalamo-cortical) and top-down (cortico-thalamic) connections between a subset of cortical and the thalamic
neurons.
In our model we observe that sleep induces both the association of patterns encoding learned images belong-
ing to the same category and a differential synaptic down-scaling. This is also reflected in a differential modula-
tion of firing rates, producing observations similar to4. We observe that such effect, probably related to energetic
optimization in biological networks, also has beneficial effects on the performances of our network in the image
recognition task.
Results
We tested the role and the mechanisms of the occurrence of SO in a thalamo-cortical network model which was
previously trained to learn and recall images (MNIST dataset). The network model included thalamic relay (tc)
and reticular (re) neurons in the thalamus, as well as pyramidal neurons (cx) and inhibitory interneurons (in) in
the cortex (Fig. 1A, see Methods) following a standard minimal structure for thalamico-cortical models17.
Figure 1B shows an example of activity time-course in the cx and tc populations during the training phase, the
retrieval phase, and the early stage of sleep phase.
Training and pre-sleep retrieval. During the training, 9 different images were presented to the network,
in a first set of runs: 3 instances for each class of digit, for a total of 3 different classes (e.g. 0, 1, 2). In a second
group of runs, 30 examples per digit constituted the training set. For each image an external stimulus (contextual
signal) induced a different subset of cx neurons to code for that specific image, with STDP shaping the
intra-cortical, the thalamo-cortical and the cortico-thalamic connectivity. In order to adopt the prescription in16
the parameters are set to make the cortical neurons fire during the training only if they receive both sensory and
contextual stimuli. This training procedure works even in the extreme case where only one neuron is used to code
each digit example. However for one or a few neurons, self sustained oscillations would not be well defined. For
this reason we chose a population of 20 cortical neurons for each newly presented example. The precise number
of such neurons is not a critical factor as we will discuss later.
After the training, images were presented again (retrieval phase) without the external stimulus. The popula-
tion of neuron responding to the image were the same as the one selected in the training phase by the external
stimulus, demonstrating the success of the retrieval.
Induction of slow-oscillation. A non-specific stimulus at low firing rate was provided to cortical neurons
only, while model parameters are modulated (see Methods for details) eliciting the spontaneous occurrence of
cortically generated Up states and of thalamo-cortical SO (see Fig. 1B).
We relied on the framework of Mean Field theory to obtain a model displaying different dynamical regimes.
An oscillatory regime, closely resembling Slow Oscillations observed in deep sleep and anesthetized states can be
induced by introducing a relatively strong recurrent excitation and spike frequency adaptation18,19. We tuned the
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
2
www.nature.com/scientificreportswww.nature.com/scientificreports/ Figure 1. Thalamo-cortical model and protocol description. (A) Sketch of the structure of the simplified
thalamo-cortical model considered, which is composed of an excitatory and an inhibitory population both for
the cortex (cx, in) and for the thalamus (tc, re). Connectivity structure is represented by solid lines. The visual
input is fed into the model through the thalamic population, mimicking the biological visual pathways. In the
training phase a lateral stimulus enhances a specific subset of cx neurons to preferentially represent the stimulus.
(B) Activity produced during training phase, pre-sleep retrieval and the first 40s of SO activity in the cx (top)
and tc (bottom) populations. Only first 180 neurons in tc population are shown for visual purposes. In the
training 3 instances of 3 classes of digits (0,1,2) are learned by the network. In the replay during sleep, thalamo-
cortical connections promotes the activation of neurons coding for similar patterns of activity, causing the
potentiation of cortico-cortical connections between neurons representing digits of the same class. A general
depression reduces the largest synaptic weights. Post SO retrieval is not shown.
0 25
1 5
) and Up state durations
parameters of the network to make it display SO frequency (between . Hz
(a few hundred of milliseconds) comparable with experimental observations in deep sleep recordings20,21.
and . Hz
The activity (Fig. 1B) during the initial stages of simulated SO displays that Up states are independently sus-
tained by populations coding for different memorized images. However, thanks to the cortico-thalamo-cortical
pathways each population tends to recruit other populations sharing similar thalamic representations. Indeed
when a population initiates an Up states, it activates thalamic patterns similar to the one responsible for the acti-
vation of the population itself during the retrieval phase. This is what we call top-down prediction. The thalamus,
in turn, activates all the other populations which coded for a similar thalamic input. Across the sleep period,
thanks to the cortico-cortical plasticity, the co-activation of populations coding for similar predicted synaptic
input becomes a more and more prominent feature.
Effects on the synaptic matrix of slow-oscillations. We first considered the case with the training
set composed of 3 classes and 3 examples per class. After the training stage the system undergoes a 600s period
of sleep. During this stage the activation of groups of neurons associated to different training examples induces,
through STDP, not only a synaptic pruning but also the creation of stronger synapses between groups of neurons
that share enough commonality in the features they received during the training.
This effect can be noticed comparing the structure of the synaptic matrix before and after sleep as in Fig. 2A,B,
that reports the change from the initial flat structure (which reflects the individual training examples) towards a
hierarchical structure (embedding the categories of learned digits).
Indeed we can observe that novel synapses are created among examples in the same digit category
(Fig. 2A-left). At the same time the system undergoes a down-scaling of the strongest synapses, those linking
neurons coding for the same image (Fig. 2A-right). The differential effect on synapses can also be observed in
Fig. 2C, where the histogram of synaptic weights after-sleep is reported. The synapses between patterns encoding
for different learned images of the "same class" and those between "different classes" (respectively orange and
green distribution in Fig. 2C) are originally drawn from the same distribution by definition (see Methods), while
after sleep they are clearly differentiated.
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
3
www.nature.com/scientificreportswww.nature.com/scientificreports/ Figure 2. SO effects on connectivity structure. (A) Synaptic weights matrix of the recurrent connectivity of cx
population, before (left) and after (right) the occurrence of sleep-like activity. The yellow squares represent high
weights emerged between neurons encoding the visual input related to the same object (single instance of 0, 1, 2
… image). Red solid lines separate the neurons encoding visual inputs related to different classes of objects (0, 1,
2 …). (B) Scatter-plot of the same synaptic weights before and after sleep. (C) Synaptic weights after sleep,
separated in three groups, synapses between neurons encoding the same object (yellow), the same class (but not
the same object, orange) and different classes (green).
Effects on the post-sleep activity. The change in structure of the synaptic weights matrix modifies the
activity expressed by the network in the retrieval phase. This can be appreciated looking at the difference of the
correlations between groups of neurons before and after sleep (see Fig. 3A).
Figure 3B reports the difference among correlations evaluated after and before sleep. It shows decorrelation
(blue squares) of populations encoding different classes, and correlation (red regions) of the ones coding the same
class. Such information is reported also in Fig. 3C, showing the correlation changes for populations in the same
class (blue) and in different classes (green). Such effect might provide benefits in retrieval and classification tasks.
The consistency of such result is testified by Fig. 4. There, the same simulation is performed for different train-
ing sets (different examples of 0,1,2 digits). All simulations show that the synapses between neurons in the same
class are the more reinforced (Fig. 4A, orange versus green) and that their internal representation has an increased
correlation (Fig. 4B, blue versus green).
Mechanistic interpretation. We propose that such effect is due to the interplay between cortico-thalamic
predictions and thalamo-cortical connections. In other words when a group of neurons undergoes an Up state it
formulates a prediction in the thalamus by activating a thalamic pattern similar to the one received during train-
ing. In turn, the thalamus projects to the cortex and activates those populations trained for similar input patterns.
This mechanism promotes the connections between populations of neurons coding for the images of the same
class through STDP. To prove this, we reproduced the same experiment switching off the cortico-thalamic predic-
tion and reported the result in Fig. 4C,D. It is evident that there is no sign of the preferential association observed
in the control condition, nor in the synaptic structure (Fig. 4C), neither in the internal representation (Fig. 4D).
To demonstrate the specific role of SO activity, we repeated the same experiment replacing the sleep like activ-
ity with an awake like asynchronous activity. We set the same adaptation strength b and →win
cx used during train-
ing and retrieval. In this case we did not observe significant changes in the synaptic matrix structure, with absence
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
4
www.nature.com/scientificreportswww.nature.com/scientificreports/ Figure 3. SO effects on internal representation. (A) Activity correlation between all pairs of populations
representing the single images before (left) and after (right) sleep. (B) Correlation difference between after
and before sleep. (C) Histogram of correlation differences for populations encoding the same class (blue) and
different classes (green).
Figure 4. Analysis of populations: synaptic weights and comparison between correlations with and without
cortico-thalamic predictions. (A) Average ratio between weights post- and pre- sleep for each simulation (top).
The different categories are separated in different colors. Yellow: synapses connecting neurons coding for the
same image, orange: different image of the same class of digits and green: different classes. (B) The average
change in correlation between post- and pre- sleep for each simulation (top) and histogram of the distribution
over all the simulations (n = 6, bottom). Blue: same class, green: different classes. (C,D) as in A-B but in absence
of cortico-thalamic connections.
of synaptic down-scaling (homeostasis) and no association of memories. This is due both to the absence of the
specific spatio-temporal structure of the activity that characterizes the SO state and the lower firing rate (indeed
asynchronous state has a lower firing rate than Up states).
Post-sleep improvement in a classification task. Finally, we evaluated the effect of sleep-like activity
on the performance of a set of classification task trials. Networks were exposed to example and test images drawn
from all the ten classes of digits in the MNIST dataset. Each simulation trial used a different test set of 250 images,
and a set of 3 training examples per digit (30 training instances in total, also randomly extracted for each classi-
fication trial). Each network was exposed to the training examples using the same protocol discussed above for
the simpler retrieval task (see Methods for details). For each test image, the classification was determined looking
for the class of the neuron responding with the higher firing rate. We note that class labels were used only during
classification and not during the training that was completely unsupervised.
± .
.58 0% to
We observed a net increase in the classification accuracy across the sleep period. Figure 5A reports in blue the
time course of accuracy increase as a function of the sleep time. After 3000s of sleep, the improvement was on
.64 0%, average performed over 24 simulations). In absence of
average .
thalamic feedback the improvement is significantly lower (Fig. 5A red line), proving that the memory association
due to the cortico-thalamo-cortical interaction is beneficial to performance in a classification task.
6 0% 0 5% (accuracy rose from
Figure 5B reports the average weights evolution as a function of sleep time. Synapses between groups of neu-
rons encoding for different instances of the same digit class (yellow solid line) were on average strongly poten-
tiated, much more than the ones connecting training examples belonging to different classes (green solid line).
Synapses interconnecting neurons representing individual training instances were down-scaled (orange solid
line). When the thalamo-cortical feedback was switched off (same colors, dashed lines) this differential effect did
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
5
www.nature.com/scientificreportswww.nature.com/scientificreports/ Figure 5. Sleep effects on a classification task. (A) Change in classification accuracy across over 30 sleep epochs
(100s each). Blue and red are respectively the conditions in which thalamus is on and off. The improvement
in accuracy is averaged over 30 simulation trials. SEM is reported in the shading. (B) Average synaptic
potentiation and depression over 30 sleep epochs. The colors indicate connections between neurons coding the
same instance (yellow), different instances of the same class (green) and instances of different classes (orange).
Dashed and solid lines represent the comparison between the conditions in which thalamus is on and off.
(C) Average synaptic depression over all the synapses. (inset) Average decrease of SO frequency across sleep
time, average over 4 simulations. (D) Scatter of single neurons activity in 8 simulations averaged over time in a
classification task before and after 3000s of sleep. Inset, average difference of activity after and before sleep as a
function of activity before sleep.
not happen. This is the same effects already reported for the case with 9 training examples (Fig. 2, simpler retrieval
task). We verified that the same qualitative results are obtained also for the cases in which each example is coded
by 10, 15 and 25 cortical neurons (not shown).
Such differential mechanism occurred together with a general synaptic depression (see Fig. 5C, thalamus on
and thalamus off in solid and dashed line respectively). We notice that the average down-scaling is very similar
for simulations executed in absence and presence of thalamic feed-back, allowing for a fair comparison of the
two conditions. As a consequence of such general synaptic depression the SO frequency decreases over time (see
Fig. 5C, inset) consistently with experimental observations12.
We observe that an optimal range of SO frequencies is important to obtain the reported results. Indeed an
extremely low Up state occurrence would make weaker the specific cortico-thalamo-cortical association. On the
other hand a very high Up state occurrence frequency would increase the probability to randomly associate differ-
ent classes of digits. Both scenarios would impair the positive effect of the sleep period on network performances.
Firing rates also underwent a differential modulation. Neurons with pre-sleep low activity became more active
after sleep (and vice versa). Figure 5D displays a scatter of time averaged single neuron activity during the classifi-
cation task, before and after the 3000s sleep period (data are drawn from 8 simulations). The distribution of indi-
vidual firing rates is rotated respect to the bisector line (red dashed line). The average difference of activity after
and before sleep is positive for low values of pre-sleep firing rates and negative for high pre-sleep rates (see inset
of Fig. 5D). This prediction is similar to what observed in4, strengthening the biological plausibility of our model.
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
6
www.nature.com/scientificreportswww.nature.com/scientificreports/ Discussion
We propose a minimal thalamo-cortical model that classifies images drawn from the MNIST set of handwritten
digits. During a training phase, spike-timing-dependent-plasticity (STDP) sculptures a pre-sleep synaptic matrix
and creates top-down synapses toward the thalamus. Then, the network is induced to produce deep-sleep-like
slow oscillations (SO)19,22, while being disconnected from sensory and lateral stimuli and driven by its internal
activity. During sleep Up states, thalamic cells are activated by top-down predictive stimuli produced by cortical
groups of neurons and respond with a forward feedback, recruiting other cortical neurons.
In our model it was of utmost importance to obtain a biologically accurate reproduction of the association/
coincidence mechanism described in16 between sensory (from thalamus) and contextual stimuli (from external
poisson processes). To do this properly would require a dedicated spiking neuronal model. The implementation
of such model in neural simulation engines is a current topic of development for the community. In this study
we approximated the coincidence mechanism through a careful subthreshold setting of both the contextual and
sensory inputs impinging in cortical neurons. In the training phase the neuron only fired whether both contextual
and sensory input was received. Our study proves the goodness of the coincidence mechanism for a fast learning
on few examples only presented once, providing an additional motivation for the development of a dedicated
model.
One delicate point is that the contextual signal facilitates the learning of each example in a different group of
neurons. Even if mixed selectivity neurons are fundamental in neural coding23, and are indeed created in our sim-
ulation during sleep, this work adopted this disjoint representation of the contextual facilitation signal, also with
the purpose of simplifying the understanding of the association mechanism itself. However we note that, in first
approximation, this mechanism would not be unreasonable in the cortex in several circumstances, for at least the
following reasons. First, we are often exposed to individual examples of objects embedded in different contexts. A
second argument in favour is that the internal state of the brain that dominates the creation of the contextual sig-
nal is never the same, even if external conditions are similar. Moreover, the coding for each internal state is sparse
and long-range connections bringing contextual information are themselves sparse. Contextual signals with low
degree of overlap are therefore plausible. The sparse selectivity of contextual signal facilitates the creation of a
wide, orthogonal internal representation of individual examples, but lacks of generalization of low levels feature.
Bottom-up sensory stimuli of examples belonging to the same class should present a high degree of overlap. We
demonstrated that sleep could help overcoming this problem, inducing specific association of examples sharing
commonality in low level features. This effect might be beneficial also to machine learning (ML) algorithms which
currently lack this generalization feature and is relevant for cerebral neural networks.
We observed differences in the synaptic structure and in the activity expressed before and after the sleep-like
period. Indeed SO induced two effects that are both beneficial and biologically plausible. The first is a reduction
in the strength of synapses inside neural populations created by the training on specific examples, the second
is a simultaneous increase of synapses that associate examples of characters belonging to the same category of
digits. The final structure of the synaptic matrix is more complex and hierarchically organized. Moreover, the
correlation of groups of neurons during post-sleep retrieval activity reflects the hierarchical structure supported
by the underlying synaptic matrix, with stronger cooperation between groups of neurons trained by different
examples belonging to the same class. Also, during the sleep phase, the network displays a rich internal dynamics
that evolves across sleep time. Indeed, the composition of neuronal groups recruited during the Up states changes
sensibly between the initial and the final stages of sleep.
We also investigated the benefits of such effects on the image classification task on the MNIST dataset, find-
ing an increase in classification performance. This improvement is due to both the described mechanisms. First
the synaptic down-scaling prevent the system from signal saturation, improving the sensibility of the network.
Second, the association due to thalamo-cortical interaction brings different groups of neurons to cooperate, lead-
ing to a more reliable result during the classification task.
Interestingly, our simulations produce dynamical change predictions that are coherent with recent biological
observations. First we observed that the SO frequency decreases during the sleep period. This is due to the asym-
metric STDP we used, and is consistent with experimental observations, where a decrease in SO frequency is
observed over night12. This also serves as a protection mechanism, to stop Up state mediated associations before
a possible catastrophic divergence to a fully connected network. Second, we found that neurons with low levels of
activity before sleep increased their firing rate after sleep and vice-versa, which is very similar to the global effect
of sleep observed in4, strengthening the biological relevance of our model.
We stress that SO activity is fundamental to achieve the results discussed in this paper, while asynchronous
state lacks the spatio-temporal structure suited to provide the memory association and the synaptic down-scaling
that we observed. Up states did not simultaneously occur in all the groups of neurons, but only populations
coding for images in the same class were likely to co-activate. The high level of simultaneous firing rates induced
stronger synapses between group of neurons coding for examples belonging to the same class, while popula-
tions of different classes remained largely unaffected. Also, the high firing rate promoted a generalized synaptic
down-scaling thanks to the asymmetric STDP rule. Such specific synaptic organizations would not be possible
during asynchronous activity.
One of the main peculiarities of this work is that, despite being biologically oriented, it investigates the effects
of SO activity on networks that perform ML relevant tasks. This follows a path of works in which different tasks
(such as image recognition and categorization) are performed in networks which are constrained to use biolog-
ically plausible mechanisms such as STDP plasticity and spike-based signal transmission24 -- 27. In this framework
the impact of the interaction between sleep-like activity and plasticity on the network performances and dynam-
ics has never been addressed. Indeed, in every day life humans and machines might be exposed, in different
spatial, temporal or social context, to instances of external objects belonging to the same class. Still, they could
not immediately realize that they are correlated, e.g. because driven by the urgent need to solve the current task.
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
7
www.nature.com/scientificreportswww.nature.com/scientificreports/ The association of examples sharing similar low-level features could be performed after the learning, e.g. during
sleep. With this work, we demonstrated a first mechanism by which sleep can be beneficial for both human and
artificial intelligence.
We acknowledge that in our model there are some strong assumptions, that make the network stereotyped and
still far from biological reality. For example cortical neuron are initially set to encode trained images in a way that
is completely disjoint using orthogonality in the contextual signal. Despite this is not completely unreasonable in
some extent (as discussed above), it is known that mixed selectivity neurons are fundamental in neural coding23
everywhere, and indeed are created in our model during sleep induced association. We plan to include mixed
selectivity in the contextual signal in future work.
Also, we plan to perform large scale simulations of Slow Wave Activity (SWA) on cortical areas with biolog-
ically plausible long-range lateral connections, and columnar organization (see Pastorelli et al.28), this way sup-
porting the integration of contextual lateral information in cortical model with retino-topical organization. We
speculate that the introduction of such spatial extension might allow our model to express the sub-state specificity
observed in4 and support the different features of Slow Oscillations and Micro Arousal states. On the long term,
such a detailed approach should enable a comparison of simulation results with experimental measures of slow
waves propagation patterns and transitions across different stages of sleep.
In the framework we proposed it will be interesting to test and compare different hypotheses of sleep func-
tions, some of which are not yet unified in a coherent framework1,4. We hope to contribute to refine these
hypotheses: they could find a conciliation path according to our preliminary simulation results, that point to the
possibility of having SO promoting both a hierarchical association of memories and differential synaptic home-
ostasis mechanisms.
Our approach is also promising in the direction of biologically plausible ML, and in the next future we aim to
obtain networks with superior recognition accuracy, and test out network on more complicated image datasets.
In15 the authors discuss the role of sleep in relation with the Integrated Information Theory (IIT) of conscious-
ness. We remark that the groups of neurons created by the pre-sleep training resemble the elementary mecha-
nisms of the conceptual framework described in IIT29. The increase in complexity among the groups resulting
from sleep, clearly visible in the final synaptic matrix structure, could be viewed as a step towards the creation of
higher order mechanisms. This should be associated to changes in the distribution of probabilities of the states
accessible to the system itself and on a change of its causal power. This model might constitute a simplified biolog-
ical computational setup to investigate this kind of conceptual framework.
Methods
Network architecture. The network, which is designed to be a minimal thalamo-cortical model17 (see
Fig. 1), is composed of two populations of cortical neurons (one excitatory cx and one inhibitory in) and two
thalamic populations (one excitatory tc and one inhibitory re).
The role of cortical inhibitory neurons in our model is to provide a shared inhibition supporting a
winner-take-all mechanism. Indeed when a part of the whole network responds to a specific input, the rest of the
network is inhibited. A classical choice to approximate a biological cortical network is to set a 4:1 ratio between
excitatory and inhibitory neurons. Keeping a fixed excitatory:inhibitory ratio is not critical for the model here
presented. In our runs, the ratio varied from 4:1 to 1:1 because for simplicity we kept the number of inhibitory
fixed in all simulations, while we increased the number of excitatory with to the number of training examples.
,
→w
1
re
re
,
→w
10
re
tc
,
=→w
60
tc
re
= −
cx
tc, →wtc
→w
in
cx, →wcx
p
,
=→w
cx
tc connections, predicting the thalamic configuration which activated a specific cortical activity pattern.
The connection probability is = .
1 0 for the populations connected by the arrows in Fig. 1. The untrained
. A subset
4
1
= −
in
cx) is plastic: their initial value and plasticity rules are specified later-on.
synaptic weights are
of synapses ( →wcx
The contextual signal is a Poissonian train of spikes which mimics a contextual signal coming from other brain
areas and selectively facilitates neurons to learn new stimuli. The top-down prediction is the signal flowing through
→cx
Pre-processing of visual input.
Images are pre-processed through the application of the histogram of
oriented gradients (HOG) algorithm. The size of original images is 28 × 28 pixel. Histograms are computed using
cells of 14 × 14 size that are applied on the images using a striding step of 7 pixels, resulting in 9 histograms per
image. Each cell provides an histogram with 9 bins, each bin assuming a real value. Each bin is then transformed
into a vector of 4 mutually exclusive truth values. In summary, each image has been transformed in a vector of
324 binary features. The resulting visual input is presented to the network through the thalamus, where each
feature provides a binary input to a different thalamic cell. Each cell receives a Poisson spike train with average
firing rate that is 30 kHz only when the element of the feature vector is 1. The specific number of thalamic neurons
used in the model is related to the specific pre-processing algorithm and the number of levels used to code the
pre-processing output.
Network size and training set. The number of thalamic (tc) neurons is the same as the dimension of the
feature vector produced by the pre-processing of visual input ( =N
). The cortical population is composed
of groups of 20 neurons for each image in the training set. In a first set of runs, the training set is composed of 9
images, with 3 different examples for 3 different digits (0, 1, 2). In this case the number of cortical neurons is
. In a second run, the number of training examples has been raised to 30 images (3 examples per digit,
=N
cx
). The number of inhibitory neurons are =N
=N
cx
Training algorithm and retrieval. Every time a new training image is presented to the network through
the thalamic pathway, the facilitation signal coming from the contextual signal provides a 2 kHz Poisson spike
train to a different set of 20 neurons, inducing the group to encode for that specific input stimulus (see the
(thalamic inhibition).
and =N
180
600
= −
,
→w
10
in
in
= −
200
in
200
re
324
th
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
8
www.nature.com/scientificreportswww.nature.com/scientificreports/ Discussion section for details about this choice). Additionally a 10 kHz Poisson spike train is provided to inhibi-
tory neurons (in) to prevent already trained neurons to respond to new stimuli in the training phase. Synaptic
poiss
. The learning mechanism is allowed by sym-
weights for Poisson inputs are
w
5
tc
metric (α = .1 0) spike-timing-dependent-plasticity (STDP) present in the →cx
cx con-
nections which shapes the weights structure. The maximum synaptic weight (see STDP eq. 4) are respectively
max
0
.
=→w
=→w
1
tc
cx
During the retrieval phase only the 30 kHz input to thalamic cell is provided, while the contextual signal is off.
. The initial values are
poiss
and
w
in
tc and →tc
0
=→w
cx
max
=→w
cx
0
=→w
cx
poiss
,
w
8
cx
cx, →cx
and
1
max
→w
tc
5 5
= .
and
150
130
15
,
1
=
=
=
,
cx
cx
cx
cx
tc
tc
Image classification task on MNIST dataset. Notwithstanding the aim to achieve biological plausibility,
our work is also oriented to perform tasks that are relevant in the machine learning (ML) scenario. Indeed in this
work we extended a model developed in a previous work, which was designed to perform an image classification
task relying on a small number of examples. In that model, the contextual signal projected toward cortical neu-
rons selected groups of target neurons during an incremental training of MNIST handwritten characters. The
training presented 25 novel characters per training step. After each training step, we measured the growth in the
recognition rate. The network reached an average recognition rate of 92% (average over 6 simulations) after the
presentation of 300 example characters (i.e. 30 examples per digit).
Slow oscillations. After the training stage, the sleep-like thalamo-cortical spontaneous slow oscillations
activity is induced for a total duration of 600s by providing a non-specific Poisson noise inside the cortex (700 Hz)
60, in eq. (1)). No external stimulus is provided to tc cells.
and increasing the strength of SFA parameter ( =b
.
Also, the synaptic weights between inhibitory and excitatory neurons in the cortex is reduced to
0 5
= − .
In this stage asymmetric STDP plasticity (α = .3 0) is active in the recurrent cx connectivity, inducing
sleep-induced modification in the synaptic weights structure. The parameters' change to obtain the slow oscillat-
ing regime were chosen relying on mean field theory framework18,19.
→w
in
cx
Simulation protocol. The simulated experimental protocol is composed of 4 phases (see Fig. 1B): the train-
ing phase, where visual patterns are learned by the network, the pre-sleep retrieval phase, in which patterns
are recalled, the sleep phase and the post-sleep phase. During post-sleep, the network is either exercised on the
retrieval of previous learned patterns or applied to the classification of novel examples. The network structure and
dynamics are compared in the pre- and post- sleep phases.
AdEx neurons. The neurons in the network are adaptive exponential (adEx) point like single compartment
neurons.
dV t
( )
dt
dW t
( )
dt
= −
= −
V t
( )
τ
m
W t
( )
τ
W
−
E
l
+
+
b
∑
k
Δ
τ
m
t
(
δ
V e
θ
V t
( )
−
V
Δ
+
I t V t
( ,
( ))
C
−
W t
( )
C
−
t
k
)
+
a V t
(
( )
−
E
l
)
(1)
where the synaptic input i is defined as
I t V t
( ,
( ))
∑=
α
syn
g
α
t V t
( )(
( )
−
syn
E
α
)
(2)
defines the excitatory (e) and the inhibitory (i)
where V t( ) is the membrane potential of the neuron and α = e i,
input. The population dependent parameters are: τm the membrane time constant, C the membrane capacitance,
El the reversal potential, θ the threshold, ΔV the exponential slope parameter, W the adaptation variable, a and b
the adaptation parameters, and αg syn the synaptic conductance, defined as
syn
)/
τ
α
(3)
We define the spiking time of the neuron when the membrane potential reaches the threshold
. αtk indicates the times of pre-synaptic spikes received by the neuron from synapse type α with
syn and its synaptic efficacy αQ . Where non specified, the parameters are the standard defined
V
spike
characteristic time τα
in the "aeif_cond_alpha" neuron in the NEST simulator.
V5
θ= + Δ
exp(
α
− −
Θ −
∑
k
syn
g
α
t Q
)
k
t
( )
=
t
k
)
(
(
t
t
STDP plasticity. We used spike-timing-dependent-plasticity (STDP) which potentiates wij when spike
occurs earlier in neuron j (spike time tj) than in neuron i (spike time ti) and viceversa. We considered STDP in its
multiplicative form30 which is described by the following equation
Δ
Δ
w
ij
w
ij
=
w
(
λ
αλ
= −
max
w e
)
−
ij
t
t
/
− −
τ
w e
j
i
ij
t
− −
j
t
i
/
τ
potentiation
depression
(4)
where wmax is the maximum weights value, α is the asymmetry parameter between potentiation and depression,
λ is the learning rate and τ is the STDP timescale.
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
9
www.nature.com/scientificreportswww.nature.com/scientificreports/ Simulation engine. We performed spiking simulation using the NEST simulation engine, the
high-performance general purpose simulator developed by the NEST Initiative, release 2.1231.
References
1. Tononi, G. & Cirelli, C. Sleep and the price of plasticity: From synaptic and cellular homeostasis to memory consolidation and
integration. Neuron 81, 12 -- 34, https://doi.org/10.1016/j.neuron.2013.12.025 (2014).
2. Tononi, G. et al. Center for sleep and consciousness - research (2018).
3. Killgore, W. D. Effects of sleep deprivation on cognition. In Kerkhof, G. A. & van Dongen, H. P. (eds) "Human sleep and cognition",
vol. 185 of Progress in Brain Research, 105 -- 129, https://doi.org/10.1016/B978-0-444-53702-7.00007-5 (Elsevier, 2010).
4. Watson, B., Levenstein, D., Greene, J., Gelinas, J. & Buzsáki, G. Network homeostasis and state dynamics of neocortical sleep. Neuron
90, 839 -- 852, https://doi.org/10.1016/j.neuron.2016.03.036 (2016).
5. Walker, M. P. & Stickgold, R. Sleep, memory, and plasticity. Annu. Rev. Psychol. 57, 139 -- 166 (2006).
6. Jadhav, S. P., Kemere, C., German, P. W. & Frank, L. M. Awake hippocampal sharp-wave ripples support spatial memory. Science 336,
1454 -- 1458 (2012).
7. Smulders, F., Kenemans, J., Jonkman, L. & Kok, A. The effects of sleep loss on task performance and the electroencephalogram in
young and elderly subjects. Biological psychology 45, 217 -- 239 (1997).
8. González-Rueda, A., Pedrosa, V., Feord, R. C., Clopath, C. & Paulsen, O. Activity-dependent downscaling of subthreshold synaptic
inputs during slow-wave-sleep-like activity in vivo. Neuron 97, 1244 -- 1252 (2018).
9. Wei, Y., Krishnan, G. P. & Bazhenov, M. Synaptic mechanisms of memory consolidation during sleep slow oscillations. Journal of
Neuroscience 36, 4231 -- 4247 (2016).
10. Wei, Y., Krishnan, G. P., Komarov, M. & Bazhenov, M. Differential roles of sleep spindles and sleep slow oscillations in memory
consolidation. PLoS computational biology 14, e1006322 (2018).
11. de Vivo, L. et al. Ultrastructural evidence for synaptic scaling across the wake/sleep cycle. Science 355, 507 -- 510, https://doi.
org/10.1126/science.aah5982 (2017).
Reviews Neuroscience 3, 679 (2002).
12. Hobson, J. A. & Pace-Schott, E. F. The cognitive neuroscience of sleep: neuronal systems, consciousness and learning. Nature
13. Sanchez-Vives, M. V., Massimini, M. & Mattia, M. Shaping the default activity pattern of the cortical network. Neuron 94, 993 -- 1001,
https://doi.org/10.1016/j.neuron.2017.05.015 (2017).
14. Vyazovskiy, V. V. & Harris, K. D. Sleep and the single neuron: the role of global slow oscillations in individual cell rest. Nature
Reviews Neuroscience 14, 443 -- 451, https://doi.org/10.1038/nrn3494 (2013).
15. Bucci, A. & Grasso, M. Sleep and dreaming in the predictive processing framework. In Metzinger, T. K. & Wiese, W. (eds) Philosophy
and Predictive Processing, chap. 6, https://doi.org/10.15502/9783958573079 (MIND Group, Frankfurt am Main, 2017).
16. Larkum, M. A cellular mechanism for cortical associations: an organizing principle for the cerebral cortex. Trends in Neurosciences
36, 141 -- 151, https://doi.org/10.1016/j.tins.2012.11.006 (2013).
17. Destexhe, A. Self-sustained asynchronous irregular states and up -- down states in thalamic, cortical and thalamocortical networks of
nonlinear integrate-and-fire neurons. Journal of computational neuroscience 27, 493 (2009).
18. Gigante, G., Mattia, M. & Del Giudice, P. Diverse population-bursting modes of adapting spiking neurons. Physical Review Letters
19. Capone, C. et al. Slow waves in cortical slices: how spontaneous activity is shaped by laminar structure. Cerebral Cortex 1 -- 17 (2017).
20. Contreras, D. & Steriade, M. Cellular basis of eeg slow rhythms: a study of dynamic corticothalamic relationships. Journal of
98, 148101 (2007).
Neuroscience 15, 604 -- 622 (1995).
21. Steriade, M., Dossi, R. C. & Nunez, A. Network modulation of a slow intrinsic oscillation of cat thalamocortical neurons implicated
in sleep delta waves: cortically induced synchronization and brainstem cholinergic suppression. Journal of Neuroscience 11,
3200 -- 3217 (1991).
22. Sanchez-Vives, M. & Mattia, M. Slow wave activity as the default mode of the cerebral cortex. Arch. Ital. Biol 152, 147 -- 155 (2014).
23. Rigotti, M. et al. The importance of mixed selectivity in complex cognitive tasks. Nature 497, 585 (2013).
24. Sacramento, J., Costa, R. P., Bengio, Y. & Senn, W. Dendritic error backpropagation in deep cortical microcircuits. arXiv preprint
25. Diehl, P. U. & Cook, M. Unsupervised learning of digit recognition using spike-timing-dependent plasticity. Frontiers in
arXiv:1801.00062 (2017).
computational neuroscience 9, 99 (2015).
26. Cayco-Gajic, N. A., Clopath, C. & Silver, R. A. Sparse synaptic connectivity is required for decorrelation and pattern separation in
feedforward networks. Nature Communications 8, 1116 (2017).
27. Nicola, W. & Clopath, C. Supervised learning in spiking neural networks with force training. Nature communications 8, 2208 (2017).
28. Pastorelli, E. et al. Gaussian and exponential lateral connectivity on distributed spiking neural network simulation. In 2018 26th
Euromicro International Conference on Parallel, Distributed and Network-based Processing (PDP), vol. 00, 658 -- 665, https://doi.
org/10.1109/PDP2018.2018.00110 (2018).
29. Tononi, G. & Koch, C. Consciousness: here, there and everywhere? Philosophical Transactions of the Royal Society B: Biological
Sciences 370, https://doi.org/10.1098/rstb.2014.0167 (2015)..
30. Gütig, R., Aharonov, R., Rotter, S. & Sompolinsky, H. Learning input correlations through nonlinear temporally asymmetric hebbian
plasticity. Journal of Neuroscience 23, 3697 -- 3714 (2003).
31. Kunkel, S. et al. Nest 2.12.0, https://doi.org/10.5281/zenodo.259534 (2017).
Acknowledgements
This work has been supported by the European Union Horizon 2020 Research and Innovation program under
the FET Flagship Human Brain Project (SGA2 grant agreement SGA2 n. 785907), System and Cognitive
Neuroscience subproject, WaveScalES experiment. This work is part of the activities performed by the INFN APE
Parallel/Distributed Computing laboratory, and we are grateful to the members of the APE lab for their strenuous
support.
Author Contributions
C.C., B.G. and P.P. conceived the experiment, C.C. and E.P. conducted the experiment, C.C. and P.P. analyzed the
results. All authors wrote and reviewed the manuscript.
Additional Information
Competing Interests: The authors declare no competing interests.
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
10
www.nature.com/scientificreportswww.nature.com/scientificreports/ Open Access This article is licensed under a Creative Commons Attribution 4.0 International
License, which permits use, sharing, adaptation, distribution and reproduction in any medium or
format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Cre-
ative Commons license, and indicate if changes were made. The images or other third party material in this
article are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the
material. If material is not included in the article's Creative Commons license and your intended use is not per-
mitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the
copyright holder. To view a copy of this license, visit http://creativecommons.org/licenses/by/4.0/.
© The Author(s) 2019
Scientific RepoRts (2019) 9:8990 https://doi.org/10.1038/s41598-019-45525-0
11
www.nature.com/scientificreportswww.nature.com/scientificreports/ |
1610.02395 | 1 | 1610 | 2016-10-08T00:39:25 | Causal-order superposition as an enabler of free will | [
"q-bio.NC"
] | It is often argued that bottom-up causation under a physicalist, reductionist worldview precludes free will in the libertarian sense. On the one hand, the paradigm of classical mechanics makes determinism inescapable, while on the other, the leading models that allow a role for quantum effects are noncommittal regarding how conscious agents are supposed to translate indeterminacy into self-formed choice. Recent developments, however, not only imply that self-formed decisions are possible, but actually suggest how they might come about. The cornerstone appears to be causality superposition rather than quantum-state entanglement, as is usually assumed, and the natural arena for applying these developments is (perhaps ironically) a framework that was built without any consideration for quantum effects. | q-bio.NC | q-bio | Causal-order superposition as an enabler of free will
Salvador Malo*
October 4, 2016
Abstract. It is often argued that bottom-up causation under a physicalist, reductionist worldview
precludes free will in the libertarian sense. On the one hand, the paradigm of classical mechanics
makes determinism inescapable, while on the other, the leading models that allow a role for quantum
effects are noncommittal regarding how conscious agents are supposed to translate indeterminacy
into self-formed choice. Recent developments, however, not only imply that self-formed decisions
are possible, but actually suggest how they might come about. The cornerstone appears to be causality
superposition rather than quantum-state entanglement, as is usually assumed, and the natural arena
for applying these developments is (perhaps ironically) a framework that was built without any
consideration for quantum effects.
The libertarian notion of free will, specifically the interpretation denoted as (freedom2) in [3]†, tends
to come under fire for assuming the possibility of actions that are self-forming, i.e. neither random
nor determined by pre-existing conditions. Under the paradigm of classical physics, determinism
follows from bottom-up causation, and self-forming actions are a nonstarter. Quantum mechanics, on
the other hand, because it allows actions to be triggered by random microscopic fluctuations, has
often been interpreted (e.g. in [6]) as merely replacing the yoke of determinism by that of chance,
which is hardly an improvement. This line of criticism has gained credence because the strongest
proposals that call attention to quantum indeterminacy as an enabler of free will (notably [5,10]) do
not explain to the satisfaction of naysayers how the harnessing of indeterminacy takes place, in other
words what makes decoherence a choice as opposed to a mere event.
Causal Order and Reductionism
When the bottom flaps of a cardboard box are tucked under their neighbors in an overlapping pattern,
the resulting floor can be rigid enough to carry lightweight objects without the help of tape. By
denoting the set of flaps of a square box as F = {F1, F2, F3, F4}, it is clear that its strength is a property
of F but not of any single Fi in isolation. (If the flaps could voice an opinion, they might claim their
individual role to be merely that of passing weight on to a neighbor, thus making no effort of their
own.)
Ignoring the subtleties of mechanics and following this abstraction to its conclusion, it could be said
that F as a whole (rather than any Fi alone) is what prevents the object from falling. Moreover, in the
causal chain that begins with an object being placed inside the box and ends with the object either
being supported or not, F is clearly a middle link. But any attempt to increase the granularity of
causation by zooming into the elements of F runs into trouble, for it is impossible to select one
* [email protected]
† To distinguish it from the harder-to-defend (freedom3)
1
particular flap Fi as the first link within that subset of the causal chain. In short, causal order within
F itself is indefinite.
From a physicalist, bottom-up causation viewpoint the above abstraction is of course untenable,
because the floor of the box is a macroscopic object of which F is but an imperfect description. The
box comprises a vast number of material particles that are acted upon by forces that only in concert
grant rigidity to F and prevent the object from falling through. A careful (even if impractical)
dissection of all the microscopic elements involved should in principle reveal a first link in the
causation chain, whether a first collision of a subatomic particle of the dropped object against another
particle from the cardboard surface, or a random fluctuation that unleashes a sequence of events that
ends in the object falling or staying put.
"So what?" might claim the engineer, for whom microscopic fluctuations are irrelevant as long as the
statistical average of the forces involved allows the object to be supported for a reasonable amount of
time. (Nobody cares whether a box lies at the mercy of particle collisions as long as it fulfills its
intended purpose.) But when human brains and other ostensibly conscious agents are subject to the
same scrutiny, such dismissive attitudes become unpalatable, and the resulting discomfiture has led
even to the suggestion that the concept of causation must be re-examined [4].
Conscious Causality
Suppose that a 3-member committee M = {M1, M2, M3} generates approval/rejection rulings R on
proposals P based either on immediate voting or on a deliberation process D, assumed to take place
behind closed doors and without communication with the outside world.
Immediate voting is algorithmic and generates R as a simple average of member votes, which in turn
reflect the pre-existing views of committee members. In contrast, D allows information exchange
between the Mi to influence each other's views. R remains undetermined until D ends, even if
individual views are polled before D begins. Moreover, the order in which Mi can influence Mj is
indefinite (as anyone knows who has participated in deliberations). Of course, this is only because
individual members play the role of homunculi within M. But it will be useful to summarize the setup
in terms that can be applied later:
1. Since each Mi is assumed to be autonomous, the order in which they exchange information is
indefinite, so R is undetermined until it is produced, even if the initial states of all Mi are
known
2. But R isn't random, for it depends on the views of M's elements.
M is therefore irreducibly the free causal agent of outcome R. However, the translation of this
metaphor to the case of an individual mind is problematic, because the autonomy of a mind's elements
is questionable, and because it isn't immediately clear what the equivalent of D should be within the
context of a physical brain. In order to reformulate it for such a context, several obstacles must be
overcome.
2
Delimiting the Agency
First, there is the question of delimitation, or the extension of the physical agent. Is it the person, the
person's brain (excluding the skull but including the spinal cord and all its nerve endings), the
neocortex, a specific network within a part of the brain, a set of microtubulesor none of the above?
Although its explanatory power as a model of consciousness is strongly contested, Integrated
Information Theory (IIT) goes a long way towards characterizing decision agents, so for the purpose
of this discussion an agent A shall be assumed to be an integrated system as defined in [7]. IIT's
definition is a formal abstraction that circumvents the issue of A's physical constitution, but it
identifies which states of a system can play the roles of cause and effect, and excludes everything else
from the discussion (a much-welcome simplification). A scaled-down version of an example from [7]
is drawn below, with input I, output O, elements Ai and connections that are either excitatory (solid)
or inhibitory (dotted). Internal feedback loops are one of its essential features. In order to forestall the
threat of homunculi, the elements Ai are assumed to be nothing more than logic gates. Can A be
characterized so that its properties are similar to those of M in the committee example?
If A is assumed to behave classically, then for elements Ai in known initial states, and for a given
input I, the production of O follows a deterministic causal chain that begins with I and terminates
with the first output O (the persistence of feedback within A may allow O to be followed by more
outputs, but that would merely extend the deterministic chain).
Shielding A from Chance
If A is assumed to behave quantum-mechanically, then several additional prospects arise, not all of
them necessarily good. There is, to begin with, a possible susceptibility to quantum fluctuations,
which in order to be useful rather than a nuisance must be exploited by the agent rather than by the
environment. In the absence of such advantage, it is clearly preferable for the Ai to be impervious to
off-network noise, thus preventing unwelcome microscopic randomness from being amplified to
macroscopic scale. Within-network quantum effects, however, are another matter, and it is ironical
that IIT is formulated entirely in classical terms, for it is ideally suited for quantum interpretations of
causality.
3
The next prospect is superposition. But what sort of superposition, if any, increases A's freedom of
action? The literature's favored candidate (quantum state entanglement) may in fact be a red herring.
In the case of the fictitious committee M described earlier, it isn't helpful for the opinions of
individuals to be entangled (so that e.g. measuring M2's view becomes redundant once M1's is
known). That would turn M into a condensatea partisan, rubberstamping, ineffectual blob. The
chief virtue of deliberation is that it prevents regrets arising from uninformed voting. Stated otherwise,
D is meant to improve decision-making, not to render individual views indistinguishable.
Returning to A, its appeal as an agency is the richness of its internal network, the complexity of which
is worth exploiting rather than smearing out. In the example above, A1 is by design the first element
to react to I. Also by design, once A4 is first acted upon, measurable output is generated which
constitutes a causal endpoint for external observers, even if internal feedback subsequently allows
additional signals to emerge as O. IIT shows in [7] that such networks are functionally equivalent to
deterministic "zombie" systems. But this is only because IIT's conceptual approach is formulated
without any reference to quantum effects. If the elements Ai were allowed to communicate in no
definite order, yet still according to the rules of the network (even if performing known operations, or
the equivalent of an Mi speaking its mind), then A would remain the irreducible cause of O without
its output being either smeared out to trivial randomness or cranked out Turing-like. What is called
for isn't state entanglement, but causal superposition, where the Ai are defined but the order in which
they communicate is not. As it happens, causal superposition is not only possible in theory [8] but
has been implemented in practice [9]. But, why is it useful?
Shielding A from Determinism
As the diffraction pattern from a double-slit experiment famously shows, a photon behaves differently
when presented with several alternative paths than when presented with only one. The alternative in
question is the path through the other slit, and the different behavior is the impact spot which (in the
aggregate) builds up a diffraction pattern. One interpretation of this effect is that a photon's actual
impact is influenced by the alternative trajectories it might have taken. What might have been
influences what in fact was. For a conscious being habituated to constant what-if considerations, this
is nothing special, but for a single photon's behavior it is perplexing, to say the least.
When speaking of causal-order superposition instead of state superposition, the alternative in question
involves cause-effect reversal rather than a state entanglement. For a process in causal-order
superposition, two distinct events are in the past of each other. Not surprisingly, the mathematics are
nontrivial, but a precise description is given in [8], which shows that this form of superposition cannot
be described by a probabilistic mixture of processes with definite causal order. Moreover, the
framework avoids paradoxes such as those stemming from closed time-like curves.
The implication is that when the signaling over a network shared by A1 and A2 is in causal-order
superposition and a qubit I is sent from one element to another, then even when information flows in
just one sense (say, from A1 to A2), the mere possibility that it "might have" gone in the opposite sense
is enough for A1 to make a better guess of A2's measurement (of I) than in the absence of causal
superposition. Stated in terms of informatics, the outcome of a logic-gate's operation on an
undetermined input can be influenced by the might-have-been operation of another logic gate in its
causal future, had it been in its causal past. This allows better-informed outputs O to be generated by
4
a system A than would be produced if its elements operated under a definite causal order. The situation
is manifestly different from that of classical feedback, where ordinary causality merely allows an
operation performed by A1 to be overwritten at a later time (based on A2's classical feedback), but
doesn't allow A2 to influence the first signal of A1!
The framework has been extended in [1] to show that absence of causal order allows perfect signaling
between 3 parties the first time around (i.e. perfect agreement of their first signals). Even more
complex examples will surely be developed as the logic and the consequences of the framework are
further explored. But the work done in [9] already proves that the above isn't mere speculation:
physical systems can be experimentally set up so that their individual elements do not (even in
principle) exchange information in a particular time order. If A is set up in such a way (and if science
has built a simple example, why wouldn't nature have built more elaborate ones?) this means that A's
decisions cannot not be determined (even in principle) by full knowledge of I and of the states of the
elements Ai at any time before O is produced. After input I is fed into a system A whose elements
communicate in a causally indefinite order, its output O can be claimed to be caused by A, and might
arguably be said to be "influenced" by I (which would require yet another definition), but is
emphatically not determined by the pre-existing state of elements Ai and/or input I.
Despite its classical formulation, IIT contributes more than just an abstract description of purportedly
conscious agents. One of its key conclusions is that "experiences" (defined in a precise way in [7],
but meaning A's internal representations of O) are intrinsic properties of integrated systems in a given
state. This conclusion relies merely on the ability of IIT to describe the neural mechanisms of
subjective perception, not on its validity as a model of consciousness in a more general sense (and
even less on whether or not it addresses the Hard Problem [2] or is subject to fading qualia arguments).
What IIT explicitly shows is how experiences are self-generated by A alone and in a manner that is
self-referential to A alone. To wit, A doesn't have to generate any measurable output, nor does it need
to consult any additional‡ input in order to form an internal representation of O.
This deserves emphasizing, for it spells out, in the context of A, an equivalent of the committee's
deliberation D. The feedback or re-entry required of integrated systems in [7] paves the way for A to
experience a given O without reference to its resultant output. This is crucial, for if O had to be written
out in order for A to experience it, then it would already be too late for the choice to be examined
before being made. The examples outlined in [1,8] and explicitly built in [9] are all defined in terms
of co-operative goal achievement, the likelihood of which is maximized under indefinite causal order.
The goal in those examples is arbitrarily imposed by the experimenters and (by necessity) rather
simple, but in more complex cases could be defined in terms of A's own subjective standards, based
for example on the stored representations of previous experiences. Causal-order superposition would
then lead to the instant production of an output O satisfying such internally-defined goal, which would
also be internally representable (perhaps later, as a classically stored memory).
To summarize, causal-order superposition applied to an integrated system A defined as in [7] but
subject to the requirements of [9] would allow A to carry out actions that are formed entirely within
‡ Additional because input I was earlier assumed have already been fed into A.
5
A, based on A's experience of possible alternatives, and neither random nor deterministic. Proof may
yet come in experimental form (as it always should), but what to call this if not libertarian free will?
References
[1] Baumeler Ä & Wolf S (2014) Perfect signaling among three parties violating predefined causal
order. Preprint at http://arxiv.org/abs/1312.5916
[2] Chalmers D (1995) Facing up to the problem of consciousness. Journal of Consciousness
Studies 2(3): 200-219
[3] Dorato M (2002) Determinism, Chance, and Freedom. In Atmanspacher H & Bishop R (Eds.)
Between Chance and Choice (Imprint Academia, Thorverton) p. 347
[4] Ellis G (2016) How can Physics Underlie the Mind? (Springer, Berlin, Heidelberg, New York)
[5] Hameroff S, Penrose R (2013) Consciousness in the universe: A review of the 'Orch OR'
Theory. Physics of Life Reviews. Elsevier 11 (1): 39-78 doi:10.1016/j.plrev.2013.08.002
[6] Harris S (2012) Free Will (Free Press) pp. 27-30
[7] Oizumi M, Albantakis L, Tononi G (2014) From the Phenomenology to the Mechanisms of
Consciousness: Integrated Information Theory 3.0. PloS Computational Biology 10(5):e1003588
[8] Brukner Časlav (2012) Quantum causality. Nature Physics 10, doi:10.1038/nphys2930
[9] Rubino G, et al. (2016) Experimental Verification of an Indefinite Causal Order. Available
online at: https://arxiv.org/abs/1608.01683
[10] Stapp H P (2007) Mindful Universe (Springer, Berlin, Heidelberg, New York)
6
|
1510.06629 | 1 | 1510 | 2015-10-22T13:57:03 | Using persistent homology to reveal hidden information in neural data | [
"q-bio.NC",
"math.AT"
] | We propose a method, based on persistent homology, to uncover topological properties of a priori unknown covariates of neuron activity. Our input data consist of spike train measurements of a set of neurons of interest, a candidate list of the known stimuli that govern neuron activity, and the corresponding state of the animal throughout the experiment performed. Using a generalized linear model for neuron activity and simple assumptions on the effects of the external stimuli, we infer away any contribution to the observed spike trains by the candidate stimuli. Persistent homology then reveals useful information about any further, unknown, covariates. | q-bio.NC | q-bio |
Using persistent homology to reveal hidden
information in neural data
Gard Spreemann1, Benjamin Dunn2, Magnus Bakke Botnan1, and Nils A. Baas1
1{gard.spreemann, magnus.botnan, nils.baas}@math.ntnu.no, Department of Mathematical Sciences,
Norwegian University of Science and Technology, 7491 Trondheim, Norway
[email protected], Kavli Institute for Systems Neuroscience, Norwegian University of Science
and Technology, 7491 Trondheim, Norway
October 10, 2018
Abstract
We propose a method, based on persistent homology, to uncover topological properties
of a priori unknown covariates of neuron activity. Our input data consist of spike train mea-
surements of a set of neurons of interest, a candidate list of the known stimuli that govern
neuron activity, and the corresponding state of the animal throughout the experiment per-
formed. Using a generalized linear model for neuron activity and simple assumptions on the
effects of the external stimuli, we infer away any contribution to the observed spike trains by
the candidate stimuli. Persistent homology then reveals useful information about any further,
unknown, covariates.
1
1 Introduction
Due to its apparent simplicity, physical space has long served as a model system for internally
generated representations in the brain [32]. In their pioneering work [26], O'Keefe and Dostro-
vsky discovered in the hippocampi of rats certain neurons, called place cells, that seemed to be
active at a level far above their baseline when the animal was located in a specific region in
space. These regions of elevated activity are known as place fields.
It has also been demon-
strated [23, 31, 30] that neurons tune not only to spatial position, but also to other external
covariates, such as for example head direction. Place fields thus do not exist only for an an-
imal's physical surroundings -- we shall rather think of them as regions in an abstract state
space, thus generalizing to the more fundamental concept of neural coding. An example of the
place fields of three cells recorded in an experiment by Buzsáki et al.
[22, 21] can be seen in
Figure 1.
Figure 1: The firing activity of three neurons as a rat explores a square box. For the two leftmost neurons
the activity is plotted against spatial position (both axes) while the activity for the rightmost one is plotted
against head direction (horizontal axis). The regions of elevated activity, the spatial and "head direction
state space" place fields, are clearly visible. In the spatial plots, the cells were least active in the blue areas,
and most in the red. White areas were not visited by the rat during the experiment. Activity and position
data for the plots come from [22, 21].
If a collection of neurons divide the animal's state space into reasonably nice place fields, it
is perhaps not surprising that the activity of these neurons somehow encode information about
the state space. A question to pose is then how much of that space can be recovered from just
recordings of the electrical activity of the neurons. We will consider this question from a novel
point of view in which the state space is partially known, and see what properties of the un-
known parts can be glanced from such recordings. This, in turn, will give us information about
unknown external covariates of neuron activity by using a generalized linear model (GLM) [20]
to infer away the contributions of any known covariates. GLMs have recently been applied in
neuroscience as a structured method to uncover dependencies as well as structure in data [27,
33, 28].
It should be pointed out that in order to compute or visualize the place fields themselves, as
done for Figure 1, the activity data for the cells must be supplemented by positional and other
state space data, e.g. head direction or other external covariates, for the animal. Without prior
knowledge of the nature of the covariates, i.e. prior knowledge of the unknown state space, it
is of course not apparent what data should be recorded. Since it is therefore too much to ask
to know the actual place fields themselves, we take a cue from topology and instead try for
only information on how the place fields fit together. Indeed, if a set of place fields (which are
themselves unknown) intersect, we should be more likely to observe the corresponding cells
2
0.000.250.500.751.00Horiz.pos.0.000.250.500.751.00Vert.pos.0.000.250.500.751.00Horiz.pos.0.000.250.500.751.00Vert.pos.0.01.53.04.56.0HeaddirectionNeuronactivity firing simultaneously, or nearly so, in time. Figure 2 illustrates the idea. This will be our guide
when we build a combinatorial topological space from neuron activity data in the form of firing
event time series (so-called spike trains). We will, in other words, transform our original time
series data -- the collection of spike trains -- into geometric data, which we will then study
with tools from topology.
Figure 2: Place cell cofiring as a proxy for place field intersections. As the animal moves along the indi-
cated path (left), we might observe the (highly idealized) firing events (center) of the corresponding place
cells. The firing events in the leftmost box are indicative of the triple intersection of the red, green and
blue place fields, and those in the rightmost box are indiciative of the double intersection of the red and
blue fields. After the space has been thoroughly explored, the intersection data obtained from the cofiring
of the spike trains, define a simplicial complex (right) by means of the nerve construction. The simplicial
complex and the space covered by place fields have the same homotopy type (in this example, that of a
point).
Cofiring is thus a proxy that gives us approximate knowledge of how all the place fields in-
tersect (doubly, triply, etc.). We turn to topology to ask what we can learn from these data. The
natural next step, when armed with intersection data, is to construct the nerve of the place fields.
The unfamiliar reader will find the formal definition in Section 2, but intuitively the process goes
as follows: for every cell, draw a point; whenever two place fields intersect, connect the corre-
sponding two points with a line segment; whenever three place fields intersect, draw a filled
triangle between the three corresponding points; whenever four place fields intersect, draw a
filled tetrahedron between the four corresponding points; and so forth. The abstract combinato-
rial construction so obtained is one example of a simplicial complex, a kind of topological space.
It is a famous theorem (Theorem 2) in algebraic topology that this simplicial complex shares an
important property with the space covered by the objects whose intersections we consider, i.e.
the state space covered by place fields in our case.
The property that the nerve of the place fields shares (under some assumptions) with the
space covered by the place fields is that of homotopy type. A reader unfamiliar with elementary
topics in algebraic topology may want to see [16], or can think of spaces with the same homo-
topy type as those that can be continuously deformed into one another without tearing. Thus, a
square, a disk and a point are of the same homotopy type, while they are of different types from
an annulus, which is again of different type from a sphere. This answers the question of what
we may hope to recover from the cells' firing information: not the full state space (i.e. not the
environment's full geometry in a setting with only spatial place cells), but rather its homotopy
type, which again says something about the kind of holes the space has. We should for example,
in the purely spatial case, at the very least be able to tell whether there is an obstacle, such as an
impassable column, somewhere in a box the rat explores (making it, in the eyes of homotopy, an
annulus rather than a box). Moreover, the circular nature of a covariate such as head direction
also influences the homotopy type of the state space, and is therefore detectable.
3
We have so far glossed over what it means for cells to cofire. Clearly a single cofiring event is
not sufficient to declare place fields as overlapping, as cells may fire spuriously with the animal
outside of place fields. Too strict a notion of cofiring is obviously not good either. Moreover,
looking to the real-world situation in Figure 1, the degree of intersection we should demand
to declare two fields as intersecting is not as clear-cut as presented in the idealized setting of
Figure 2. Persistent homology, which we recap in Section 2, offers a way to consider all degrees of
cofiring and intersections simultaneously and as one mathematical object. While not as sharp
an invariant as homotopy type, (persistent) homology still captures the number of holes in the
space under consideration, their dimensionality, and to some extent also their size.
In summary, the correlations of spike trains allow us to build combinatorial spaces that re-
flect topological properties of some partially unknown animal state space. In the purely spatial
setting, several papers [7, 8, 15] have already demonstrated the feasibility of recovering proper-
ties of the animal's physical environment in this way. We now propose a new method wherein
topological properties of a partially unknown state space are uncovered by successively ac-
counting for known covariates using a GLM.
1.1 Contributions
Already in [26] it was clear that spatial position is not the only influence on the firing of certain
neurons. Other covariates, confirmed or suspected, include [23, 31, 30]: head direction, theta
wave phase, behavioral tasks (eating, biting, sleeping, grooming, etc.) and tactile or olfactory
sensory stimuli.
In addition to these external influences, neurons are connected to a set of
neighboring neurons, and are excited or inhibited by the activity of those neighbors.
In this paper, we consider a general setting of neural data governed by a list of suspected
possible covariates (also referred to as stimuli or influences). We further assume that the re-
searcher supplies spike train recordings of the relevant neurons, and, in addition, measurements
of the suspected physical covariates throughout the experiment. We then set out to answer the
following questions:
Q0: What is the (persistent) homology encoded by the covariates, i.e. what is the (persistent)
homology of the corresponding state space?
Q1: Does the list of suspected covariates adequately explain the observations?
Q2: If not, do the covariates missing from the list of candidates encode non-trivial homological
information?
The purely spatial version of question Q0 has already been examined by others [7, 8, 15].
Standard spectral methods, for example using Wigner's semicircle law [34] on the correlations
of neurons, may be sufficient to answer question Q1. Still, we will in this paper approach that
question using persistent homology, as done in [15]. We believe that question Q2 has not been
considered before, and that its answers can be useful to neuroscientists. The technique we
outline in this paper seeks to answer that question. Moreover, the method is quite general and
may be useful in other applications both inside and outside of neuroscience.
1.2 Outline
Section 2 reviews the necessary theory of persistent homology at the level of elementary applied
topology, recapping also some of the major results that we use. We then briefly describe the
model we use for neuron activity.
4
Section 3 describes the process of inferring away the contributions of covariates until the
list of suspected such is empty, whereupon information about the homology of any hidden
covariates is revealed. Demonstrations of the efficacy of the technique with simulated neuronal
data under various conditions follow in Section 4.
Finally, Section 5 sketches some possible applications of our technique outside of neuro-
science. We also attempt to place it on a firmer and more general mathematical ground.
1.3 Related work
Curto and Itskov [7] were the first to employ tools from algebraic topology to study neural data
in general, and in particular to attempt reconstruction of the homology of an animal's physical
environment from such data.
Dabaghian, Mémoli, Frank, and Carlsson [8] form a model of neuron activity, and use persis-
tent homology very much along the lines sketched above to study qualitative properties of this
model. The properties include the time taken to form a homologically correct representation of
the environment, and the robustness of this representation with regard to model parameters.
Arai, Brandt, and Dabaghian [1] improve on this model, add theta wave phase-precession, and
study its effect on the persistent homology learning of the environment.
Giusti, Pastalkova, Curto, and Itskov [15] very recently showed how persistent homology
can help determine whether the information encoded in the spike trains of neurons encodes for
something geometric, or is just random, thus answering our question Q1.
Our work is in part based on the same approach as [15], but seeks also to answer ques-
tions Q0 and Q2 within one framework, and should also be applicable in a very general setting
not specific to neuroscience.
2 Background
In this section we survey the necessary tools from persistent homology in particular and topo-
logical data analysis (TDA) in general, and fix basic notation. Thorough introductions can be
found in [13, 10, 11]. Familiarity with elementary algebraic topology is assumed (see [16] for an
introduction).
The p-simplices of a simplicial complex K with vertex set V will be written like [i0, i1, . . . ip] ⊆
K, where each i0, . . . , ip ∈ V are implicitly distinct. The geometric realization of K will be de-
noted K. Unless otherwise is noted, Hk(K) denotes the k'th simplicial homology of K, com-
puted with coefficients in Z/2Z. All simplicial complexes we consider are assumed to be finite.
As hinted at in Section 1, our basic data will be intersection information, or a proxy thereof.
The following basic construction that lies at the heart of TDA encodes such data as a simplicial
complex.
Definition 1. Let X be a topological space. The nerve of a collection of sets U = {Ui ⊆ X i ∈ I}
is the simplicial complex NU defined by
[i0, i1, . . . , ip] ⊆ NU ⇐⇒ Ui0 ∩ Ui1 ∩ · · · ∩ Uip (cid:54)= ∅.
The nerve construction's importance to TDA stems from its ability to discretely encode the
homotopy type of a topological space, as the nerve theorem shows. The theorem exists in a more
general form, but for the purpose of our paper it suffices to state it for metric spaces.
5
property that for all subsets J ⊆ I, the intersection(cid:84)
Theorem 2. Let X be a metric space, and let U = {Ui
NU have the same homotopy type.
i ∈ I} be a finite closed cover of X with the
j∈J Uj is either empty or contractible. Then X and
A cover satisfying the above condition is said to be good. Throughout, we let B(x; r) denote
the closed ball of radius r centered at x ∈ Rn. One aspect of TDA is the study of point clouds
-- finite point sets in Euclidean space -- by applying Theorem 2 to a nerve of closed balls.
Definition 3. Let P = {p1, p2, . . . , pN} ⊆ Rn be a point cloud. The Cech complex of P at scale ε is
the nerve
C(P; ε) = N{B(p; ε/2) p ∈ P}.
We write [i0, . . . , ik] ⊆ C(P; ε) for the k-simplex corresponding to the intersection B(pi0; ε) ∩
· · · ∩ B(pik; ε).
Since a finite set of closed balls is a good cover of its union, Theorem 2 ensures that the
denoting
purely combinatorial nerve reflects the homotopy type of that union. Thus, with Hsing
singular homology with coefficients in Z/2Z,
∗
(cid:91)
p∈P
∼= Hsing
∗
(cid:0)
(cid:1)
(cid:0) C(P; ε)(cid:1) .
Hsing
∗
B(p; ε/2)
C(P; ε)
∼= H∗
(1)
In equation (1), the first isomorphism is a consequence Theorem 2, while the second is a stan-
dard result in algebraic topology. Computing the right hand side amounts to doing linear alge-
bra.
2.1 Persistent homology
If a point cloud is assumed to have been sampled from some unknown subspace X of Rn, then
it is reasonable to apply equation (1) in order to reconstruct Hsing
(X). This of course begs the
question: what is the "correct" scale at which to view the point cloud, or, explicitly in the case
of the Cech complex, what (if any) is the "right" ball radius? Persistent homology sidesteps the
question by ecompassing all scales in one unifying construction.
∗
We define the necessary constructions in a rather compact language below. The unfamiliar
reader may want to read the wordier survey [12].
Definition 4. A persistence module V (over R) is a collection of finite-dimensional k-vector spaces
{V(t)
t ∈ R} and linear maps {V(s ≤ t) : V(s) → V(t) s, t ∈ R, s ≤ t} satisfying
V(s ≤ s) = idV(s) and V(s(cid:48) ≤ t) ◦ V(s ≤ s(cid:48)) = V(s ≤ t) for all s ≤ s(cid:48) ≤ t.
interval I (cid:54)= ∅ ⊆ R to be a set such that if s, t ∈ I, then s(cid:48) ∈ I for all s(cid:48) satisfying s ≤ s(cid:48) ≤ t.
Definition 5. The interval persistence module χI on an interval I ⊆ R is given by
In other words, V is functor from R to finite dimensional k-vector spaces. We define an
(cid:40)
χI(s) =
k
0
if s ∈ I
otherwise
where any two nontrivial vector spaces are connected by the identity map.
6
Direct sums of persistence modules are defined point-wise, i.e. (V ⊕ W)(s) = V(s) ⊕ W(s)
and similarly for the maps, and those that can only be written as trivial direct sums are called
indecomposable. It is easy to see that interval persistence modules are indecomposable. An exis-
tence theorem [6] guarantees that there for every V exists a multiset B(V) of intervals in R such
that
(cid:77)
I∈B(V)
V ∼=
χI.
That this decomposition is essentially unique is ensured by the Azumaya -- Krull -- Remak -- Schmidt
theorem [2]. In particular, (the isomorphism class of) a persistence module V is completely de-
termined by the multiset B(V). This is the famous barcode description of V. Moreover, an inter-
val in R can be canonically identified with a point in (R ∪ {±∞})2 by means of its endpoints.
Hence, a persistence module is completely described by a multiset of points in (R ∪ {±∞})2;
we shall refer to this description as the persistence diagram of V.
Definition 6. A finite filtration K of simplicial complexes
Kε0 ⊆ Kε1 ⊆ · · · ⊆ KεN
for −∞ = ε0 < ε1 < · · · < εN gives rise to the k'th persistent homology module PHk(K) of K
defined by
PHk(K)(s) = Hk(Kmax{εi εi≤s})
with the linear maps induced by inclusions. This is commonly known as a sublevel persistence
module.
We will often refer to PH∗(K) as a persistent homology module without specifying the ho-
mology dimension. Observe that as K is a finite filtration of simplicial complexes, there will
only be a finite number of intervals in B(PH∗(K)).
It is often useful to reduce the information content of persistent homology to an even sim-
pler, if incomplete, descriptor. In line with the k'th Betti number being defined as the rank of
the k'th homology group, we define the following.
Definition 7. Let V be a persistence module. Its Betti curve is the function β : R → N defined
by β(s) = dim (V(s)). When referring to persistent homology modules PH∗(K), we write βk
for the Betti curve of PHk(K).
There are obviously non-isomorphic persistence modules that give rise to the same Betti
curve, so this descriptor discards information. Its advantage is that, as functions, Betti curves
are easy to compare, especially if considered as bounded functions R → R. This becomes partic-
ularly useful if the persistence diagrams contain a great number of points. See also Sections 2.3
and 2.4.
2.2 Some filtered simplicial complexes
Finite filtrations of simplicial complexes abound in TDA. One natural example is the Cech fil-
tration
of a finite point cloud P given a sequence of radii ε1 < ε2 < · · · < εN. We write C(P) for the
filtration.
C(P; ε1) ⊆ C(P; ε2) ⊆ · · · ⊆ C(P; εN)
7
The Cech filtration of a point cloud is of theoretical importance since Theorem 2 applies not
only to each of its individual complexes (as in equation (1)), but also functorially so [4] (respect-
ing maps). In many applications, including the one in this paper, the data we are given do
not consist of points in Euclidean space, but rather of relationships (such as distances) between
points in an unknown metric space. Many simplicial complexes can be built from such data, for
example on graphs [17].
Definition 8. Let G = (V, E) be a graph. The flag complex or clique complex of G is the largest
simplicial complex fl(G) whose 1-skeleton is E.
If a graph G = (V, E) has edge weights w : E → R, defining
Gε = (V, w−1((−∞, ε]))
gives a filtration of flag complexes
V ⊆ fl(Gε1 ) ⊆ fl(Gε2 ) ⊆ · · · ⊆ fl(GεN )
The terms "flag complex" and " Cech complex", and their corresponding notation, will also
for a sequence ε1 < ε2 < · · · < εN.
interchangeably refer to the filtered versions.
Flag complexes can be constructed, and be useful, for any underlying graph. However, they
play a special role when that graph is a metric space. Indeed, if P = {p1, . . . , pN} ⊆ Rn is a
point cloud, then the complete graph G on the vertices {1, . . . , N}, with edge weights
w(i, j) = (cid:107)pi − pj(cid:107)2
gives a (filtration of) flag complex(es) called the Vietoris -- Rips complex of P. We write VR(P) for
the entire filtration and VR(P; ε) for the complex at scale ε in the filtration.
We obviously have the inclusion C(P; ε) ⊆ VR(P; ε) at any scale ε. Conversely, if the points
p0, . . . , pk ∈ P are arranged at the vertices of the standard k-simplex in Euclidean space and
scaled so that they are ε apart pairwise, then [0, 1, . . . , k] ⊆ C(P;√2ε). The standard simplices
are in fact the worst cases [11], and so
C(P; ε) ⊆ VR(P; ε) ⊆ C(P;√2ε)
for any scale ε. This sandwiching of C(P) and VR(P) implies that homological features that per-
sist under the morphism Hk(VR(P; ε)) → Hk(VR(P;√2ε)) correspond to features of Hk( C(P); ε)),
and vice versa. The persistent homology of the Vietoris -- Rips complex thus contains important
information from that of the Cech complex, while using only distance data.
2.3 Comparing persistence modules
We now define a family of metrics on multisets in (R ∪ {±∞})2, which in turn pull back to (ex-
tended) (pseudo)metrics on persistence modules. Such metrics allow us to compare persistence
modules with one another in a quantitative way.
Definition 9. Let A and A(cid:48) be finite multisets in (R∪ {±∞})2, i.e. persistence diagrams, and let
∆ denote the diagonal with countably infinite multiplicity at every point. Write M for the set of
all multiset bijections A ∪ ∆ → A(cid:48) ∪ ∆. For q = 1, 2, . . . , ∞, let dE
q denote the extended Euclidean
q-metric on (R ∪ {∞})2. Then for p = 1, 2, . . . define the Vaseršteın distance
(cid:17)p
(cid:33)1/p
(cid:32)
dp,q(A, A(cid:48)) = inf
f∈M
dE
q (a, f (a)
(cid:16)
∑
a∈A
8
and for p = ∞ the bottleneck distance
d∞,q(A, A(cid:48)) = inf
f∈M
sup
a∈A
dE
q (a, f (a)).
The computation of these distances is equivalent to a maximum bipartite matching problem,
and so can be performed in O
gorithm [19, 24]. For the purpose of our paper we will only consider p = 1 and q = 2.
(cid:0)max(A,A(cid:48))3(cid:1) time using the (improved) Kuhn -- Munkres al-
We remark that the bottleneck distance d∞,q is of great theoretical importance because of its
role in stability theorems [5]. For example, if P(cid:48) is an ε-perturbation of a point cloud P in Euclidean
space, then d∞,∞(PH∗( C(P)), PH∗( C(P(cid:48)))) ≤ ε and d∞,∞(PH∗(VR(P)), PH∗(VR(P(cid:48)))) ≤ ε. In
particular, persistent homology is stable with respect to perturbation of the input data.
2.4 Statistics of persistence modules and Betti curves
Later in the paper, we will build filtered simplicial complexes from measurements, and we may
wonder how persistent homology depends on the statistical properties of the measurements.
The survey of Kahle [18] covers much of what is known about some special cases of random
simplicial complexes, namely those that are flag complexes of Erdos -- Rényi random graphs1 and
those that are Vietoris -- Rips complexes of random points from Euclidean space (the latter will
hereafter be referred to as random geometric complexes). While most results known are asymptotic
in the number of vertices in the complex, and thus of little direct relevance to our setting, some
qualitative conclusions can be drawn also about the finite case assuming the the size of the
vertex set is not too small.
The first qualitative observation is that non-bounding cycles, i.e. homology generators, are
likely to occur only in ER complexes when the edge probability parameter is in a certain range,
and this range becomes narrower as the homology dimension grows. Within the "allowed
range", however, a large Betti number may occur. As the homology dimension increases, the
allowed range shrinks while the peak Betti number within the range grows. Figure 3 illustrates
the behavior. The second observation is that in the random geometric setting, large Betti numbers
Figure 3: Persistent homology of a realization of an Erdos -- Rényi complex with 120 vertices. Increasing
homology dimension results in an increase in peak Betti number and a narrowing of the edge probability
range where non-vanishing Betti numbers are likely. This behavior is a signature of ER complexes.
1An (n, p)-Erdos -- Rényi random graph has n nodes, and each possible edge appears independently with probability
p. We will refer to flag complexes of such graphs as ER complexes with parameters n and p.
9
0.000.120.240.360.480.60Birth&filtrationscale0.000.120.240.360.480.60DeathPH1050100150200250300350Bettinumber0.00.10.20.30.40.50.6Filtrationscale05001000150020002500Bettinumberβ0β1β2β3 are much less likely. Intuitively, the triangle inequality puts heavy constraints on where points
may be placed in space for the cycles they are a part of not to become bounding at comparatively
small filtration scales. Highly persistent non-bounding cycles are thus geometrically fragile and
very sensitive to point coordinates, and thus unlikely to occur by chance. As homology dimen-
sion increases, the peak Betti number goes down, as one would expect from intuition, since
high-dimensional cycles are even more susceptible to being filled in by the triangle inequality.
Figure 4 illustrates the behavior.
Figure 4: Persistent homology of a random geometric complex, the Vietoris -- Rips complex of 120 points
sampled uniformly at random from the unit cube in R4.
The takeaway for us is that the Betti curves of ER complexes behave substantially different
from geometric complexes and, of course, from those of complexes arising from structured
data. Betti curves can therefore act as an indicator of the absence of meaningful information;
contrast Figure 3 and Figure 4. Experiments show that around 100 vertices seems to be more
than sufficient for this indicator to be robust.
2.4.1 Quantitatively testing for randomness
If G is a complete weighted graph on V, we write sh(G) for (a realization of) the complete graph
on V having the edge weights of G randomly shuffled. Except in degenerate cases fl(sh(G))
should be a good realization of an ER random complex. The discussion above then suggests
two ad hoc measures for how consistent a flag complex is with an ER random complex by
comparing PH∗(fl(G)) and PH∗(fl(sh(G))).
For various p and q we define
δk(G) = dp,q (PHk(fl(G)), PHk(fl(sh(G)))) ,
i.e. the persistence module metric between PH(fl(G)) and its shuffled version. We further let
βk and β(cid:48)k denote the Betti curve of PHk(fl(G)) and of PHk(fl(sh(G))), respectively, and define
the ratio
∆k(G) =
maxs βk(s)
maxs β(cid:48)k(s)
whenever it exists. ∆k(G) thus compares the peak Betti number of fl(G) to that of its shuffled
version. If, for as large a k as is computationally feasible, δk(G) is close to zero and ∆k(G) is
close to one, we have an indication that fl(G) resembles an ER random complex.
10
0.000.120.240.360.480.60Birth&filtrationscale0.000.120.240.360.480.60DeathPH1051015202530Bettinumber0.00.10.20.30.40.50.6Filtrationscale020406080100120Bettinumberβ0β1β2β3 2.5 Model for neuron activity
We need a model for neuron activity for two reasons. Most importantly, the process of infer-
ring away the contribution that specific covariates have on the neuron activity, as sketched in
Section 1 and detailed in Section 3, requires such a model. In addition, we prefer to work with
synthetic data so that we are in complete control of all "experimental" parameters when testing
the feasibility of our technique.
It has recently been shown [29, 28, 9] that statistical physics' kinetic Ising model, a simple
generalized linear model (GLM), does a good job modelling networks of neurons. In its original
language, the model governs the discrete time evolution of a set of spins under the influence of
both an external driving field and the neighboring spins (as defined either by a discretization of
Euclidean space in the original applications, or in general by a weighted graph). In general, the
model may be defined as follows.
Definition 10. A set of ±1-valued random variables Si(k), with 1 ≤ i ≤ N and k = 1, 2, . . . ,
are said to obey the (discrete time) kinetic Ising model with couplings J ∈ RN×N and external fields
E1, . . . , En : N → R if their conditional probabilities are
(cid:16)
(cid:17)(cid:17)
P (Si(k + 1) = si(k + 1) S1(k) = s1(k), . . . , SN(k) = sN(k))
(cid:16)
(cid:16)
=
exp
si(k + 1)
Ei(k) + ∑N
j=1 Ji,jsj(k)
2 cosh
Ei(k) + ∑N
j=1 Ji,jsj(k)
(cid:17)
.
For ease of notation, we set s1(0) = · · · = sN(0) = −1. We refer to the expression Fi(k) =
Ei(k) + ∑N
j=1 Ji,jsj(k) as the system's Hamiltonian (at time step k).
In the neuroscience interpretation of the model, the probability of neuron i firing or not at
time step k, as signified by Si(k) taking the value +1 or −1 respectively, is governed by the
external field Ei(k) and by whether or not the neighboring neurons {j Ji,j (cid:54)= 0} fired during
time step k − 1. We will in Section 3 make vital assumptions about the external fields E1, . . . , EN.
The function x (cid:55)→ exp(x)/ cosh(x) is a sigmoidal, and Figure 5 shows the firing probability
of a cell (i.e. the probability of s = 1 in Definition 10) as a function of the Hamiltonian.
Figure 5: Firing probability as a function of the Hamiltonian in the GLM.
Figure 6 shows the place field of a cell whose firing is generated by the GLM. In the ter-
minology of Definition 10, the external field is in this case predominantly a spatial Gaussian,
together with a smaller head direction tuning Gaussian. The cells are randomly, but sparsely,
coupled.
11
−2−1012Fi(k)0.00.20.40.60.81.0Firingprobability Figure 6: Example of the place field of one cell in a population of 100 whose activities are governed by the
GLM. Colors have qualitatively the same meaning as in Figure 1.
3 Uncovering hidden information
We suppose that a neuroscientist provides a list of candidate stimuli, such as for example spatial
preference, head direction preference, theta phase preference, and so forth. We further suppose
that the influence of such stimuli can be described by certain functions of distances on some
simple manifolds. We will specificically consider
• boxes (0, 1)d of any dimension d, with the Euclidean metric,
• boxes with any number number of d-disks removed so long as these do not disconnect the
manifold,
• circles, with their usual Riemannian metric,
• spheres, with their usual Riemannian metric,
• products of the above, with the product metric.
If the candidate list contains L stimuli with corresponding manifolds M1, . . . , ML from the list
above, then a complete description of the relevant state of the animal at any time is assumed to
be a point in the product M = M1 × · · · × ML. An experiment is then performed wherein the
neurons are recorded together with samples of the trajectory α : R → M of the animal through
its state space.
The neuronal activity recorded is temporally binned with width δt (in this paper δt ≈ 10
ms). Thus the primary input data to our method consist of a spike train for each neuron, i.e. N
binary vectors of the form
together with samples of the state space trajectory
si = (si(1), si(2), . . . , si(T)) ∈ {−1, 1}T,
α(1), α(2), . . . , α(T) ∈ M.
As a concrete example, we might imagine a hypothetical situation where the researcher
believes that only spatial position and head direction govern neuron activity. In this case, an ex-
periment might be performed wherein a rat explores a box, while its position and the direction
of its head are recorded at the same regular time instances as neuron activity. Supposing that 100
12
neurons are recorded over the course of 10 minutes2, the data we are given then consist of 100
binary spike train vectors of length 60000 together with 60000 samples of the rat state as points
in the state space (0, 1)2 × S1. Our goal, as outlined in Section 1.1, is then to determine whether
the two suspected stimuli -- spatial position and head direction -- describe the observed activ-
ity, and, most importantly, if not, what the homological properties of any remaining unknown
stimuli are.
3.1 Spike train order complex
Having obtained the necessary spike trains, we compute a measure of the degree of cofiring
between all pairs of neurons. The Pearson correlation of the corresponding spike trains is a
sensible choice, and for s ∈ RT and s(cid:48) ∈ RT(cid:48) we thus define
(cid:113)
(s(k) − s)(s(cid:48)(k) − s(cid:48))
corr(s, s(cid:48)) =
∑min(T,T(cid:48))
(cid:113)
(2)
,
k=1
∑min(T,T(cid:48))
k=1
(s(k) − s)2
∑min(T,T(cid:48))
k=1
(s(cid:48)(k) − s(cid:48))2
where
s =
1
min(T, T(cid:48))
min(T,T(cid:48))∑
k=1
s(k).
To reduce the Pearson correlation's sensitivity to errors in the binning process and slight
systematic timing errors in measurements, we will average over a small number of time bins.
For a vector s = (s(1), . . . , s(T)) ∈ RT, we will write its i'th left shift as
s[i] = (s(1 + i), . . . , s(T)) ∈ RT−i.
Then for τ ≥ 0, define
corrτ(s, s(cid:48)) =
1
τ + 1
max
(cid:32) τ∑
i=0
(cid:33)
corr(s[i], s(cid:48)),
τ∑
i=0
corr(s, s(cid:48)[i])
.
Finally, to follow the convention that simplicial complexes are filtered in order of increasing
distance, we define for a set of spike trains s1, . . . , sN the spike train distances
Dτ(si, sj) = max
k,l∈{1,...,N}
corrτ(sk, sl) − corrτ(si, sj).
We will throughout this paper write D = D0 and only briefly discuss specific choices of the
averaging time scale τ in Section 4.
While Dτ is not a metric when τ (cid:54)= 0 (clearly the triangle inequality need not hold), it does
provide some semblance of a closeness notion for spike trains, and should be a useful indi-
rect measure of how much place fields intersect. An obvious next step is therefore to take the
flag complex of the graph with vertices 1, . . . , N corresponding to the neurons, and whose edge
weights are w(i, j) = D(si, sj). However, as pointed out in [15], the chemical and biological pro-
cesses that go into a neuron building its action potential and firing, and the physical processes
that constitute the measurement of that event, are of a highly non-linear nature. What we ob-
serve, then, is merely a highly transformed version of the real cofiring relationship of neurons.
2This is a realistically sized data set according to computational neuroscience data sharing website http://crcns.
org.
13
In other words, if we think of C(i, j) as a true low-level measure of the cofiring relationship of
neurons i and j, then D(si, sj) = f (C(i, j)) for some unknown function f : R → R. On biological
grounds it is reasonable to assume that the only thing that is known, and can ever be known,
about f is that it is monotonic.
Following [15], we rescale our data so as not to prescribe meaning to the actual values of the
distance D, since it may well be the case that no such meaning exists.
Definition 11. Let G = (E, V) be the complete graph on the vertices V = {1, . . . , n} with edge
weights w. Let ϕ : {1, . . . , (n
(breaking ties arbitrarily), and let (cid:101)G be the complete graph on V with the edge weights
2)} → E be a bijection that sorts the weights
w(ϕ(1)) < w(ϕ(2)) < · · ·
(cid:101)w(i, j) =
ϕ−1(i, j)
.
The order complex of G is the flag complex OC(G) = fl((cid:101)G).
(n
2)
If we take the order complex of the spike trains and their distances, we discard all informa-
tion except precisely that which survives the unknown monotone transformation f , namely the
ordering of the edge weights.
3.2 Removing the contribution of stimuli to reveal hidden information
The likelihood of data
observed under the GLM is
s = {si(k) 1 ≤ i ≤ N, 1 ≤ k ≤ T}
(cid:16)
(cid:16)
(cid:16)
L(s) =
exp
N∏
i=1
T−1∏
k=1
si(k + 1)
Ei(k) + ∑N
j=1 Ji,jsj(k)
2 cosh
Ei(k) + ∑N
j=1 Ji,jsj(k)
(cid:17)(cid:17)
.
(cid:17)
We now assume that the external field part of the Hamiltonian decomposes into a sum of Gaus-
sians on the various factors of the state space M corresponding to candidate covariates.
As before, M = M1 × · · · × ML, where each Mi is assumed to be one of the manifolds listed
earlier (Section 5 discusses a generalization). Projections onto the factors are written pri : M →
Mi, and we denote by di the metric on Mi. We define the Gaussians Vl,q : Ml → R by
Vl,q(x) = exp
dl(x, cl,q)
2σ2
l,q
(cid:16)
−
(cid:16)
(cid:17)2
(cid:17)
for 1 ≤ l ≤ L and 1 ≤ q ≤ Ql, and assume that the external part of the Hamiltonian for neuron
i at time step k can be written (recall that α is the animal's path through its state space)
Ei(k) =
L∑
Ql∑
l=1
q=1
Ai,l,q
for cl,q ∈ Ml, Ai,l,q ∈ R and σl,q ∈ R.
14
Vl,q ◦ prl ◦α
(k)
We may use the language of Section 1 if for each i and l there is only one q for which Ai,l,q (cid:54)=
0. Fix an i and an l and let q be the only index for which Ai,l,q (cid:54)= 0. Then cl,q is the center of a place
field corresponding to covariate number l, σl,q is a measure of its width, while Ai,l,q specifies the
peak strength with which it influences the firing of neuron i. In other words, neuron i has a place
field specified by cl,q and σl,q. When the above relationship between the indices of A•,l,• does
not hold, we allow each covariate to govern place cell activity through a linear combination of
Gaussian fields, which will be necessary for the inference process described next.
The log-likelihood of the observed data is
log(L(s)) =
N∑
T−1∑
i=1
k=1
(cid:34)
(cid:16)
(cid:32) L∑
(cid:32) L∑
Ql∑
l=1
q=1
Ai,l,q
Ql∑
l=1
q=1
si(k + 1)
(cid:32)
− log
2 cosh
(cid:17)
Vl,q ◦ prl ◦α
(cid:16)
(k) +
(cid:17)
Ai,l,q
Vl,q ◦ prl ◦α
N∑
j=1
Ji,jsj(k)
(k) +
N∑
j=1
(cid:33)
(cid:33)(cid:33)(cid:35)
Ji,jsj(k)
,
(3)
and it is an easy calculus exercise to show that
(cid:40)
l=1 Ql → R
RN ∑L
(A•,•,•, J•,•) (cid:55)→ log(L(s))
is a convex function. We can therefore perform likelihood maximization by means of convex
optimization to infer the Ai,l,q's and Ji,j's that best fit the observed data. With these coefficients
in hand, we can selectively remove the (expected) contribution of each stimulus on the candi-
date list. The residual data after removal no longer consists of binary spike trains, but instead
real-valued time series. Low and high values correspond to improbable and probable firing
events, respectively, while values near zero reflect lack of knowledge of the probability of either
outcome.
Throughout this paper all the inference will be performed with 252 = 625 spatial basis
functions, and 25 circular basis functions, both with means uniformly distributed. For example,
if the state space is M = M1 × M2 with M1 = (0, 1)2 and M2 = S1, then Q1 = 252 and Q2 = 25,
and with the c1,q's forming a regular grid on (0, 1)2 and the c2,q's uniformly distributed on S1.
The assumptions on the external fields above thus allow us to remove the contributions of
specific stimuli towards the firing activity, suggesting the following algorithm:
1. Perform an experiment as described above, yielding spike trains s1, . . . , sN and samples
of the state space path α.
2. Compute the spike train distances D(si, sj) for 1 ≤ i, j ≤ N, and let them be the edge
weights on a complete graph G on N vertices.
3. Compute PH∗(OC(G)).
4. Is PH∗(OC(G)) consistent with the persistent homology of an ER random complex per
Section 2.4.1? If so, go to step 4a. If not, go to step 5.
(a) There is nothing more to learn from our method. We are done.
5. Is the list of candidate stimuli exhausted? If so, go to step 5a. If not, go to step 6.
(a) PH∗(OC(G)) may reveal information about the homology of the state space corre-
sponding to unknown covariate(s). We are done.
15
6. Using the inference process described above, remove the contributions of a selected stimu-
lus. This yields real-valued time-series that from now on replace the spike trains s1, . . . , sN.
Go to step 2.
Figure 7 summarizes the above procedure.
Figure 7: Our main contribution summarized.
As a concrete illustration, we return to the example from earlier: assume that place cell
activity is in fact governed only by spatial position and head direction, but suppose that head
direction tuning is unknown to the researchers. The candidate covariate list therefore consists
only of spatial position. For an experiment conducted with the animal exploring a box, the state
space is (0, 1)2. After having performed steps 2 through 6 once, the second iteration leads us to
step 5a; there are no more stimuli to account for, yet we see persistent homology inconsistent
with random data. Examination of the persistence diagrams of PH∗(OC(G)) reveals that the
unknown covariates correspond to a state space with the same homology as a circle. Together
with other evidence, this may lead the researchers to suspect head direction tuning as a hidden
covariate. New experiments may then be performed to investigate this, and our method may
be applied also to the new data.
We finally point out that it is not essential to our method that we necessarily capture the
correct homology of the full state space. It is the state space after the removal of known stimuli
that matters (and even then, the weaker information of whether the remaining state space is
homologically trivial or not may be useful).
4 Computational results
We test the efficacy of our method using synthetically generated data in order to be in complete
control of all "experimental parameters", and because publicly available real data often come
from experiments without a topological focus (for example the spatial environment tends to
be homologically trivial). For the results presented here, neuron activity was simulated from
the same GLM as used for the inference process by appropriate selection of peak field strength
coefficients A•,•,•, centers c•,• and widths σ•,•. In addition, a constant negative term (typically
16
ExperimentorsimulationSpiketrainsPersistenthomologyPersistencediagramsandBetticurvesRandom?NomoretolearnYesStimulilistexhausted?NoInformationabouttopologyofhiddenstimuliYesi)Pickstimulus;E.g.spatialposition,headdirection,thetaphasepreference,...ii)Inferitscontributiontospiketrainsiii)RemovecontributionNo −1) was added to the external fields to make the overall firing rate low outside of place fields,
as is the case for many real cells. This term essentially just lowers the noise floor of our data,
and should be of no deeper significance to us.
Even if one of the experiments below does not call for spatial or head direction tuning, the
state space will always consist of at least a factor for the physical environment and a factor S1
for the head direction. If the physical environment is denoted B, then the B × S1 factor of the
state space is explored as follows: If at some time step the animal's state is (x, y, θ) ∈ B × S1,
then the next state is found by choosing a θ(cid:48) randomly and uniformly within 0.02 of θ in S1.
If (x + 5 · 10−4 cos θ(cid:48), y + 5 · 10−4 sin θ(cid:48), θ(cid:48)) is within B × S1, this point is the next state. If not,
new angles are drawn until the new state is valid. We note that qualitatively similar results
are obtained if this slightly realistic random walk is replaced by ordinary Brownian motion or
uniform random sampling.
Except in Section 4.6, we do not average the correlation measure, i.e. we work with the spike
train distance D0.
Some additional computational experiments have been relegated to Appendix A.
4.1 Recovering spatial homology
While the central point of this paper is the uncovering of the homological properties of un-
known stimuli underlying neural activity, we begin by providing an example of how persistent
homology of the order complex of spike train distances can recover the homology of a spatial
environment. This is analogous to some of the results presented in [8].
State space is now the unit square punctured by four disks of radius 0.15. The disks are
centered at (0.27, 0.27), (0.27, 0.72), (0.72, 0.27) and (0.72, 0.72). We refer to the punctured box
as B below. For technical reasons, data were generated with the state space having an extra
circle factor, but every cell had its activity coefficients corresponding to this factor set to zero.
Thus only spatial tuning is present in this example.
In the notation of Section 3.2, N = Q1 = Q2 = 100, M1 = B, M2 = S1, L = 2, and Ji,j = 0 for
all 1 ≤ i, j ≤ N. The c1,q's form a regular grid on B, and
if i = q
otherwise,
Ai,1,q =
2
0
(cid:40)
while Ai,2,q = 0 for all i, q. Figure 8a shows the spatial tuning of a single neuron.
Figure 8b shows that we correctly recover the homology of B, the only part of M detected by
the neural activity. In other words, this example illustrates how we can detect and even count
the obstructions in a space from neuron activity alone.
4.2 Proof of concept
The simplest possible setting wherein our inference scheme, laid out in Section 3.2, is useful, is
perhaps one where both spatial and head direction tuning govern neuron activity, but where
the researcher believes only one of those to be real.
For this computation, the spatial component of the state space is a unit square punctured in
its center by a single disk of radius 0.2, denoted by B. The head direction component is S1.
In the notation of Section 3.2, N = Q1 = Q2 = 100, M1 = B, M2 = S1, L = 2, and Ji,j = 0 for
17
(a)
(b)
Figure 8: Results of the experiment in Section 4.1. a: The activity of a single cell as a function of spatial
position. b: Persistent homology of the order complex. All higher dimensions of homology are trivial or
close to trivial.
all 1 ≤ i, j ≤ N. The c1,q's form a regular grid on B, and
if i = q
otherwise,
Ai,1,q =
2
0
(cid:40)
Similarly, the c2,q's are uniformly spread out over S1. To avoid artificially coupling head direc-
tion and spatial tuning through the ordering of their place field centers, we let σ be a random
permutation of {1, . . . , N} and then let
(cid:40)
Ai,2,q =
2
0
if i = σ(q)
otherwise,
for all i, q.
We reiterate that we assume that the researcher is unaware of head direction tuning as a real
influence on place cell activity in this example; he believes spatial position is the only relevant
stimulus. After conducting an experiment, the researcher sees the neurons' spatial dependence
examplified in Figure 9a. The head direction dependence in Figure 9b is not known to the
researcher.
Persistent homology of the order complex of the correlations observed can be seen in Fig-
ure 10. Note that we do not observe the actual homology of the state space M, which is a
thickened torus. This is not entirely satisfactory, but at least the observed persistence diagrams
indicate there is homologically nontrivial information present in the neuron activity.
Satisfied that the persistence diagrams are consistent with his hypothesis about the relevant
covariates, the researcher proceeds to the next step, namely removing the effect of the spatial
covariate (the only one he is aware of). We maximize the expression in equation (3) in terms of
the Ai,l,1's using L-BFGS-B3, and subtract from the observed spike trains the expected activities
predicted from only spatial influence.
The result of the removal on spatial activity tuning can be seen in Figure 11. It is as expected,
and obviously does not provide new information to the researcher. The persistence diagram in
3As provided by SciPy.
18
0.00.20.40.60.81.0x0.00.20.40.60.81.0y−1.0−0.8−0.6−0.4−0.20.00.20.40.60.81.00.000.020.040.060.080.10Birth0.000.020.040.060.080.10DeathPH1 (a)
(b)
Figure 9: Activity of a single neuron in the experiment from Section 4.2. a: Spatial dependence. This is the
only view of activity visible to the researcher. b: Head direction dependence. The researcher is not privy
to this information, as he is unaware that head direction preference is real.
Figure 10: Observed persistent homology of the order complex of the correlations for the experiment from
Section 4.2.
(a)
(b)
Figure 11: Average activity of a single neuron in the experiment from Section 4.2 after removal of spatial
tuning. a: Spatial dependence. This is the only view of activity visible to the researcher. b: Head direc-
tion dependence. The researcher is not privy to this information, as he is unaware that head direction
preference is real.
19
0.00.20.40.60.81.0x0.00.20.40.60.81.0y−1.0−0.8−0.6−0.4−0.20.00.20.40.60.81.00123456θ−1.0−0.50.00.51.0Averageneuronstate0.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH20.00.20.40.60.81.0x0.00.20.40.60.81.0y−1.0−0.8−0.6−0.4−0.20.00.20.40.60.81.00123456θ−1.0−0.50.00.51.0Averageneuronstate Figure 12: Observed persistent homology of the order complex of the correlations for the experiment from
Section 4.2 after removal of spatial tuning.
Figure 12, however, shows that there is still homologically non-trivial information contained in
the observed data. This should hopefully lead the researcher to suspect that there are further,
hidden, influences on neuron activity, and, most importantly, that this/these are of a circular
nature.
Guided by this, the researcher might consider head direction tuning. He therefore sets up a
new experiment where also this is tracked, so that also its influence may be removed from the
data. Doing so results in the activity plot in Figure 13.
(a)
(b)
Figure 13: Average activity of a single neuron in the experiment from Section 4.2 after removal of both spatial
and head direction tuning. a: Spatial dependence. b: Head direction dependence.
The persistence diagrams in Figure 14 now lack clear suggestions of non-trivial homology.
Moreover, the Betti curves are highly indicative of an ER random complex (compare Figure 3).
To verify this last claim, we also permuted the neuron correlations randomly; Figure 15 shows
the result. This should serve as a firm indication that all stimuli underlying the neuron activity
in the data have been accounted for.
20
0.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH20.00.20.40.60.81.0x0.00.20.40.60.81.0y−1.0−0.8−0.6−0.4−0.20.00.20.40.60.81.00123456θ−1.0−0.50.00.51.0Averageneuronstate Figure 14: Observed persistent homology of the order complex of the correlations for the experiment from
Section 4.2 after removal of both spatial and head direction tuning.
Figure 15: After all relevant covariates have been removed from the data in the experiment from Sec-
tion 4.2, the Betti curves of the order complex of the correlations are consistent with those of an ER random
complex (here formed by randomly shuffling the correlations). C.f. Figure 3.
21
0.000.080.160.240.320.40Birth&filtrationscale0.000.080.160.240.320.40DeathPH1050100150200250Bettinumber0.000.080.160.240.320.40Birth&filtrationscale0.000.080.160.240.320.40DeathPH20100200300400500600700Bettinumber0.000.050.100.150.200.250.30Filtrationscale0100200300400500600700Bettinumberβ0β0(shuffled)β1β1(shuffled)β2β2(shuffled) 4.2.1 Alternative scenario
One may also want to consider an alternative hypothetical scenario wherein head direction
tuning is the only suspected stimulus. For readability reasons we do not include that scenario
in full here. The interesting part is the persistent homology after the removal of head direction
tuning. Figure 16 shows that we recover the correct PH1 also in this case.
Figure 16: Persistent homology of the order complex after removing the effect if head direction tuning in
an alternative version of the experiment in Section 4.2 (see Section 4.2.1).
4.3 Effect of couplings
In the preceding experiments, cells were never coupled. To illustrate that our method also copes
with such "internal stimuli", we repeated the experiment from Section 4.2 with the change that
every cell is given a weak but random coupling to a every other cell. Specifically, we kept all
simulation parameters as before, but let every Ji,j be drawn independently and uniformly from
[−0.1, 0.1]. The couplings are thus weak compared to the external stimuli (which peak at 2 in
the centers of fields), but should nevertheless introduce noise to the data.
Figure 17 shows the results before any covariate removal. Observe that the random cou-
plings introduce significant noise in the spatial dependence of the activity compared to that in
Figure 9. Figure 18 shows that we are able to carry out the same procedure as in Section 4.2 also
in the presence of couplings.
4.4 Effect of theta wave phase preference
We simplistically model theta wave phase preference as each neuron preferentially firing near
a randomly chosen phase of a 7 Hz sinusoidal wave in time. The experimental parameters are
the same as in Section 4.2, except now L = 3, and the state space gains an extra factor M3 = S1.
The c3,q's are uniformly spread out over S1, and for a random permutation4 σ of {1, . . . , N} the
field strength coefficients are
(cid:40)
Ai,3,q =
2
0
if i = σ(q)
otherwise,
4Present for the same reason as for the head direction tuning in Section 4.2.
22
0.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH1 (a)
(b)
Figure 17: Prior to removal of any covariates in the experiment from Section 4.3. a: Spatial dependence
for the activity of a single neuron. b: PH1 persistence diagram.
(a)
(b)
Figure 18: Removing external and internal covariates in the experiment from Section 4.3. a: Persistence
diagram after spatial dependence and internal couplings are removed. b: Persistence diagram and Betti
curve after all external (space and head) and internal (couplings) influences are removed.
23
0.00.20.40.60.81.0x0.00.20.40.60.81.0y−1.0−0.8−0.6−0.4−0.20.00.20.40.60.81.00.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth&filtrationscale0.000.080.160.240.320.40DeathPH1050100150200250Bettinumber for all i, q.
Theta phase preference is thus, as far as topology is concerned, precisely the same as head
direction preference, except that the M3 factor of state space is explored by always moving for-
ward in time (modulo 1/14) instead of by a random walk. We therefore expect that theta phase
preference will contribute to homology in the same way as head direction tuning. Figure 19
confirms this. Again it should be pointed out that we are not observing homology consistent
with the three dimensional torus S1 × S1 × S1. While this may seem unsatisfactory, it is a quite
natural effect of one covariate suppressing the expression of the homology of the others, i.e.
one of the radii of the torus dominating over the others. This illustrates well why the inference
process and sequential removing of covariates really is necessary; the fact that there are three
circular factors in the state space cannot be glanced directly from the observed data.
(a)
(b)
Figure 19: Observations from the experiment in Section 4.4 before removal of any covariates. a: Activity
of a single neuron as a function of theta wave phase. b: Persistent homology of the order complex of the
spike train distances.
Figure 20 shows that we obtain the expected results when left with only theta phase prefer-
ence and when all covariates are removed.
(a)
(b)
Figure 20: Persistent homology of the order complex of the residual spike trains in the experiment in
Section 4.4 after removing some external covariates. a: After removing spatial and head direction tuning.
b: After removing all covariates, including theta phase preference.
24
0.000.010.020.030.040.050.060.07tmod1/14−1.0−0.50.00.51.0Averageneuronstate0.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth&filtrationscale0.000.080.160.240.320.40DeathPH1050100150200250Bettinumber 4.5 Sensitivity to field strength variations
The preceding experiments illustrate the efficacy of our technique for fixed choices of peak field
strengths. We now wish to demonstrate that our method is robust over a wide range of field
strengths. To this end, we repeat the experiment from Section 4.2, but now for varying choices
of peak strength for both the spatial and head direction fields.
For conducting this experiment, we need a numerical measure of the outcome that is more
succinct than a persistence diagram or a Betti cruve. Since we generate data with head direc-
tion and spatial tuning enabled, we hope to see the same outcome as in the experiment from
Section 4.2:
• Before removal of any covariates, we should see a prominent PH1 generator, i.e. one that
stands out in the persistence diagram by having a longer lifetime than most. At the same
time, the persistence diagrams and their associated Betti curves should be inconsistent
with those of ER random complexes.
• This should remain the case when inferring away one of either head direction or spatial
tuning.
• After inferring away both head direction and spatial tuning, the persistence diagrams and
Betti curves should be highly consistent with those of ER random graphs.
As a very crude measure of the presence of "highly persistent" or "prominent" PH1 genera-
tors, we simply consider the ratio of the of the lifetime of the most persistent one to the lifetime
of the second-most persistent one. We refer to the ratio as ρ1 throughout this section. Note that
our aim is that ρ1 measures how easy a single prominent generator is to distinguish from noise.
In the case of multiple highly persistent PH1 generators, such as for the persistent homology
of a torus, it reports a false negative. However, since the experiment in Section 4.2 failed to
capture this true homology of the state space, we feel confident this measure is sufficient. We
also verified this by manual inspection of several of the persistence diagrams computed.
To compare with random ER complexes, we use both δ1 and ∆1. Recall that the closer δ1(G)
is to zero and ∆1(G) is to one, the more consistent the flag complex built on the data in G is with
an ER complex.
Figure 21 summarizes the relevant measures of success. Observe that we meet our criteria
for expected outcome for parameters that lie outside the blue band in Figure 21a and that at
the same time are not too weak in either the spatial or the head direction tuning strength (see
e.g. Figure 21g). The latter requirement is no surprise, as we with weak field strengths observe
neurons that barely respond to external stimuli. Further investigation reveals that as the head
direction field strength is made weaker (i.e. as we approach the band of failure from above in
Figure 21a), the most persistent PH1 lives for a shorter and shorter time, until it becomes indis-
tinguishable from the noise near the diagonal. For even weaker head direction field strengths,
i.e. below the band, we are essentially in the domain of activity governed entirely by the spatial
fields.
Note also that when we in Figure 21a are within the blue band, Figures 21d and 21g show
that the observed data are not consistent with an ER complex. Inferring away the spatial in-
fluence is therefore something one may still wish to do, whereupon one uncovers homology
strongly consistent with a circle everywhere except for with weak head direction tuning, as
seen in Figure 21b.
25
(a)
(d)
(g)
(b)
(e)
(h)
(c)
(f)
(i)
Figure 21: Summary of the behavior of our method over varying field strengths, as detailed in Section 4.5.
In all the plots, the horizontal axis denotes the peak spatial field strength and the vertical axis denotes the
peak head direction field strength. Arrows in the color bar indicate that the colors are clipped above or
below a value. For a spike train distance graph G obtained using the given peak field strengths, the first
row shows ρ1(G), the second row shows ∆1(G), and the third row shows δ1(G). In the first column G
comes from the original observations, in the second column the spatial influence is removed, and in the
third column both spatial and head direction influence are removed.
26
0.00.51.01.52.02.53.00.00.51.01.52.02.53.01.02.54.00.00.51.01.52.02.53.00.00.51.01.52.02.53.01.02.54.00.00.51.01.52.02.53.00.00.51.01.52.02.53.01.02.54.00.00.51.01.52.02.53.00.00.51.01.52.02.53.00.00.51.00.00.51.01.52.02.53.00.00.51.01.52.02.53.00.00.51.00.00.51.01.52.02.53.00.00.51.01.52.02.53.00.00.51.00.00.51.01.52.02.53.00.00.51.01.52.02.53.005100.00.51.01.52.02.53.00.00.51.01.52.02.53.005100.00.51.01.52.02.53.00.00.51.01.52.02.53.00510 4.6 Homology from internal couplings
In the preceding experiments, internal couplings have been absent, or, as in Section 4.3, not
themselves been the focus of our attention. We now illustrate that our method is also capable
of detecting the homology of (the flag complex of) the graph defining the neighbor relations of
the neurons.
We generate data with only spatial fields and internal couplings. Specifically, N = Q1 = 100,
M = M1 = (0, 1)2, the c(cid:48)1,qs form a regular grid on (0, 1)2, and the peak field strengths are
(cid:40)
Ai,1,q =
if i = q
1
0 othewise
(for technical reasons, just as in Section 4.2, the explored state space still contains an extra S1
factor, but the corresponding field strengths are set to zero). The symmetric matrix J describes
a circle on all N nodes with edges in both directions with weights 2. The indices defining the
edges of the circle are chosen randomly to avoid an unnatural coupling to the spatial fields
through the ordering.
(a)
(b)
Figure 22: Persistent homology of the order complex of the observed spike train distances Dτ in the
experiment from Section 4.6. a: τ = 1. b: τ = 10.
Figure 22a shows an unexpected result consistent with (at least) two circles. Manual in-
spection of the spike train distances reveals that this is an artifact of the GLM being coupled
across only two consecutive time steps. When we use a spike train distance without tempo-
ral average (i.e. D0), we are unable to resolve any coupling interactions across an even num-
ber of neurons. The 1-skeleton of the order complex of the spike train distances thus breaks
into two circles, corresponding to the even and odd parity edges of the neighborhood graph
(cid:54)= 0}). This undesired behavior vanishes when the correla-
({1, . . . , N},{(i, j) Ji,j
tion time average is greater, for example τ = 10, as is shown in Figure 22b.
(cid:54)= 0 or Jj,i
5 Discussion and sketches of a general framework
We believe that the core aspects of the method presented in this paper are applicable outside
of neuroscience. In a general setting, we imagine a point cloud C = {p1, . . . , pN} of points on
a manifold M. The exact assumptions on M have not been worked out, but we believe that a
27
0.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH10.000.080.160.240.320.40Birth0.000.080.160.240.320.40DeathPH1 compact, connected, homogenous Riemannian manifold without boundary suffices.5 Write the
metric of M as d. The points themselves, and their pairwise distances, are a priori unknown.
Estimates of the latter are obtained by performing a random walk on M while observing N
Bernoulli processes S1, . . . , SN (producing time series s1, . . . , sN). If we denote by X(k) the ran-
dom variable of the random walk's position at time step k, the vital assumption is that the
parameter (the "success" probability) of Bernoulli process Si at time step k is
(4)
for some monotonically decreasing sufficiently integrable (unknown) function f : R+ → [0, 1].
If the random walk has progressed long enough that the distribution of X(k) is close to uniform,
then we are in a setting where we believe our methods are applicable.
P (Si(k) = 1 X(k) = x) = f (d(x, pi))
The question to ask is whether the persistent homology of the flag complex (or order com-
plex) of the graph with edge weights Dτ(si, sj) for 1 ≤ i, j ≤ N closely approximates the persis-
tent homology of VR(C).
In the above setup, define gk : M × M → R by
(cid:90)
gk(y, z) =
f (d(y, x)) f (d(z, x)) dρk(x)
M
with ρk the probability density function for the random walk at time step k. The function gk
arises naturally as the probability
P(Si(k) = 1 ∩ Sj(k) = 1) = gk(pi, pj)
and is the essential part of the estimated Pearson correlation corr(si, sj) from equation (2). In the
case that X(k) is fully uniform, i.e. dρk = dx, proving correctness of our recovered persistent
homology amounts to showing the existence of a monotonically decreasing h : R+ → R that
makes
commute. In the Euclidean situation (M = Rn and dropping the assumptions on M) this is an
easy calculus exercise, but the details need to be worked out for the more general situations.
The actual persistent homology of (say, the Vietoris -- Rips complex of) C, i.e. the level of ques-
tion Q0, may be of interest in applications outside of neuroscience. While there are perhaps not
many situations in which observations of the Bernoulli processes are available, but the values
on the right hand side of equation (4) are not, one example might be sensor network coverage
problems [14] with sensors of severely reduced capabilities. Instead of being able to sense the
strength of their neighbors, we imagine that the sensors carry simple devices which trigger with
a frequency monotonically related to the distances to other sensors.
Work should be done to give rigorous bounds on the persistence modules based on the
statistical properties of the Pearson correlations corr(si, sj), especially in the settings when the
random walk distribution is not yet truly uniform.
At the level of questions Q1 and Q2 we encounter other systems described by a GLM (such
as in [3]) or, in principle, by other models that accommodate an inference process like that in
Section 3.2. Our method may be able to shed light on the homological properties of parts of the
5Some of the state spaces considered in Section 4 are of course not covered by these assumptions. Numerical evi-
dence still strongly suggests that our method works well in practice also in the situations considered.
28
M×MRgR+dh external influences on such a system, or, as in the experiment in Section 4.6, unknown internal
couplings.
Future applications of our method on actual data should demonstrate its usefulness in iden-
tifying hidden topological information in data, both inside and outside of neuroscience.
6 Acknowledgments
This paper is a product of a cooperative project where also Yasser Roudi (Kavli Institute for
Systems Neuroscience, Norwegian University of Science and Technology) takes part.
GS would like to thank Geir-Arne Fuglstad for valuable discussions on matters of probabil-
ity.
All persistent homology computations were performed using Nanda's Perseus software [25].
A Additional computational results
Some additional computational results have been relegated here for the sake of readability.
A.1 Comparison with multi-dimensional scaling
One may well ask whether the use of persistent homology actually contributes anything useful
to our proposed method. To investigate this, we considered whether multi-dimensional scaling
(MDS) could successfully, i.e. near-isometrically, embed the "distances" from the various stages
of the experiment in Section 4.2 in low-dimensional Euclidean space.
Figure 23 shows that MDS provides the same useful information as persistent homology
before removal of covariates and after removal of only spatial tuning. After removing head
direction tuning, however, MDS is unable to detect the annular nature of the remaining spatial
covariate (compare in particular Figure 23c and Figure 16). This illustrates the essential role
played by persistent homology in our work.
We point out that embedding in R3 does not seem to qualitatively improve the situation.
29
(a)
(c)
(b)
(d)
Figure 23: MDS embedding of the neuron distances from the experiment in Section 4.2. a: Original data.
b: After removal of spatial tuning. c: After removal of head direction tuning. d: After removal of both
covariates.
30
−0.2−0.10.00.10.2−0.2−0.10.00.10.2−0.2−0.10.00.10.2−0.2−0.10.00.10.2−0.2−0.10.00.10.2−0.2−0.10.00.10.2−0.2−0.10.00.10.2−0.2−0.10.00.10.2 References
[1] Mamiko Arai, Vicky Brandt, and Yuri Dabaghian. "The Effects of Theta Precession on
Spatial Learning and Simplicial Complex Dynamics in a Topological Model of the Hip-
pocampal Spatial Map". In: PLoS computational biology 10.6 (2014).
[2] Gorô Azumaya. "Corrections and supplementaries to my paper concerning Krull -- Remak --
Schmidt's theorem". In: Nagoya Mathematical Journal 1 (1950), pp. 117 -- 124.
[3] Stanislav Borysov, Yasser Roudi, and Alexander Balatsky. "US stock market interaction
network as learned by the Boltzmann Machine". In: ArXiv e-prints (2015). arXiv:1504 .
02280 [q-fin.ST].
[4] Frédéric Chazal and Steve Yann Oudot. "Towards persistence-based reconstruction in Eu-
clidean spaces". In: Proceedings of the twenty-fourth annual symposium on Computational ge-
ometry. ACM. 2008, pp. 232 -- 241.
[5] David Cohen-Steiner, Herbert Edelsbrunner, and John Harer. "Stability of persistence di-
agrams". In: Discrete & Computational Geometry 37.1 (2007), pp. 103 -- 120.
[6] William Crawley-Boevey. "Decomposition of pointwise finite-dimensional persistence
modules". In: Journal of Algebra and Its Applications 14.5 (2015).
[7] Carina Curto and Vladimir Itskov. "Cell groups reveal structure of stimulus space". In:
PLoS Computational Biology 4.10 (2008). DOI: 10.1371/journal.pcbi.1000205.
[8] Yuri Dabaghian, Facundo Mémoli, L. Frank, and Gunnar Carlsson. "A Topological Paradigm
for Hippocampal Spatial Map Formation Using Persistent Homology". In: PLoS computa-
tional biology 8.8 (2012).
[9] Benjamin Dunn, Maria Mørreaunet, and Yasser Roudi. "Correlations and functional con-
nections in a population of grid cells". In: PLoS computational biology 11.2 (2015).
[10] Herbert Edelsbrunner. A Short Course in Computational Geometry and Topology. Springer,
2014.
[11] Herbert Edelsbrunner and John Harer. Computational Topology: An Introduction. American
Mathematical Society, 2010. ISBN: 978-0-8218-4925-5.
[12] Robert Ghrist. "Barcodes: the persistent topology of data". In: Bulletin of the American
Mathematical Society 45.1 (2008), pp. 61 -- 75.
[13] Robert Ghrist. Elementary Applied Topology. 2014.
[14] Robert Ghrist and Abubakr Muhammad. "Coverage and hole-detection in sensor net-
works via homology". In: Proceedings of the 4th international symposium on Information pro-
cessing in sensor networks. IEEE Press. 2005.
[15] Chad Giusti, Eva Pastalkova, Carina Curto, and Vladimir Itskov. "Clique topology re-
veals intrinsic geometric structure in neural correlations". In: Proceedings of the National
Academy of Sciences of the United States of America (2015). DOI: 10.1073/pnas.1506407112.
[16] Allen Hatcher. Algebraic Topology. Cambridge University Press, 2002. ISBN: 0521-79540-0.
Jakob Jonsson. Simplicial Complexes of Graphs. Vol. 1928. Lecture Notes in Mathematics.
[17]
Springer, 2008.
[18] Matthew Kahle. "Topology of random simplicial complexes: a survey". In: Contemporary
Mathematics 620 (2014), pp. 201 -- 222.
[19] Harold Kuhn. "The Hungarian method for the assignment problem". In: Naval research
logistics quarterly 2.1-2 (1955), pp. 83 -- 97.
31
[20] Peter McCullagh and John A. Nelder. Generalized linear models. Vol. 37. Monographs on
Statistics and Applied Probability. CRC press, 1989.
[21] Kenji Mizuseki, Anton Sirota, Eva Pastalkova, and György Buzsáki. "Theta oscillations
provide temporal windows for local circuit computation in the entorhinal-hippocampal
loop". In: Neuron 64.2 (2009), pp. 267 -- 280.
[22] Kenji Mizuseki, Anton Sirota, Eva Pastalkova, Kamran Diba, and György Buzsáki. Multi-
ple single unit recordings from different rat hippocampal and entorhinal regions while the animals
were performing multiple behavioral tasks. CRCNS.org. DOI: 10.6080/K09G5JRZ.
[23] Robert Muller. "A quarter of a century of place cells". In: Neuron 17.5 (1996), pp. 813 -- 822.
DOI: 10.1016/S0896-6273(00)80214-7.
James Munkres. "Algorithms for the assignment and transportation problems". In: Journal
of the Society for Industrial & Applied Mathematics 5.1 (1957), pp. 32 -- 38.
[24]
[25] Vidit Nanda. Perseus. URL: http://www.sas.upenn.edu/~vnanda/perseus.
[26]
John O'Keefe and Jonathan Dostrovsky. "The hippocampus as a spatial map. Preliminary
evidence from unit activity in the freely-moving rat". In: Brain research 34.1 (1971), pp. 171 --
175.
Jonathan W. Pillow et al. "Spatio-temporal correlations and visual signalling in a complete
neuronal population". In: Nature 454.7207 (2008), pp. 995 -- 999.
[27]
[28] Yasser Roudi, Benjamin Dunn, and John Hertz. "Multi-neuronal activity and functional
connectivity in cell assemblies". In: Current opinion in neurobiology 32 (2015), pp. 38 -- 44.
[29] Yasser Roudi and John Hertz. "Mean field theory for nonequilibrium network reconstruc-
tion". In: Physical review letters 106.4 (2011).
[30] Alon Rubin, Michael Yartsev, and Nachum Ulanovsky. "Encoding of head direction by
hippocampal place cells in bats". In: The Journal of Neuroscience 34.3 (2014), pp. 1067 -- 1080.
[31] Etienne Save, Ludek Nerad, and Bruno Poucet. "Contribution of multiple sensory infor-
mation to place field stability in hippocampal place cells". In: Hippocampus 10.1 (2000),
pp. 64 -- 76.
[32] Edward C. Tolman. "Cognitive maps in rats and men". In: Psychological review 55.4 (1948),
pp. 189 -- 208.
[33] Wilson Truccolo, Uri T. Eden, Matthew R. Fellows, John P. Donoghue, and Emery N.
Brown. "A point process framework for relating neural spiking activity to spiking his-
tory, neural ensemble, and extrinsic covariate effects". In: Journal of neurophysiology 93.2
(2005), pp. 1074 -- 1089.
[34] Eugene Wigner. "On the distribution of the roots of certain symmetric matrices". In: An-
nals of Mathematics (1958), pp. 325 -- 327.
32
|
1701.01219 | 1 | 1701 | 2017-01-05T06:07:48 | Is neuroscience facing up to statistical power? | [
"q-bio.NC",
"stat.AP"
] | It has been demonstrated that the statistical power of many neuroscience studies is very low, so that the results are unlikely to be robustly reproducible. How are neuroscientists and the journals in which they publish responding to this problem? Here I review the sample size justifications provided for all 15 papers published in one recent issue of the leading journal Nature Neuroscience. Of these, only one claimed it was adequately powered. The others mostly appealed to the sample sizes used in earlier studies, despite a lack of evidence that these earlier studies were adequately powered. Thus, concerns regarding statistical power in neuroscience have mostly not yet been addressed. | q-bio.NC | q-bio | Is neuroscience facing up to statistical power?
Geoffrey J. Goodhill
Queensland Brain Institute and School of Mathematics and Physics,
The University of Queensland, St Lucia, QLD 4072, Australia.
Email: [email protected]
January 6, 2017
Abstract
7
1
0
2
n
a
J
5
]
.
C
N
o
i
b
-
q
[
1
v
9
1
2
1
0
.
1
0
7
1
:
v
i
X
r
a
It has been demonstrated that the statistical power of many neuroscience studies is very low, so that
the results are unlikely to be robustly reproducible. How are neuroscientists and the journals in which
they publish responding to this problem? Here I review the sample size justifications provided for all
15 papers published in one recent issue of the leading journal Nature Neuroscience. Of these, only
one claimed it was adequately powered. The others mostly appealed to the sample sizes used in
earlier studies, despite a lack of evidence that these earlier studies were adequately powered. Thus,
concerns regarding statistical power in neuroscience have mostly not yet been addressed.
Introduction
It is well-documented that the biomedical sciences are beset by bad statistical practices, and that this
is one of the reasons for the current 'reproducibility crisis' [1, 2]. Prominent amongst these problems
are n values that provide only low statistical power. Genuine effects that do actually exist are missed,
and many effects that are found to be significant are likely to be just random chance. Neuroscience
is no exception to this rule [3].
Indeed, due to the particular challenges of this field, studies which
can be completed within traditional parameters of time, cost and ethical approval are often restricted
to low n values. Furthermore this data is then analysed post-hoc from many different perspectives in
the hope of finding significant results, further increasing the probability of false positives.
The purpose of the present article is not to review again these problems, which are well documented.
Rather, I consider how the community of authors, reviewers and journal editors in neuroscience is
responding to this clearly visible challenge. The leading neuroscience journal Nature Neuroscience
provides a good opportunity to do this, since (unlike most neuroscience journals) it requires authors
to provide answers to some basic statistical questions about the design of their experiments.
It
is therefore of interest to see what answers have been forthcoming in recently published papers.
Presumably, because the papers were published, the authors, reviewers and editors all thought these
answers were acceptable.
Here I reproduce the statements regarding sample size from all 15 papers published in the August
2016 issue, and find that all of them except one essentially confess they are probably statistically
underpowered. I do not explicitly identify which papers these came from, because my goal is not to
cast doubt on any specific work: this is simply a (somewhat) random subset to illustrate a very broad
issue. Furthermore, there is no reason to think these problems are any different in other journals
(though a recent study has argued that statistical power is negatively correlated with journal impact
factor [4]). What makes Nature Neuroscience attractive for analysis in this regard, besides its current
ranking as the highest-impact primary research journal in the field, is that it takes the trouble to require
authors to make explicit comments about certain statistical matters.
1
Statements of powerlessness
These are ordered thematically, and do not reflect the ordering within the issue.
1. The sample size for each experiment was determined based on published studies using similar
experimental designs together with pilot experiments from our laboratory. This allowed us to deter-
mine the sample size required for each experiment to ensure a statistical power of 0.8 and an alpha
level of 0.05.
Here the authors clearly address the issue of statistical power. Although potentially one might want
to see the evidence for the claim, this statement provides reassurance that these results are likely to
be reproducible.
2. Sample size was predetermined on the basis of published studies, experimental pilots and in-
house expertise.
It is encouraging that pilot studies were undertaken, but it is unclear how these pilots or the in-house
expertise were used to determine statistical power.
I now group several statements together, since they are all very similar.
3. Sample sizes for each condition of this study are similar to those generally employed in the
field. . . and were not predetermined by a sample size calculation.
4. No statistical methods were used to predetermine sample sizes, but our sample sizes are similar
to those generally employed in the field.
5. Sample size choice was based on previous studies, not predetermined by a statistical method.
6. No statistical tests were used to predetermine sample sizes, but our sample sizes are similar to
those in previous studies
7. No statistical methods were used to predetermine sample sizes, but our sample sizes are similar
to those generally employed in the field.
8. Group sample sizes were chosen on the basis of previous studies.
9. No statistical methods were used to predetermine sample sizes, but our sample sizes are similar
to those previously reported.
10. No statistical methods were used to pre-determine sample sizes but our sample sizes are larger
to those reported in previous publications.
The obvious problem with all these statements is that they do not address whether any of these pre-
vious studies demonstrated they were adequately powered (that previous work produced significant
results says nothing about statistical power, a basic point that appears not to be widely appreciated).
In addition, unless exactly the same experiments were performed, the variability and effects sizes are
likely to be different, meaning that the sample size required to achieve adequate power will also be
different.
11. No statistical methods were used to predetermine sample sizes, but the tissues were randomly
chosen in each age group. . . and uniformly processed. Also, our samples sizes are similar to those
of the discovery set of a similar experimental design in a previous publication
Besides the problems mentioned above, this statement conflates statistical power with other issues
of experimental design.
12. No statistical methods were used to predetermine sample sizes. Sample size was decided on the
basis of our previous experience in the field and was not pre-determined by a sample size calculation.
The sample size are similar to those generally employed in the field and is justified by the high rate
of exclusion due to the difficulty of the combined methodological approaches
Here the authors appeal to practical limitations on sample sizes. These limitations are real and
worthy of acknowledgement. However this does not provide information pertinent to the statistical
power, and thus reproducibility, of the results.
2
13. No estimates of statistical power were performed before experiments; animal numbers were min-
imized to conform to ethical guidelines while accurately measuring parameters of animal physiology.
Here the authors appeal to ethical limitations on sample sizes. Again, while real and worthy of
acknowledgement, the same arguments apply as mentioned above.
Finally, we come to perhaps the two most worrying statements.
14. No statistical tests were used to predetermine sample sizes, but our sample sizes are similar
to those generally employed in the field. Normal distribution of data was assumed, but not formally
tested.
15. Normality of the data distributions was assumed, but not formally tested.
The last makes no statement about sample sizes at all, despite this supposedly being a requirement
of the journal. More importantly, both statements explicitly state that the authors do not know whether
the statistical tests they applied were actually appropriate.
For comparison the statements in the July 2016 issue were very similar. One article justified sample
sizes in terms of a power calculation, while the remainder bar one (which simply stated that 'no
statistical methods were used to predetermine our sample sizes') appealed to similarity with sample
sizes used in previous studies, in one case in a different species.
Discussion
All of the statements reviewed above were approved by the authors, reviewers, and journal editors,
and one must therefore conclude that they reflect currently accepted practice in the field. It is widely
known and understood that statistical power is a key issue affecting reproducibility, yet 14/15 of these
statements (93%) do not address statistical power. Most of them appeal to precedent for sample
sizes, despite the facts that most neuroscience studies are underpowered [3], and that new exper-
iments will most likely have different variances and effect sizes from previous work.
It is clearly a
step forward that Nature Neuroscience requires authors to explicitly answer some key questions re-
garding statistical analysis. However, that the journal is willing to accept answers which are clearly
inadequate, and even sometimes admissions that the statistical analysis performed was quite possi-
bly wrong, suggests that the journal is still contributing to the problem rather than the solution. For
comparison the relatively new journal eNeuro requires authors to provide the statistical power of each
experiment reported. However this is merely the observed power, which provides little or no addi-
tional information beyond the observed p value [5], and thus this policy does not help matters much
either.
I am not attempting to single out these 14 papers as being of any more concern than any other work in
the field, rather they simply provide a revealing window on community standards at the highest level.
Neuroscientists seem willing to accept that work in the field generally uses low n values. Sometimes
this might be reasonable: for instance the effect size of the difference in phenotype between a wild-
type and knockout may be very large (though even in cases such as these power calculations are
rarely provided). Certainly many important findings have been robustly reproduced, even though
the statistical power of the original (or indeed subsequent) results was not established (e.g.
[6]).
However, in many experiments the effects are subtle, and low n values and thus power mean that, on
average, reproducibility will be low.
This is a very difficult problem (which I hasten to add I am also struggling with in my own research).
Neuroscience experiments are often intrinsically long-term and low-throughput. For instance uncov-
ering the function of a disease-related gene in a mouse model, or studying the neural correlates of
consciousness in an awake behaving primate model, can require large resources and many years of
work to obtain a single main result. Many of the latest techniques are extremely technically challeng-
ing, and therefore (as alluded to in one of the statements above) a large proportion of experiments
3
fail. This can lead to a big mismatch between the n values at the start and end of the experiment.
Funding is tight, and increasing n by even one animal for a particular experiment can have costly
implications in time and money. Increasing ethical pressures on the use of animals in research (as
alluded to above) add additional constraints on the n values that are practically achievable. How-
ever there is clearly also a cultural component to sample sizes in neuroscience: for instance work
in organisms such as C. elegans and Drosophila, to which some of the above constraints are less
applicable, also do not usually consider statistical power.
What can be done? Clearly better education for neuroscientists (at all levels) regarding statistical
issues is important to address the lack of statistical scepticism that apparently plagues the field, and
books such as [7] should be more widely read and understood (for instance many neuroscientists
still appear to think that obtaining p < 0.05 means it is 95% likely they have discovered something
true, no matter how small the n value [8]). From the perspective of a field such as statistical genetics,
one solution seems obvious: neuroscientists should collaborate in larger teams and share data, so
that many small and weakly powered results can be replaced with a few strongly powered results.
However, while data sharing in general should and is being broadly encouraged, there are problems
with this general model for the community of neuroscientists at large (for an interesting discussion
see [9]). How would everyone agree what were the right experiments to do? How would cutting-
edge and often highly non-trivial methodologies be standardized between labs? How would this
model not be detrimental to the entreprenurial spirit that has fuelled so many important discoveries
in neuroscience? The risk is that progress in neuroscience, where publication rates are already low
compared to some other fields of biomedical science (for the reasons mentioned above), could be
reduced to an unviable level.
Another approach is to replace the ubiquitous current statistical paradigm, of null hypothesis sig-
nificance testing, with estimation [10]. This can be done by confidence intervals and/or Bayesian
approaches. Now, instead of there being a 'bright line of truth' at an arbitrarily chosen probability
threshold relating to the null hypothesis (which, after all, is not what one is actually interested in), the
focus is on determining degrees of confidence in quantities such as the effect size and the difference
between means. Binary statements about 'significance' versus 'nonsignificance' are replaced with
graded confidence variations which are easier to interpret intuitively. However, this has yet to catch
on in neuroscience.
Staying within the confines of null hypothesis significance testing, I have argued that neuroscientists
seem willing so far to accept the current situation regarding (lack of) statistical power. Perhaps that
is indeed the best that can presently be achieved, given the current statistical paradigm and practical
constraints in the field. Perhaps we should just accept that most studies will be likely underpow-
ered and reproducibility will likely be a recurring issue, at least until some more mature stage of
development is reached. However if this is the case, it would be helpful if authors, reviewers and
journal editors more clearly acknowledged that underpowered studies lead to weak and potentially
irreproducible results.
4
References
[1] Ioannidis JP (2005). Why most published research findings are false. PLoS Med, 8:e124.
[2] Colquhoun D (2014). An investigation of the false discovery rate and the misinterpretation of
p-values. R Soc Open Sci, 1:140216.
[3] Button KS, Ioannidis JP, Mokrysz C, Nosek BA, Flint J, Robinson ES, Munafo MR (2013). Power
failure: why small sample size undermines the reliability of neuroscience. Nat Rev Neurosci.,
14, 365-376.
[4] Szucs D,
Ioannidis JPA (2016). Empirical assessment of published effect sizes
literature.
neuroscience
psychology
and
the
power
and
http://biorxiv.org/content/early/2016/08/25/071530
cognitive
recent
in
[5] Lenth, RV (2001). Some practical guidelines for effective sample size determination. Am Stat,
55, 187-193.
[6] Bliss TVP & Lomo T (1973). Long-lasting potentiation of synaptic transmission in the dentate
area of the anaesthetized rabbit following stimulation of the perforant path. J. Physiol., 232,
331-356.
[7] Reinhardt, A. (2015). Statistics Done Wrong: The Woefully Complete Guide. No Starch Press.
[8] Halsey LG, Curran-Everett D, Vowler SL, Drummond GB (2015). The fickle P value generates
irreproducible results. Nat Methods, 12, 179-85.
[9] Mainen ZF, Hausser M & Pouget A (2016). A better way to crack the brain. Nature, 539, 159-
161.
[10] Cumming, G. & Calin-Jageman, R. (2017). Introduction to the new statistics: estimation, open
science, and beyond. Routledge.
5
|
1511.04338 | 2 | 1511 | 2018-03-14T12:44:16 | The sensorimotor loop as a dynamical system: How regular motion primitives may emerge from self-organized limit cycles | [
"q-bio.NC",
"cond-mat.dis-nn",
"cs.RO"
] | We investigate the sensorimotor loop of simple robots simulated within the LPZRobots environment from the point of view of dynamical systems theory. For a robot with a cylindrical shaped body and an actuator controlled by a single proprioceptual neuron we find various types of periodic motions in terms of stable limit cycles. These are self-organized in the sense, that the dynamics of the actuator kicks in only, for a certain range of parameters, when the barrel is already rolling, stopping otherwise. The stability of the resulting rolling motions terminates generally, as a function of the control parameters, at points where fold bifurcations of limit cycles occur. We find that several branches of motion types exist for the same parameters, in terms of the relative frequencies of the barrel and of the actuator, having each their respective basins of attractions in terms of initial conditions. For low drivings stable limit cycles describing periodic and drifting back-and-forth motions are found additionally. These modes allow to generate symmetry breaking explorative behavior purely by the timing of an otherwise neutral signal with respect to the cyclic back-and-forth motion of the robot. | q-bio.NC | q-bio | Frontiers in Robotics and AI
Research Article
7 November 2018
The sensorimotor loop as a dynamical
system: How regular motion primitives may
emerge from self-organized limit cycles
Bulcs ´u S´andor 1,∗, Tim Jahn 1, Laura Martin 1 and Claudius Gros 1
1Institute for Theoretical Physics, Goethe University Frankfurt, Frankfurt am Main,
Germany
Correspondence*:
Bulcs´u S´andor
Institute for Theoretical Physics, Goethe University Frankfurt, Max-von-Laue Str. 1,
Frankfurt am Main, 60438, Germany, [email protected]
Theory and Applications of Guided Self-Organization in Real and Synthetic
Dynamical Systems
ABSTRACT
We investigate the sensorimotor loop of simple robots simulated within the LPZRobots
environment from the point of view of dynamical systems theory. For a robot with a cylindrical
shaped body and an actuator controlled by a single proprioceptual neuron we find various types
of periodic motions in terms of stable limit cycles. These are self-organized in the sense, that
the dynamics of the actuator kicks in only, for a certain range of parameters, when the barrel
is already rolling, stopping otherwise. The stability of the resulting rolling motions terminates
generally, as a function of the control parameters, at points where fold bifurcations of limit cycles
occur. We find that several branches of motion types exist for the same parameters, in terms of
the relative frequencies of the barrel and of the actuator, having each their respective basins of
attractions in terms of initial conditions. For low drivings stable limit cycles describing periodic
and drifting back-and-forth motions are found additionally. These modes allow to generate
symmetry breaking explorative behavior purely by the timing of an otherwise neutral signal with
respect to the cyclic back-and-forth motion of the robot.
Keywords: Sensorimotor loop, Adaptive behavior, Self-organization, Limit cycles, Period tripling, Embodiment, Explorative behavior,
Symmetry breaking
1 INTRODUCTION
Robots moving through an environment need to take the physical laws into account. This can be achieved
either via classical control theory (de Wit et al., 2012), or by considering the full sensorimotor loop as
an overarching dynamical system (Ay et al., 2012). This distinction could be cast, alternatively, into
open-loop control, e.g. via central pattern generators (Ijspeert, 2008), and closed-loop schemes using
feedback to control the states of an internal dynamical system (Dorf and Bishop, 1998). The presence of
such feedback mechanisms capable of amplifying local instabilities are key components leading to the
emergence of self-organization (Der and Martius, 2012). A closely related notion is that of embodiment
(Ziemke, 2003), for which no need arises for an explicit modeling of the interactions between the robot and
its surroundings. The agent situated in a given environment can be treated, in an embodied approach, as an
overarching dynamical system, incorporating both the external dynamics (body-environment interaction),
1
8
1
0
2
r
a
M
4
1
]
.
C
N
o
i
b
-
q
[
2
v
8
3
3
4
0
.
1
1
5
1
:
v
i
X
r
a
S´andor et al.
The sensorimotor loop as a dynamical system
as well as the internal (controller-body) processes. Thus, combining the closed-loop control with the
embodied approach leads to movements generated through self-organizing processes. These, may in
turn be guided by generic, e.g. information theoretical objective functions (Martius et al., 2013), such
as predictive information (Ay et al., 2008), resulting in explorative or even playful behavior (Der and
Martius, 2012).
Similar objective functions, such as the free energy (Friston, 2010), can also be considered for the brain
as a whole (Baddeley et al., 2008) and in the context of adaptive behavior (Friston and Ao, 2011). Distinct
control mechanisms for neural networks can also be derived from other information theoretical generating
functionals, such as the relative information entropy (Triesch, 2007), the mutual information (Toyoizumi
et al., 2005), the Fisher information (Echeveste and Gros, 2014), and the recently introduced active
information storage measure (Lizier et al., 2012; Dasgupta et al., 2013). Starting from first principles
Hebbian learning rules have also been derived (Echeveste et al., 2015).
A parallel approach for studying the power of embodiment is provided by evolutionary robotics. Robots,
selected through evolutionary processes (Nolfi and Floreano, 2000) take environmental feedback naturally
into account, as they would otherwise not be positively selected. The notion of an acting agent in a
reacting environment becomes blurry, to a certain extent, when the full sensorimotor loop is considered,
with the motion coming to a standstill without a fully functional feedback cycle. Within other approaches
to embodiment, the physical constraints acting on compliant real-world robots are studied (Pfeifer et al.,
2007), or the flow of information, e.g. in terms of transfer entropy, through the sensorimotor loop (Schmidt
et al., 2013). A related question is how to ground actions generically, i.e. without ´a priori knowledge, in
sensorimotor perceptions (Olsson et al., 2006), or how to select actions from universal and agent-centric
measures of control (Klyubin et al., 2005).
Abstracting from the sensorimotor loop, one may regard, from the point of view of dynamical system
theory (Beer, 2000), motions as organized sequences of movement primitives in terms of attractor
dynamics (Schaal et al., 2000), which the agent needs first to acquire by learning attractor landscapes
(Ijspeert et al., 2002, 2013). These may be used later on for encoding the transients leading to periodic
motions (Ernesti et al., 2012), or may furthermore self-organize into complex behaviors (Tani and Ito,
2003). In this context the fully embodied approach may serve as an algorithmic first step to generate a
palette of motion primitives. One may also observe that all regular motions are, per definition, attractors in
terms of stable limit cycles in the overarching sensorimotor loop, which may be controlled either actively
(Laszlo et al., 1996), or passively in terms of limit-cycle walking (Hobbelen, 2008). As an alternative
approach for creating and controlling limit cycles one could use prototype dynamical systems, a concept
recently proposed for the study of complex bifurcation scenarios (S´andor and Gros, 2015).
In the present study we examine in detail the notion of periodic movements as stable limit cycles, using
the LPZRobots package (Der and Martius, 2012; Martius et al., 2013) for simulating robots (current
development version), which are geometrically simple enough to allow for an at least partial modeling
in terms of dynamical system theory (Gros, 2015). Our robots, see Figs. 1 and 2, are controlled by a
single proprioceptual neuron with a time dependent threshold b = b(t). We find a region of parameters in
which the motion is fully embodied, and where the movement vb = vb(t) of the robot and the threshold
dynamics are mutually fully interdependent, vanishing when one of them, either b(t) or vb(t), is clamped.
In engineering terms the engine db/dt powering the motion of the robot is turned on dynamically through
the feedback of its very motion.
We also find that a set of qualitatively distinct movements can arise for identical settings of the
parameters in terms of stable limit cycles, having their own distinct basins of attraction in phase space.
Control signals may hence switch between different motion primitives without the need to interfere with
the parameter setting of the sensorimotor loop. Most modes found lead to regular motions with finite
average velocities. We discovered however also a particular mode corresponding to a cyclic back-and-
forth movement, without an average translational motion of the robot. When the parameter settings are
changed in this mode, the robot will enter a rolling motion, either to the left or to the right, depending on
This is a provisional file, not the final typeset article
2
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 1. Screenshots from the LPZRobots simulation package, of the one- and two rod barrel robots
used (left- and right panel respectively).
the timing of the signal with respect to the phase of the cycle, allowing, as a matter of principle, for a truly
explorative behavior.
A central result of the present study is that even very simple controller dynamics (a single differential
equation, in our case) may lead via the sensorimotor loop to surprisingly rich repertoires of regular motion
primitives, which may be selected in turn through higher-order decision processes. This is due to the self-
stabilization of motion patterns within the sensorimotor loop. Goal oriented behavior would in this context
be achieved not by optimizing motion directly, but by selecting from the many attracting states generated
by an embodied controller within the overall sensorimotor loop.
2 METHODS
We start by describing the one-neuron controller used together with the actuator in terms of a damped
spring, and the actual setup of the robot.
2.1 RATE ENCODING NEURONS WITH INTERNAL ADAPTION
In this paper we consider actuators controlled by simple rate encoding neurons, characterized by a
sigmoidal transfer function
y(x, b) =
1
1 + ea(b−x) ,
b = εa(2y − 1)
(1)
between the membrane potential x and the firing rate y, where a is the gain, taken to be fixed, and b = b(t)
a time-dependent threshold. The dynamics b for the threshold in (1) would lead to b → x and y → 1/2
for any constant input x(t) = x, with a relaxation time being inversely proportional to the adaption rate ε.
This adaption rate can also be motivated by information-theoretical considerations for the distribution of
the firing rates (Triesch, 2005; Markovi´c and Gros, 2010).
2.2 DAMPED SPRING ACTUATORS
Our robots are controlled by actuators regulating the motion of the ball of mass m on a rod, as illustrated
in Fig. 2, from its actual position x on the rod, to its target position
xt = 2R
Frontiers in Robotics and AI
,
(2)
3
(cid:18)
y(x, b) − 1
2
(cid:19)
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 2. Left: Illustration of the proprioceptual single-neuron controlled damped spring actuator. The
input x of the neuron (described by Eq. (1)) is given by the actual position x ∈ [−R, R] of the ball of mass
m moving on the rod, while the output y being proportional, via Eq. (2), to the target position xt of the
ball. The PID controller then simulates the dynamics of a damped spring, with constant k and damping γ,
between the current and the target positions of the mass.
Right: Sketch of the one-rod robot composed of a barrel of mass M and radius R, with a mass m moving
along a rod, as illustrated in the left panel. Slipping is not allowed, the robot moves hence with a velocity
vb = Rω = R ϕ, where ϕ measures the angle of the rod with respect to the horizontal.
where R is the radius of the barrel containing the rod and where y(x, b) is the sigmoidal (1). We note
that the input and the output of the neuron are, via (2), of the same dimensionality, namely positions. The
force F = mx moving the ball is evaluated by the PID controller
F = gP (xt − x) + gI
(xt − x)dt + gD
d(xt − x)
dt
,
(3)
(cid:90) t
0
provided by the LPZRobots simulation environment (Der and Martius, 2012), characterized by the
standard PID-control parameters gP , gI and gD.
For our simulations we considered the case gI = 0, for which the PID controller reduces to a damped
spring, see Fig. 2,
mx = −k(x − xt) − γ
d(x − xt)
dt
,
(4)
with k = gP and γ = gD.
• Eq. (4) represents only the contribution of the actuator to the force moving the ball along the rod. The
gravitational pull acting on the mass m, and the centrifugal force resulting from the rolling motion of
the barrel on the ground are to be added to the RHS of Eq. (4).
• The target position xt = xt(t) is time-dependent through (2) and (1).
• Eq. (4) is strictly dissipative, due to the damping γ > 0. The same holds for the rolling motion of
the barrel on the ground, which is also characterized by a finite rolling friction. Thus, the dynamics
db/dt of the threshold in (1) can be considered as an engine, providing, by adjusting continuously the
target position xt of the ball, and hence the length of the spring, the energy dissipated by the physical
motions.
This is a provisional file, not the final typeset article
4
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 3. The phase diagram of the one-rod barrel, as a function of the gain a and of the adaption rate
ε. The results are obtained using the LPZRobots package, apart from the red solid line separating the off-
and the on mode, which follows from (5), for a fixed but otherwise arbitrary angle ϕ = ϕ0. The dashed
vertical and horizontal gray lines indicate the cuts used for the phase diagrams presented in Fig. 5. The
number of stable limit cycles found in the respective parameter regions are denoted by nr.
Non-rolling modes: The red dots/line indicate the locus of a Hopf bifurcation, where a stable non-rolling
limit cycle (on mode) emerges from the trivial non-rolling fixpoint (off mode). In the off mode the 'engine'
db(t)/dt, see Eq. (1), kicks in only when the barrel is already moving.
Rolling modes: Shown are the regions containing nr = 1 (enclosed by the solid gray line) and nr = 2
(enclosed by the solid black line) attracting limit-cycles corresponding to a barrel moving with a finite
velocity (cid:104)vb(cid:105). Note, that the robot is able to move also in the off mode (of the engine). The stationary and
the drifting back-and-forth modes, discussed in Fig. 6, have been omitted, in order to avoid overcrowding.
2.3 MOTION OF A MASS ON A FIXED ROD
As an example we consider a robot, for which we keep the angle ϕ between the rod and the horizontal
fixed, ϕ = ϕ0, by preventing it from rolling. We are then left with a self-coupled motion of a ball along
a rod, as illustrated in the left panel of Fig. 2, resulting in a dynamics similar to the one of a self coupled
neuron (Markovi´c and Gros, 2012; Gros et al., 2014). Using Ω2 = k/m and Γ = γ/m we find in this case
x = v
v = −Ω2(x − xt) − Γ(v − xt) − g sin(ϕ0)
xt = 2 R a y(1 − y)(v − b)
b = 2 ε a(y − 1/2)
,
(5)
when combining Eqs. (1), (2) and (4). The gravitational term −g sin(ϕ0) can be transformed away via
x → x − g/Ω2 sin ϕo ,
b → b − g/Ω2 sin ϕo ,
(6)
and does hence not influence the phase diagram, which is shown in Fig. 3 for Ω2 = 200, Γ = 2Ω and
g = 9.81. We have used standard numerical methods (Clewley, 2012).
We find a Hopf bifurcation line separating the stability regions for the trivial fixpoint and for a limit
cycle, denoted respectively as off and on modes. This behavior is similar to the one observed for a self
coupled neuron with intrinsic adaption (Markovi´c and Gros, 2012; Gros et al., 2014).
Frontiers in Robotics and AI
5
1.21.41.61.82.0a0.00.51.0εoffmodeonmodenr=1nr=2 S´andor et al.
The sensorimotor loop as a dynamical system
Figure 4. The motion x(t) of the mass along the rod of the one-rod barrel (top row), together with
the corresponding phase-plane trajectories (x(t), vb(t)) (bottom row), compare Fig. 2. The gain and
the adaption rate are a = 1.9 and ε = 0.25 respectively. Shown are the 0:1, 1:1 and 1:3 modes
(left/middle/right column). Note, that the velocity vb(t) of the barrel vanishes for the 0:1 mode, oscillating
but remaining otherwise positive for the 1:1 and the 1:3 mode. (click for movie).
3 RESULTS
In Fig. 1 the screenshots of the one- and two-rod robots simulated with the LPZRobots package (current
development version) (Der and Martius, 2012; Martius et al., 2013) are presented. Throughout the
simulations the control parameters Γ = 2Ω and Ω2 = 200 for the actuator, Λ = 1 for the mass ratio
m/M (ball to barrel), R = 1 for the radius for the barrel, and Ψ = 0.3 for the coefficient of the rolling
friction have been held constant, varying only the adaption rate for the threshold of the neuron, and the
gain a. For the simulations a stepsize of 0.001 was used. In the Figures (and in the rest of the paper) the
parameters will be presented in dimensionless units, with SI units being implied: seconds/meter for the
time and length respectively and g = 9.81 m/s2 for the gravitational acceleration. Our barrel has a radius
of 1 m and a moving mass of 1 kg, rolling typically at speeds of (1 − 4) m/s ≈ (3 − 12) km/h. A table of
the parameters is given in the Supplementary Material.
3.1 ONE-ROD BARREL
The overall phase diagram of the one-rod barrel shown in Fig. 3 contains regions of non-rolling fixpoints or
limit cycles, and regions where one or more attracting limit cycles corresponding to a continuously rolling
barrel are present, in part additionally. Depending on the initial conditions the system will eventually settle
into one of the attracting states.
3.1.1 Coexisting modes as behavioral primitives. Standard robot control aims at achieving a predefined
outcome and for this purpose it is indispensable,
that identical robot actions lead also to identical
movements. This is not necessarily the case for robots controlled by self-organized processes, as
investigated here.
This is a provisional file, not the final typeset article
6
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 5. The average speed (cid:104)vb(cid:105) of the one-rod barrel for the 1:1 (green/orange dots) and for 1:3 (blue
crosses) mode. The vertical dashed line denotes the locus of the Hopf bifurcation line shown in Fig. 3.
In the off mode (on mode) the attracting state for the non-rolling mode is a stable fixpoint (limit cycle)
respectively. Presumably existing unstable limit cycles are indicated by dashed lines (labeled with question
marks).
Left: For a gain a = 1.9. The colored region for very small adaption rates ε indicates a region with both
stable and drifting back-and-forth modes, further described in Fig. 6.
Right: For an adaption rate ε = 0.25.
In Fig. 4 we illustrate the time series and the corresponding phase-space plots of the dominant modes
of the one-rod barrel shown in Fig. 1. The simulation parameters a = 1.9 for the gain, and the ε = 0.25
adaption rate are close to the Hopf bifurcation line shown in Fig. 3, but in the on mode. Which means,
that the ball moves both for fixed horizontal or vertical rods.
The first of the three coexisting stable limit cycles, illustrated in Fig. 4, corresponds to the non-moving
barrel with the ball oscillating vertically along the rod (first column). For the second, 1:1 mode, the average
rolling frequency of the barrel and of the oscillation of the ball along the rod match (second column). For
the 1:3 mode the corresponding ratio of frequencies is however 1:3 (third column).
The occurrence of several distinct limit cycles for identical parameters can be interpreted in terms of
behavioral primitives, potentially allowing an agent to switch rapidly between different types of motions,
by shortly destabilizing the currently active limit cycle.
Note that the self-coupled neuron, controlling the dynamics of the ball along the horizontally fixed rod,
has only two possible stable states (a fixpoint and a limit cycle). Considering however the fully embodied
rolling robot, coexisting states are arising, which can lead to different behavioral patterns purely as a result
of the environmental context. An external force applied to the robot can qualitatively change its behavior,
indicating the sign of multifunctionality (Williams and Beer, 2013).
3.1.2 Embodiment as self-organized motion. Most robots are autonomously active in the sense, that
the motion is not essentially dependent on the feedback of the environment. For the case of self-organized
motion, as considered here, there would be, on the other side, no motion when the sensorimotor loop
would be interrupted.
We present in the left plot of Fig. 5 the evolution of the self-sustained rolling modes, in terms of the
indicates, as a guide to the eye, that the velocity increases roughly ∝ √
averaged measured velocity, for a = 1.9 and as a function of adaption rate ε. The dashed black line
ε for the 1:1 mode. The two
branches are stable for ε ∈ [0.018, 0.55] and ε ∈ [0.19, 0.61] respectively for the 1:1 and the 1:3 mode, and
Frontiers in Robotics and AI
7
onmodeoff modemodeoffon S´andor et al.
The sensorimotor loop as a dynamical system
Figure 6. The time evolution of the position x (colored lines) of the mass along the rod, and of the
(rescaled) speed vb of the barrel (black lines), for a = 1.9 and ε = 0.019/0.017/0.015 (left/center/right),
all in the off mode (compare Fig. 5). The respective average velocities are (cid:104)vb(cid:105) = 0.63/0.00/0.25 for the
1:1 mode (left), the stationary back-and-forth mode (middle) and the drifting back-and-forth mode (right).
(click for movie).
terminate (presumably) through saddle node bifurcations of limit cycles. We have indicated this scenario
by adding by hand in Fig. 5, as guides to the eye, the respective unstable branches.
The locus of the Hopf bifurcation shown in Fig. 3, at ≈ 0.05, is indicated in (the left panel of) Fig. 5
by the dashed vertical line, separating the off from the on mode. In the off and on modes the non-rolling
attractors are a fixpoint and a limit cycle respectively. Note that self-sustained rolling modes exist in
the off mode as well, where the 'engine' db(t)/dt of the barrel only kicks in, through amplifying local
fluctuations (damped oscillations around the fixpoint), when the barrel is already moving. This underlines
the embodied nature of the motion, which arises in a truly self-organized fashion (in term of dynamical
systems theory (Gros, 2015)) through the bidirectional feedback between environment and both the body
and the controller of the robot.
However, in the absence of feedback mechanisms (such as centrifugal- and Coriolis-forces), the neuron
controlled actuator could only generate a single regular rolling motion, similar to the ones achieved by
sending motor signals generated by some central pattern generators (Der and Martius, 2012). This is not
the case for our robot, which exhibits, as shown in Fig. 5 (and in Fig. 6, see discussion below) a wide
spectrum of possible rolling modes.
3.1.3 Avoided pitchfork bifurcations of limit cycles
In the right panel of Fig. 5 we present the measured
mean velocity (cid:104)vb(cid:105) of the ball for ε = 0.25, as a function of the gain a. The Hopf bifurcation between the
off- and on- non-rolling modes occurs at aH ≈ 1.83, compare Fig. 3.
For 1.23 < a < 1.83 the ball hence is moving in the off mode, with the engine kicking in only through
the feedback from the environment, which we interpret as self-organized embodied motion, with the
environment being an essential component of the overarching dynamical system.
Comparing both panels of Fig. 5 one can notice that the low-velocity mode (green dots) connects either
to the the 1:1 mode (as in the left panel) or to the 1:3 mode (as in the right panel). The reason for the
apparent discrepancy lies in the fact, that the respective bifurcation line is oblique in the phase space
plane (a, ε). The evolution of these modes suggests in any case, that the low-velocity mode connects to
the two higher-velocity modes via an avoided pitchfork transition of limit cycles (Gros, 2015).
3.1.4 Explorative motion via noise induced directional switching. Our robot contains a single
dynamical variable, the threshold b(t), generating self-stabilizing motions via the sensorimotor loop. The
This is a provisional file, not the final typeset article
8
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 7. Two superimposed runs for the time evolution of the speed vb of the barrel (black lines), for
a = 1.9. In the first run the adaption rate ε is changed discontinuously at time t1 = 45 from ε = 0.017
(corresponding to the stationary back-and-forth mode, see Fig. 6) to ε = 0.02 (corresponding to the 1:1
rolling mode). In the second run identical initial conditions have been used and an identical change is
made to the adaption rated ε, but now at time t2 = 53. In both runs (dashed and dotted lines respectively),
the barrel settles into the 1:1 rolling motion, albeit in opposite directions (to the right/left with (cid:104)vb(cid:105) > 0
and (cid:104)vb(cid:105) < 0 respectively). (click for movie).
palette of modes generated is, despite this apparent simplicity, surprisingly large and may be used to
generate higher order behavior.
There are three dominant branches, the 0:1, 1:1 and 1:3 modes (in terms of the ratios of the respective
barrel- and mass frequencies), compare Figs. 4 and 5, which are stable for a wide range of parameters. We
found in addition also a parameter region for which different types of motions arise from minute changes
of control parameters, such as the adaption rate ε.
In Fig. 6 the motion x(t) of the ball along the rod and the velocity vb(t) of the barrel are given for three
closely spaced adaption rates ε = 0.019, 0.017 and 0.015, for which three qualitatively different types of
motions are found (which have partially, but not completely overlapping stability regions).
• For ε = 0.019 the standard 1:1 rolling motion is recovered, with an average velocity (cid:104)vb(cid:105) = 0.63.
• For ε = 0.017 a new mode arises, for which the ball rolls back and forth forever. The motion is
exactly symmetric with respect to the left and to the right, and the average velocity (cid:104)vb(cid:105) = 0.0 of the
barrel hence vanishes exactly.
• For ε = 0.015 the ball also rolls back and forth, but asymmetrically, giving rise to a drifting motion
with small but finite average velocity of (cid:104)vb(cid:105) = 0.25.
The occurrence of a limit cycle corresponding to a symmetric back-and-forth rolling motion, sandwiched
between symmetry breaking modes, gives rise to an interesting venue for the generation of explorative
behaviors, as the robot will be sensitive to finite but otherwise very small perturbations influencing
its internal control parameters. This behavior is illustrated in Fig. 7. Depending on the timing of the
perturbation with respect to the back-and-forth rolling cycle, the robot will settle into a left- or into a right-
moving motion (in the 1:1 or in the back-and-forth drifting mode respectively for increasing/decreasing
ε). It is hence possible to break spatial symmetries, in general, purely via the timing of a perturbation.
The perturbation itself, here acting on the adaption rate ε, does not need to carry any information about
the direction of motion.
Frontiers in Robotics and AI
9
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 8. Left: The average speed (cid:104)vb(cid:105) of the two-rod barrel for the 1:1/1:3/1:5 (orange dots, blue crosses,
dark-cyan stars) modes. The gain is a = 1.9, all other parameters are identical to the ones used for the one-
rod barrel. The respective time series and phase-space plots are presented in Fig. 9. The filled symbols
denote examples of additional higher order modes, of which the 1:21,1:13 and 1:9 (pink star, maroon
pentagon, green rhombus) are illustrated in Fig. 10. Right: A blow up, showing the relative location of the
1:8,1:7,1:6 and 1:11 modes found at ε=1.009,1.122,1.263 and 1.370 respectively.
3.2 TWO-ROD BARREL
Adding a second actuator perpendicular to the first one, a neuron controlled ball moving along a rod, one
can increase the complexity of the robot (see the right picture of Fig. 1). Both actuators work, in our
setup, independently, with the crosstalk being provided exclusively by the environmental feedback. Both
actuators are identical to the rod used for the single-rod barrel, with each rod having its own adapting
threshold bα(t) and membrane potential xα(t), with α = 1, 2. The adaption rate ε, the gain a, and all other
parameters are identical for the two rods.
In Fig. 8 we show in the right panel the stability range, for a = 1.9 and as a function of the adaption rate
ε, of the three most dominant rolling modes (1:1, 1:3 and 1:5) of the two-rod barrel. A large variety of
higher order 1:M modes (with M being an integer) is found in addition. We did not carry out a systematic
search of their stability range, which becomes progressively smaller with increasing M, and present here
only exemplary parameter settings for which the respective modes have been found by trial-and-error (by
randomly kicking the barrel). A blow-up is given in the right panel of Fig. 8. Most values of M found
are odd, but not exclusively. We cannot exclude, at this stage, that an infinite cascade M → ∞ of higher
order limit cycles may possibly occur.
The time series and the respective phase space trajectories (x1(t), x2(t)) of the 1:1, 1:3 and of the 1:5
modes are presented in Fig.9. As one can see in the time series plots, the two independent actuators,
being only coupled through the dynamics of the barrel, self-organize themselves in a constant phase-shift,
necessary for a consistent rolling. In the reduced phase space (x1, x2) the trajectories exactly close on
themselves, needing respectively 1,3 and 5 revolutions around the origin (0, 0) to close, for respectively
the 1:1, 1:3 and for the 1:5 limit cycles. In Fig. 10 we show the corresponding phase-space trajectories of
the M=9, 13 and 21 limit cycles. These modes have progressively slower average velocities vb , compare
Fig. 8, and smaller basins of attractions, being otherwise regular stable limit cycles. Whether they arise
through a bifurcation cascade of limit cycles (S´andor and Gros, 2015), or via some other mechanism, is
however beyond the scope of the present study.
This is a provisional file, not the final typeset article
10
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 9. Time series x1(t) and x2(t) of the balls along the two rods of the two-rod barrel (top row),
and the respective phase plots (x1(t), x2(t)). Shown are the 1:1/1:3/1:5 modes (left/middle/right column)
for ε = 1.0/0.5/1.5, compare Fig. 8, needing respectively 1/3/5 revolutions around the origin (x1, x2) =
(0, 0) in order to close. (click for movie).
4 DISCUSSION
It is, in a certain sense, a trivial statement, that the environment is part of the dynamical system a biological
or artificial agent lives in. Little of the environmental dynamics is however in general accessible, or known,
from the perspective of a robot, and it is hence often more suitable, as in closed-loop control (Dorf and
Bishop, 1998), to consider the sensorimotor loop as a sequence of stimulus-response reactions of the
agent, eliciting at every step the subsequent environmental signal. Here we have considered simple barrel-
shaped robots in a simulated environment, for which the sensorimotor loop constitutes truly a dynamical
system, capable of generating, even in a simple setup, a very rich palette of dynamical modes and hence
a wide range of qualitatively different types of motions.
The dominant rolling modes found are 1:M attractors, where the actuators cycle M=1,3,5,.. times during
one revolution ϕ → ϕ + 2π of the barrel. These modes coexist with non-rolling modes, having their own
respective basins of attractions, emerging from the mutual feedback of robot and environment. There exist,
in addition, regions of phase space with stationary rolling modes (rolling periodically back and forth), and
drifting back-and-forth modes. We have also found preliminary indications of rolling modes living on
two- or higher dimensional tori, with incommensurate revolution frequencies, which we did however not
investigate in detail in the present study. There may additionally exist further attracting states, yet not
discovered when performing numerical simulations within the LPZRobots environment.
All modes found are attracting dynamical states and hence robust against noise. This robustness varies
however, with the dominant 1:1 being the most stable, and higher order modes, like the 1:3 or the 1:21
limit cycles, being relatively less stable. There is, in addition, the need to overcome the dissipation, which
is present in the simulated environment, by an appropriate energy intake of the actuator. As for all robots
the question then arises, whether the observed behavior can be considered as dominantly driven, in the
Frontiers in Robotics and AI
11
S´andor et al.
The sensorimotor loop as a dynamical system
Figure 10. Examples of higher order limit cycles found for the two-rod barrel, closing (within numerical
accuracy, viz the thickness of the lines) after 9/13/21 revolutions around the origin (x1, x2) = (0, 0)
(left/middle/right). The gain is a = 1.9 and the respective adaption rates are ε = 0.61, 0.76 and 0.92,
compare Fig. 8.
sense of actuator overpowering, or as self-organized, via an inherent and essential feedback loop through
the environment (in this context see (Egbert et al., 2010) for an analogous discussion in the context of
bacterial sensorimotor system involving chemotaxis).
Actuator-controlled behavior would generally lead, in our perspective, to rather stereotypical movements
modes. The fact that our robots show a very large variety of modes upon changing the adaption rate
ε, viz the reaction time 1/ε of the actuator, indicates self-organization. These modes are also partially
overlapping with several rolling modes possibly coexisting for the same settings. It is then a question of
starting conditions, into which behavior the robot then settles.
We have also investigated the dynamics of the actuators employed, a damped-spring ball moving
along a rod, when the rolling motion dϕ/dt → 0 of the barrel is turned off. In this setting the
environmental feedback from the rolling motion is not present. We find parameter regions where the
engine is autonomously active and parameter regions, where the engine shuts itself off. In the later region
the engine may be kicked in again, when the barrel is given a kick, and allowed to roll normally. In this
case the environmental feedback is hence essential, and the motion of the robot is a consequence of self-
organizing processes in the combined phase space of the internal degrees of freedom of the robot and of
the physical environment.
Thus the behavior of the robot can not be attributed to merely one of the subsystems, but it is a property
of the coupled brain-body-environment system, a result also found in the context of minimally cognitive
agents (Beer, 2003; Beer and Williams, 2015). Since we are not aiming here for the presence of higher
level cognitive processes, our work can be seen as a purely dynamical systems approach for understanding
embodiment directly within the sensorimotor loop.
Our work has been performed with the LPZRobots simulation package, which has been used extensively
to investigate the emergence of 'playful' behavior and sensorimotor intelligence in terms of intermittent
chaotic motion patterns (Der and Martius, 2012; Martius et al., 2013). In this context our investigation
is embedded in the long-standing effort (Taga et al., 1991; Kelso, 1994; Pfeifer et al., 2007; Der and
Martius, 2015) to reduce the demanding problem of programming robots by investigating the emergence
of self-organized motions within the sensorimotor loop.
CONFLICT-OF-INTEREST STATEMENT
The authors declare that the research was conducted in the absence of any commercial or financial
relationships that could be construed as a potential conflict of interest.
This is a provisional file, not the final typeset article
12
S´andor et al.
The sensorimotor loop as a dynamical system
AUTHOR CONTRIBUTIONS
Most data and figures where produced by B. S´andor, the paper written by C. Gros and B. S´andor, with
T. Jahn and L. Martin adding data and material.
ACKNOWLEDGMENTS
We thank Georg Martius for extensive discussions and for helping setting-up the LPZRobots simulation
environment.
REFERENCES
Ay, N., Bernigau, H., Der, R., and Prokopenko, M. (2012).
dynamical system approach to autonomous robot behavior. Theory in Biosciences 131, 161 -- 179
Ay, N., Bertschinger, N., Der, R., Guttler, F., and Olbrich, E. (2008). Predictive information and
explorative behavior of autonomous robots. The European Physical Journal B 63, 329 -- 339
Baddeley, R., Hancock, P., and Foldi´ak, P. (2008).
Information theory and the brain (Cambridge
Information-driven self-organization: the
Beer, R. D. (2000). Dynamical approaches to cognitive science. Trends in cognitive sciences 4, 91 -- 99
Beer, R. D. (2003). The dynamics of active categorical perception in an evolved model agent. Adaptive
University Press)
Information processing and dynamics in minimally cognitive
Behavior 11, 209 -- 243
Beer, R. D. and Williams, P. L. (2015).
agents. Cognitive science 39, 1 -- 38
Biology 8, e1002628
Clewley, R. (2012). Hybrid models and biological model reduction with pydstool. PLoS Computational
Media)
Dasgupta, S., Worgotter, F., and Manoonpong, P. (2013).
Information dynamics based self-adaptive
reservoir for delay temporal memory tasks. Evolving Systems 4, 235 -- 249
de Wit, C. C., Siciliano, B., and Bastin, G. (2012). Theory of robot control (Springer Science & Business
Der, R. and Martius, G. (2012). The Playful Machine: Theoretical Foundation and Practical Realization
of Self-Organizing Robots, vol. 15 (Springer Science & Business Media)
Der, R. and Martius, G. (2015). A novel plasticity rule can explain the development of sensorimotor
intelligence. arXiv preprint arXiv:1505.00835
Dorf, R. C. and Bishop, R. H. (1998). Modern control systems (Pearson (Addison-Wesley))
Echeveste, R., Eckmann, S., and Gros, C. (2015). The fisher information as a neural guiding principle for
independent component analysis. Entropy 17, 3838 -- 3856
Echeveste, R. and Gros, C. (2014). Generating functionals for computational intelligence: The fisher
information as an objective function for self-limiting hebbian learning rules. Frontiers in Robotics and
AI 1
Egbert, M. D., Barandiaran, X. E., and Di Paolo, E. A. (2010). A minimal model of metabolism-based
chemotaxis. PLoS computational biology 6, e1001004
Ernesti, J., Righetti, L., Do, M., Asfour, T., and Schaal, S. (2012). Encoding of periodic and their transient
motions by a single dynamic movement primitive. In 2012 12th IEEE-RAS International Conference
on Humanoid Robots (Humanoids 2012) (IEEE), 57 -- 64
Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience 11,
127 -- 138
methods in medicine 2012
Friston, K. and Ao, P. (2011). Free energy, value, and attractors. Computational and mathematical
Gros, C. (2015). Complex and adaptive dynamical systems: A primer (Springer)
Gros, C., Linkerhand, M., and Walther, V. (2014). Attractor metadynamics in adapting neural networks.
In Artificial Neural Networks and Machine Learning -- ICANN 2014 (Springer). 65 -- 72
Frontiers in Robotics and AI
13
S´andor et al.
The sensorimotor loop as a dynamical system
Hobbelen, D. G. (2008). Limit cycle walking (TU Delft, Delft University of Technology)
Ijspeert, A. J. (2008). Central pattern generators for locomotion control in animals and robots: a review.
Neural Networks 21, 642 -- 653
Ijspeert, A. J., Nakanishi, J., Hoffmann, H., Pastor, P., and Schaal, S. (2013). Dynamical movement
primitives: learning attractor models for motor behaviors. Neural computation 25, 328 -- 373
Ijspeert, A. J., Nakanishi, J., and Schaal, S. (2002). Learning attractor landscapes for learning motor
primitives. In Advances in Neural Information Processing Systems (MIT Press), 1547 -- 1554
Kelso, J. (1994). The informational character of self-organized coordination dynamics. Human Movement
Science 13, 393 -- 413
Klyubin, A. S., Polani, D., and Nehaniv, C. L. (2005). Empowerment: A universal agent-centric measure
of control. In Evolutionary Computation, 2005. The 2005 IEEE Congress on (IEEE), vol. 1, 128 -- 135
Laszlo, J., van de Panne, M., and Fiume, E. (1996). Limit cycle control and its application to the animation
of balancing and walking. In Proceedings of the 23rd annual conference on Computer graphics and
interactive techniques (ACM), 155 -- 162
Lizier, J. T., Prokopenko, M., and Zomaya, A. Y. (2012). Local measures of information storage in
complex distributed computation. Information Sciences 208, 39 -- 54
Markovi´c, D. and Gros, C. (2010). Self-organized chaos through polyhomeostatic optimization. Physical
Markovi´c, D. and Gros, C. (2012). Intrinsic adaptation in autonomous recurrent neural networks. Neural
Review Letters 105, 068702
Computation 24, 523 -- 540
behaviors. PLOS ONE 8, e63400
Self-organizing Machines (MIT Press)
Martius, G., Der, R., and Ay, N. (2013).
Information driven self-organization of complex robotic
Nolfi, S. and Floreano, D. (2000). Evolutionary Robotics: The Biology, Intelligence, and Technology of
Olsson, L. A., Nehaniv, C. L., and Polani, D. (2006). From unknown sensors and actuators to actions
grounded in sensorimotor perceptions. Connection Science 18, 121 -- 144
Pfeifer, R., Lungarella, M., and Iida, F. (2007). Self-organization, embodiment, and biologically inspired
robotics. Science 318, 1088 -- 1093
S´andor, B. and Gros, C. (2015). A versatile class of prototype dynamical systems for complex bifurcation
cascades of limit cycles. Scientific Reports 5, 12316
Schaal, S., Kotosaka, S., and Sternad, D. (2000). Nonlinear dynamical systems as movement primitives.
In IEEE International Conference on Humanoid Robotics. 1 -- 11
Schmidt, N. M., Hoffmann, M., Nakajima, K., and Pfeifer, R. (2013). Bootstrapping perception using
information theory: case studies in a quadruped robot running on different grounds. Advances in
Complex Systems 16, 1250078
Taga, G., Yamaguchi, Y., and Shimizu, H. (1991). Self-organized control of bipedal locomotion by neural
oscillators in unpredictable environment. Biological cybernetics 65, 147 -- 159
Tani, J. and Ito, M. (2003). Self-organization of behavioral primitives as multiple attractor dynamics: A
robot experiment. Systems, Man and Cybernetics, Part A: Systems and Humans, IEEE Transactions on
33, 481 -- 488
Toyoizumi, T., Pfister, J.-P., Aihara, K., and Gerstner, W. (2005). Generalized bienenstock-cooper-
munro rule for spiking neurons that maximizes information transmission. Proceedings of the National
Academy of Sciences of the United States of America 102, 5239 -- 44
Triesch, J. (2005). A gradient rule for the plasticity of a neurons intrinsic excitability. In Artificial Neural
Networks: Biological InspirationsICANN 2005 (Springer). 65 -- 70
Triesch, J. (2007). Synergies between intrinsic and synaptic plasticity mechanisms. Neural Computation
19, 885 -- 909
Williams, P. and Beer, R. (2013). Environmental feedback drives multiple behaviors from the same neural
circuit. In Advances in Artificial Life, ECAL 2013 (MIT Press), vol. 12, 268 -- 275
Ziemke, T. (2003). Whats that thing called embodiment. In Proceedings of the 25th Annual meeting of
the Cognitive Science Society (Mahwah, NJ: Lawrence Erlbaum), 1305 -- 1310
This is a provisional file, not the final typeset article
14
|
1010.2530 | 1 | 1010 | 2010-10-12T22:29:48 | Emergent complex neural dynamics | [
"q-bio.NC",
"cond-mat.dis-nn",
"physics.bio-ph"
] | A large repertoire of spatiotemporal activity patterns in the brain is the basis for adaptive behaviour. Understanding the mechanism by which the brain's hundred billion neurons and hundred trillion synapses manage to produce such a range of cortical configurations in a flexible manner remains a fundamental problem in neuroscience. One plausible solution is the involvement of universal mechanisms of emergent complex phenomena evident in dynamical systems poised near a critical point of a second-order phase transition. We review recent theoretical and empirical results supporting the notion that the brain is naturally poised near criticality, as well as its implications for better understanding of the brain. | q-bio.NC | q-bio |
Emergent complex neural dynamics: the brain at the edge∗
Dante R. Chialvo
Department of Physiology, David Geffen School of Medicine, UCLA,
Los Angeles, CA 90024, USA and Facultad de Ciencias Medicas,
Universidad Nacional de Rosario, Rosario 2000, Argentina.
A large repertoire of spatiotemporal activity patterns in the brain is the basis for adaptive be-
haviour. Understanding the mechanism by which the brains hundred billion neurons and hundred
trillion synapses manage to produce such a range of cortical configurations in a flexible manner
remains a fundamental problem in neuroscience. One plausible solution is the involvement of uni-
versal mechanisms of emergent complex phenomena evident in dynamical systems poised near a
critical point of a second-order phase transition. We review recent theoretical and empirical results
supporting the notion that the brain is naturally poised near criticality, as well as its implications
for better understanding of the brain.
Understanding the brain is among the most challenging problems to which a physicist can be attracted. As a
system with an astronomical number of elements, each one known to have plenty of nonlinearities, the brain exhibits
collective dynamics that in many aspects resemble some of the classic problems well studied in statistical physics.
The contradiction, and the provoking point in these notes, is that only a minority of the publications in the field
today are concerned with the understanding of the brain dynamics as a collective process. To the contrary, the great
majority of the work explains the brain through explicit or implicit connectionist paradigms. In our opinion there is
a need to reflect and recognize to what degree these collectivist-connectionist views imply more than just a semantic
difference, and that its adequate resolution holds the key to resolve some of the more puzzling questions about the
brain. We review key results on emergent complex neural dynamics over the past few years. From the outset it
should be noted that the intentionally provoking nature of these notes naturally induces a strong bias regarding cited
publications; consequently this is neither a fair, nor historically correct, exhaustive or updated review of the relevant
literature. Another cautionary note is that being a subject at the fringe of disciplines, physicists and biologists
alike, will encounter boring passages on their most familiar topics. Nevertheless, for the sake of clarity, and with the
forgiveness of the readers, we will proceed to (even excessively) define each issue at hand.
What are the issues?
Understanding human behavior and cognition requires the description of the laws of the underlying neural collective
phenomena, the patterns of spatio-temporal brain activity. Formal approaches to study collective phenomena are one
of the classical topics at the center of statistical physics, with recent novel and successful applications in diverse
areas such as genetics, ecology, computer science, social and economic settings [1–13]. While in all these fields there
is a clear transfer of methods and ideas from statistical physics, a similar flow has only recently started to impact
neuroscience.
The main issue addressed here belongs to the under the rug class of problems in the field, namely, how the very
large conglomerate of interconnected neurons produce a repertoire of given behaviors in a flexible and self organized
way. Although colloquial explanations abound, when detailed models are constructed to account for that, each of
the three emphasized condiments are systematically violated. Either 1) the model is a low dimensional version of
the neural structure of interest or 2) it produces a single behavior (hardwired in the system), and consequently 3)
it cannot flexibly do anything else. It is only by arbitrarily changing the neuronal connections that most of current
models can play a repertoire of behaviors. This requires a kind of supplementary brain holding the key to which
connections need to be rewired. Of course, no proof of such supplementary brain has been offered, and this is the
question that is screaming to be answered and seldom is even being asked.
Approaches to this problem, for a variety of historical and conceptual reasons, are still drawn from connectionist
paradigms which restrict the dynamics to be generated by circuits, and consequently funnel our efforts in the same
sterile direction. While collective properties have been mentioned for a long time, its relevance remains secondary to
most of us. Even Hopfield’s call [14] three decades ago (in his “Neural networks and physical systems with emergent
collective computational abilities” seminal paper), seems to have been forgotten, perhaps displaced by the appeal and
initial excitement of computational ideas:
∗ Nature Physics 6, 744-750 (2010)
2
“Much of the architecture of regions of the brains of higher animals must be made from a proliferation of simple
local circuits with well-defined functions. The bridge between simple circuits and the complex computational properties
of higher nervous systems may be the spontaneous emergence of new computational capabilities from the collective
behavior of large numbers of simple processing elements.”
Emergence
It is accepted that almost all macroscopic phenomena -from superconductivity to gravity and from economics to
photosynthesis- are the consequence of an underlying collective dynamics of their microscopic components. In neuro-
science, it is the macroscopic behavior (cognitive, emotional, motor, etc) aspect that will be ultimately understood
as the emergent phenomena of an underlying neuronal collective. However, neurons being nonlinear elements, makes
such understanding far from straightforward. It would be fair to say that while the problem is cast in terms most
familiar to biology the solution is written in terms very familiar to physics.
Let’s recall what emergent phenomena are. Emergence refers to the unexpected collective spatiotemporal patterns
exhibited by large complex systems. In this context, “unexpected” shows our inability (mathematical and otherwise)
to derive such emergent patterns from the equations describing the dynamics of the individual parts of the system. As
discussed at length elsewhere [1, 15], complex systems are usually large conglomerates of interacting elements, each
one exhibiting some sort of nonlinear dynamics. Without entering into details, it is also known that the interaction
can also be indirect, for instance through some mean field. Usually energy enters into the system, thus some sort of
driving is present. The three emphasized features, ( i.e., large number of interacting nonlinear elements) are necessary,
although not sufficient, conditions for a system to exhibit emergent complex behavior at some point.
As long as the dynamics of each individual element is nonlinear, other details are not important [1, 16]; for instance,
they can be humans, driven by food and other energy resources, from which some collective political or social structure
eventually arises. Whatever the type of structure that emerges, it is unlikely to appear if one of the three above-
emphasized properties is absent. For instance, it is well established that a small number of isolated linear elements
are unable to produce unexpected behavior (indeed this is the case in which everything can be mathematically
anticipated).
Spontaneous brain activity is complex
It is evident, from the very early electrical recordings a century ago, that the brain is spontaneously active, even in
absence of external inputs. However obvious this observation could appear, it was only recently that the dynamical
features of the spontaneous brain state started to be studied in any significant way.
Recent work on brain rhythms at small and large brain scales showed that spontaneous healthy brain dynamics is
not composed by completely random activity patterns or by periodic oscillations [17]. Careful analysis of the statistical
properties of neural dynamics under no explicit inputs has identified complex patterns of activity previously neglected
as background noise dynamics. The fact is that brain activity is always essentially arrhythmic regardless of how it is
monitored, whether as electrical activity in the scalp (EEG), by techniques of functional magnetic resonance imaging
(fMRI), in the synchronization of oscillatory activity [18, 19], or in the statistical features of local field potentials
peaks [20].
It has been pointed out repeatedly [21–25] that, under healthy conditions no brain temporal scale takes primacy
over average, resulting in power spectral densities decaying of “1/f noise”. Behavior, the ultimate interface between
brain dynamics and the environment, also exhibits scale invariant features as shown in human cognition [26–28] human
motion [29] as well as animal motion [30]. The origin of the brain scale free dynamics was not adequately investigated
until recently, probably (and paradoxically) due to the ubiquity of scale invariance in nature [1]. Currently, there is
increasing interest and the potential significance of a renewed interpretation of the brain spontaneous patterns is at
least double. Its presence provides important clues about brain circuits organization, in the sense that our previous
ideas cannot easily accommodate these new findings. Also, the class of complex dynamics observed seems to provide
the brain with previously unrecognized robust properties. These aspects will be reviewed, on two different scales, in
the next sections.
Emergent complexity is always critical
The commonality of scale-free dynamics in the brain naturally leads one to ask what physics knows about very
general mechanisms able to produce such dynamics. Attempts to explain and generate nature’s non- uniformity
3
FIG. 1: Neuronal avalanches are complex. Size distribution of neuronal avalanches in mature cortical cultured networks follows
a power law with an exponent close to 3/2 (dashed line) and exhibits finite size scaling. The relative probability of observing
an avalanche covering a given number of electrodes is shown for three sets of grid sizes (insets with n=15, 30 or 60 sensing
electrodes, equally spaced at 200µm). The statistics is taken from data recordings lasting a total of 70 hours and accumulating
58000 (+/− 55000) avalanches per hour (mean +/− SD). (Re-plotted from Figure 4 of [42]).
included several mathematical models and recipes, but few succeeded in creating complexity without embedding the
equations with complexity. The important point is that including the complexity in the model will only result in a
simulation of the real system, without entailing any understanding of complexity. The most significant efforts were
those aimed at discovering the conditions in which something complex emerges from the interaction of the constituting
non-complex elements [1, 31]. Initial inspiration was drawn from work in the field of phase transitions and critical
phenomena (see Box 1). Precisely, one of the novelties of critical phenomena is the fact that out of the short-range
interaction of simple elements eventually long-range spatiotemporal correlated patterns emerge. As such, critical
dynamics have been documented in species evolution [1], ants collective foraging [32, 33] and swarm models [34],
bacterial populations [35], traffic flow in highways [1] and on the Internet [10], macroeconomic dynamics [7], forest
fires [8], rainfall dynamics [11–13] and flock formation [36]. Same rationale lead to the conjecture [1, 37, 38] that also
the complexity of brain dynamics was just another signature of an underlying critical process. Since at the point near
the transition exist the largest number of metastable states, the brain can then be accessing the largest repertoire of
behaviors in a flexible way. That view claimed that the most fundamental properties of the brain only are possible
staying close to that critical instability (see Box 2) independently of how such state is reached or maintained.
Small scale: Cortical quakes
Beggs and Plenz [39, 40] reported the first convincing evidence that neuronal populations could exhibit critical
dynamics. They described a novel type of electrical activity for the brain cortex called “neuronal avalanches”. These
collective neuronal patterns sit halfway in between two previously well-known cortical patterns: the oscillatory or
wave-like highly coherent activity on one side and the asynchronic and incoherent spiking on the other. Typically,
each avalanche engages a variable number of neurons. What is peculiar is the statistical pattern that these avalanches
follow. On the average, one observes many more small avalanches than big ones (e.g., each neuronal avalanche has a
large chance to engage only a few neurons and a very low probability to spread and activate the whole cortical tissue
(see Fig. 1)). In these experiments, a number of properties suggestive of criticality were estimated. This included
a scale-free distribution of avalanche sizes following an inverse power law with an exponent close to 3/2 that agrees
exactly with the theoretical expectation for a critical branching process, previously worked out by Zapperi et al[41].
The avalanches lifetimes statistics followed also an inverse power law with an exponent close to 2, which agrees with
the theoretical expectation for a cascade of activity [41, 43, 44].
The initial scale invariance for the avalanches has already been replicated [45, 46]. Furthermore, similar findings
have been reported on a wide variety of diverse settings, including in vivo monkey’s cortex [47] and adult cats [48]. In
addition, the functional significance of the avalanches was highlighted by the fact that they were observed during the
4
FIG. 2: Large scale emergent brain networks: The analysis of interactions during spontaneous human brain spatiotemporal
patterns reveals emergent networks. Top sequence of images shows in red/blue the increases/decreases over the mean fMRI
BOLD during four minutes of consecutive brain resting (Single-subject consecutive data, starting from the top-left corner
where each row is 1 minute, and images are taken at 2.5 sec.
intervals). The bottom images are the results of computing
linear correlations between the activity of a small region within the networks of interest and the rest of the brain (stronger
correlations indicated with brightest colors). These networks correspond to the six mayor systems in the brain: visual, auditory,
sensorimotor, default mode, executive control and dorsal attention (From Chialvo, unpublished data. The MNI152 coordinates
for the slices are z=18 for the top sequence and z=0,8,44,24,26,44 for the bottom figures, left to right).
earliest time of the development of superficial layers in the cortex [49, 50] requiring the presence of a neuro-modulator
(dopamine) and a certain balance between excitatory and inhibitory transmission [39, 49, 50].
The precise neuronal mechanisms leading to the observed scale-free avalanches is yet uncertain, despite modeling
efforts underway (see Box 3), because similar statistics can be generated by several mechanisms other than critical
dynamics. Yet, no convincing alternative experimental analysis or evidence has been presented up to now. Numerical
evidence recently reported [52] suggesting non-critical alternatives gives inverse power law exponents one order of
magnitude larger (> 20) than the 3/2 experimentally observed.
The stumbling block of the discussions concerning the origin of the avalanches has been the limitation to replicate
only probability densities, either of sizes or durations. However, the debate can be placed in a more rigorous context
if other invariants are analyzed. In that direction, recent results [53] provided novel experimental evidence for five
fundamental properties of neuronal avalanches consistent with criticality. These were: (1) time scales separation
between the dynamics of the triggering event and the avalanche itself. This is demonstrated by the fact that the inverse
power law density of avalanches sizes and lifetimes remain invariant to slow driving; (2) stationary of avalanches size
statistics despite wide avalanching rate fluctuations, excluding non-homogeneous Poisson processes; (3) the avalanche
probabilities preceding and following main avalanches obey Omori’s law for earthquakes; (4) the average size of
avalanches following a main avalanche decays as an inverse power law; (5) avalanches spread spatially on a fractal.
Overall, these results support the notion that neuronal avalanches are the manifestation of criticality in the brain
and exclude, in some cases explicitly, the majority of the mechanisms discussed in the literature as alternatives to
criticality.
Large scale
Probably the first report concerning mesoscopic patterns in connection with behavior was Kelso et al [55] brain
imaging analysis pursuing their previous observation that human hand movements exhibit abrupt phase transitions for
increasing cycling frequency [54]. They were able to show, using magneto-encephalographic techniques, spontaneous
transitions in neuromagnetic field patterns in the human brain. These transitions happened at a critical value of a
systematically varied behavioral parameter supporting “the thesis that the brain is a pattern forming system that
1 5
can switch flexibly from one coherent state to another”. Similar considerations and concerns were expressed in these
early days, by commenting that: “In summary, higher brain functions in humans such as perception, learning and
goal directed movement are often hypothesized to depend on the collective dynamics of large numbers of interacting
neurons distributed throughout the cortex. But typical signs of cooperative phenomena are not accessible through
single neuron investigations. On the other hand, from studies of nonequilibrium systems it is well known that at
critical points, spatial and temporal patterns form in a so-called self-organized fashion” [55].
A large body of work needs to be omitted here to be able to fast forward to present day, when it is recognized that
the brain is spontaneously creating and reshaping complex functional networks of correlated dynamics responding to
the neural traffic between regions. These networks had been recently studied, using functional magnetic resonance
imaging in humans. The flurry of activity in this area could be well gauged by the words chosen by the author [56] of
a recent review stating: “Commenting on the wealth of existing data on anatomical and functional cortical networks
organization may seem like “carrying coals to Newcastle”. Extensive reviews [57, 58] cover the statistical physics
approaches that are increasingly being use to analyze this large body of complex data.
For the purpose of this note, it is relevant to limit our attention to the study of spontaneous “resting” fMRI
dynamics. [59, 60]. Brain “rest” is defined -more or less unsuccessfully- as the state in which there is no explicit brain
input or output, or overt external stimulation. The subject is scanned while lying with eyes closed, and instructed
to avoid falling asleep. Each of the thousands of signals (so called BOLD, for blood oxygenation level dependent)
obtained from these experiments reflects the amount of neural activity on each small region of typically dozen of cubic
millimeters, allowing one to map the entire activity of the brain. An example is presented as a sequence in Figure 2,
which shows for graphical purposes only one of the many slices that are recorded. From careful visual inspection of
the data it appears already that there are important spatiotemporal correlations, somehow resembling the image of
passing clouds. The fascinating point here is that from rather simple linear cross correlations of the BOLD signals a
few collective groups emerge. This is shown by the clusters in the bottom panels, which were found to closely match
the same regions responding to a wide variety of different activation conditions [59, 60]. Thus, at rest the “passing
clouds” (i.e., the collective spatiotemporal dynamics) visit the same brain regions that are activated during any given
active behavior. The relevance of these findings is further highlighted by the fact that these networks are identifiable
with great consistency across subjects [61–63], even during sleep [64] or anesthesia [65].
A natural question arising at this point is what kind of known dynamical scenario corresponds to these brain resting
patterns. This was tackled initially in three recent reports. In the first Kitzbichler et al [66] analyzed functional MRI
and MEG data recorded from normal volunteers at resting state using phase synchronization between diverse spatial
locations. They reported a scale invariant distribution for the length of time that two brain locations on the average
remained locked. This distribution was also found in the Ising and the Kuramoto model [67] at the critical state.
In the second report Fraiman et al [68] compared the paradigmatic 2-dimensional Ising model at various tempera-
tures with the resting brain data. Correlations networks were prepared by computing cross-correlation between the
Ising states at all lattice points and placing links between those points (nodes) with correlation larger than a certain
threshold. Similar computations were conducted for the brain fMRI data. After comparing the most descriptive net-
works properties, the authors concluded that while the Ising networks at sub- and supercritical temperatures greatly
differ from the brain networks, those derived at the critical temperature are “indistinguishable from each other”. The
example shown in Figure 3, shows one of the properties compared, the distribution of the number of links (i.e., degree)
for these networks. Notice the fat tails in the brain data, (reported earlier in refs[69, 70]) which are only replicated
when the Ising model is posed at critical temperature. In addition, calculation of the fraction of sites with positive and
negative correlations, (a variable related to magnetization elsewhere) showed values close to one, in agreement with
previous results by Baliki et al [71] which suggested this balance as an index of healthy brain function at rest. Overall,
these results show that networks derived from correlations of fMRI signals in human brains are indistinguishable from
networks extracted from Ising models at critical temperature.
The third effort directed to shed light on the mechanism underlying resting fMRI dynamics is due to Expert et al
[72] who examined the two point correlation function after successive steps in spatial coarse graining, a renormalization
technique widely used in critical phenomena. Their results show spatial self-similarity which in addition to temporal
1/f frequency behavior of the power spectrum are indicative of critical dynamics.
Since the initial fMRI work [69, 70, 73] progress has been made to use these approaches to evaluate the integrity
of brain function under normal [74] and pathological conditions [75, 76] including Alzheimer [77], squizophrenia [78]
and epilepsy [79, 80]. Even the impact of long enduring chronic pain seems to alter brain dynamics beyond the feeling
of pain itself [71], thus motivating further work to better understand the fundamental mechanisms behind the brain
resting state large-scale organization.
6
FIG. 3: Complex networks derived from the brain fMRI data mimic those from the Ising model only at the critical temperature.
Plots correspond to the degree (k) distribution of the network derived from the fMRI brain resting data (bottom panel) and
from the Ising model (top three panels) at T = 2, T = 2.3 and T = 3, for different values of averages degrees (From Fraiman
et al [68]).
Summary and Outlook
We have reviewed recent key results of emergent collective neural dynamics. It is important to note that, at present
time, no theory (in the sense of the initial paragraphs) comprehensibly can accommodate these results without invoking
criticality. One motivation for neuroscience to look at the physical laws governing other complex systems is the hope
that universality will give the field an edge. Instead to search for ad-hoc laws for the brain, under the pretense that
biology is special, most probably a good understanding of universal laws might provide a breakthrough since brains
must share some of the fundamentally laws of nature. A major difference between the preceding decade and now is
that, as presented in this note, there is spatiotemporal brain data to confront theories with; a playground awaiting
for physicists to take up the challenge of explaining the underlying mechanism of the collective.
Box 1: What is special about being critical?
Given the claims that brain dynamics can be critical, let’s recall what is special about such a state. The generality
of the scenario of ferromagneticparamagnetic phase transition is used, without implying that the brain would reach
criticality in this way. As an iron magnet is heated, the magnetization decreases until it reaches zero beyond a critical
temperature Tc. Individual spins orientations are, at high temperatures, changing continuously in small groups. As
a consequence, the mean magnetization, expressing the collective behavior, vanishes. At low temperature the system
will be very ordered exhibiting large domains of equally oriented spins, a state with negligible variability over time. In
between these two homogeneous states, at the critical temperature Tc, the system exhibits very different properties
both in time and space. The temporal fluctuations of the magnetization are known to be scale invariant. Similarly,
the spatial distribution of correlated spins show long-range (power law) correlations. It is only close enough to Tc that
large correlated structures (up to the size of the system) emerge, even though interactions are with nearest neighbor
elements. In addition, the largest fluctuations in the magnetization are observed at Tc. At this point the system
is at the highest susceptibility, a single spin perturbation has a small but finite chance to start an avalanche that
reshapes the entire systems state, something unthinkable on a non-critical state. Many of these dynamical properties,
once properly translated to neural terms, exhibit striking analogies to brain dynamics. Neuro-modulators, which are
known to alter brain states acting globally over nonspecific targets, could be thought as control parameters, as is
temperature in this case.
101102103Degree (k)101102103Frequency101102103Degree (k)101102103Degree (k)101102103Degree (k)101102103Frequency< k > ~ 713< k > ~ 127< k > ~ 26T = 2T = 2.3 T = 3Brain 7
FIG. 4: Complex is critical: Three snapshots of the spin configurations at one moment in time for three temperatures (subcrit-
ical, critical and supercritical) from numerical simulations of the Ising model (d=2). Only at the critical temperature systems
exhibiting a second-order phase transition show the highly heterogeneous correlated domains commonly seen as complex, while
both sub- and supercritical conditions result in homogeneous states.
Box 2: Why we need a brain at all?
It is self-evident that the brains we see today are those that inherited an edge useful to survive. In light of this,
how consistent with Darwinian constraints could it be to suggest that the brain should evolve to be near a critical
point? The answer, in short, is that brains should be critical because the world in which they must survive is to some
degree critical as well. Lets see the alternatives: in a subcritical world, everything would be simple and uniform (as in
the left panel of Box 1) and there would be nothing to learn; a brain would be completely superfluous. At the other
extreme, in a supercritical world, everything would be changing continuously (as in the right panel of Box 1); under
these circumstances there would not be sufficient regularity as to make learning possible or valuable. Thus, brains
are only needed to navigate a complex, critical world, where surprising events still have a finite chance of occurring.
In other words, animals need a brain because the world is critical [1, 15, 31, 37]. Furthermore, a brain not only has
to remember, but also has to forget and adapt. In a subcritical brain, memories would be frozen. In a supercritical
brain, patterns change continuously so that no long-term memory would be possible. To be highly susceptible, the
brain itself has to be in an in-between, critical state.
Which generic features of systems at criticality should be expected in brain experiments?
1. At relatively large scale:
Cortical long-range correlations in space and time,
Correlation length divergence
Near-zero magnetization or equivalently, the presence of anti-correlated cortical states.
2. At relatively small scale:
Cortical circuits exhibit neuronal avalanches, cascades of activity obeying inverse power law statistics as well as
long-range correlations.
3. At behavioral level:
Adaptive human behavior should be bursty appearing unstable, as it was always at the “edge of failure”.
Life-long learning continuously “raises the bar” to more challenging tasks, making performance critical as well.
Box 3: Models
At small scale, explicit models have been already presented in which criticality can be self organized either by
some form of Hebbian learning [81, 82] or by the inclusion of activity-dependent depressive synapses [83]. Some novel
properties endowed by criticality have been also studied, such as the widest dynamical range and optimal sensitivity to
sensory stimuli shown by Kinouchi and Copelli [84]. Concerning learning, de Arcangelis and Herrmann [85] extending
Bak & Chialvo [37, 38] earlier network model, found recently that avalanches of activity are able to shape a network
with complex connectivity as well as learning logical rules. A handicap of these models is the absence of ongoing
activity, an important brain feature. This aspect is covered by Marro and colleagues [86] networks which show
spontaneous unstable dynamics and nonequilibrium phases in which the global activity wanders irregularly among
attractors resulting in 1/f noise as the system falls into the most irregular behavior. At large scale although brain
data sets with unprecedented spatiotemporal resolution are now available, there is no model able to mimic a phase
transition at such brain scale. Nevertheless, the challenge to construct data driven models at this level is being taken,
as shown by Honey et al [87] recent efforts to model fMRI resting state.
T<TCT>TCT~TCCriticalSubCriticalSuperCritical Acknowledgments
8
DRC is with the National Research and Technology Council (CONICET) of Argentina. This work was supported
by CONICET and by NIH NINDS NS58661. Thanks to E. Tagliazucchi and F. Lebensohn-Chialvo for reading the
manuscript.
[1] Bak P. How nature works. Oxford, UK: Oxford University Press (1997).
[2] Jensen HJ. Self-Organized Criticality. Cambridge University Press (1998).
[3] Buchanan M, Ubiquity. London, UK, Weidenfeld and Nicolson (2000).
[4] Christensen K & Moloney NR, Complexity and Criticality, Imperial College Press, (2005).
[5] Aldana M, Balleza E, Kauffman SA, Resendis O. Robustness and evolvability in genetic regulatory networks. J. of Theo-
retical Biology, 245, 433-448 (2007).
[6] Balleza E, Alvarez-Buylla E, Chaos A, Kauffman SA, Shmulevich I, Aldana M. Critical dynamics in genetic regulatory
networks: examples from four kingdoms, PLoS ONE, 3, 2456, (2008).
[7] Lux T & Marchesi M. Scaling and criticality in a stochastic multi-agent model of a financial market. Nature 397, 498-500
(1999).
[8] Malamud BD, Morein G, Turcotte DL. Forest fires: An example of self-organized critical behavior. Science 281,1840-1842
(1998).
[9] Nykter M, Price ND, Aldana M, Ramsey SA, Kauffman SA, Hood LE, Yli-Harja O, Shmulevich I. Gene expression
dynamics in the macrophage exhibit criticality. Proc Natl Acad Sci U S A. 105(6):1897-1900 (2008).
[10] Takayasu M, Takayasu H, Fukuda K. Dynamic phase transition observed in the internet traffic flow. Physica A 277, 248-255
(2000)
[11] Peters O. & Neelin D, Critical phenomena in atmospheric precipitation. Nature Phys. 2, 393-396 (2006).
[12] Peters O, Hertlein C, Christensen K, A complexity view of rainfall. Phys. Rev. Lett. 88, 018701-1 (2002).
[13] Peters O & Christensen K, Rain: relaxations in the sky. Phys. Rev. E 66, 036120-1 (2002).
[14] Hopfield JJ, Neural networks and physical systems with emergent collective computational capabilities, Proc Natl Acad
Sci U S A 79, 2554 (1982).
[15] Bak P & Paczuski M. Complexity, contingency, and criticality. Proc Natl Acad Sci U S A 92, 6689-6696. (1995).
[16] Anderson P. More is different. Science 4393-396 (1972)
[17] Buzsaki G. Rhythms of the Brain (New York, NY: Oxford University (2006).
[18] Linkenkaer-Hansen K, Nikouline VV, Palva JM, Ilmoniemi RJ. Long-range temporal correlations and scaling behavior in
human brain oscillations. J Neurosci 21:1370-1377 (2001).
[19] Stam CJ & de Bruin EA. Scale-free dynamics of global functional connectivity in the human brain. Hum. Brain Mapp. 22,
97-109 (2004)
[20] Plenz D &Thiagarajan TC. The organizing principles of neuronal avalanches: Cell assemblies in the cortex? Trends
Neurosci 30,101-110 (2007).
[21] Bullock TH, Mcclune MC, Enright JT. Are the electroencephalograms mainly rhythmic? Assessment of periodicity in
wide-band time series. Neuroscience 121, 233-252 (2003).
[22] Logothetis, NK. The neural basis of the blood-oxygen-level-dependent functional magnetic resonance imaging signal. Philos.
Trans. R. Soc. Lond. B Biol. Sci. 357, 1003-1037 (2002).
[23] Eckhorn, R. Oscillatory and non-oscillatory synchronizations in the visual cortex and their possible roles in associations of
visual features. Prog. Brain Res. 102, 405-426 (1994).
[24] Miller KJ, Sorensen LB, Ojemann JG, den Nijs M. Power law scaling in the brain surface electric potential. PLoS Comput.
Biol. 5, e1000609. 10.1371/journal.pcbi.1000609 (2009).
[25] Manning JR, Jacobs J, Fried I, Kahana MJ. Broadband shifts in local field potential power spectra are correlated with
single-neuron spiking in humans. J. Neurosci. 29, 13613-13620 (2009).
[26] Gilden DL. Cognitive emissions of 1/f noise. Psychol. Rev. 108, 33-56.(2001).
[27] Maylor EA, Chater N, Brown GD. Scale invariance in the retrieval of retrospective and prospective memories. Psychon.
Bull. Rev. 8, 162-167 (2001).
[28] Ward LM, Dynamical Cognitive Science, London: The MIT Press (2002).
[29] Nakamura T, Kiyono K, Yoshiuchi K, Nakahara R, Struzik ZR, Yamamoto Y. Phys. Rev. Lett. 99, 138103 (2007).
[30] Anteneodo C & Chialvo DR. Unraveling the fluctuations of animal motor activity. Chaos. 19(3), 033123 (2009).
[31] Bak P, Tang C, Wiesenfeld K. Self-organized criticality: An explanation of the 1/f noise. Phys. Rev. Lett. 59, 381 (1987).
[32] Beckers R, Deneubourg J-L, Goss S, Pasteels JM. Collective decision making through food recruitment. Insectes Sociaux,
37, 258-267 (1990).
[33] Beekman M, Sumpter DJT, Ratnieks FLW Phase transition between disordered and ordered foraging in Pharaohs ants.
Proc Natl Acad Sci USA 98, 9703-9706 (2001).
[34] Rauch EM, Chialvo DR, Millonas MM. Pattern formation and functionality in swarm models. Phys Lett A, 207, 185-193
(1995).
[35] Nicolis G & Prigogine I. Self-Organization in nonequilibrium systems: From dissipative structures to order through fluc-
9
tuations. Wiley, New York (1977).
[36] Cavagna A., et al, Scale-free correlations in starling flocks. Proc Natl Acad Sci USA 107, 11865-11870 (2010).
[37] Bak P & Chialvo DR, Adaptive learning by extremal dynamics and negative feedback. Phys. Rev. E, 63, 031912 (2001).
[38] Chialvo, DR & Bak P. Learning from mistakes. Neuroscience, 90, 1137 (1999).
[39] Beggs JM & Plenz D. Neuronal avalanches in neocortical circuits. J Neurosci 23, 11167-11177 (2003).
[40] Beggs JM & Plenz D. Neuronal avalanches are diverse and precise activity patterns that are stable for many hours in
cortical slice cultures. J Neurosci 24, 5216-5229 (2004).
[41] Zapperi S, Baekgaard LK, Stanley HE. Self-organized branching processes: meanfield theory for avalanches, Phys. Rev.
Lett. 75, 4071-4074 (1995).
[42] Chialvo DR. Critical brain networks. Physica A 340, 756-765 (2004).
[43] Eurich CW, Herrmann JM, Ernst UA. Finite-size effects of avalanche dynamics. Phys. Rev. E. 66, 066137 (2002).
[44] Teramae JN & Fukai T. Local cortical circuit model inferred from power-law distributed neuronal avalanches. J Comput
Neurosci 22(3), 301-12 (2007).
[45] Mazzoni A, Broccard FD, Garcia-Perez E, Bonifazi P, Ruaro ME, Torre V. On the dynamics of the spontaneous activity
in neuronal networks. PLoS ONE 2(5), e439 (2007).
[46] Pasquale V, Massobrio P, Bologna LL, Chiappalone M, Martinoia S. Self-organization and neuronal avalanches in networks
of dissociated cortical neurons. Neuroscience 153(4),1354-69 (2008).
[47] Petermann T, Thiagarajan TC, Lebedev MA, Nicolelis MA, Chialvo DR, Plenz D. Spontaneous cortical activity in awake
monkeys composed of neuronal avalanches. Proc Natl Acad Sci U S A 106, 15921-15926 (2009).
[48] Hahn G, Petermann T, Havenith MN, Yu S, Singer W, Plenz D, Nikoli DJ. Neuronal avalanches in spontaneous activity
in vivo. J of Neurophysiol. In press (2010).
[49] Stewart CV & Plenz D. Inverted U profile of dopamine-NMDA mediated spontaneous avalanche recurrence in superficial
layers o rat prefrontal cortex. J. Neurosci. 23, 8148-8159. (2006).
[50] Gireesh ED & Plenz D. Neuronal avalanches organize as nested theta- and beta/gamma oscillations during development
of cortical layer 2/3. Proc.Nat.Acad.Sci. USA 105, 7576-7581 (2008).
[51] Bedard C, Kroger H, Destexhe A. Does the 1/f frequency scaling of brain signals reflect self-organized critical states? Phys
Rev Lett 97(11):118102, (2006).
[52] Touboul J & Destexhe A. Can power-law scaling and neuronal avalanches arise from stochastic dynamics? PLoS One
5(2):e8982 (2010).
[53] Plenz D & Chialvo DR. Scaling properties of neuronal avalanches are consistent with critical dynamics. arXiv:0912.5369v1
[q-bio.NC] (2009).
[54] Kelso JAS. (1984). Phase transitions and critical behavior in human bimanual coordination. Am. J. Physiol. Regul. Integr.
Comp. 15, R1000-R1004 (1984).
[55] Kelso JAS, Bressler SL, Buchanan S, DeGuzman GC, Ding M, Fuchs A, Holroyd T (1992) Phase transition in brain and
human behavior. Phys Lett A 169, 134-144 (1992).
[56] Werner G. Fractals in the nervous system: conceptual implications for Theoretical neuroscience. Frontiers in Fractal
Physiology 1, 15 (2010).
[57] Bullmore E & Sporn O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nat.
Rev. Neurosci. 10,186-198 (2009).
[58] Sporns O, Chialvo DR, Kaiser M, Hilgetag CC. Organization, development and function of complex brain networks. Trends
Cog. Sci. 8(9), 418-425 (2004).
[59] Fox MD & Raichle ME. Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging.
Nat. Rev. Neurosci. 8, 700-711 (2007).
[60] Smith SM et al. Correspondence of the brains functional architecture during activation and rest. Proc. Natl. Acad. Sci. U.
S. A. 106, 13040-13045 (2009).
[61] Xiong J, Parsons L, Gao J, Fox P. Interregional connectivity to primary motor cortex revealed using MRI resting state
images. Hum Brain Mapp 8, 151-156 (1999).
[62] Cordes D et al. Mapping functionally related regions of brain with functional connectivity MR imaging. Am J Neuroradiol
21, 1636-1644 (2000).
[63] Beckmann CF, De Luca M, Devlin JT, Smith SM Investigations into resting-state connectivity using independent compo-
nent analysis. Philos Trans R Soc London 360, 1001-1013 (2005).
[64] Fukunaga M et al. Large-amplitude, spatially correlated fluctuations in BOLD fMRI signals during extended rest and early
sleep stages. Magn Reson Imaging 24, 979-992 (2006).
[65] Vincent J L et al. Intrinsic functional architecture in the anaesthetized monkey brain. Nature 447: 83-87 (2007).
[66] Kitzbichler MG, Smith ML, Christensen SR, Bullmore E. Broadband criticality of human brain network synchronization.
PLoS Comput Biol 5:e1000314. (2009).
[67] Kuramoto Y. Chemical oscillations, waves and turbulence. Springer, Berlin (1984).
[68] Fraiman D, Balenzuela P, Foss J, Chialvo DR. Ising-like dynamics in large-scale functional brain networks. Phys. Rev. E,
79, 061922 (2009).
[69] Eguiluz VM, Chialvo DR, Cecchi GA, Baliki M, Apkarian AV. Scale-free brain functional networks. Phys. Rev. Lett. 94,
018102 (2005).
[70] van den Heuvel MP, Stam CJ, Boersma M, Hullshof Pol HE. Small-world and scale-free organization of voxel-based
resting-state functional connectivity in the human brain. NeuroImage 43, 528-539 (2008).
[71] Baliki MN, Geha PY, Apkarian AV, Chialvo DR. Beyond feeling: chronic pain hurts the brain, disrupting the default-mode
10
network dynamics. J. of Neurosc. 28, 1398-1403 (2008).
[72] Expert P, Lambiotte R, Chialvo DR, Christensen K, Jensen, HJ, Sharp DJ, Turkheimer F. (2010) Self-similar correlation
function in brain resting state fMRI. arXiv:1003.3682v1 [q-biol.NC].
[73] Salvador R et al. Neurophysiological architecture of functional magnetic resonance images of human brain. Cerebral Cortex
15, 1332-1342 (2005).
[74] Damoiseaux JS, et al Proc. Natl. Acad. Sci. U.S.A. 103, 13848 (2006).
[75] Broyd SJ, et al. Default-mode brain dysfunction in mental disorders: A systematic review. Neuroscience and Biobehavioral
Reviews 33, 279-296 (2009).
[76] Zhang D & Raichle ME. Disease and the brains dark energy. Nat. Rev. Neurol. 6, 15-28 (2010).
[77] He Y, Chen Z, Evans A. Structural insights into aberrant topological patterns of large-scale cortical networks in Alzheimers
disease. J. of Neurosc. 28, 4756. (2008).
[78] Garrity AG, et al. Aberrant “Default Mode” functional connectivity in schizophrenia. Am. J. Psychiatry 164, 450 (2007).
[79] Laufs H, et al. Temporal lobe interictal epileptic discharges affect cerebral activity in default mode brain regions. Hum.
Brain. Mapp. 28, 1023 (2007).
[80] Osorio I, et al. Epileptic seizures: Quakes of the brain? Phys. Rev. E. 82, 021919 (2010).
[81] Bienenstock E & Lehmann D. Regulated criticality in the brain? Advances in Complex Systems, 1, 361-384 (1998).
[82] Magnasco MO, Piro O, Cecchi GA. Self-Tuned Critical Anti-Hebbian Networks. Phys. Rev. Lett. 102, 258102 (2009).
[83] Levina A, Herrmann JM, Geisel T. Dynamical synapses causing self-organized criticality in neural networks. Nature Phys
3, 857-860 (2007).
[84] Kinouchi O & Copelli M, Optimal dynamical range of excitable networks at criticality. Nature Physics, 2, 348-352. (2006).
[85] de Arcangelis L & Herrmann J. Learning as a phenomenon occurring in a critical state. Proc. Natl. Acad. Sci. USA, 107,
3977-3981 (2010).
[86] de Franciscis S, Torres JJ, Marro J. Unstable dynamics, nonequilibrium phases and criticality in networked excitable media.
arXiv:1007.4675v2 [cond-mat.dis-nn] (2010).
[87] Honey CJ, Kotter R, Breakspear M, Sporns O. Network structure of cerebral cortex shapes functional connectivity on
multiple time scales. Proc. Natl. Acad. Sci. USA 104, 10240-10245 (2007)
|
1512.04293 | 1 | 1512 | 2015-12-14T13:02:56 | Active subthreshold dendritic conductances shape the local field potential | [
"q-bio.NC"
] | The main contribution to the local field potential (LFP) is thought to stem from synaptic input to neurons and the ensuing subthreshold dendritic processing. The role of active dendritic conductances in shaping the LFP has received little attention, even though such ion channels are known to affect the subthreshold neuron dynamics. Here we used a modeling approach to investigate the effects of subthreshold dendritic conductances on the LFP. Using a biophysically detailed, experimentally constrained model of a cortical pyramidal neuron, we identified conditions under which subthreshold active conductances are a major factor in shaping the LFP. We found that particularly the hyperpolarization-activated inward current, I$_{\rm h}$, can have a sizable effect and cause a resonance in the LFP power spectral density. To get a general, qualitative understanding of how any subthreshold active dendritic conductance and its cellular distribution can affect the LFP, we next performed a systematic study with a simplified model. We found that the effect on the LFP is most pronounced when (1) the synaptic drive to the cell is asymmetrically distributed (i.e., either basal or apical), (2) the active conductances are distributed non-uniformly with the highest channel densities near the synaptic input, and (3) when the LFP is measured at the opposite pole of the cell relative to the synaptic input. In summary, we show that subthreshold active conductances can be strongly reflected in LFP signals, opening up the possibility that the LFP can be used to characterize the properties and cellular distributions of active conductances. | q-bio.NC | q-bio | Title:
Active subthreshold dendritic conductances shape the local field potential
Abbreviated title:
Active dendritic conductances shape the LFP
Author names and affiliation:
Torbjørn V Ness, Department of Mathematical Sciences and Technology, Norwegian Uni-
versity of Life Sciences, 1432 Ås, Norway.
Michiel W H Remme, Institute for Theoretical Biology, Humboldt University Berlin, 10115
Berlin, Germany.
Gaute T Einevoll, Department of Mathematical Sciences and Technology, Norwegian Uni-
versity of Life Sciences, 1432 Ås, Norway; Department of Physics, University of Oslo, 0316
Oslo, Norway.
Corresponding author:
Gaute T Einevoll, Department of Mathematical Sciences and Technology, Norwegian Univer-
sity of Life Sciences, 1432 Ås, Norway; [email protected]
5
1
0
2
c
e
D
4
1
]
.
C
N
o
i
b
-
q
[
1
v
3
9
2
4
0
.
2
1
5
1
:
v
i
X
r
a
1
Active subthreshold dendritic conductances shape the
local field potential
Torbjørn V Ness1, Michiel W H Remme2, and Gaute T Einevoll1, 3∗
1Department of Mathematical Sciences, Norwegian University of Life
Sciences, Ås, Norway
2Institute for Theoretical Biology, Humboldt University Berlin, Berlin,
3Department of Physics, University of Oslo, Oslo, Norway
Germany
Key points summary
• The local field potential (LFP), the low frequency part of extracellular potentials recorded
in neural tissue, is often used for probing neural circuit activity. Interpreting the LFP
signal is difficult, however.
• While the cortical LFP is thought to mainly reflect synaptic inputs onto pyramidal neu-
rons, little is known about the role of the various subthreshold active conductances in
shaping the LFP.
• By means of biophysical modeling we obtain a comprehensive qualitative understanding
of how the LFP generated by a single pyramidal neuron depends on type and spatial
distribution of active subthreshold currents.
• For pyramidal neurons, the h-type channels likely play a key role and can cause a distinct
resonance in the LFP power spectrum.
• Our results show that the LFP signal can give information about the active properties of
neurons and imply that preferred frequencies in the LFP can result from those cellular
properties instead of, e.g., network dynamics.
∗Corresponding author: [email protected]
2
Abstract
The main contribution to the local field potential (LFP) is thought to stem from synap-
tic input to neurons and the ensuing subthreshold dendritic processing. The role of active
dendritic conductances in shaping the LFP has received little attention, even though such
ion channels are known to affect the subthreshold neuron dynamics. Here we used a mod-
eling approach to investigate the effects of subthreshold dendritic conductances on the LFP.
Using a biophysically detailed, experimentally constrained model of a cortical pyramidal
neuron, we identified conditions under which subthreshold active conductances are a major
factor in shaping the LFP. We found that particularly the hyperpolarization-activated inward
current, Ih, can have a sizable effect and cause a resonance in the LFP power spectral den-
sity. To get a general, qualitative understanding of how any subthreshold active dendritic
conductance and its cellular distribution can affect the LFP, we next performed a system-
atic study with a simplified model. We found that the effect on the LFP is most pronounced
when (1) the synaptic drive to the cell is asymmetrically distributed (i.e., either basal or
apical), (2) the active conductances are distributed non-uniformly with the highest channel
densities near the synaptic input, and (3) when the LFP is measured at the opposite pole of
the cell relative to the synaptic input. In summary, we show that subthreshold active con-
ductances can be strongly reflected in LFP signals, opening up the possibility that the LFP
can be used to characterize the properties and cellular distributions of active conductances.
Abbreviations list
LFP, Local Field Potential; PSD, Power Spectral Density
Introduction
The local field potential (LFP), the low-frequency part ((cid:46) 500 Hz) of the extracellular potential
recorded in the brain, has experienced a rejuvenated interest in the last decade. This is partly due
to the development of novel silicon-based microelectrodes with tens or hundreds of electrode
contacts covering large volumes of brain tissue, the realization that the LFP offers a unique
window into neural population activity, and the possibilities offered by the signal in steering
neuroprosthetic devices (Buzsáki et al., 2012; Einevoll et al., 2013a). The interpretation of
the LFP signal in terms of neural activity is challenging, however, as in general thousands of
neurons contribute to the recorded signal (Lindén et al., 2011). Careful mathematical modeling
is thus needed to take full advantage of the opportunities that the LFP signal offers (Einevoll et
al., 2013a).
Fortunately, the biophysical origin of the LFP signals seems well understood in the context
of volume conduction theory (Nunez and Srinivasan, 2006), and a forward-modeling scheme
linking transmembrane neural currents to recorded electrical potentials, including LFPs, has
3
been established (Rall and Shepherd, 1968; Holt and Koch, 1999). The forward-modeling
scheme has been successfully used to study the cortical LFP signal generated by single neu-
rons (Lindén et al., 2010) as well as cell populations (Pettersen et al., 2008; Lindén et al., 2011;
Łeski et al., 2013; Tomsett et al., 2015). These studies showed, for example, that the lower
frequencies of the synaptic input current give the largest contributions to the LFP for two rea-
sons. First, the lowest frequencies give the largest transmembrane current dipole moments due
to the filtering properties of dendrites (Lindén et al., 2010). Second, increased correlations in
synaptic input to a population of neurons can strongly amplify the LFP signal (Lindén et al.,
2011; Einevoll et al., 2013a,b), and this amplification is strongest for the lowest input frequen-
cies (Łeski et al., 2013). Further, the amplification of the LFP due to synaptic input correlations
requires the input to be asymmetrically distributed over the dendrites (Lindén et al., 2011; Łeski
et al., 2013).
However, these studies were typically carried out using neuron models with passive den-
drites. Neurons express a variety of active conductances that are distributed in specific ways
across both soma and dendrites (Stuart et al., 2007). For example, the h-type current is most
densely concentrated in the distal apical dendrites (Williams and Stuart, 2000; Magee, 1998),
whereas the M-type slow potassium current is mostly localized close to the soma (Hu et al.,
2007). It is well established that such active conductances affect the subthreshold processing of
synaptic input (see, e.g., Remme and Rinzel 2011; Major et al. 2013). This is highly relevant
for understanding the LFP signal, since, in cortex, the LFP is thought to mainly reflect synaptic
currents and their associated return currents (Mitzdorf, 1985; Einevoll et al., 2007, 2013a).
Here we use a modeling approach to investigate specifically the role of voltage-dependent
conductances in shaping the LFP in the subthreshold regime. We focused on pyramidal cells
since they are thought to be the main contributor to the LFP in cortical recordings (Einevoll et
al., 2013a) and used the state-of-the-art cortical layer 5 pyramidal neuron model by Hay et al.
(2011) as a starting point for our investigation. With this highly detailed model we established
that subthreshold active conductances can indeed have a sizable impact on the LFP signal. Using
a simplified neuron model that used generalized descriptions of active conductances, so-called
quasi-active currents (Koch, 1984; Remme, 2014), we were able to obtain an understanding of
how any active conductance can affect the LFP, depending on the characteristics of the current
and its distribution across the neuron.
Methods
Pyramidal neuron model
Numerical simulations for Figures 1 and 2 were carried out using a model of a cortical layer 5
pyramidal cell that was published by Hay et al. (2011). This model has a detailed morphology
4
and includes ten active ionic conductances fitted to experimental data by multi-objective opti-
mization with an evolutionary algorithm. For simulations of a passive model, the active conduc-
tances were removed from the model. Simulations with a so-called "frozen" h-type conductance
were performed by keeping the gating variable of the h-current constant, yielding an additional
passive conductance. The original model showed a voltage gradient from soma to distal apical
dendrites. In order to simplify the interpretation of the simulation results, we adjusted the leak
reversal potential of each compartment such that we could set the resting potential uniformly to
a specified chosen potential (Carnevale and Hines, 2006).
The neuron model received either excitatory synaptic input or white-noise current input.
Synaptic inputs were modeled as steps in the synaptic conductance followed by an exponential
decay with a time constant of 2 ms and used a reversal potential of 0 mV. The white-noise
current input consisted of a sum of sinusoids with identical amplitudes but random phases for
each integer frequency from 1 Hz to 500 Hz (see Lindén et al. 2010). The resulting white-noise
signal was scaled to obtain current fluctuations with a standard deviation of 8 pA. Injection
of this input into the distal apical dendrite at a distance of 1094 µm away from the soma of
the active cell held at −60 mV yielded local membrane potential fluctuations with a standard
deviation of 1.5 mV . The exact same white-noise input was used in all simulations (i.e., so-
called "frozen" noise).
Quasi-active approximation of voltage-dependent ion currents
Voltage-dependent membrane currents often behave in a near-linear fashion for small perturba-
tions around a holding potential. This can be exploited by making linear approximations, so-
called "quasi-active" models, of the nonlinear ionic currents (Mauro et al., 1970; Koch, 1984;
Hutcheon and Yarom, 2000; Remme, 2014). In this way one can reduce the parameter space
while retaining key dynamical features of the system. Results in Figures 2 -- 7 used quasi-active
currents to simplify the original, nonlinear cortical pyramidal cell model (see above) and to
allow for a systematic study of the effects of active conductances on the LFP.
We here briefly describe the derivation of a quasi-active description of a single cellular com-
partment. The compartment includes one active current Iw that depends on the membrane poten-
tial V (t) and is described by Iw(t) = ¯gww(t)(V (t)−Ew), with peak conductance ¯gw, reversal po-
tential Ew, and gating variable w(t). The passive leak current is given by IL(t) = gL(V (t)− EL),
with conductance gL and reversal potential EL. Finally, the axial current, i.e., the net current
entering or leaving to neighboring cellular compartments, is denoted by Iaxial(t). The voltage
of the compartment then evolves according to
(1)
dt = − ¯gww(t)(cid:0)V (t)− Ew
dV (t)
cm
(cid:1)− gL
(cid:0)V (t)− EL
(cid:1) + Iaxial(t),
5
where the term cm
of the gating variable w(t) is given by
dt
dV (t)
is the capacitive current with membrane capacitance cm. The dynamics
dt = w∞
with voltage-dependent activation time constant τw
τw
(cid:0)V(cid:1)dw(t)
(cid:0)V(cid:1)− w(t),
(cid:0)V(cid:1) and activation function w∞
(cid:0)V(cid:1).
(2)
The quasi-active description is obtained by linearizing V and w around resting-state values
VR and w∞(VR), respectively, by means of Taylor expansions. Defining the variable m(t) ≡
(w(t)− w∞(VR))/ ∂
∂V w∞(VR), we can write the linearized equation describing the voltage dy-
namics of a single compartment:
(cid:16)
(cid:0)V (t)−VR
(cid:17)
(cid:1) + µm(t)
cm
dV (t)
dt = −gL
γR
+ Iaxial(t),
(3)
where γR ≡ 1+ ¯gww∞(VR)/gL, i.e., the ratio between the total membrane conductance (at VR) and
the leak conductance. The parameter µ ≡ ( ¯gw/gL)(VR − Ew) ∂
∂V w∞(VR) determines whether the
quasi-active current functions as a positive feedback (when µ < 0; i.e., a regenerative current)
amplifying voltage deviations from the holding potential VR, or as a negative feedback (when
µ > 0; i.e., a restorative current) counteracting changes in the voltage. When µ = 0, the quasi-
active current is frozen and functions as a static passive current (throughout the text we will refer
to this as the "passive-frozen" case). The dynamics of the linear gating variable m is described
by,
τw(VR)
dm(t)
dt = V (t)−VR − m(t).
(4)
The description of the ionic currents in a single compartment is easily extended to a multi-
compartmental model where each compartment can have its own set of parameters to describe
the passive and quasi-active currents.
For the simulations with a single linearized h-type current or persistent sodium current (see
Figure 2), we kept the passive parameters, as well as the peak conductance and activation time
constant (at the specified holding potential) of the relevant active current, the same as in the
original detailed model.
To systematically study the effect of the cellular distribution of a quasi-active current on the
LFP (see Figure 3), we used three different channel density distributions: (1) linearly increas-
ing with distance from the soma, (2) linearly decreasing with distance from the soma, and (3) a
uniform distribution. The slopes of the increasing (decreasing) distributions were set such that
the most distal tip of the apical dendrite had a sixty-fold larger (smaller) density compared to
that of the soma (in line with experimental estimates for Ih distributions: Mishra and Narayanan
2015; Lörincz et al. 2002; Nusser 2009; Kole et al. 2006), and the total membrane conduc-
tance of the quasi-active current (i.e., summed over all compartments) was the same as the total
6
passive leak conductance gL. The passive leak conductance was set uniformly to 50 µS/cm2
for all cases. For w∞(VR) = 0.5 the distance-dependent quasi-active peak conductance was
¯gw(x) = 5.29 + 0.242x µS/cm2 for the linear increase, and ¯gw(x) = 143− 0.109x µS/cm2 for
the linear decrease, where the distance x was measured in µm and had a maximum of 1291 µm.
Note that the three distributions had the same total quasi-active membrane conductance summed
over the neuronal membrane.
The parameters µ and γR vary along the cell for the non-uniform channel distributions.
We introduced µ∗ ≡ µ(x)gL/ ¯gw(x), such that µ∗ = (VR − Ew) ∂
∂V w∞(VR) is independent of the
distribution of the quasi-active conductance and can be specified as a single constant. We used
µ∗ = −0.5 for the regenerative conductance, µ∗ = 0 for the passive-frozen conductance, and
µ∗ = 2 for the restorative conductance. For the uniform distribution this gives the same values
as those used in Remme and Rinzel (2011), namely µ = −1, 0, 4 for the regenerative, passive-
frozen and restorative cases, respectively. The activation time constant τw(VR) of the quasi-
active conductance was set to 50 ms (unless specified otherwise) in order to have dynamics
similar to the h-type conductance. The intracellular resistivity was Ra = 100 Ωcm, and the
specific membrane capacitance cm =1 µF/cm2.
Calculation of extracellular potentials
Extracellular potentials recorded inside the brain are generated by transmembrane currents from
cells in the vicinity of the electrode contact (Nunez and Srinivasan, 2006). The biophysical
origin of the recorded signals is well understood in the context of volume conduction theory.
Extracellular potentials originating from a simulated multi-compartmental neuron model can be
computed by first obtaining the transmembrane currents In(t) from each compartment n at posi-
tion(cid:126)rn. Next, the extracellular potential φ ((cid:126)r,t) at position(cid:126)r resulting from these transmembrane
currents can be calculated (Holt and Koch, 1999; Lindén et al., 2014):
(cid:90)
φ ((cid:126)r,t) =
1
4πσ
N
∑
n=1
In(t)
d(cid:126)rn
(cid:126)r−(cid:126)rn,
(5)
where σ is the conductivity of the extracellular medium. This corresponds to the so-called line-
source formula assuming the transmembrane currents to be evenly distributed along the axes of
cylindrical neural compartments, see Lindén et al. (2014) for a detailed description.
All simulations and computations of the extracellular potentials were carried out using LFPy
(Lindén et al., 2014), an open-source Python package that provides an interface to NEURON
(Carnevale and Hines, 2006). The time step of the neural simulation was 0.0625 ms. For all
simulations the first 1000 ms was discarded to avoid initialization effects. All simulation code
used to produce the figures in this study are available upon request.
7
Results
To examine the contribution of subthreshold, active conductances on the extracellular potential
we started by considering a previously published, detailed model of a layer 5 cortical pyramidal
neuron endowed with a large variety of active conductances (Hay et al., 2011; see Methods).
Note that the term LFP commonly refers to the low-pass filtered version (below 300-500 Hz)
of the extracellular potential (to filter away spikes). Since there is no spiking activity in the
present use of the model, we simply refer here to the unfiltered version of the extracellular
potential as the LFP.
Active subthreshold conductances shape the LFP
We first determined the LFP in response to a single excitatory synaptic input. We compared
the LFP signal of the active neuron model with a passive version of the model (i.e., all active
conductances removed). The synaptic input was provided either to the distal apical dendrite
(Figure 1A, top row) or to the soma (Figure 1A, bottom row) while the entire cell was hyper-
polarized to −80 mV (Figure 1A, left column) or depolarized to −60 mV (Figure 1A, right
column). The LFP was calculated at five different positions outside the neuron (marked by
cyan dots). The difference between the LFPs of the active and passive models depended on the
holding potential of the cell, the synaptic input position, as well as the extracellular recording
position. In particular, we observed for hyperpolarized potentials that the active and passive
LFPs were markedly different for apical synaptic input, while being very similar for somatic
input.
In order to more fully characterize the effects of active currents on the LFP we next applied
white-noise current input instead of a single synaptic input, and determined the power spectral
density (PSD) of the resulting LFP (i.e., the square of the Fourier amplitude of the signal as
a function of frequency). The white-noise input current had equal signal power at all integer
frequencies in the considered frequency range (1 -- 500 Hz; see also Lindén et al., 2010). At most
extracellular recording positions, a low-pass filtering was observed in the LFP of both the active
and the passive model. This results from the intrinsic dendritic filtering (see, e.g., Pettersen et
al., 2012; Lindén et al., 2010; Pettersen and Einevoll, 2008) and increases with distance from
the synaptic input site (Figure 1B). The most striking difference between the active and the
passive case was again found for the cell in a hyperpolarized state with apical input (Figure 1B,
top left), which showed a strong resonance (i.e., band-pass filter with a peak around 17 -- 22 Hz)
for the active model LFP, but not for the passive one. Note that the cell in a depolarized state and
apical input shows qualitatively the same effect, though less prominent (Figure 1B, top right).
While the main effect of the active conductances in the case of apical input was to reduce
the LFP power at the lowest frequencies, the effect from active conductances on somatic input
8
was qualitatively different. For the cell with a depolarized resting potential, white-noise current
resulted in an amplification of the low frequencies (Figure 1B, bottom right). In all situations
considered, a difference between the active and passive model was only apparent below about
30 -- 50 Hz. At higher frequencies, capacitive currents dominated the transmembrane return
currents both in the active and passive models (Lindén et al., 2010). The exact cross-over
frequency where the capacitive currents become dominant over the ionic membrane currents,
depends on neuronal properties (e.g., ion channel densities), as well as the positions of the input
and the recording electrode.
Which ionic currents in the model are responsible for the large difference between the active
and passive model for hyperpolarized potentials and with apical input? One obvious candidate
is the hyperpolarization-activated inward current Ih. This current operates at subthreshold volt-
ages, is concentrated in the apical dendrites, and dampens the lowest frequency components of
the membrane potential response to input (Magee, 1998; Hu et al., 2009, 2002; Kole et al., 2006;
Almog and Korngreen, 2014). Indeed, we found that when we extended the passive model with
only the h-type conductance, the LFP was indistinguishable from the full active model when
the cell was in a hyperpolarized state and received apical input (Figure 1C).
Two effects are involved in the reduction of the LFP signal power at low frequencies in the
model with Ih. First, with an additional active current comes an increased membrane conduc-
tance, which shifts the transmembrane return currents associated with the input current closer to
the input position on the neuron. This gives smaller current-dipole moments and thus generally
smaller LFPs (Pettersen and Einevoll, 2008; Lindén et al., 2010; Pettersen et al., 2012). This is
still a passive-membrane effect and can simply be incorporated into a passive model by rescal-
ing the leak conductance corresponding to adding a "frozen" h-type conductance. The second
effect stems from the dynamical properties of an active conductance such as Ih, which actively
counteract voltage fluctuations at low frequencies. While the extension of the passive model
with a "frozen Ih" was found to strongly dampen the low-frequency components of the LFP
compared to the full active case (Figure 1C, gray dotted lines), it failed to reproduce the reso-
nance of the models with the dynamic Ih conductance (blue dashed and red). This demonstrates
that the resonance observed in the LFP was an effect caused by the dynamics of the h-type chan-
nels, similar to previously reported Ih-induced resonances in the membrane potential (Hutcheon
and Yarom, 2000; Hu et al., 2009; Zhuchkova et al., 2013).
Linear models capture the effects of active conductances on the LFP in the
subthreshold regime
Above, we established that the h-type conductance is key in shaping the LFP for subthresh-
old input to the apical dendrites in the model by Hay et al. (2011). Because in reality there is
considerable variation between neurons regarding their biophysical properties, and perhaps not
9
all relevant subthreshold currents were captured by this specific model, we next set out to per-
form a systematic study of the effects that any subthreshold active conductance can have on the
LFP. For this we made use of the so-called "quasi-active" approximation of voltage-dependent
currents. Active membrane conductances in general exhibit nonlinear behavior, but for small
deviations of the membrane potential around a holding potential, active conductances often be-
have in a near-linear fashion. The dynamics of an active conductance can then be approximated
by linearizing the current around a holding potential (Mauro et al., 1970; Koch, 1984; Remme
and Rinzel, 2011; Remme, 2014; see Methods). This retains the voltage-dependent current dy-
namics for potentials not too far away from that potential, while strongly reducing the parameter
space of the model and allowing an intuitive understanding of the model behavior.
Linearized active conductances can be divided into two classes: restorative and regenera-
tive. Restorative conductances dampen the lowest frequency components of the membrane po-
tential response to synaptic input, also leading to narrower synaptic voltage responses (Remme
and Rinzel, 2011). The intrinsic low-pass filtering of the cellular membrane (Koch, 1999), in
combination with the active dampening of low frequencies (i.e., high-pass filtering) by restora-
tive conductances, can thus cause a resonance in the membrane potential PSD (Remme and
Rinzel, 2011; Zhuchkova et al., 2013; Hu et al., 2009). Examples of restorative currents are
the h-current and the M-type slow potassium current. In contrast, regenerative conductances
amplify the low-frequency components of the membrane potential response to synaptic input,
making the voltage responses to synaptic input broader in time as they propagate along the den-
drites. This will only add to the intrinsic low-pass filtering of the cellular membrane, and as
such, regenerative conductances lack the easily recognizable feature of the restorative conduc-
tances. Examples of regenerative conductances are the persistent sodium current INaP and the
low-voltage activated calcium current (Remme and Rinzel, 2011).
It is well established that quasi-active approximations of active currents can accurately cap-
ture the subthreshold effects on the membrane potential (Koch, 1984; Remme and Rinzel, 2011).
To test whether this conclusion also holds for the LFP, we started with the original cortical
pyramidal cell model from Hay et al. (2011), and simplified it to two separate cases with only
a single active conductance remaining in each. For the first case, we kept only the restorative
Ih conductance, with its increasing conductance along the apical dendrite away from the soma.
The resting potential was set uniformly to the hyperpolarized potential of −80 mV and the cell
was stimulated with apical white-noise input. In the second case we kept only the regenerative
persistent sodium current, solely present in the soma, set the resting potential uniformly to the
depolarized potential of −60 mV, and applied somatic white-noise input. The Ih-only model
yielded the resonance in the LFP that was discussed above (Figure 2A, blue dashed lines),
while the INaP-only model gave an amplification of the low-frequency components of the LFP
(Figure 2B, pink dashed lines). We then simulated the models with linearized versions of the
original conductances and found the LFP from the active and linearized conductances to be
10
indistinguishable (Figure 2, blue versus green, pink versus orange). In fact, for all tested com-
binations of resting potentials, input positions, and LFP recording positions, the quasi-active
approximation was found to faithfully reproduce the signature of the active conductance in the
membrane potentials, the transmembrane currents, as well as in the LFP. This applied as long as
the perturbations of the membrane potential around the resting potential were not too large. For
Ih, fluctuations of 10 -- 15 mV are typically within the appropriate range (data not shown, but see,
e.g., Remme and Rinzel, 2011). We conclude that systematic studies with a single quasi-active
conductance are justified.
Spatially asymmetric distributions of ion channels and inputs enhance the
effect of active currents on the LFP
We next turned to a systematic exploration of how active subthreshold currents affect the LFP.
For this we used a pyramidal cell model with a single quasi-active current. The key parameter
characterizing the quasi-active current is µ∗, the sign of which determines whether the current
is regenerative (µ∗ < 0) or restorative (µ∗ > 0), while µ∗ = 0 for a passive-frozen conductance
(i.e., the added conductance has no dynamics and thus simply adds to the passive leak con-
ductance). The quasi-active current was distributed across the cell either uniformly, linearly
increasing with distance from soma, or linearly decreasing with distance from soma (Figure 3
-- left, middle and right columns, respectively). In all cases the models had the same total mem-
brane conductance summed over the neuronal membrane. We applied white-noise current input
either to the apical dendrite (panel B) or to the soma (panel C) and calculated the LFP-PSD for
each case in the apical region and in the somatic region. A large difference between the results
from the passive-frozen model and the restorative or regenerative models would imply that the
dynamics of the active conductance in question shapes the LFP. Indeed, we found for most
scenarios, that the regenerative current caused an amplification of the low-frequency compo-
nents of the LFP (Figure 3, red lines) compared to the passive-frozen case (black line), while a
restorative current dampened the low frequencies, leading to the previously discussed resonance
(Figure 3, blue lines). However, as discussed below, there were exceptions.
We found that the impact of the quasi-active current was always largest in the LFP recorded
on the opposite side of the cell compared to the position of the input, i.e., next to the soma for
apical input (panel B, bottom row) and next to the apical dendrite for somatic input (panel C,
top row). This can be understood since the LFP is a distance-weighted sum of all transmem-
brane currents, so that transmembrane currents close to the recording electrode will contribute
more to the measured LFP signal. For recordings on the opposite side relative to the input,
the transmembrane return currents have been strongly affected by the active conductance after
propagating along the apical dendrite. The effect was also largest when the quasi-active current
was non-uniformly distributed and concentrated at the position of the input (panel B, middle
11
column; panel C, right column), and smallest when the input was to a region of low density
(panel B, right column; panel C, middle column). This also follows from the reasoning above:
when the input is to a region with a low density of an active current, the return currents close
to the input will be hardly affected by the active current. As the signal propagates along the
dendrites to regions with higher densities of the active current, it will be increasingly affected
by the active current, however, the signal amplitude will also gradually decrease. Therefore, the
LFP is less affected by the presence of a non-uniformly distributed active current when the input
is provided to a region with low densities compared to when input is provided to a high-density
region.
Besides the type of an active current (i.e., restorative or regenerative) and how it is dis-
tributed across the neuron, also its activation time constant is key for the effects on the LFP. The
dynamics of an active current will only express its effect on the membrane potential, transmem-
brane currents and LFP for frequencies that are sufficiently low relative to the activation time
constant (τw(VR)). This is seen by considering different activation time constants for the quasi-
active conductance (Figure 4). Increasing the activation time constant decreased the frequency
at which the regenerative/restorative responses became similar to the passive-frozen case. For
the resonance this means that its peak becomes broader and shifts to lower frequencies for larger
activation time constants.
Resonance is expressed for asymmetrically distributed synaptic input
In the above simulations we considered a single white-noise current input to characterize the
effect of active currents on the LFP. We next examined whether the observed effects also hold
when the cell receives a barrage of synaptic input as is typical for in vivo conditions (Destexhe
et al., 2003).
We considered a cell model with a single quasi-active conductance that increases linearly in
density with distance from the soma. The model received input from 1000 excitatory synapses
that were distributed across the neuronal membrane and randomly activated according to inde-
pendent Poisson statistics. We considered three synaptic distributions: (1) uniformly distributed
across the distal apical tuft dendrites located more than 900 µm from the soma (Figure 5A),
(2) uniformly distributed across the apical dendrites more than 600 µm from the soma (Fig-
ure 5B), and (3) uniformly across the entire cell (Figure 5C). For each distribution the exact
same 1000 synapse positions and input spike trains were used in the simulations for the regen-
erative, passive-frozen, and restorative conductances. Hence, the only difference between the
simulations with a given synapse distribution was the nature of the quasi-active conductance.
The resonance for the restorative model was clearly retained for LFP signals recorded close
to the soma when the synaptic inputs were distributed asymmetrically, being concentrated in
the apical dendrites (Figure 5A,B). However, for uniformly distributed input (Figure 5C) the
12
resonance was practically undetectable. Similarly, for the regenerative model, increases of
low frequencies in the LFP-PSD were only reliably observed for the models with asymmetric
synaptic input distributions (Figure 5A,B). Note that we obtained the same results when keeping
the synapse densities fixed across the three synapse distributions, instead of keeping the total
synapse number constant (results not shown). The dependence of the modulation of the LFP
on the input distribution can be understood by considering the results in Figure 3: the case with
uniform input approximately corresponds to a linear combination of the LFPs stemming from
apical and somatic input (Figure 3, middle column). In the somatic region the net LFP signal
will be strongly dominated by the transmembrane currents resulting from the somatic input,
which are hardly affected by the active conductances since these are concentrated in the distal
apical dendrites.
Resonance is a spatially stable feature in the LFP
The LFP signals measured in in vivo experiments are generated by many neurons located within
a distance from a few hundred micrometers up to a few millimeters away from the recording
electrode (Lindén et al., 2011; Łeski et al., 2013). An effect of a quasi-active conductance on
the LFP generated by a single neuron will only carry over to the LFP generated by an entire
population if the effect is spatially "robust". For example, the resonance of the single-neuron
LFP will only be observed in the population-LFP if it is present in a sizable part of the volume
surrounding the cell. If it would only be recorded at special electrode positions, then it would
be expected to average out when considering an entire population. The linearity of the LFP
signal generation implies that the summed contributions of many single neuron-LFP signals
with similar shape, would result in a population LFP with a similar shape. It is therefore key to
study how the LFP varies in the space around the neuron.
We calculated the LFP at a dense grid of positions around a pyramidal cell receiving apical
white-noise input (Figure 6). The model expressed a restorative quasi-active conductance that
increased in density with distance from the soma. Our previous results (see Figure 3) demon-
strated that such a model can show strong resonances in the LFP. The maximum amplitude
of the LFP-PSD on the position-grid around the cell had the shape of a dipole (panel A), as
expected following the requirement of charge conservation (Lindén et al., 2010). This dipolar
nature of the LFP pattern was also demonstrated by comparing two normalized LFP time traces
at opposite sides of the zero-crossing region of the dipole and noting that they are almost mirror
images (panel B; depicted traces correspond to first and third electrode contacts from the top in
the first column in panel A).
For two columns of electrode positions (marked in panel A) we examined the LFP-PSD in
detail (panel C). The LFP signals again demonstrated that the resonance was most pronounced
on the side of the cell opposite to the input (compare lightest trace to the darker ones). Note that
13
the electrodes positioned in the zero-crossing region of the dipole exhibited more variable PSD
shapes, for example, a strong band-stop at around 20 Hz. This is because different frequency
components of the LFP will have slightly different zero-crossing regions (see also Lindén et
al. 2010). We focused on the LFP resonance and determined the frequency of maximum LFP
power for the grid of positions around the neuron. The peak frequency was found to be stable
at around 20 Hz, except in the zero-crossing region of the dipole (panel D).
The so-called Q-value is commonly used to characterize resonances. In the neurophysi-
ological literature it is often defined as the magnitude of a variable (typically the membrane
potential) at the resonance frequency divided by its value at the lowest frequency considered
(Hutcheon et al., 1996). We found that the LFP Q-value was stable and large at the soma region
of the cell (panel E), i.e., on the side opposite to the input, in accordance with observations
from Figures 1 -- 3. Note that the largest LFP Q-values were in fact observed in the zero-crossing
region. However, in this region it does not reflect a resonance caused by a restorative current,
but rather the very variable LFP pattern in this region. The lower signal power in this region
(Figure 6A), in combination with the fact that the position of the zero-crossing region will de-
pend on various parameters, such as the exact position of the synaptic input, means that the very
large Q-values seen here will not carry over to the population LFP.
Spatial profile of the membrane potentials, transmembrane currents and
LFP
Above we demonstrated an amplification and a dampening of low-frequency LFP components
caused by regenerative and restorative conductances, respectively. These effects were qualita-
tively similar to the effects of such active conductances on the membrane potential as reported
by, for example, Remme and Rinzel (2011). However, as LFP signals reflect transmembrane
currents, it is a fundamentally different measure of neural activity than the membrane poten-
tial (Pettersen et al., 2012; Einevoll et al., 2013a). It can therefore not be expected a priori that
quasi-active conductances always have qualitatively similar effects on the LFP and the mem-
brane potential. We thus next studied the differences in the effects of active currents on the
membrane potential, transmembrane currents and LFP in a simplified setting, i.e., white-noise
current input injected into a single, long neurite with uniform passive membrane parameters
and a single uniformly distributed quasi-active conductance.
Focusing first on the membrane potential (Vm, Figure 7A), we indeed observe that the
low-frequency components are always amplified by regenerative currents (red), whereas they
are always dampened by restorative conductances (blue), compared to the passive-frozen case
(black). The effects increase with "exposure" to the quasi-active current, i.e., how far the input
has propagated along the neurite (Remme and Rinzel, 2011). In addition, the membrane poten-
tial is increasingly low-pass filtered with distance, resulting in a stronger membrane-potential
14
resonance with distance from the input site.
The position dependence of transmembrane currents is more complicated (Im, panel B). We
first consider the passive-frozen case: at the input site (panel B, leftmost) the transmembrane
current is approximately equal for all frequencies, as expected due to the white-noise current
input. From charge conservation it follows that the current that enters the cell must also exit the
cell, i.e., that transmembrane currents across the neuron sum to zero (Koch, 1999). Since high-
frequency components propagate less than low-frequency components, they dominate close to
the input. Hence, the transmembrane currents show a high-pass filtered PSD at positions close
to the white-noise input (panel B, second column). The reverse is true far away from the input
(panel B, rightmost), where there is a low-pass effect similar as for the membrane potential.
This results from the stronger propagation of the low frequencies, which therefore dominate
over the high frequencies at larger distances (Koch, 1999; Lindén et al., 2010). Note that at
intermediate distances, a "passive resonance" (panel B, third column) can be observed through
the combination of a low-pass and a high-pass filter.
A restorative current functions as a negative feedback, counteracting membrane potential
changes, such that voltage signals do not spread as far down the neurite as in the passive-
frozen case (panel A; compare blue and black traces). Therefore the return currents resulting
from the current injections move closer to the input site, leading to an amplification of the
transmembrane currents close to the input site, and a decrease far away from the input site,
relative to the passive-frozen case (panel B).
Regenerative currents amplify membrane potential deviations along the neurite (panel A;
compare red and black traces) and thus shift the return currents further away from the input
compared to the passive-frozen case. This leads to a decrease in the transmembrane currents
close to the input site, and an amplification far away from the input site compared to the passive-
frozen case (panel B).
Because the LFP signal is a distance-weighted sum of all transmembrane currents, the de-
scribed effects in the transmembrane currents are all present in the LFP measured very close to
the neurite (panel C; 20 µm). For electrodes further away from the neurite (panel D; 300 µm)
the LFP is no longer dominated by transmembrane currents from any single part of the neurite.
Since the restorative currents move the transmembrane return currents closer to the input site,
a smaller current dipole moment is generated, thus resulting in a decrease of the LFP-power at
most recording positions. The opposite is found for the model with a regenerative conductance.
As we demonstrated in Figure 4, these effects hold for the low-frequency components of the
LFP, because the dynamics of the active current only affect the cell response for frequencies
that are slow compared to the activation time constant.
The resulting transmembrane currents from a single white-noise input into a neurite also
helps explain why we needed asymmetric input to retain the effects that an active current has
on the LFP-PSD (see Figure 5). For a single input location, the effects of an active current on
15
the low-frequency components of the transmembrane currents close to, and far away from the
white-noise input were opposite (Figure 7B, second versus fourth panel). Hence, for input that is
localized to one region of a cell, the contributions from active currents will sum "constructively"
and largely retain the characteristic LFP signatures for a single individual synaptic input. For
uniform input, however, an active current will give a combination of amplifying and dampening
effects which will sum "destructively", decreasing the total effect of the active current on the
shape of the LFP.
Discussion
In the present paper we explored the role played by active conductances in shaping the local
field potential (LFP) generated by neurons. It thus goes beyond previous principled studies on
the LFP generated by synaptically activated passive neurons (Lindén et al., 2010, 2011; Łeski et
al., 2013). While the various effects of active conductances in forming the LFP have previously
been explored in comprehensive network simulations of thousands of spiking neurons (Reimann
et al., 2013), the present work was instead tailored to probe subthreshold effects in detail.
With a morphologically detailed and experimentally constrained cortical pyramidal cell
model, we demonstrated that active conductances can strongly shape the LFP stemming from
subthreshold input. The precise effects depended on the position of the input, the position of the
extracellular electrode, and the membrane potential of the cell (Figure 1). The subthreshold ac-
tive conductances had distinct effects on the power spectral density (PSD) of the LFP, providing
either amplification or attenuation of the low frequencies, in the latter case leading to a reso-
nance in the LFP-PSD. The effects on the LFP were observed for various spatial distributions
of the active conductances across the cell, but generally required an asymmetric distribution of
synaptic input onto the cell (Figure 5). We found that the effect of a subthreshold active conduc-
tance on the LFP was maximized for (i) an asymmetric distribution of the active conductance,
(ii) when the input was targeted to regions where the active conductance was most strongly ex-
pressed, and (iii) the LFP was recorded on the opposite side of the cell with respect to the input
(Figure 3).
Peaks in the LFP-PSD are commonly observed experimentally (see, e.g., Roberts et al.,
2013; Hadjipapas et al., 2015), and such peaks are usually interpreted as the result of network
oscillations, i.e., oscillatory firing activity driving the LFP-generating neurons (Buzsáki and
Draguhn, 2004). Importantly, our work shows that such peaks may also be due to subthreshold
restorative conductances molding the transmembrane return currents. In particular, in our sim-
ulations with the pyramidal cell model by Hay et al. (2011) we found that the h-type current
had a prominent role in shaping the LFP and could cause a strong resonance in the LFP-PSD
(Figures 1C, 2A). The h-type current is strongly expressed in cortical and hippocampal pyrami-
dal neurons and has a particularly asymmetric distribution across the cell, increasing in density
16
along the apical dendrites and peaking in the distal apical tuft dendrites (Magee, 1998; Williams
and Stuart, 2000; Harnett et al., 2015). The h-type current can be expected to impact the LFP
resulting from synaptic input that is predominantly targeting the apical dendritic tuft of popula-
tions of cortical or hippocampal pyramidal neurons. Indeed, various input pathways to pyrami-
dal neurons target specific domains of the cell (see, e.g., Petreanu et al., 2009). An intriguing
consequence of our work is that the LFP can contain information on the spatial distribution of
subthreshold active channels, as well as on the location of synaptic input to the LFP generating
cells. For example, given a known apical concentration of h-type conductances, a resonance in
the LFP due to Ih would suggest that the cells receive asymmetric input. Vice versa, if asym-
metrical input is provided, an absence of a resonance in the LFP (measured at multiple locations
along the cell axis) suggests that the cell does not strongly express restorative currents.
A direct approach to test our model predictions would be through in vitro cortical slice
experiments. The asymmetric input can be provided by targeting the apical dendrites using, e.g.,
glutamate uncaging. The polarized distribution of Ih that pyramidal neurons typically display,
should then give rise to a notable resonance in the LFP-PSD, in particular when measured close
to the pyramidal cell soma (Figure 6). Addition of an Ih blocker (e.g., ZD7288) is expected to
remove any resonance caused by Ih.
In most of our simulations we described voltage-dependent channels with so-called quasi-
active currents, which are linear approximations of the voltage-dependent currents (Mauro et al.,
1970; Koch, 1984; Hutcheon and Yarom, 2000; Remme, 2014). The linear descriptions high-
lights that there are basically two types of active currents: regenerative and restorative, which
have opposite effects on the voltage response, amplifying or counteracting voltage deflections,
respectively. The quasi-active description not only permits the use of various analytical tech-
niques for linear systems, but also strongly reduces the number of parameters to describe an
active current. This simplified description enables the systematic study of the effect of sub-
threshold input on the voltage response of active, dendritic neurons and allows a thorough un-
derstanding of the dynamics of the system (Koch, 1984; Bressloff, 1999; Coombes et al., 2007;
Goldberg et al., 2007; Remme et al., 2009, 2010; Remme and Rinzel, 2011). We showed here
that the quasi-active current descriptions can also be used to accurately capture the LFP signal
(Figure 2). Because the h-type current was dominating the subthreshold effects on the LFP in
the model by Hay et al. (2011), a single quasi-active current was sufficient to approximate the
LFP signal of the full, nonlinear model. The approach enabled us to establish an overview of
the effects of active currents on the LFP signal. We achieved this by using the detailed neuron
morphology with a single quasi-active current, where the current was either restorative or re-
generative, had a range of activation time constants (Figure 4), and was distributed in various
ways across the cell (Figure 3).
In the in vivo situation, the measured LFP will reflect the activity in populations of neu-
rons (Einevoll et al., 2013a). A natural extension of the present work will be to investigate
17
the effect of active dendritic currents on the LFP in a population of neurons receiving synaptic
inputs (Lindén et al., 2011; Łeski et al., 2013). For the case of passive dendrites the key factor
in determining the magnitude and spread of the population LFP was found to be the amount of
correlation between the numerous synaptic inputs driving the population (Lindén et al., 2011;
Einevoll et al., 2013a; Łeski et al., 2013). The synaptic correlation level will be expected to
also be key for determining the population LFP in the case of active dendritic conductances,
but insight into its specific influence will require detailed modeling studies. A key simplify-
ing feature of the present study was the observation that linearized quasi-active conductances
accounted very well for the salient effects of active subthreshold conductances on the single-
neuron LFP. This approximation will also be applicable when studying effects of subthreshold
active conductances on population LFPs, thus assuring linearity of the model system and greatly
simplifying the analysis as in Lindén et al. (2011) and Łeski et al. (2013).
In the present study we have not considered suprathreshold active conductances, i.e., the
active channels underlying spike generation. While the present biophysical modeling scheme
is equally applicable to spikes (Holt and Koch, 1999; Gold et al., 2006; Pettersen and Einevoll,
2008), the contributions to the LFP from such spikes may be expected to be negligible for cor-
tical networks in the in vivo situation, at least for the low frequencies of LFP. In Pettersen et al.
(2008) the extracellular potential generated by a synaptically-activated population of 1040 pyra-
midal neurons mimicking a layer-5 population in rat somatosensory (barrel) cortex was modeled
by the present biophysical scheme. In order for the model predictions to be in accordance with
measured extracellular potentials from the rat barrel cortex, both for the LFP (frequencies (cid:46)
500 Hz) and for multi-unit activity (frequencies (cid:38) 500 Hz), only 4% (40 of 1040) of the model
neurons were tuned to fire an action potential following the stimulus input. In this situation the
modeled LFP signal was observed to be completely dominated by the synaptic input currents
and their associated return currents, with negligible contributions from the spikes themselves.
However, in situations with strongly correlated firing, for example during sharp-wave ripples
in hippocampus, the spikes have been found to have sizable contributions to the extracellular
potentials for frequencies down to 100 Hz (Schomburg et al., 2012; see also Reimann et al.,
2013; Taxidis et al., 2015). For further discussion on the biophysical origin of the extracellular
potentials in this "high-gamma" range, see for example Ray and Maunsell (2011) and Scheffer-
Teixeira et al. (2013). Note also that as the extracellular potential is given as a linear sum of
contributions from the various transmembrane currents surrounding the contact, putative con-
tributions from spikes can be estimated by adding computed spike signatures (Holt and Koch,
1999; Gold et al., 2006; Pettersen and Einevoll, 2008) to the present results for the LFP.
Even though the LFP has been measured for more than half a century, the interpretation of
the recorded data has so far largely been qualitative (for a review, see Einevoll et al., 2013a).
We believe that the present investigation on the role of active dendritic conductances in shaping
the recorded signal, will turn out to be an important contribution towards the goal of making
18
combined modeling and measurement of the LFP signal a practical research tool for detailed
probing of neural circuit activity.
Competing Interests
The authors declare no competing financial interests.
Funding
The research leading to these results has received funding from the European Union Seventh
Framework Program (FP7/2007-2013) under grant agreement 604102 (Human Brain Project,
HBP), the Research Council of Norway (NevroNor, Notur, nn4661k), the Norwegian node of
the International Neuroinformatics Coordinating Facility (INCF, NFR 214842/H10), and the
Einstein Foundation Berlin.
Author contributions
All authors took part in designing the work, interpreting data, and writing the manuscript. TVN
wrote the simulation code. TVN and MWHR made the figures. All authors approved the final
version of the manuscript and agree to be accountable for all aspects of the work in ensuring
that questions related to the accuracy or integrity of any part of the work are appropriately
investigated and resolved. All persons designated as authors qualify for authorship, and all
those who qualify for authorship are listed.
19
References
Almog M & Korngreen A (2014). A quantitative description of dendritic conductances and its
application to dendritic excitation in layer 5 pyramidal neurons. J Neurosci 34, 182 -- 96.
Bressloff PC (1999). Resonantlike synchronization and bursting in a model of pulse-coupled
neurons with active dendrites. J Comput Neurosci 6, 237 -- 249.
Buzsáki G, Anastassiou CA & Koch C (2012). The origin of extracellular fields and currents --
EEG, ECoG, LFP and spikes. Nat Rev Neurosci 13, 407 -- 20.
Buzsáki G & Draguhn A (2004). Neuronal Oscillations in Cortical Networks. Science 304,
1926 -- 1930.
Carnevale NT & Hines ML (2006). The NEURON Book. Cambridge University Press, Cam-
bridge, UK.
Coombes S, Timofeeva Y, Svensson CM, Lord GJ, Josi´c K, Cox SJ & Colbert CM (2007).
Branching dendrites with resonant membrane: a "sum-over-trips" approach. Biol Cybern 97,
137 -- 149.
Destexhe A, Rudolph M & Paré D (2003). The high-conductance state of neocortical neurons
in vivo. Nat Rev Neurosci 4, 739 -- 751.
Einevoll GT, Kayser C, Logothetis N & Panzeri S (2013a). Modelling and analysis of local
field potentials for studying the function of cortical circuits. Nat Rev Neurosci 14, 770 -- 785.
Einevoll GT, Lindén H, Tetzlaff T, Łeski S & Pettersen KH (2013b). Local field potentials -
biophysical origin and analysis. In Quiroga RQ & Panzeri S, editors, Principles of Neural
Coding, chapter 3, pp. 37 -- 60. CRC Press, Boca Raton, FL.
Einevoll GT, Pettersen KH, Devor A, Ulbert I, Halgren E & Dale AM (2007). Laminar pop-
ulation analysis: estimating firing rates and evoked synaptic activity from multielectrode
recordings in rat barrel cortex. J Neurophysiol 97, 2174 -- 2190.
Gold C, Henze Da, Koch C & Buzsáki G (2006). On the origin of the extracellular action
potential waveform: A modeling study. J of Neurophysiol 95, 3113 -- 28.
Goldberg JA, Deister CA & Wilson CJ (2007). Response properties and synchronization of
rhythmically firing dendritic neurons. J Neurophysiol 97, 208 -- 219.
Hadjipapas A, Lowet E, Roberts M, Peter A & De Weerd P (2015). Parametric variation of
gamma frequency and power with luminance contrast: A comparative study of human MEG
and monkey LFP and spike responses. NeuroImage 112, 327 -- 340.
20
Harnett MT, Magee JC & Williams SR (2015). Distribution and Function of HCN Channels in
the Apical Dendritic Tuft of Neocortical Pyramidal Neurons. J Neurosci 35, 1024 -- 1037.
Hay E, Hill S, Schürmann F, Markram H & Segev I (2011). Models of neocortical layer 5b
pyramidal cells capturing a wide range of dendritic and perisomatic active properties. PLoS
Comp Biol 7, 1 -- 18.
Holt GR & Koch C (1999). Electrical interactions via the extracellular potential near cell bodies.
J Comput Neurosci 6, 169 -- 184.
Hu H, Vervaeke K & Storm JF (2002). Two forms of electrical resonance at theta frequencies,
generated by M-current, h-current and persistent Na+ current in rat hippocampal pyramidal
cells. J Physiol 545, 783 -- 805.
Hu H, Vervaeke K, Graham LJ & Storm JF (2009). Complementary theta resonance filtering by
two spatially segregated mechanisms in CA1 hippocampal pyramidal neurons. J Neurosci 29,
14472 -- 83.
Hu H, Vervaeke K & Storm JF (2007). M-channels (Kv7/KCNQ channels) that regulate synap-
tic integration, excitability, and spike pattern of CA1 pyramidal cells are located in the peri-
somatic region. J Neurosci 27, 1853 -- 1867.
Hutcheon B, Miura RM & Puil E (1996). Models of subthreshold membrane resonance in
neocortical neurons. J Neurophysiol 76, 698 -- 714.
Hutcheon B & Yarom Y (2000). Resonance, oscillation and the intrinsic frequency preferences
of neurons. Trends Neurosci 23, 216 -- 22.
Koch C (1984). Cable Theory in Neurons with Active, Linearized Membranes. Biol Cybern 33,
15 -- 33.
Koch C (1999). Biophysics of Computation Oxford Univ Press, Oxford.
Kole MHP, Hallermann S & Stuart GJ (2006). Single Ih channels in pyramidal neuron dendrites:
properties, distribution, and impact on action potential output. J Neurosci 26, 1677 -- 87.
Łeski S, Lindén H, Tetzlaff T, Pettersen KH & Einevoll GT (2013). Frequency dependence of
signal power and spatial reach of the local field potential. PLoS Comp Biol 9, 1 -- 23.
Lindén H, Hagen E, Łeski S, Norheim ES, Pettersen KH & Einevoll GT (2014). LFPy: a tool
for biophysical simulation of extracellular potentials generated by detailed model neurons.
Front Neuroinform 7, 1 -- 15.
21
Lindén H, Pettersen KH & Einevoll GT (2010).
Intrinsic dendritic filtering gives low-pass
power spectra of local field potentials. J Comput Neurosci 29, 423 -- 44.
Lindén H, Tetzlaff T, Potjans TC, Pettersen KH, Grün S, Diesmann M & Einevoll GT (2011).
Modeling the Spatial Reach of the LFP. Neuron 72, 859 -- 72.
Lörincz A, Notomi T, Tamás G, Shigemoto R & Nusser Z (2002). Polarized and compartment-
dependent distribution of HCN1 in pyramidal cell dendrites. Nat Neurosci 5, 1185 -- 1193.
Magee JC (1998). Dendritic hyperpolarization-activated currents modify the integrative prop-
erties of hippocampal CA1 pyramidal neurons. J Neurosci 18, 7613 -- 24.
Major G, Larkum ME & Schiller J (2013). Active properties of neocortical pyramidal neuron
dendrites. Annu Rev Neurosci 36, 1 -- 24.
Mauro A, Conti F, Dodge F & Schor R (1970). Subthreshold behavior and phenomenological
impedance of the squid giant axon. J Gen Physiol 55, 497 -- 523.
Mishra P & Narayanan R (2015). High-conductance states and A-type K+ channels are poten-
tial regulators of the conductance-current balance triggered by HCN channels. J Neurophys-
iol 113, 23 -- 43.
Mitzdorf U (1985). Current Source-Density Method and Application in Cat Cerebral Cortex:
Investigation of Evoked Potentials and EEG Phenomena. Physiol Rev 65.
Nunez PL & Srinivasan R (2006). Electric Fields of the Brain, Oxford University Press, New
York.
Nusser Z (2009). Variability in the subcellular distribution of ion channels increases neuronal
diversity. Trends Neurosci 32, 267 -- 274.
Petreanu L, Mao T, Sternson SM & Svoboda K (2009). The subcellular organization of neocor-
tical excitatory connections. Nature 457, 1142 -- 1145.
Pettersen KH, Lindén H, Dale AM & Einevoll GT (2012). Extracellular spikes and CSD.
In Brette R & Destexhe A, editors, Handbook of Neural Activity Measurement, chapter 4,
pp. 92 -- 135. Cambridge University Press, Cambridge, UK.
Pettersen KH & Einevoll GT (2008). Amplitude variability and extracellular low-pass filtering
of neuronal spikes. Biophys J 94, 784 -- 802.
Pettersen KH, Hagen E & Einevoll GT (2008). Estimation of population firing rates and current
source densities from laminar electrode recordings. J Comput Neurosci 24, 291 -- 313.
22
Rall W & Shepherd G (1968). Theoretical reconstruction dendrodendritic of field potentials
and in olfactory bulb synaptic interactions. J Neurophysiol 31, 884 -- 915.
Ray S & Maunsell JHR (2011). Different origins of gamma rhythm and high-gamma activity
in macaque visual cortex. PLoS Biol 9, 1 -- 15.
Reimann M, Anastassiou C, Perin R, Hill SL, Markram H & Koch C (2013). A Biophysically
Detailed Model of Neocortical Local Field Potentials Predicts the Critical Role of Active
Membrane Currents. Neuron 79, 375 -- 390.
Remme MWH (2014). Quasi-active approximation of nonlinear dendritic cables In Jaeger D &
Jung R, editors, Encyclopedia of Computational Neuroscience, pp. 1 -- 5. Springer New York.
Remme MWH, Lengyel M & Gutkin BS (2009). The role of ongoing dendritic oscillations in
single-neuron dynamics. PLoS Comput Biol 5, e1000493.
Remme MWH, Lengyel M & Gutkin BS (2010). Democracy-independence trade-off in oscil-
lating dendrites and its implications for grid cells. Neuron 66, 429 -- 437.
Remme MWH & Rinzel J (2011). Role of active dendritic conductances in subthreshold input
integration. J Comput Neurosci 31, 13 -- 30.
Roberts M, Lowet E, Brunet N, TerWal M, Tiesinga P, Fries P & DeWeerd P (2013). Robust
Gamma Coherence between Macaque V1 and V2 by Dynamic Frequency Matching. Neu-
ron 78, 523 -- 536.
Scheffer-Teixeira R, Belchior H, Leao RN, Ribeiro S & Tort ABL (2013). On High-Frequency
Field Oscillations (>100 Hz) and the Spectral Leakage of Spiking Activity. J Neurosci 33,
1535 -- 1539.
Schomburg EW, Anastassiou Ca, Buzsáki G & Koch C (2012). The spiking component of
oscillatory extracellular potentials in the rat hippocampus. J Neurosci 32, 11798 -- 811.
Stuart G, Spruston N & Häusser M (2007). Dendrites Oxford University Press, Oxford, UK.
Taxidis J, Anastassiou CA, Diba K, Koch Correspondence C & Koch C (2015). Local Field
Potentials Encode Place Cell Ensemble Activation during Hippocampal Sharp Wave Ripples.
Neuron 87, 590 -- 604.
Tomsett RJ, Ainsworth M, Thiele A, Sanayei M, Chen X, Gieselmann MA, Whittington MA,
Cunningham MO & Kaiser M (2015). Virtual electrode recording tool for extracellular po-
tentials (vertex): comparing multi-electrode recordings from simulated and biological mam-
malian cortical tissue. Brain Struct Funct 220, 2333 -- 2353.
23
Williams SR & Stuart GJ (2000). Site independence of EPSP time course is mediated by
dendritic I(h) in neocortical pyramidal neurons. J Neurophysiol 83, 3177 -- 3182.
Zhuchkova E, Remme MWH & Schreiber S (2013). Somatic versus dendritic resonance: dif-
ferential filtering of inputs through non-uniform distributions of active conductances. PloS
One 8, e78908.
24
Figure 1: Active conductances can shape the extracellular signature of synaptic inputs. A:
A single synaptic input is provided to a cortical layer 5 pyramidal cell model. The extracellular
response is shown at five positions (cyan dots) for two cases: the active model that includes
various voltage-dependent conductances (red; see Methods), or a passive model from which the
active conductances have been removed (black). The position of the input is marked by the
yellow star: at the distal apical dendrite (top panels) or at the soma (bottom panels). The cell's
resting potential was held uniformly at a hyperpolarized potential of -80 mV (left panels) or at
a depolarized potential of -60 mV (right panels). The synaptic peak conductance was 0.001 µS.
The plots show the x,z-plane of the cell; the soma and the electrodes are positioned at y=0.
B: As in panel A, but using white-noise current input (see Method section) instead of synaptic
input, and displaying the response as the power spectral density (PSD). The PSD is calculated
from 1000 ms long simulations. C: Apical input (marked by star) to the cell model held at a
hyperpolarized potential (-80 mV). The extracellular potential is shown at two positions (cyan
dots) for the active (red) and passive (black) model and for two additional versions of the model:
the passive model supplemented by Ih (dashed blue), and the passive model supplemented by
frozen Ih (dotted gray; see Methods).
25
C-80 mVactivepassivepassive + Ihpassive + frozen IhFrequency (Hz)Frequency (Hz)LFP-PSD (μV2/Hz)LFP-PSD (μV2/Hz)010-410-210101001-410-610101001-210B-60 mV-80 mVLFP-PSD (μV2/Hz)Apical inputSomatic inputAApical inputSomatic input-60 mV-80 mV30 ms10 nVpassive200 μmFreq. (Hz)active-21010-71100 Figure 2: Quasi-active conductances accurately mimic the responses of nonlinear conduc-
tances. A: Apical white-noise current input (yellow star) is provided to a cortical pyramidal
cell model held at a hyperpolarized potential (-80 mV). The LFP-PSD is shown at two locations
(cyan dots) for a passive model (black), a passive model with Ih (dashed blue), and a passive
model that includes a linearized or quasi-active Ih (green; see Methods). B: As in panel A, but
with somatic white noise-current input and using the following three models: a passive model
(black), a passive model with INaP (dashed magenta), and a passive model with linearized INaP
(orange).
26
A-80 mVpassive + linearized Ihpassivepassive + IhFrequency (Hz)Frequency (Hz)LFP-PSD (μV2/Hz)LFP-PSD (μV2/Hz)-110-510-310101001-510-7101-310B-60 mVFrequency (Hz)Frequency (Hz)LFP-PSD (μV2/Hz)LFP-PSD (μV2/Hz)-210-610-210101001-410-610101001-410passive + linearized INaPpassivepassive + INaP10100 Figure 3: The type of active conductance, its distribution across the cell membrane, and
the location of the input shape the LFP. A: A single quasi-active conductance was distributed
across the pyramidal cell model in three ways: uniformly (left), linearly increasing (center), or
linearly decreasing (right) with distance from the soma. The density of the linearly increas-
ing conductance increased sixty-fold from the soma to the most distal apical dendrite, while
the linearly decreasing conductance had a sixty-fold decrease over the same range. B: Apical
white-noise current input (yellow star) was provided to the cell model and the LFP-PSD was
determined at two positions (cyan dots). The quasi-conductance was either regenerative (red),
passive-frozen (black) or restorative (blue). The three cases had the same total resting mem-
brane conductance and the total passive leak conductance was equal to the total quasi-active
conductance. C: As panel B, but with white-noise current input provided to the soma.
27
Linear decreaseUniformLinear increaseABCregenerativepassive-frozenrestorativeDistance from somaChannel densityLFP-PSD (μV2/Hz)-210-610-410LFP-PSD (μV2/Hz)-210-610-410Frequency (Hz)110100Frequency (Hz)110100110100110100LFP-PSD (μV2/Hz)-210-610-410LFP-PSD (μV2/Hz)-210-610-410Frequency (Hz)110100110100-210-610-410-210-610-410-210-610-410-210-610-410-210-610-410-210-610-410-210-610-410-210-610-410110100110100110100110100110100110100Distance from somaDistance from somaFrequency (Hz)Frequency (Hz)Frequency (Hz) Figure 4: Activation time constants of active conductances determine the frequencies for
which they affect the LFP. A: Apical white-noise current input (yellow star) to the cortical
pyramidal cell model with a single linearly increasing quasi-active conductance (as in middle
column of Fig. 3B). B: The quasi-active conductance is either restorative (blue), passive-frozen
(black) or regenerative (red), and the PSD is shown for the somatic membrane potential Vm
(left panel), the somatic transmembrane currents Im (middle panel), and the LFP (right panel)
at a position close to the soma (cyan dot). The activation time constant of the quasi-active
conductance is τw(VR) = 5 ms. C: As panel B, but with τw(VR) = 50 ms. D: As panel B, but
with τw(VR) = 500 ms.
28
regenerativepassive-frozenrestorativeτw(VR)= 500 msD110100110100110100Frequency (Hz)Frequency (Hz)Frequency (Hz)PSD (μV2/Hz)-110-510-310-1010-610-810-310-510-410-610τw(VR)= 50 msC110100110100110100PSD (μV2/Hz)-110-510-310-1010-610-810-310-510-410-610BLFPτw(VR)= 5 msVmIm110100110100110100PSD (μV2/Hz)-110-510-310-1010-610-810-310-510-410-610AFrequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz) Figure 5: Resonance is retained for asymmetric input from distributed synapses. A:
The pyramidal cell model expressed a single linearly increasing quasi-active conductance (see
Figure 4A) that was either restorative (blue), passive-frozen (black), or regenerative (red). 1000
excitatory conductance-based synapses (green dots) with a peak conductance of 0.0001 µS were
distributed across the distal apical tuft more than 900 µm away from soma. The synapses
were activated by independent Poisson processes with a mean rate of 5 spikes per second.
Simulations were run for 20 seconds and the LFP-PSD was calculated using Welch's method.
B: As panel A, but with synapses distributed above the main bifurcation, 600 µm away from the
soma. C: Synapses were distributed uniformly across the entire cell.
29
ABCFrequency (Hz)110100Frequency (Hz)110100Frequency (Hz)110100LFP-PSD (normalized)10.010.110.010.110.010.1regenerativepassive-frozenrestorative Figure 6: Resonance in the LFP resulting from a restorative conductance is a spatially
stable feature. A: Apical white-noise current input (yellow star) is provided to the cortical
pyramidal cell model with a single quasi-active conductance with increasing density with dis-
tance from soma. The resulting LFP-PSD was calculated at a dense 2D-grid surrounding the
cell. The maximum power of the LFP is denoted in shades of gray. The green contour line
shows where the LFP power is 10−7 µV2/Hz. B: Excerpts of LFP time traces at two extracel-
lular positions, corresponding to the two electrodes at the opposite side of the zero crossing
region of the dipole in panel A (i.e., the first and third electrode contact in the left column). C:
LFP-PSD computed at the eight electrode positions marked by shades of red and blue in panel
A. D: The frequency for which the LFP-PSD shows the maximum power (see panel A). E: The
Q-value of the LFP-PSD, defined as the maximum power (see panel A) divided by the power at
1 Hz.
30
ABDEMaximum LFP-PSDFrequency at maximum LFP-PSDQ-value of LFP-PSD10 msLFP (normalized)40302010020151051-410-610-510-710LFP-PSD (μV2/Hz)110100Frequency (Hz)-810-410-610C500 μmμV2/HzHz Figure 7: Voltage, transmembrane current, and LFP for a semi-infinite neurite with a
quasi-active conductance. White-noise current input (yellow star in schematic above panel
C) was applied to the end of a semi-infinite neurite with a single quasi-active conductance that
was either restorative (blue), passive-frozen (black) or regenerative (red). The neurite had length
2000 µm, diameter 2 µm, intracellular resistivity Ra = 100 Ωcm, and the passive leak and quasi-
active conductance density were both uniformly set to 50 µS/cm2. A: The membrane potential
PSD (Vm) is shown at the positions marked by the orange dots in the schematic (at 0, 177, 531
and 973 µm distance from the input). B: As in panel A, but showing the transmembrane current
PSD at the positions of the orange dots. C: As in panel A, but showing the LFP-PSD for the
positions marked by the cyan dots (20 µm away from the neurite). D: As in panel C, but showing
the LFP-PSD for electrodes at 300 µm away from the neurite, marked by the cyan dots.
31
ABPSD (mV2/Hz)PSD (nA2/Hz)PSD (μV2/Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)110100110100110100110100110100110100110100110100110100110100110100110100-310-510-410-310-510-410-310-510-410-310-510-410-210-410-310-510-710-610-610-810-710-610-810-710210210210210010010010010110110110110310310310310LFPDVmImFrequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)CPSD (μV2/Hz)110100110100110100110100010-210-110-210-410-310-210-410-310-210-410-310LFPFrequency (Hz)Frequency (Hz)Frequency (Hz)Frequency (Hz)regenerativepassive-frozenrestorativeμmμm50 50 |
1708.03666 | 1 | 1708 | 2017-08-11T19:12:46 | Attention but not musical training affects auditory streaming | [
"q-bio.NC"
] | While musicians generally perform better than non-musicians in various auditory discrimination tasks, effects of specific instrumental training have received little attention. The effects of instrument-specific musical training on auditory grouping in the context of stream segregation are investigated here in three experiments. In Experiment 1a, participants listened to sequences of ABA tones and indicated when they heard a change in rhythm. This change is caused by the manipulation of the B tones' timbre and indexes a change in perception from integration to segregation, or vice versa. While it was expected that musicians would detect a change in rhythm earlier when their own instrument was involved, no such pattern was observed. In Experiment 1b, designed to control for potential expectation effects in Experiment 1a, participants heard sequences of static ABA tones and reported their initial perceptions, whether the sequence was integrated or segregated. Results show that participants tend to initially perceive these static sequences as segregated, and that perception is influenced by similarity between the timbres involved. Finally, in Experiment 2 violinists and flautists located mistuned notes in an interleaved melody paradigm containing a violin and a flute melody. Performance did not depend on the instrument the participant played but rather which melody their attention was directed to. Taken together, results from the three experiments suggest that the specific instrument one practices does not have an influence on auditory grouping, but attentional mechanisms are necessary for processing auditory scenes. | q-bio.NC | q-bio | Running head: ATTENTION AFFECTS AUDITORY GROUPING
1
Attention but not musical training affects auditory grouping
Sarah A. Sauvé & Marcus T. Pearce
Queen Mary University of London
Author note
Correspondence concerning this article should be addressed to
Sarah Sauvé, School of Electronic Engineering and Computer Science, Queen Mary
University of London, Mile End Road, E1 4NS, London, United Kingdom
ATTENTION AFFECTS AUDITORY GROUPING
2
Abstract
While musicians generally perform better
than non-musicians
in various auditory
discrimination tasks, effects of specific instrumental training have received little attention. The
effects of instrument-specific musical training on auditory grouping in the context of stream
segregation are investigated here in three experiments. In Experiment 1a, participants listened
to sequences of ABA_ tones and indicated when they heard a change in rhythm. This change
is caused by the manipulation of the B tones' timbre and indexes a change in perception from
integration to segregation, or vice versa. While it was expected that musicians would detect a
change in rhythm earlier when their own instrument was involved, no such pattern was
observed. In Experiment 1b, designed to control for potential expectation effects in Experiment
1a, participants heard sequences of static ABA_ tones and reported their initial perceptions,
whether the sequence was integrated or segregated. Results show that participants tend to
initially perceive these static sequences as segregated, and that perception is influenced by
similarity between the timbres involved. Finally, in Experiment 2 violinists and flautists
located mistuned notes in an interleaved melody paradigm containing a violin and a flute
melody. Performance did not depend on the instrument the participant played but rather which
melody their attention was directed to. Taken together, results from the three experiments
suggest that the specific instrument one practices does not have an influence on auditory
grouping, but attentional mechanisms are necessary for processing auditory scenes.
Keywords: musical training, stream segregation, timbre, attention
ATTENTION AFFECTS AUDITORY GROUPING
3
Attention but not musical training affects auditory grouping
Coined by Albert Bregman in his 1990 book, auditory scene analysis is the process by
which we analyse the auditory world around us. The auditory system offers a 360o view of the
world and provides information about objects that cannot be seen. Auditory streaming is the
perceptual decomposition of sound input into its component sources, and has been the main
conceptual approach to studying auditory scene analysis. It has been investigated in the context
of sound attributes such as pitch (van Noorden, 1975), location (Jones & Macken, 1995),
periodicity (Vliegen, Moore, & Oxenham, 1999) and timbre (Iverson, 1995) among others.
Over the decades, a number of paradigms have been developed to explore and understand
auditory streaming; two influential ones will be summarized here.
The first paradigm was pioneered by van Noorden (1975), in his doctoral research. It
is a subtle but clever modification of the Miller & Heise (1950) paradigm: instead of alternating
sounds in an ABAB pattern, van Noorden alternated sounds in an ABA- pattern, where '-' is a
silence. This creates a triplet pattern when A and B are perceived as integrated, or coming from
the same source, and an even pattern where the A stream is twice as fast as the B stream when
A and B are perceived as segregated, or coming from two different sources. van Noorden
(1975) explored the influence of pitch, tempo and loudness on this rhythmic perception and
found that as pitch, tempo and loudness difference increased, perception tended towards
segregation. In other words, the more different A and B are, the more likely they are to be
perceived as coming from different sources. van Noorden further defined segregation
parameters with fission and temporal coherence boundaries. While the fission boundary
defines the difference below which integration is inevitable (the pitch, tempo or loudness are
too similar or slow to lead to segregation), the temporal coherence boundary defines the
difference above which segregation is inevitable (the pitch, tempo or loudness are too
dissimilar or fast to allow for integration). Between these two boundaries, perception is bi-
ATTENTION AFFECTS AUDITORY GROUPING
4
stable, meaning that either integration or segregation are possible and depend on other factors
(Denham & Winkler, 2006). This paradigm has been used to help researchers understand how
various sound attributes contribute to our perception of the world around us (Rose & Moore,
2000; Singh & Bregman, 1997; Vliegen, Moore, & Oxenham, 1999). For example, that
streaming occurs before temporal integration of the auditory scene (Yabe et al., 2001) and that
both spectral content and location interact in the processing of ambiguous auditory scenes
(Shinn-Cunningham, Lee, & Oxenham, 2007).
Another prominent paradigm, using slightly more complex stimuli, was introduced by
Dowling (Dowling, 1973) and was named the interleaved melody paradigm. Here, the notes of
two melodies are presented in an alternation, such that melody 'ABCDEF' and melody 'abcdef'
become 'AaBbCcDdEeFf'. Dowling found that as pitch overlap decreased, participants were
more easily able to detect, or segregate, each individual melody. Trained musicians could
tolerate more pitch overlap than non-musicians. The concept can similarly be applied with
many other parameters including loudness and timbre (Hartmann & Johnson, 1991), where it
is easier to track a melody if the two interleaved melodies are of different loudness, or played
by different instruments.
Timbre is a complex auditory parameter and timbral perception has been investigated
in detail using both synthesized tones and real instrumental sounds (Alluri & Toiviainen, 2010;
Caclin, McAdams, Smith, & Winsberg, 2005; McAdams, Winsberg, Donnadieu, De Soete, &
Krimphoff, 1995).
The most common method of investigating
timbre has been
multidimensional scaling, or MDS. Based on dissimilarity ratings between pairs of timbres,
sounds are mapped into a multi-dimensional space representing perceptual distance. In research
to date, three dimensions seems to provide an optimal representation of perceptual timbre space;
though the first two are fairly stable across experiments, the third is less well established. The
first two represent log rise time (the attack), and spectral centroid while the third dimension
ATTENTION AFFECTS AUDITORY GROUPING
5
that emerges is usually a spectro-temporal feature such as spectral flux or spectral irregularity.
One of the biggest issues with this research however is that in most cases the rated sounds are
synthesized (though see Kendall & Carterette, 1991, and Lakatos, 2000, for examples of MDS
using natural stimuli). Besides this, our perceptual system is not used to hearing synthetic
sounds such as these and may process them differently than natural sounds (Gillard, J. &
Schutz, M., 2012). Therefore, it is important to complement studies using controlled
synthesized tones with investigations using natural sounds.
The role of musical training has been extensively studied in the context of auditory
skills, including auditory streaming (François, Jaillet, Takerkart, & Schön, 2014; Zendel &
Alain, 2009). As a result of training, musicians are more sensitive to changes in auditory stimuli
based on pitch, time and loudness for example (Marozeau, Innes-Brown, & Blamey, 2013;
Marozeau, Innes-Brown, Grayden, Burkitt, & Blamey, 2010), with discrimination thresholds
being lower in musicians than in non-musicians. One problem with treating musicians as a
single category is that differences between instrumentalists may be missed (Tervaniemi, 2009).
Pantev and colleagues (Pantev, Roberts, Schulz, Engelien, & Ross, 2001) found that certain
instrumentalists were more sensitive to the timbre of their own instrument than to others, as
measured by auditory evoked fields (AEF). Violinists and trumpet players were presented with
trumpet, violin and sine tones while MEG was recorded. Both instrumentalists presented
stronger AEFs for complex over sine tones, and stronger AEFs still for their own instrument.
In a similar study (Shahin, Roberts, Chau, Trainor, & Miller, 2008), professional violinists and
amateur pianists as well as young piano students and young non-musicians were presented with
piano, violin and sine tones while reading or watching a movie and EEG was recorded. Gamma
band activity (GBA) was more robust in professional musicians for their own instruments and
young musicians showed more robust GBA to piano tones after their one year of musical
training. Furthermore, Drost, Rieger, & Prinz, (2007) found that pianists and guitarists'
ATTENTION AFFECTS AUDITORY GROUPING
6
performance on a performance task was negatively affected by auditory interference, but only
if it was their own instrument. Taking a step further and using more ecological stimuli,
Margulis, Mlsna, Uppunda, Parrish, & Wong, (2009) explored neural expertise networks in
violinists and flautists as they listened to excerpts from partitas for violin and flute by J. S.
Bach. Increased sensitivity to syntax, timbre and sound-motor interactions were seen for
musicians when listening to their own instrument.
More recently, pianists, violinists and non-musicians listened to music during fMRI
scanning (Burunat et al., 2015). The authors investigated the effects of musical training on
callosal anatomy and interhermispheric functional symmetry and found that symmetry was
increased in musicians, and particularly in pianists, in visual and motor networks. They
concluded that motor training, including differences between instrumentalists, affects music
perception as well as production. Other research has investigated differences between types of
musical training. For example, one study used EEG to show that conductors have improved
spatial perception, when compared to non-musicians and pianists (Nager, Kohlmetz,
Altenmuller, Rodriguez-Fornells, & Münte, 2003). Another line of research investigates
pianists' formation of action-effect mappings due to the design of their instrument (Baumann
et al., 2007; Drost, Rieger, Brass, Gunter, & Prinz, 2005; Repp & Knoblich, 2009; Stewart,
Verdonschot, Nasralla, & Lanipekun, 2013).
However, such specific effects of instrumental training have not yet been observed in
auditory streaming, where an effect would be seen by a change in streaming threshold. We
hypothesise that with increased sensitivity to a particular timbre, it would take less time to
detect two separate auditory objects when one's own instrument is one of these objects. This
is the basis of the first experiment reported in this paper.
The objective of this research is to test the hypothesised increased sensitivity in
streaming the instrument(s) which a musician plays. Three experiments will be presented. The
ATTENTION AFFECTS AUDITORY GROUPING
7
first is a classic ABA_ streaming paradigm, the second is a control study that examines the
effects of prior expectation and the third is an interleaved melody paradigm, designed to
corroborate findings in Experiments 1a and 1b using more musically realistic stimuli.
Experiment 1a
The ABA_ paradigm (van Noorden, 1975) is used here and timbre is manipulated
instead of pitch. While the timbre of a standard sequence remains static throughout a given
trial, a target sequence morphs from one timbre to another, creating a qualitative change from
a galloping ABA_ rhythm to the perception of two simultaneous, isochronous A_A_A and
B___B___B patterns as the standard and target sequences' timbres become more and more
different, or vice versa as the timbres become more similar. The point of change in rhythmic
perception reflects the detection of a new sound object, or, in the other direction, the merging
together of two sound objects. The sound objects (standard and target streams) are defined
solely by their timbre, as pitch, length and loudness are controlled. Based on previous work
(Sauvé, Stewart, & Pearce, 2014), detection of a sound object defined by one's own
instrumental timbre is predicted to occur sooner than for other instrumental timbres, when the
participants' instrument is the target (i.e. it is 'new to the mix' and captures attention) and later
than for other instrumental timbres when the instrumentalists' timbre is the standard (i.e. it
already holds attention and delays perception of the arrival of a new sound object). This
previous study compared seven different instrumental timbres in the same ABA_ paradigm,
while additionally exploring the effect of attention on streaming by manipulating participants'
attentional focus. Results guided the design of the current study by providing target effect
sizes, refining the test timbres and allowing the elimination of the attention manipulation, as it
was confirmed to have a significant impact on the perception of auditory streams.
Method
ATTENTION AFFECTS AUDITORY GROUPING
8
Participants. Participants were 20 musicians (13 females, average age 34.45; SD =
7.59; range 21-69) recruited from universities and the community. Their average Gold-MSI
score (Müllensiefen, Gingras, Musil, & Stewart, 2014) for the musical training subscale was
40.15 (SD = 4.23); 5 were violinists, 6 were cellists, 5 were trumpet players and 4 were
trombone players.
Stimuli. All four timbral sounds (violin, trumpet, trombone, cello) were chosen from
the MUMS library (Opolko & Wapnick, 2006) with pitches spanning an octave (all 12 pitches
between A220 to G#415.30). The files were adjusted to equal perceptual length of 100ms and
equal loudness, based on the softest sound. A 10ms fade out was applied to each timbral sound.
All editing was done in Audacity and the final product was exported as a CD quality wav file
(44,100 Hz, 16 bit). See Appendix A for full details.
Using a metronome in Max/MSP, the standard sequence was presented by playing a
selected timbre with an inter-onset interval of 220ms. The target sequence was presented using
another metronome at a rate of onset of 440ms, beginning 110ms after the standard sequence
to create the well-known galloping ABA_ pattern (van Noorden, 1975). The target sequence
was a series of 100ms sound files representing a 30s morph between the standard timbre and
Figure 1. Illustration of ABA_ paradigm, ascending and descending, modifying timbre
only.
ATTENTION AFFECTS AUDITORY GROUPING
9
the target timbre, achieved using a slightly modified Max/MSP patch entitled 'convolution-
workshop'. This patch is distributed by Cycling '74 with Max/MSP. The target sequence
morphed from standard to target timbre in the ascending condition, creating a galloping to even
rhythm change, and from target to standard timbre in the descending condition, creating an
even to galloping rhythm change (see Figure 1). Each trial ended when the participant indicated
a change in perception or after 30s if participants did not reach a change in perception.
Procedure. The experiment was coded and run in Max/MSP, with output presented
through headphones and input taken from mouse clicks. Participants were first presented with
a practice block with instructions and an opportunity to listen to each timbre and rhythm
separately. Up to four practice trials were included in the block and questions were welcomed.
Participants then began the first of two experimental blocks.
For each trial, participants indicated by clicking a button on the screen at which point
the galloping sequence became perceived as two separate streams of standard and target
tones, or the opposite for descending presentation. This point was recorded as the percent of
time passed in the trial, which equates to the percent of morphing at that time. Each trial
lasted a maximum of 30s, at which point the trial ended automatically and a value of '-1' was
recorded, indicating that the participant had not reached a change in perception on that trial.
Trials were presented in two blocks, and participants were instructed to indicate a change in
rhythm as soon as it was perceived for the ascending block and to hold on to the original
rhythm as long as possible for the descending block. Together, this gives two measures of
the fission boundary (van Noorden, 1975). The fission boundary was measured instead of the
temporal coherence boundary due to its higher sensitivity for detecting timbral effects in
perception, and due to confirmation that the fission and temporal coherence boundaries are
separate phenomena that can be manipulated by instruction (Sauvé et al., 2014). For every
block, every timbre modulated to every other timbre once for a total of 12 trials (4 timbres
ATTENTION AFFECTS AUDITORY GROUPING
10
each modulating to the 3 other timbres), each separated by 4s and each at a different pitch, to
reduce trial to trial expectancy and habituation. Participants were randomly assigned to one
of two different orders to control for any order effects.
Once both blocks were completed, participants filled out the musical training sub-scale
of the Goldsmiths Musical Sophistication Index (Müllensiefen et al., 2014).
Analysis. Effect sizes and confidence intervals were used in the analysis of
Experiments 1a and 1b, in addition to traditional methods. These methods are based on
Cumming (2012; 2013), who advocates wider use of effect sizes and confidence intervals in
the research community to increase integrity, accuracy and the use of replication. According
to Cumming, the low occurrence of null results in the literature and a pressure towards new
studies and away from replication translates into misrepresentation and inhibition of scientific
knowledge. Cumming advocates the use of effect sizes, confidence intervals, and meta-analysis
in place of null hypothesis significance testing (NHST). This method is preferred because
confidence intervals give more information both about the current effect size, and about
potential future replications by offering a range of potential values for a measure, rather than
one indicator of significance or non-significance. For more information about effect size and
confidence interval methods, see Cummings' book, The New Statistics (2012) or the
corresponding article for a shorter summary (2013).
Results
Percentage of time passed (degree of morphing) is the dependent variable analysed; for
descending trials the percentage was subtracted from 100 so that ascending and descending
conditions can be compared directly. A low percentage indicates early streaming in the
ascending condition and late integration in the descending condition while a high percentage
indicates late streaming in the ascending condition and early integration in the descending
ATTENTION AFFECTS AUDITORY GROUPING
11
condition. Furthermore, trials in the ascending condition where the percentage exceeded 100
were replaced with 100 and trials in the descending condition where the percentage was
negative were replaced with 0. These are all cases where the participant listened to the trial for
more than 30 seconds and still did not hear a change in rhythm. Five participants' data were
removed because they did not hear a change in rhythm in more than half of the trials, in either
or both blocks (two violinists, two cellists and a trumpet player). The difference between mean
percentage for ascending and descending conditions was 1.2, 95% CI [-4.4, 6.8]. As the CIs
include zero, the difference was not significant. However, mean percentage of time passed was
significantly higher for the first block of trials than the second, with a difference of 10.6 [2.6,
18.6] for the ascending and 10.7 [3.1, 18.1] for the descending conditions. As both CIs do not
include zero, the difference is significant.
Effects of specific instrumental training were investigated next. Data were grouped by
instrumentalist and then sub-grouped by standard timbre. For violinists, mean percent time
passed when violin was the standard timbre was 56.5 [50.7, 62.3], mean percent for cello was
59.8 [53.5, 66.1], mean percent for trumpet was 65.8 [51.8, 79.8] and mean percent for
trombone was 64.6 [50.8, 78.4]. See Table 1 for details of all instrumentalists. Data were then
sub-grouped by target timbre. When violin was the target timbre, mean percent for violinists
was 62.1 [50.0, 74.2], mean percent for cellists was 48.2 [36.9, 59.5], mean percent for
trumpeters was 54.5 [44.5, 64.5] and mean percent for trombonists was 48.4 [38.1, 58.7]. See
Table 1 for details of all target timbres. Figure 2 displays results graphically.
Thresholds for an instrumentalists' own timbre were hypothesised to be lower when
their own instrument was the target and higher when it was the standard. However, interpreting
the CIs above does not reveal any reliable pattern of results. If more than half the margins of
error (MOE), which is one half of the CI, overlap when comparing between subject groups, the
difference is not considered significant. While two comparisons attain significance (trombone
ATTENTION AFFECTS AUDITORY GROUPING
12
Table 1. Mean percent of trial duration by standard and target timbre, and by instrumentalist, with
95% confidence interval margins of error (MOE).
Mean Duration ± MOE
Violin
Cello
Trumpet
Trombone
i
e
r
b
m
T
d
r
a
d
n
a
t
S
e
r
b
m
T
i
t
e
g
r
a
T
Violinist
56.5 ± 5.8
59.8± 6.3
65.8 ± 14.0
64.6 ± 13.8
Cellist
50.7 ± 12.0
46.3 ± 12.1
55.7 ± 11.9
47.8 ± 11.0
Trumpeter
53.7 ± 11.8
50.8 ± 6.6
50.0 ± 10.3
59.6 ± 12.1
Trombonist
49.4 ± 13.9
57.0 ± 9.1
48.0 ± 9.9
39.5 ± 9.5
Violinist
62.1 ± 12.1
62.1 ± 13.4
62.9 ± 9.1
62.0 ± 9.8
Cellist
48.2 ± 11.3
51.3 ± 11.4
46.8 ± 13.9
53.2 ± 11.6
Trumpeter
54.5 ± 10.0
47.3 ± 9.0
67.1 ± 11.4
49.2 ± 9.8
Trombonist
48.4 ± 10.3
45.3 ± 12.3
45.4 ± 11.4
53.4 ± 7.4
players have a lower threshold than trumpet players for the trombone sound as standard, and
trombone players have a lower threshold than trumpet players for the trumpet sound as target),
this is not enough to establish a pattern. Comparison of confidence intervals cannot be done
so easily for within-subject measures, therefore a mixed effects model was applied, where
A
B
Figure 2. Percent target timbre contained in the morphing stream at the point of a change in percept
as a function of instrumentalist, and standard (A) and target (B) timbres. Error bars represent 95%
CIs.
ATTENTION AFFECTS AUDITORY GROUPING
13
instrument played and standard, or target timbre,
predicted threshold. The instrument played had no
effect on perceptual threshold, χ2 (3) = 3.83 and χ2 (3)
= 3.82, p = .28 for standard and target models
respectively.
Effects of instrumental family were also
investigated. Performance by instrumental group was
analysed for string pair and brass pair trials (i.e. where
the standard and target timbres were both string or both
Figure 3. Percent target timbre of
the morphed stream at the point of
a change in percept for brass and
string instrumental family groups.
Error bars represent 95% CIs.
brass instruments). String players performed with a mean percentage of 54.5 [46.0, 63.0] on
string pairs and 62.4 [52.2, 72.6] on brass pairs. Brass performed with a mean percentage of
56.0 [46.7, 65.3] on string pairs and 58.4 [48.2, 78.6] on brass pairs (see Figure 3). Interpreting
the CIs indicates that there was no difference between string players and brass players;
however, a mixed effects model to investigate within group differences found that instrument
played had an effect on threshold, χ2 (1) = 3.54, p = .05, where string players had a lower
discrimination threshold for string instruments than for brass instruments.
Trials where participants did not hear a change in rhythm were examined separately.
Most participants only had a few trials where this happened, if at all. As noted above, for five,
this case was more prominent and their data were removed (it is interesting to note that the
mean age for these five participants is 51.8 (SD = 12.7) and every participant was at or above
the average age for all participants). Every type of instrumentalist was represented in this group
of trials; all for the ascending block and all but cellists for the descending blocks. The
frequency of each of the standard and target timbres was different within each direction by
timbre type condition (i.e. the number of times a trial had cello as the standard or target timbre,
ATTENTION AFFECTS AUDITORY GROUPING
14
versus the other instruments, in the ascending or descending block), but no single timbre was
consistently more or less represented. When looking at pairs of timbres, the cello-trumpet and
trombone-trumpet pairs were most commonly still perceived as an even percept by the end of
a trial in the descending condition and the violin-trombone pair was the most commonly still
perceived as a galloping percept by the end of a trial in the ascending condition.
Discussion
This experiment was designed to corroborate neuroscientific measures showing that
instrumentalists are more sensitive to their own instrument's timbre than to others (Pantev et
al., 2001). Accordingly, in the ABA_ paradigm, we hypothesised a lower timbre discrimination
threshold for instrumentalists hearing their own instrument when their instrument is the target
timbre, and a higher discrimination threshold when their instrument is the standard timbre.
Results show no reliable effect of instrument played on the perception of timbral stream
segregation when looking at individual target instruments. Though thresholds for an
instrumentalists' timbre were slightly lower than for other timbres when looking at standards,
contrary to the hypothesis, none of these differences were significant. Similarly for target
timbres, no threshold differences were significant, though the largest effect was seen in trumpet
players, where the threshold when trumpet was the target was higher than for other instruments.
There was a small effect of instrument played when comparing performance on instrumental
families: string players detected the difference between two brass instruments later than for
two string instruments. They were not better than brass players at detecting the difference
between two string instruments, nor did brass players show an advantage for brass instruments.
Thresholds for string instruments were overall lower than for brass instruments. Perhaps the
two string instruments were more different than the two brass instruments, thus making them
overall easier to distinguish (this is supported by timbre dissimilarity ratings collected in
ATTENTION AFFECTS AUDITORY GROUPING
15
Experiment 1b). The effect of order is unexpected and could be the result of a familiarization
with the task that led to greater sensitivity in the second block.
How can such results be explained when the literature reviewed, particularly Pantev's
work (2001), suggests an effect of instrumental training on perception? Let us first place the
question in a more generalized context. Imagine a trained musician is listening to an orchestral
work. Just like most listeners, they clearly hear the melody. What if they were asked to listen
to the bass line? Or another instrument? If instrumentalists are more sensitive to their own
instrument's timbre, then it would be expected that they could more easily and more accurately
pick out (and perhaps transcribe, for potential experimental purposes) their own instrument
than any other. However, according to the present results, they could also pick any instrument
out of the auditory scene and transcribe it just as well. This would suggest that ability to pick
out and transcribe a particular line in a polyphonic work is not related to the instrument one
plays, but rather to general musical training, and to where their attention is directed. It would
be interesting to conduct a transcription experiment along these lines in future research.
However, a reasonable explanation of the present results is that listeners simply heard what
they paid attention to, though it is only a proposition here and cannot be supported or countered
with the current data. The possibility of attention directing perception will be further explored
in Experiment 1b and Experiment 2.
One of the basic claims of auditory streaming is that coherence is the default percept
(Bregman, 1978; Bregman, 1990; Rogers & Bregman, 1998). However, if this were the case,
then initial segregation in the descending condition of this experiment would not be possible.
The fact that participants were told what they would be hearing (even to galloping for
descending blocks and galloping to even for ascending blocks) could have influenced their
perception of the stimuli by setting up a specific expectation. Therefore, an experiment to
control for this was designed and is reported next.
ATTENTION AFFECTS AUDITORY GROUPING
16
Experiment 1b
This experiment was designed to control for the possible expectation effect of the
instructions given in Experiment 1a. Participants were presented with 10s of ABA_ pattern
where the timbres are unchanging and maximally different (the same as the beginning of a
descending block trial in Experiment 1a) and were asked to report whether they heard an even
or a galloping pattern. If participants tend to hear these stimuli as even, then there is cause to
revisit the default coherence concept; alternatively, if participants tended to hear the stimuli as
galloping, then the instructions given in Experiment 1a likely set up an expectation which
strongly influenced perception, enough to hear an even pattern at first hearing. Participants
were also asked to indicate which of the two timbres was most salient. If the standard timbre
(the faster stream) is chosen most often then timing tends to attract attention more than timbre;
if the standard and target timbres are chosen approximately equally often, then it is the timbre
itself that is most salient in capturing focus.
Method
Participants. Data was collected in two groups: first, undergraduate and graduate
musicians and, second, individuals with various other backgrounds recruited from universities
in London and the community. The first group of participants were the same 20 participants
as in Experiment 1a (the same five participants' data was excluded here); they completed both
paradigms. The second group was tested separately and included a wider range of backgrounds
to control for effects of musical training in the first group. This second group consisted of 20
individuals (7 males, mean age 22.5 years; SD = 4.33; range = 18-32; mean Gold-MSI score =
23.3, SD = 11.9, range = 7-46) recruited through volunteer email lists, credit scheme and
acquaintances. Participants in the first group were entered in a draw for an Amazon voucher
while participants in the second group were either entered in a draw for an Amazon voucher or
given course credit as part of a university credit scheme.
ATTENTION AFFECTS AUDITORY GROUPING
17
Stimuli. The stimuli were the same as Experiment 1a, except that there were seven
timbres (piano, violin, cello, trumpet, trombone, clarinet, bassoon) and there was no morphing.
One timbre was presented at 220ms and the other at 440s with a 110ms offset and the total
length of one trial was 10s.
Procedure. This paradigm was also presented in Max/MSP. After reading the
information sheet and giving written consent, instructions were presented on the screen along
with examples of the even and galloping patterns, each accompanied by an illustration to help
clearly distinguish the two rhythms. Five practice trials were provided and were compulsory,
giving a chance for questions and clarification before beginning the data collection.
When ready to begin, for each trial participants indicated as they were listening which
percept they heard first using the keyboard, pressing 'H' (horse) for the galloping pattern and
'M' (morse) for the even pattern (terminology from Thompson, Carlyon, & Cusack, 2011). At
the end of the trial, they clicked on the timbre that was most salient to them (the appropriate
two were displayed at each trial). Every possible timbre pair was explored, for a total of 21
trials.
Participants then completed the musical training sub-scale of the Gold-MSI
(Müllensiefen et al., 2014).
Timbre dissimilarity ratings. Timbre dissimilarity ratings were collected separately
using Max/MSP. 15 listeners of varying backgrounds, none of which participated in the
reported experiments, rated the similarity of pairs of timbres on a 7-point Likert scale where 1
was the least similar timbre pair and 7 was the most similar timbre pair, with other pairs rated
between these numbers. The participants could listen to seven musical tones at any time. These
were the same as in Sauve et al. (2014) (piano, violin, cello, trumpet, trombone, clarinet,
bassoon). Participants clicked a button to begin a trial: two timbres were presented for
ATTENTION AFFECTS AUDITORY GROUPING
18
A
B
Figure 4. A. Timbre dissimilarity ratings (1-7 Likert scale; 1 is very dissimilar, 7 is very similar).
When the initial percept is even, timbres are less similar (2.30 [2.22, 2.38]) and when the initial
percept is galloping, timbres are more similar (2.90 [2.80, 3.00]). B. Timbre dissimilarity ratings
presented in a heat map, where red is most dissimilar and green is most similar.
comparison and participants rated the similarity between the sounds. There was no time limit
and participants submitted each rating on their own time, completing the trial. Pairs of timbres
were presented randomly. Results are shown in Figure 4.
Results
A comparison of the two groups revealed no significant difference between the initial
percept for musicians and for non-musicians; difference in proportions were .03 [-.04, .10].
Therefore the remaining analysis was performed on aggregated data.
The mean of the initial percept, where even was coded as 0 and galloping was coded as
1, was .35 [.32, .39]. Interpreting the CIs in Figure 5 indicates that this is significantly different
from chance (.5). Because the mean of initial percept is closer to zero than it is to one, the
initial percept is dominantly even.
A 'matching' variable was created, where if the timbre identified as salient matched the
standard timbre, a value of 1 was assigned and if it did not, a value of 0 was assigned. The
ATTENTION AFFECTS AUDITORY GROUPING
19
mean of the matching variable was .69 [.65, .72].
Once again, interpreting the CIs in Figure 5
indicates that this is significantly different from
chance (.5), confirmed by an exact binomial test,
p < .01. Therefore, the most salient timbre is
most often the standard timbre.
The
influence
of
timbre was
investigated using timbral dissimilarity ratings
to assess whether more similar timbre pairs
Figure 5. Mean of initial (left) and
matching (right) variables, both
significantly different from chance.
would encourage integration while less similar pairs would encourage segregation. This
pattern was observed in the data. The average dissimilarity rating over all trials where
segregation was the initial percept was lower, 2.30 [2.22, 2.38], than when integration was the
initial percept, 2.90 [2.80, 3.00], confirmed by t (499) = -8.11, p < .01.
Discussion
This experiment was designed to investigate whether the instructions in Experiment 1a
enabled the possibility of initial segregation in the descending blocks by setting up the
expectation for segregation, as according to streaming theory, integration is always the default
percept until enough evidence is gathered for the existence of two separate streams (Bregman,
1990).
Results indicate that the even percept is the most common initial percept, which is
contrary to the streaming theory discussed above. However, this experiment does not rule out
the possibility that the build-up of evidence for two streams simply happened very quickly. A
reliable neural streaming marker is needed to investigate this question at the millisecond level.
While some such markers have been suggested (Alain, Arnott, & Picton, 2001; Fujioka,
ATTENTION AFFECTS AUDITORY GROUPING
20
Trainor, & Ross, 2008; Sussman, Ritter, & Vaughan, 1999), none of them constitute direct
measures of streaming.
Furthermore, the initial percept depended on how similar pairs of timbres were. It
would presumably take longer for the brain to find evidence for two streams if the sources were
more similar, and less time if they were less similar. A similar pattern for pitch was found by
Deike et al. (2012), where participants were presented with ABAB sequences and asked to
indicate as quickly as possible whether they heard one or two streams. The separation between
A and B tones varied from 2 to 14 semitones. Results showed that the larger the pitch
separation between A and B tones, the more likely participants were to hear the sequence as
segregated in the first place. Predictability was also found to influence degree of segregation
(Bendixen, Denham, & Winkler, 2014): when degree of predictability between two interleaved
sequences was high, an integrated percept was supported, while when the predictability within
each interleaved sequence alone was high, a predominantly segregated percept was induced.
This is contrary to the integration-by-default concept proposed by Bregman (1990). However,
auditory scene analysis is complex and we have not addressed the role of context, which has
been shown to speed or slow the buildup of evidence for perceptual segregation (Sussman-Fort
& Sussman, 2014).
Attentional mechanisms were probed by asking participants which timbre was most
salient. Results show that the standard timbre was most often the most salient timbre. In
feedback, some participants described it as more driving and therefore more attention-drawing.
This suggests that rhythm is a more salient feature than timbre, adding interesting evidence to
discussions about the relative salience of different features in the perception of polyphonic
music (Esber & Haselgrove, 2011; Prince, Thompson, & Schmuckler, 2009; Uhlig, Fairhurst,
& Keller, 2013).
ATTENTION AFFECTS AUDITORY GROUPING
21
In terms of the influence of instructions in Experiment 1a, it seems that they did
influence participants' perception; otherwise, initial segregation on trials with similar pairs of
timbres would not be possible. It is already known that attention influences perception in this
paradigm (Sauvé et al., 2014), and this experiment suggests that prior expectation about the
number of streams also has an impact.
Experiment 2
Experiment 1a aimed to behaviourally test the hypothesis that instrumentalists are more
sensitive to their own instrument's timbre than to others. Experiment 1b was designed to
control for the effect of expectation. However, neither of these paradigms are particularly
ecologically valid; the ABA_ pattern is especially synthetic and though the sounds are recorded
and not synthesized, the way they are combined is not reminiscent of actual music. Experiment
2 was designed with the same goal as Experiment 1a and to allow results to be extended towards
more ecological musical listening. The interleaved melody paradigm introduced by Dowling
(1973) was selected to achieve this goal. The task was to detect one or multiple mistunings, as
intonation is a developed skill in many instrumentalists. In the original interleaved melody
paradigm, Dowling asked participants to identify the melodies being played and found that this
was more likely to occur when pitch overlap between the two melodies was minimal. With
increased sensitivity, more pitch overlap is possible; for example, musicians are able to identify
melodies with more pitch overlap than non-musicians. Similarly, it is hypothesized that
instrumentalists should identify mistunings more accurately for their own instrument overall,
and with more pitch overlap as well.
Method
ATTENTION AFFECTS AUDITORY GROUPING
22
Participants. Participants were 15 musicians, 8 flautists and 7 violinists, recruited from
music schools and conservatoires in London and in Canada. If desired, they were entered in a
draw for one of two Amazon vouchers.
Stimuli. Melodies were two excerpts from compositions by J. S. Bach: BWV 772-786
Invention 1, mm13 and BWV 772-786, Invention 9, mm14-15.1 (only the first beat of mm15).
They are in different meters (4/4 and 3/4 respectively) and different keys (A minor and F minor
respectively), but have similar ranges (perfect 12th - octave + perfect 5th - and diminished 12th
- octave + diminished 5th - respectively) and similar median pitches (C#4 and B4 respectively).
The 4/4 melody was played on a violin and the 3/4 melody on a flute.
A violinist and a flautist were recorded using a Shure SM57 microphone, recorded into
Logic and exported as CD quality audio files. These original recordings were verified by a
separate violinist and flautist for good tuning and corrections to tuning were made using
Melodyne Editor by Celemony. Melodies were recorded at notated pitch and for every
necessary transposition to create each overlap condition as tuning in a solo instrument changes
slightly as a function of key, especially in Baroque music (just intonation).
Using Melodyne Editor, 50 cent sharp mistunings were inserted. Each trial contained
either zero, one or two mistunings. The location of each mistuning is presented in Table 2.
Though it is recognized that sharp or flat tuning may be perceived differently and depends on
the context (Fujioka, Trainor, Ross, Kakigi, & Pantev, 2005), only one direction was used here
for simplicity. The tempo and note length of the melodies were quantized, and the melodies
interleaved, so that the onset of the first note of the second melody fell exactly between the
onsets of the first and second notes of the first melody, the second between the second and the
third, and so on.
1
2
3
4
5
6
7
8
9
10
11
12
Flute
Flute
1.3 / 3.3
Flute / Violin
Flute
2.4 / 4.2
Violin / Violin
Flute
Violin
None
3.1
2.2
None
Violin
Flute
Violin
Violin
Flute
1.4 / 2.1
Flute / Violin
Flute
1.2 / 3.4
Violin / Flute
Violin
3.2 / 4.3
Flute / Flute
Violin
None
None
Flute
2.4 / 4.1
Flute / Flute
Flute
2.3 / 4.1
Violin / Violin
Violin
Flute
Flute
Flute
Both
Flute
Violin
Violin
Flute
Violin
3.2 / 4.2
Violin / Flute
Violin
Both
5th
5th
2nd
2nd
2nd
2nd
3rd
3rd
3rd
3rd
5th
5th
5th
5th
-
-
ATTENTION AFFECTS AUDITORY GROUPING
23
Table 2. Experimental design: details of metrical and instrument location of mistunings (where there
are two mistunings, these are separated by a backslash), the higher melody, where attention was
directed and the amount of pitch overlap between the mean pitch of the two melodies for each trial,
including practice and control trials.
Trial
Location
(metrical)
Location
Top
Attentional
Pitch overlap
(instrument)
melody
focus
Practice 1
4.1
Violin
Violin
Violin
Practice 2
1.3 / 2.4
Flute / Flute
Flute
Control 1
Control 2
2.3
3.1
Violin
Flute
-
-
-
-
Twelve experimental trials were created, along with two practice trials and two control
trials. Five variables were manipulated: metrical mistuning location, instrumental mistuning
location, top melody, attentional focus and pitch overlap. The mistunings were either on strong
or weak beats; location is indicated by beat (first number) and subdivision (second number) i.e.
4.2 = beat 4, second subdivision (sixteenth note). The mistunings were either in the violin or
the flute melody, the top (also the first tone heard) melody was either the violin or the flute
melody and the participants' focus was directed at either the violin melody, the flute melody,
or both. Pitch overlap was either a 2nd, a 3rd or a 5th, where the distance between the central (in
ATTENTION AFFECTS AUDITORY GROUPING
24
terms of range) pitches of each melody matched these intervals. The instrumental mistuning
location, top melody and attentional focus were manipulated so that they sometimes match and
sometimes do not (i.e. the mistuning may not be in the same melody to which the participant
is asked to attend). This was intended to assess whether a mistuning in the non-attended
melody influences identification of mistunings in the attended melody.
The control trials were single melodies, designed to ensure that participants were able
to detect mistunings in a simpler listening situation. In a pilot study, a 50 cent mistuning in a
single melody was always detected.
Procedure. This experiment was carried out online, using the survey tool Qualtrics.
Once presented with the information sheet and detailed instructions, participants could give
informed consent. The two original melodies (with no mistunings) were both presented for
participants via SoundCloud to familiarize themselves with the tunes, and in every subsequent
trial in case participants wanted to refresh their memory. Each page of the survey contained
the two original melodies, the current trial (also via SoundCloud) and a click track. Participants
clicked on the beats where they heard a mistuning; this was set up using Qualtrics' hot spot
tool. There was one click track for trials where focus was on one instrument and two, stacked
vertically and labelled with the corresponding instrument, when participants were instructed to
listen to both (see Figure 6). The word 'none' under the click track was also a selection option
if participants detected no mistuning.
Participants started with two
practice trials, always in the same
order. Then, trial 11 was always
Figure 6. Single (left) and double (right) click tracks
presented to participants alongside the relevant audio
files.
presented first because it was one
of the trials with the least amount
ATTENTION AFFECTS AUDITORY GROUPING
25
of overlap (and, therefore, presumably easier) and all other trials followed in random
presentation. Finally, the two control trials were presented, always in the same order.
Participants finally selected their primary instrument, either violin or flute, and had the
option to submit their email address for the Amazon voucher draw.
Results
Initial inspection of the data showed a high rate of false alarms. Participants were first
screened by performance on the control trials; only participants who had correctly identified
the mistunings in both control trials, without false alarms, were included in analysis. This left
12 participants; 6 violinists and 6 flautists.
A mixed effects binomial logistic regression was performed, with musicianship
(violinist or flautist), metrical mistuning location, instrumental mistuning location, top melody,
attentional focus and pitch overlap as predictors for accuracy and random intercepts on
participants. Accuracy was simply defined by the number of correctly identified mistunings.
Only attentional focus and top melody were strong significant predictors, z (1) = -4.85 and z
(1) = 3.65 respectively, both p < 0.01 while instrumental mistuning location was moderately
significant, z (1) = 2.15, p = .03. There were no significant interactions (see Table 3 for details).
Accuracy when attention was directed to the violin line was highest, at .28 [.22, .35], to
the flute line was .25 [.20, .32] and to both was lowest, at .10 [.06, .17]. Accuracy when the
violin line was on top was lower than when the flute was on top, at .17 [.13, .22] and .29 [.24,
.34] respectively. Accuracy when the mistuning was in the violin line was .38 [.31, .46] and in
the flute line was .20 [.16, .25].
Discussion
The interleaved melody paradigm was designed to examine whether musical training
on a particular instrument increases timbral sensitivity to that instrument, using mistuning
ATTENTION AFFECTS AUDITORY GROUPING
26
Table 3. Details of the mixed effects binomial logistic regression, where accuracy is predicted by
fixed effects as described in the text and participant number as random effects on intercepts.
Predictor
Intercept
Musicianship
Metrical mistuning location
Instrumental mistuning location
Top melody
Attentional focus
Pitch Overlap
Random Intercepts
Participant
Estimate
p-value
-2.43
< .01
0.23
0.30
0.47
0.87
-0.94
0.04
.32
.07
.03
< .01
< .01
.61
Variance
0.02
detection in real melodies rather than rhythm judgements for artificial tone sequences, as in
Experiment 1a. Contrary to the hypothesis, results converge with Experiment 1a and 1b:
musical training does not have an influence on timbre sensitivity, and support the alternate
hypothesis proposed in Experiment 1a: attention influences perception. Similarly to the
hypothetical orchestral line transcription described before, in this paradigm detection of
mistunings, which first requires the separation of the melody from its context, did not depend
on the instrument in which the mistuning appeared, but rather which line the listener's attention
was directed to. The idea that attention influences perception is certainly not new (Carlyon,
Cusack, Foxton, & Robertson, 2001; Dowling, 1990; Snyder, Gregg, Weintraub, & Alain,
2012; Spielmann, Schröger, Kotz, & Bendixen, 2014) but the above results suggest that
attentional focus is more important than specific musical training in driving auditory stream
segregation, leading to the lack of effect of specific instrumental musical training.
ATTENTION AFFECTS AUDITORY GROUPING
27
General Discussion
Though previous literature would suggest that instrumentalists are more sensitive to
their own instrument's timbre (Margulis et al., 2009; Pantev et al., 2001), behavioural evidence
for this claim was not found here. We instead propose that behaviour is guided by attention
rather than musical training, consistent with literature exploring the effects of attention on
auditory scene analysis (Andrews & Dowling, 1991; Bigand, McAdams, & Forêt, 2000; Jones,
Alford, Bridges, Tremblay, & Macken, 1999; Macken, Tremblay, Houghton, Nicholls, &
Jones, 2003). This interpretation was supported in both Experiments 1b and 2. In Experiment
1b, trials where rhythm captured attention more often also resulted in a segregated percept
(from post-hoc analysis). In Experiment 2, attentional focus was a predictor of response
accuracy for identifying mistunings, where accuracy depends on participants successfully
streaming the relevant melody. Furthermore, performance when participants were asked to
identify mistunings in both lines at once was particularly poor, highlighting the importance of
attentional focus for successful task completion.
It is interesting to consider why the present results diverge from those found in
cognitive-neuroscientific studies which have found instrument-specific effects of musical
training. It may be that methods such as EEG, MEG and fMRI provide more sensitive measures
that are capable of picking up on small effects of instrument-specific training which are not
expressed in behavioural measures such as those used here. Greater sensitivity of neural over
behavioural measures has been observed in research on processing dissonant and mistuned
chords (Brattico et al., 2009) and harmonic intervals varying in dissonance (Schön et al., 2005).
Alternatively, it may be that the instrument-specific effects observed in previous research were
actually driven by greater attention to an instrumentalists' own instrument. Further research is
required to disentangle these alternative accounts.
ATTENTION AFFECTS AUDITORY GROUPING
28
Let us now look at the ABA- paradigm more closely. Despite listeners most often
initially perceiving maximally different timbres as segregated, there is still a fairly large
proportion of trials heard as integrated. This was explained above by timbre similarity, but it
may not be the only factor; based on personal listening, and participant feedback, the stimuli
are clearly bistable, suggesting that timbre alone may not be enough to fully segregate two
sounds played with same pitch, loudness and length. In a musical sense, this is very useful and
is often employed by composers wanting to create instrumental chimerae or even simply
writing passages involving the entire sections of the orchestra playing the same line. This
suggests that timbre is a less important feature in perception of polyphonic music, with pitch,
rhythm and loudness taking precedence. Relative importance of these four parameters for
auditory streaming could be evaluated by combining parameters to see which causes streaming
first. Some questions concerning salience and combining musical parameters in a streaming
paradigm have been investigated (Dibben, 1999; Prince, Thompson, & Schmuckler, 2009; van
Noorden, 1975) but a clear map of relationships between parameters has not yet been
established, largely due to the complexity of polyphonic music. It might be a different situation
for non-musical, or 'environmental' sounds, and both would be interesting to investigate
further.
According to Horváth et al. (2001), predictive representations for both galloping and
even patterns are held in parallel, but this was only tested where auditory stimuli were ignored.
This explanation relies on predictive regularity to explain auditory scene analysis, as does the
auditory event representation system model (Schröger et al., 2014). This model attempts to
explain how auditory streams are formed right from the beginning rather than through a gradual
buildup of evidence (Bregman, 1978), or through bistability (Pressnitzer, Suied, & Shamma,
2011). It builds chains of potential perceptual representations that compete for dominance; as
new sounds are fed in, certain representations are validated and others are deleted until there is
ATTENTION AFFECTS AUDITORY GROUPING
29
only one 'winner'. In terms of the ABA- paradigm, the model output is not conclusive as both
percepts are valid and stable with respect to the model and so it does not help explain how
auditory streams are formed in this paradigm. It seems attention is necessary to explain the
creation of auditory streams when the stimuli could be interpreted in multiple ways, as is
demonstrated in the experiments presented above.
To summarize, two streaming paradigms designed to investigate timbre sensitivity
show that task performance depends not on sensitivity to a particular timbre due to instrument-
specific musical training, but on allocation of attention to the appropriate, in the case of
Experiment 2, or simply the chosen, in the case of Experiment 1a, auditory object.
Acknowledgements
Thank you to Lauren Stewart who provided access to MUMS and to Christoph Reuter and
Andrew Staniland for help with constructing the morphing timbre stimuli.
ATTENTION AFFECTS AUDITORY GROUPING
30
References
Alain, C., Arnott, S. R., & Picton, T. W. (2001). Bottom–up and top–down influences on
auditory scene analysis: Evidence from event-related brain potentials. Journal of
Experimental Psychology: Human Perception and Performance, 27(5), 1072–1089.
http://doi.org/10.1037/0096-1523.27.5.1072
Alluri, V., & Toiviainen, P. (2010). Exploring Perceptual and Acoustical Correlates of
Polyphonic Timbre. Music Perception: An Interdisciplinary Journal, 27(3), 223–242.
http://doi.org/10.1525/mp.2010.27.3.223
Andrews, M. W., & Dowling, W. J. (1991). The development of perception of interleaved
melodies and control of auditory attention. Music Perception, 8(4), 349–368.
Baumann, S., Koeneke, S., Schmidt, C. F., Meyer, M., Lutz, K., & Jancke, L. (2007). A
network for audio–motor coordination in skilled pianists and non-musicians. Brain
Research, 1161, 65–78. http://doi.org/10.1016/j.brainres.2007.05.045
Bendixen, A., Denham, S. L., & Winkler, I. (2014). Feature Predictability Flexibly Supports
Auditory Stream Segregation or Integration. Acta Acustica United with Acustica,
100(5), 888–899. http://doi.org/10.3813/AAA.918768
Bigand, E., McAdams, S., & Forêt, S. (2000). Divided attention in music. International
Journal of Psychology, 35(6), 270–278. http://doi.org/10.1080/002075900750047987
Brattico, E., Pallesen, K.J., Varyagina, O., et al. (2009). Neural discrimination of
nonprototypical chords in music experts and laymen: an MEG study. Journal of
Cognitive Neuroscience 21(11), 2230–44. doi:10.1162/jocn.2008.21144.
Bregman. (1978). Auditory Streaming is Cumulative. Journal of Experimental Psychology:
Human Perception and Performance, 4(3), 380–387.
Bregman, A. S. (1990). Auditory Scene Analysis: The Perceptual Organization of Sound.
MIT Press.
Burunat, I., Brattico, E., Puoliväli, T., Ristaniemi, T., Sams, M., & Toiviainen, P. (2015).
Action in Perception: Prominent Visuo-Motor Functional Symmetry in Musicians
during Music Listening. PLoS ONE, 10(9), e0138238.
http://doi.org/10.1371/journal.pone.0138238
Caclin, A., McAdams, S., Smith, B. K., & Winsberg, S. (2005). Acoustic correlates of timbre
space dimensions: a confirmatory study using synthetic tones. The Journal of the
Acoustical Society of America, 118(1), 471–482.
Carlyon, R. P., Cusack, R., Foxton, J. M., & Robertson, I. H. (2001). Effects of attention and
unilateral neglect on auditory stream segregation. Journal of Experimental Psychology:
Human Perception and Performance, 27(1), 115–127. http://doi.org/10.1037/0096-
1523.27.1.115
Cumming, G. (2012). Understanding the New Statistics: Effect Sizes, Confidence Intervals,
and Meta-analysis. Routledge.
ATTENTION AFFECTS AUDITORY GROUPING
31
Cumming, G. (2013). The New Statistics Why and How. Psychological Science,
0956797613504966. http://doi.org/10.1177/0956797613504966
Deike, S., Heil, P., Böckmann-Barthel, M., & Brechmann, A. (2012). The Build-up of
Auditory Stream Segregation: A Different Perspective. Frontiers in Psychology, 3, 461.
http://doi.org/10.3389/fpsyg.2012.00461
Denham, S. L., & Winkler, I. (2006). The role of predictive models in the formation of
auditory streams. Journal of Physiology, Paris, 100(1-3), 154–170.
http://doi.org/10.1016/j.jphysparis.2006.09.012
Dibben, N. (1999). The Perception of Structural Stability in Atonal Music: The Influence of
Salience, Stability, Horizontal Motion, Pitch Commonality, and Dissonance. Music
Perception: An Interdisciplinary Journal, 16(3), 265–294.
http://doi.org/10.2307/40285794
Dowling, W. J. (1990). Expectancy and attention in melody perception. Psychomusicology: A
Journal of Research in Music Cognition, 9(2), 148–160.
http://doi.org/10.1037/h0094150
Drost, U. C., Rieger, M., Brass, M., Gunter, T. C., & Prinz, W. (2005). When hearing turns
into playing: Movement induction by auditory stimuli in pianists. The Quarterly
Journal of Experimental Psychology Section A, 58(8), 1376–1389.
http://doi.org/10.1080/02724980443000610
Drost, U. C., Rieger, M., & Prinz, W. (2007). Instrument specificity in experienced
musicians. The Quarterly Journal of Experimental Psychology, 60(4), 527–533.
http://doi.org/10.1080/17470210601154388
Esber, G. R., & Haselgrove, M. (2011). Reconciling the influence of predictiveness and
uncertainty on stimulus salience: a model of attention in associative learning.
Proceedings of the Royal Society of London B: Biological Sciences, 278(1718), 2553–
2561. http://doi.org/10.1098/rspb.2011.0836
François, C., Jaillet, F., Takerkart, S., & Schön, D. (2014). Faster Sound Stream
Segmentation in Musicians than in Nonmusicians. PLoS ONE, 9(7), e101340.
http://doi.org/10.1371/journal.pone.0101340
Fujioka, T., Trainor, L. J., & Ross, B. (2008). Simultaneous pitches are encoded separately in
auditory cortex: An MMNm study. NeuroReport: For Rapid Communication of
Neuroscience Research, 19(3), 361–366.
http://doi.org/10.1097/WNR.0b013e3282f51d91
Fujioka, T., Trainor, L. J., Ross, B., Kakigi, R., & Pantev, C. (2005). Automatic encoding of
polyphonic melodies in musicians and nonmusicians. Journal of Cognitive
Neuroscience, 17(10), 1578–1592. http://doi.org/10.1162/089892905774597263
Gillard, J., & Schutz, M. (2012). Improving the efficacy of auditory alarms in medical
devices by exploring the effect of amplitude envelope on learning and retention. In
Proceedings of the International Conference on Auditory Display (pp. 240–241).
Atlanta, GA.
ATTENTION AFFECTS AUDITORY GROUPING
32
Hartmann, W. M., & Johnson, D. (1991). Stream segregation and peripheral channeling.
Music Perception, 9(2), 155–183.
Horváth, J., Czigler, I., Sussman, E., & Winkler, I. (2001). Simultaneously active pre-
attentive representations of local and global rules for sound sequences in the human
brain. Brain Research. Cognitive Brain Research, 12(1), 131–144.
Jones, D., Alford, D., Bridges, A., Tremblay, S., & Macken, B. (1999). Organizational factors
in selective attention: The interplay of acoustic distinctiveness and auditory streaming
in the irrelevant sound effect. Journal of Experimental Psychology: Learning, Memory,
and Cognition, 25(2), 464–473. http://doi.org/10.1037/0278-7393.25.2.464
Kendall, R. A., & Carterette, E. C. (1991). Perceptual Scaling of Simultaneous Wind
Instrument Timbres. Music Perception: An Interdisciplinary Journal, 8(4), 369–404.
http://doi.org/10.2307/40285519
Lakatos, S. (2000). A common perceptual space for harmonic and percussive timbres.
Perception & Psychophysics, 62(7), 1426–1439.
Macken, W. J., Tremblay, S., Houghton, R. J., Nicholls, A. P., & Jones, D. M. (2003). Does
auditory streaming require attention? Evidence from attentional selectivity in short-term
memory. Journal of Experimental Psychology: Human Perception and Performance,
29(1), 43–51. http://doi.org/10.1037/0096-1523.29.1.43
Margulis, E. H., Mlsna, L. M., Uppunda, A. K., Parrish, T. B., & Wong, P. C. M. (2009).
Selective neurophysiologic responses to music in instrumentalists with different
listening biographies. Human Brain Mapping, 30(1), 267–275.
http://doi.org/10.1002/hbm.20503
McAdams, S., Winsberg, S., Donnadieu, S., De Soete, G., & Krimphoff, J. (1995). Perceptual
scaling of synthesized musical timbres: Common dimensions, specificities, and latent
subject classes. Psychological Research, 58(3), 177–192.
http://doi.org/10.1007/BF00419633
Müllensiefen, D., Gingras, B., Musil, J., & Stewart, L. (2014). The Musicality of Non-
Musicians: An Index for Assessing Musical Sophistication in the General Population.
PLoS ONE, 9(2), e89642. http://doi.org/10.1371/journal.pone.0089642
Nager, W., Kohlmetz, C., Altenmuller, E., Rodriguez-Fornells, A., & Münte, T. F. (2003).
The fate of sounds in conductors' brains: an ERP study. Cognitive Brain Research,
17(1), 83–93. http://doi.org/10.1016/S0926-6410(03)00083-1
Pantev, C., Roberts, L. E., Schulz, M., Engelien, A., & Ross, B. (2001). Timbre-specific
enhancement of auditory cortical representations in musicians. NeuroReport: For Rapid
Communication of Neuroscience Research, 12(1), 169–174.
http://doi.org/10.1097/00001756-200101220-00041
Pressnitzer, D., Suied, C., & Shamma, S. A. (2011). Auditory scene analysis: The sweet
music of ambiguity. Frontiers in Human Neuroscience, 5.
http://doi.org/10.3389/fnhum.2011.00158
ATTENTION AFFECTS AUDITORY GROUPING
33
Prince, J. B., Thompson, W. F., & Schmuckler, M. A. (2009). Pitch and time, tonality and
meter: how do musical dimensions combine? Journal of Experimental Psychology.
Human Perception and Performance, 35(5), 1598–1617.
http://doi.org/10.1037/a0016456
Repp, B. H., & Knoblich, G. (2009). Performed or observed keyboard actions affect pianists'
judgements of relative pitch. The Quarterly Journal of Experimental Psychology,
62(11), 2156–2170. http://doi.org/10.1080/17470210902745009
Rogers, W. L., & Bregman, A. S. (1998). Cumulation of the tendency to segregate auditory
streams: Resetting by changes in location and loudness. Perception & Psychophysics,
60(7), 1216–1227. http://doi.org/10.3758/BF03206171
Rose, M. M., & Moore, B. C. (2000). Effects of frequency and level on auditory stream
segregation. The Journal of the Acoustical Society of America, 108(3 Pt 1), 1209–1214.
Sauvé, S., Stewart, L., & Pearce, M.T. (2014). The Effect of Musical Training on Auditory
Grouping. In Proceedings of the ICMPC-APSCOM 2014 Joint Conference (pp. 293–
296). Seoul, Korea.
Schön, D., Ystad, S., Regnault, P., et al. (2005). Sensory Consonance: An ERP Study.
Music Perception 23(2), 105–17.
Schröger, E., Bendixen, A., Denham, S. L., Mill, R. W., Bőhm, T. M., & Winkler, I. (2014).
Predictive regularity representations in violation detection and auditory stream
segregation: From conceptual to computational models. Brain Topography, 27(4), 565–
577. http://doi.org/10.1007/s10548-013-0334-6
Shahin, A. J., Roberts, L. E., Chau, W., Trainor, L. J., & Miller, L. M. (2008). Music training
leads to the development of timbre-specific gamma band activity. NeuroImage, 41(1),
113–122. http://doi.org/10.1016/j.neuroimage.2008.01.067
Shinn-Cunningham, B. G., Lee, A. K. C., & Oxenham, A. J. (2007). A sound element gets
lost in perceptual competition. PNAS Proceedings of the National Academy of Sciences
of the United States of America, 104(29), 12223–12227.
http://doi.org/10.1073/pnas.0704641104
Singh, P. G., & Bregman, A. S. (1997). The influence of different timbre attributes on the
perceptual segregation of complex-tone sequences. Journal of the Acoustical Society of
America, 102(4), 1943–1952. http://doi.org/10.1121/1.419688
Snyder, J. S., Gregg, M. K., Weintraub, D. M., & Alain, C. (2012). Attention, awareness, and
the perception of auditory scenes. Frontiers in Psychology, 3.
http://doi.org/10.3389/fpsyg.2012.00015
Spielmann, M. I., Schröger, E., Kotz, S. A., & Bendixen, A. (2014). Attention effects on
auditory scene analysis: Insights from event-related brain potentials. Psychological
Research, 78(3), 361–378. http://doi.org/10.1007/s00426-014-0547-7
Stewart, L., Verdonschot, R. G., Nasralla, P., & Lanipekun, J. (2013). Action–perception
coupling in pianists: Learned mappings or spatial musical association of response codes
ATTENTION AFFECTS AUDITORY GROUPING
34
(SMARC) effect? The Quarterly Journal of Experimental Psychology, 66(1), 37–50.
http://doi.org/10.1080/17470218.2012.687385
Sussman, E., Ritter, W., & Vaughan, H. G. J. (1999). An investigation of the auditory
streaming effect using event-related brain potentials. Psychophysiology, 36(1), 22–34.
http://doi.org/10.1017/S0048577299971056
Sussman-Fort, J., & Sussman, E. (2014). The effect of stimulus context on the buildup to
stream segregation. Frontiers in Neuroscience, 8.
http://doi.org/10.3389/fnins.2014.00093
Tervaniemi, M. (2009). Musicians-Same or Different? Annals of the New York Academy of
Sciences, 1169(1), 151–156. http://doi.org/10.1111/j.1749-6632.2009.04591.x
Thompson, S. K., Carlyon, R. P., & Cusack, R. (2011). An objective measurement of the
build-up of auditory streaming and of its modulation by attention. Journal of
Experimental Psychology: Human Perception and Performance, 37(4), 1253–1262.
http://doi.org/10.1037/a0021925
Uhlig, M., Fairhurst, M. T., & Keller, P. E. (2013). The importance of integration and top-
down salience when listening to complex multi-part musical stimuli. NeuroImage, 77,
52–61. http://doi.org/10.1016/j.neuroimage.2013.03.051
Van Noorden, L. (1975). Temporal Coherence in the Perception of Tone Sequences.
Technical University Eindhoven, Eindhoven.
Vliegen, J., Moore, B. C. J., & Oxenham, A. J. (1999). The role of spectral and periodicity
cues in auditory stream segregation, measured using a temporal discrimination task.
Journal of the Acoustical Society of America, 106(2), 938–945.
http://doi.org/10.1121/1.427140
Yabe, H., Winkler, I., Czigler, I., Koyama, S., Kakigi, R., Sutoh, T., … Kaneko, S. (2001).
Organizing sound sequences in the human brain: The interplay of auditory streaming
and temporal integration. Brain Research, 897(1-2), 222–227.
http://doi.org/10.1016/S0006-8993(01)02224-7
ATTENTION AFFECTS AUDITORY GROUPING
35
Appendix A – Stimuli experiment 1
Timbre (original file
name)
Cello (CelA3_3.84sec)
Cello (CelC#4_2.44sec)
Cello (CelD4_2.77sec)
Cello (CelE4_2.67sec)
Cello (CelF4_2.56sec)
Cello (CelF#4_2.12sec)
Trombone
(TTbnG3_2.17sec)
Trombone
(TTbnG#3_2.22sec)
Trombone
(TTbnB3_2.54sec)
Trombone
(TTbnD4_2.81sec)
Trombone
(TTbnD#4_3.54sec)
Trombone
(TTbnF4_3.01sec)
Trumpet
(CTptG#3_6.06sec)
Trumpet
(CTptA#3_2.75sec)
Trumpet
(CTptC4_7.44sec)
Trumpet
(CTptD#4_3.54sec)
Trumpet
(CTptE4_7.42sec)
Trumpet
(CTptF#4_6.55sec)
Pitch
Length (ms)
Peak Amplitude
(dB)
Fadeout (ms)
A3
C#4
D4
E4
F4
F#4
G3
G#3
B3
D4
D#4
F4
G#3
A#3
C4
D#4
E4
F#4
114
-16
10
113
111
-12
-15
-16
-12.5
-16
-13
-12
-15
-14
10
10
ATTENTION AFFECTS AUDITORY GROUPING
36
G3
A3
A#3
B3
C4
C#4
Violin (VlnG3_8.79sec)
Violin (VlnA3_8.98sec)
Violin
(VlnA#3_8.58sec)
Violin (VlnB3_9.67sec)
Violin (VlnC4_7.69sec)
Violin
(VlnC#4_7.12sec)
114
-16
-15
-16
-12
-14
-16
10
|
1203.1076 | 1 | 1203 | 2012-03-06T00:25:23 | On the Relation between Encoding and Decoding of Neuronal Spikes | [
"q-bio.NC",
"stat.AP"
] | Neural coding is a field of study that concerns how sensory information is represented in the brain by networks of neurons. The link between external stimulus and neural response can be studied from two parallel points of view. The first, neural encoding refers to the mapping from stimulus to response, and primarily focuses on understanding how neurons respond to a wide variety of stimuli, and on constructing models that accurately describe the stimulus-response relationship. Neural decoding, on the other hand, refers to the reverse mapping, from response to stimulus, where the challenge is to reconstruct a stimulus from the spikes it evokes. Since neuronal response is stochastic, a one-to-one mapping of stimuli into neural responses does not exist, causing a mismatch between the two viewpoints of neural coding. Here, we use these two perspectives to investigate the question of what rate coding is, in the simple setting of a single stationary stimulus parameter and a single stationary spike train represented by a renewal process. We show that when rate codes are defined in terms of encoding, i.e., the stimulus parameter is mapped onto the mean firing rate, the rate decoder given by spike counts or the sample mean, does not always efficiently decode the rate codes, but can improve efficiency in reading certain rate codes, when correlations within a spike train are taken into account. | q-bio.NC | q-bio |
On the Relation between Encoding and Decod-
ing of Neuronal Spikes
Shinsuke Koyama
Department of Mathematical Analysis and Statistical Inference, Institute of Statistical
Mathematics, Tokyo, 190-8562, Japan
Keywords: Neural coding, statistical inference, asymptotic theory
Abstract
Neural coding is a field of study that concerns how sensory information is represented
in the brain by networks of neurons. The link between external stimulus and neural re-
sponse can be studied from two parallel points of view. The first, neural encoding refers
to the mapping from stimulus to response, and primarily focuses on understanding how
neurons respond to a wide variety of stimuli, and on constructing models that accurately
describe the stimulus-response relationship. Neural decoding, on the other hand, refers
to the reverse mapping, from response to stimulus, where the challenge is to reconstruct
a stimulus from the spikes it evokes. Since neuronal response is stochastic, a one-to-one
mapping of stimuli into neural responses does not exist, causing a mismatch between
the two viewpoints of neural coding. Here, we use these two perspectives to investigate
the question of what rate coding is, in the simple setting of a single stationary stimu-
lus parameter and a single stationary spike train represented by a renewal process. We
show that when rate codes are defined in terms of encoding, i.e., the stimulus parame-
ter is mapped onto the mean firing rate, the rate decoder given by spike counts or the
sample mean, does not always efficiently decode the rate codes, but can improve effi-
ciency in reading certain rate codes, when correlations within a spike train are taken
into account.
1 Introduction
Sensory and behavioral states are represented by neuronal responses. Determining
which code is used by neurons is important in order to understand how the brain car-
ries out information processing (Dayan & Abbott, 2001; Rieke et al., 1997). Coding
schemes used by neurons can be divided approximately into two categories. In rate
coding, the stimulus is mapped onto the firing rate, defined as the average number of
spikes per unit time. A variation in the number of emitted spikes in response to the
same stimulus across trials, is then considered noise. In temporal coding, on the other
hand, the stimulus is encoded in moments of the spike pattern that have higher order
than the mean (Theunissen & Miller, 1995).
While neural codes are characterized in terms of these encoding views (i.e., how
the neurons map the stimulus onto the features of spike responses), these are often in-
vestigated and validated using decoding. From the decoding viewpoint, rate coding is
operationally defined by counting the number of spikes over a period of time, with-
out taking into account any correlation structure among spikes. Any scheme based on
such an operation is equivalent to decoding under the stationary Poisson assumption,
because the number of spikes over a period of time, or the sample mean of interspike
intervals (ISIs), is a sufficient statistic for the rate parameter of a homogeneous Poisson
process. In this manuscript, a decoder based on counting the number of spikes, or on
taking the sample mean of ISIs, is labeled as "rate decoder". Similarly, temporal cod-
ing can be defined by decoding the stimulus using a statistical model with a correlation
structure between spikes (such as the MI model, introduced below). If such a decoder
improves on the performance of the rate decoder, it indicates that significant informa-
tion about the stimulus is carried in the temporal aspect of spike trains (Jacobs et al.,
2009; Pillow et al., 2005).
A simple statistical model with a correlation structure has been introduced in the
2
literature, taking the intensity function of a point process to be a product of two factors:
λ(t, s∗(t)) = φ(t)g(t − s∗(t)),
(1)
where s∗(t) represents the last spike time preceding t. This statistical model with the
intensity function (1) has been called the multiplicative intensity (MI) model by Aalen
(1978) and the multiplicative inhomogeneous Markov interval model by Kass & Ventura
(2001). φ(t) is the free firing rate, which depends only on the stimulus, and g(t− s∗(t))
is the recovery function, which describes the dependency of the last spike time preced-
ing t and hence allows the MI model to have a correlation structure between spikes.
Note that Eq.(1) becomes the intensity function of an inhomogeneous Poisson process
if the recovery function is constant in time. It has been reported that the MI model
enhances decoding performance in real data analysis (Jacobs et al., 2009), which en-
courages use of the MI model to test temporal codes.
Although neural codes can be defined in terms of either encoding or decoding, the
resulting codes generally differ from one another. Here, we investigate the relation
between the two viewpoints of neural coding in terms of rate and temporal coding
schemes. Specifically, we consider, for the sake of analytical tractability, a simple set-
ting of a single stationary stimulus parameter and a single stationary spike train rep-
resented by a renewal process, and investigate the extent to which decoders of each
scheme decode neural codes that are defined in terms of encoding. Our main claim is
that when rate codes are defined in terms of encoding, i.e., the stimulus parameter is
mapped onto the mean firing rate, the rate decoder does not always efficiently decode
the rate codes, whereas the temporal decoder can improve efficiency in reading certain
rate codes.
In order to deduce our results, we develop, in section 2, a statistical theory based on
asymptotic estimation, i.e., inference from a large number of ISIs. However, care must
be taken when results based on asymptotic analysis are translated into non-asymptotic
cases, which are certainly relevant in more realistic coding contexts. This will be ad-
dressed in section 3.
3
2 Theory
2.1 Definition of encoding and decoding
We suppose, for simplicity, that neural spikes are described by a stationary renewal
process. The response of single neurons is then described by an ISI density, p(xθ),
where x ∈ [0,∞), and θ ∈ Θ ⊂ (−∞,∞) is a one-dimensional stimulus parameter.
The renewal assumption is not exactly true for actual neural data, but often provides
a reasonable approximation (Troy & Robson, 1992). Let µ = E(xθ) be the mean
parameter, E(·θ) being the expectation with respect to p(xθ).
Consider first the rate encoding scheme. Since the early work of Adrian & Zotterman
(1926), there has been a search for a functional relationship between stimulus parame-
ters and the average firing rate, which is often described as a function of the stimulus
parameters. This motivates us to formulate rate encoding as a one-to-one mapping from
θ to µ(θ). The variation in x around the mean µ is then regarded as noise. In short, the
rate encoding scheme can formally be defined as follows:
Definition 1 If there exists a one-to-one and differentiable mapping θ 7→ µ(θ), the
scheme is rate encoding.
The assumption of differentiability in µ(θ) with respect to θ is required for analytical
purposes, but is also reasonable physiologically because it shows that a small change in
θ results in a small, smooth change in µ(θ).
Temporal encoding, on the other hand, intuitively means that the stimulus is encoded
in statistical structures of ISIs beyond the firing rate. Since it allows for many alterna-
tives, we do not explicitly define temporal encoding here, but instead give an example
below. Let p(xµ, κ) be a dispersion model, where µ is the mean and κ is the dispersion
parameter that characterizes moments of the ISIs of higher order than the mean. If the
stimulus parameter is mapped onto the dispersion parameter, θ 7→ κ(θ), this scheme
can be categorized under temporal encoding (Kostal, Lansky & Pokora, 2011).
For decoding, we assume an ISI density, q(xφ), φ ∈ Φ ⊂ (−∞,∞), which is
chosen according to the decoding schemes introduced below. We suppose that decod-
ing is performed by the maximum likelihood estimation (MLE) with q(xφ). In rate
decoding, one usually counts the number of spikes over a period of time, without taking
4
into account any dependency among spikes. This is equivalent to decoding under the
Poisson assumption, because the number of spikes is a sufficient statistic for the rate pa-
rameter of a homogeneous Poisson process. Thus, q(xφ) is taken to be the exponential
distribution, q(xφ) = φ exp(−φx), for rate decoding.
In temporal decoding, on the other hand, where a temporal dependency of spike
timing relative to the last spike is considered, we take q(xφ) to be the MI model. Here,
the ISI distribution of the MI model is constructed as follows. Since we only take into
account stationary renewal processes, the rate factor in Eq.(1) is reduced to a constant,
and then the intensity function, λ(x), of the MI model becomes
λ(x) = φg(x),
where φ ∈ [0,∞) is the free firing rate and g(x)(≥ 0) is the recovery function 1. The
ISI distribution of the MI model is then obtained as
q(xφ) = φg(x) exp[−φG(x)],
(2)
where
G(x) = Z x
0
g(u)du.
In order for the MI model to be well behaved as a decoder, we assume that the variance
of G(x) is finite. It is obvious from the factorization theorem (Schervish, 1995) that
G(x) is a sufficient statistic for φ. Note that Eq.(2) becomes an exponential distribution
with firing rate φ if g(x) = 1, x ≥ 0. The two decoding schemes are summarized as
follows:
Definition 2 In rate decoding, θ is decoded with q(xφ) being the exponential distribu-
tion via the MLE. In temporal decoding, θ is decoded with q(xφ) being the MI model
via the MLE.
We use the MI model in temporal decoding for the following reasons. First, the
inhomogeneous version of the MI model given by Eq.(1) is useful in practice, as it
can be easily fitted to data by well-established statistical methods (Kass & Ventura,
1 Since the units of λ(x) are those of firing rate (i.e., spikes per unit time), by convention, we let φ
also have units of firing rate, leaving g(x) dimensionless (Kass & Ventura, 2001).
5
2001; DiMatteo et al., 2001). In fact, Jacobs et al. (2009) demonstrated the importance
of temporal coding by using this model. Second, generalized linear models (GLMs)
(McCullagh & Nelder, 1989; Paninski, 2004; Paninski et al., 2007; Truccolo et al., 2005),
which have been used extensively for statistical analysis of neural data, include the MI
model as a special case. Specifically, the GLM corresponds to the MI model when the
spiking history term contains only the last spike and a log-link function is used (e.g.,
soft-threshold integrate-and-fire models (Paninski et al., 2008)).
In order to investigate the extent to which decoders of each scheme decode neural
codes that are defined in terms of encoding, in section 2.2, we introduce a correlation
quantity ρ2
θ given by Eq.(4), which measures decoding performance with q(xφ).
2.2 Correlation quantity
We shall assume that p(xθ) and q(xφ) satisfy the traditional regularity assumptions
needed for standard asymptotics (Schervish, 1995). We first define a correlation quan-
tity that measures a "similarity" between two models. Let
and
sp(x, θ) =
∂ log p(xθ)
∂θ
sq(x, φ) =
∂ log q(xφ)
∂φ
be the score functions of p(xθ) and q(xφ), respectively. For a given θ, the parameter
of the decoder model, φ, is taken to be a function φ(θ) of θ satisfying
E[sq(x, φ(θ))θ] = 0.
We define the square correlation coefficient ρ2
θ as
ρ2
θ ≡
Cov[sp(x, θ), sq(x, φ(θ))θ]2
Var[sp(x, θ)θ]Var[sq(x, φ(θ))θ]
=
E[sp(x, θ)sq(x, φ(θ))θ]2
JθE[sq(x, φ(θ))2θ]
,
(3)
(4)
6
where Var[·θ] and Cov[·θ] represent, respectively, the variance and the covariance with
respect to p(xθ), and Jθ is the Fisher information defined by
Jθ ≡ E[sp(x, θ)2θ].
Note that we used E[sp(x, θ)θ] = 0 in deriving the right-hand side of Eq.(4). The
θ is related to the coefficient of determinant, R2, used in
square correlation coefficient ρ2
a simple regression analysis (Rawlings et al., 1998).
θ has the following geometrical property. In a linear space of square integral func-
ρ2
tions, the inner product and norm are defined to be
hsp, sqiθ = E(spsqθ),
kskθ = hs, si1/2
θ = E(s2θ)1/2.
The square correlation coefficient is then rewritten as
ρ2
θ = h
sp(x, θ)
ksp(x, θ)k
,
sq(x, φ(θ))
ksq(x, φ(θ))ki2
θ = cos2ϕ,
(5)
where ϕ is the angle between sp(x, θ) and sq(x, φ(θ)) with respect to h,iθ. Thus, ρ2
if sq(x, φ(θ)) is parallel to sp(x, θ), while ρ2
θ = 1
θ = 0 if sq(x, φ(θ)) is orthogonal to sp(x, θ).
θ, in terms of statistical infer-
ence (Lemma 3) and information theory (Lemma 4), which will provide useful insights
In the following, we will give two interpretations of ρ2
for translating the meaning of ρ2
θ into the context of neural decoding.
2.1 Asymptotic efficiency
Let x1, x2, . . . , xn be independent and identically distributed random variables from
p(xθ), and φn = φn(x1, x2, . . . , xn) be the MLE of q(xφ) based on x1, x2, . . . , xn.
Then, φn → φ(θ) as n → ∞, where φ(θ) satisfies Eq.(3) (White, 1982). For the
inference of θ from φn, we assume that dφ(θ)/dθ 6= 0. An estimator of θ would, thus,
be transformed from φn as θn = φ−1( φn). We also assume that θn is an unbiased
estimator of θ. The performance of the unbiased estimator is evaluated by its variance,
and the ratio of it to its lower bound is called the efficiency (Schervish, 1995). The
7
following lemma holds under the above conditions.
Lemma 3 ρ2
θ gives the asymptotic efficiency of θn.
Proof: Under suitable regularity conditions, it is proven that φn is asymptotically nor-
mal (White, 1982):
√n( φn − φ(θ)) → N(0, v)
in distribution,
where
v = E[sq(x, φ(θ))2θ]E(cid:20) ∂sq(x, φ(θ))
∂φ
By the delta method (Schervish, 1995), we obtain
θ(cid:21)−2
.
(cid:12)(cid:12)(cid:12)(cid:12)
√n(θn − θ) → N(0, v/c2)
in distribution,
where
c =
dφ(θ)
dθ
= −E[sp(x, θ)sq(x, φ(θ))θ]E(cid:20) ∂sq(x, φ(θ))
dφ
(6)
(7)
θ(cid:21)−1
(cid:12)(cid:12)(cid:12)(cid:12)
is derived by differentiating Eq.(3) with respect to θ. Since the lower bound of the
asymptotic variance is given by the inverse of the Fisher information (i.e., the Cram´er-
θ /v. Using
Rao lower bound), the asymptotic efficiency is defined by the ratio c2J −1
Eqs.(6) and (7), we obtain c2J −1
✷
θ.
θ /v = ρ2
2.2 Information-theoretic quantity
We next connect ρ2
a neuron is subjected to a stimulus chosen from a probability distribution, p(θ).
θ to an information-theoretic measure. Consider a situation in which
In
information theory, the amount of information about the stimulus transferred through a
noisy channel is quantified by the mutual information (Cover & Tomas, 1991):
I = −Z p(x) log p(x)dx +Z Z p(θ′)p(xθ′) log p(xθ′)dθ′dx.
(8)
The amount of information that can be gained by decoding depends on the probability
distribution used in a decoder.
In order to introduce this information, we revisit an
information-theoretic interpretation of the mutual information. Suppose that the neuron
8
is subjected to a set of stimuli, and consider how many stimuli can be encoded in its
response. If each stimulus is encoded in a sequence of n(≫ 1) ISIs, the upper bound
on the number of stimuli that can be encoded almost error-free is enI. In decoding, if
the true model, p(xθ), is used to build a decoder, the upper bound of the number of
stimuli that can be decoded almost freely from errors is the same, enI. If, on the other
hand, the inaccurate model, q(xθ), is used, then the upper bound is typically smaller,
enI ∗, where I ∗ ≤ I was derived in Merhav et al. (1994) as
I ∗ = I ∗(β∗) = −Z p(x) logZ p(θ′)q(xφ(θ′))β ∗
dθ′dx+Z Z p(θ′)p(xθ′) log q(xφ(θ′))β ∗
(9)
dθ′dx,
with β∗ being the value that maximizes I ∗(β). Thus, the normalized quantity, I ∗/I, is
regarded as an information gain obtained by using q(xφ) in decoding. See Latham & Nirenberg
(2005); Oizumi et al. (2010) for more details and use of I ∗ in the context of neural de-
coding. The following lemma connects I ∗/I with ρ2
θ.
Lemma 4 Suppose that the mean and variance of p(θ′) are given by θ and ǫ2, respec-
tively. For ǫ ≪ 1, the information ratio is given by
I ∗
I
= ρ2
θ + O(ǫ).
(10)
Proof: For an integrable function, f (x), that is twice differentiable, it follows that
Z f (x)p(θ′)dθ′ = f (θ) +
f ′′(θ)
2
ǫ2 + O(ǫ3).
By using this, we obtain
I ∗(β) = βE[sp(x, θ)sq(x, φ(θ))θ]ǫ2 −
β2
2
E[sq(x, φ(θ))2θ]ǫ2 + O(ǫ3).
(11)
The optimal β∗ is obtained by maximizing Eq.(11) with respect to β as
β∗ =
E[sp(x, θ)sq(x, φ(θ))θ]
E[sq(x, φ(θ))2θ]
+ O(ǫ).
(12)
9
Substituting Eq.(12) into Eq.(11) leads to
I ∗ ≡ I ∗(β∗) =
E[sp(x, θ)sq(x, φ(θ))θ]2
2E[sq(x, φ(θ))2θ]
ǫ2 + O(ǫ3).
(13)
In the same manner, the mutual information (8) is evaluated as
I =
1
2
Jθǫ2 + O(ǫ3).
From Eqs.(13) and (14), we obtain Eq.(10).
2.3 Properties of ρ2
θ
Lemma 5 ρ2
θ has the following properties:
(14)
✷
θ ≤ 1.
i) 0 ≤ ρ2
ii) ρ2
θ achieves unity when the MLE of q(xφ) is a complete sufficient statistic for θ2.
i) is obvious from Eq.(5). To prove ii), let φ = φ(x) denote the MLE of q(xφ).
Proof:
Let f1( φ) and f2( φ) be unbiased estimators of θ. Then, we have E[f1( φ)−f2( φ)θ] = 0
for all θ ∈ Θ. Since φ is a complete statistic, it follows that f1( φ) = f2( φ), a.s. [pθ]
for all θ. Thus, all unbiased estimators of θ, which are functions of φ, are equal, a.s.
[pθ]. Now, suppose that there is an unbiased estimator f (x) of θ with finite variance,
and define
θ = E[f (x)θ, φ],
(15)
which forms an estimator of θ, since φ is sufficient for θ and thus the conditional ex-
pectation given φ does not depend on θ. θ is unbiased because
E(θθ) = E[E[f (x)θ, φ]θ] = E[f (x)θ] = θ.
Thus, θ defined by Eq.(15) is equal with the one defined in Lemma 3, a.s. [pθ]. It
2 A statistic φ is complete if for every measurable, real-valued function f , E[f ( φ)θ] = 0 for all θ ∈ Θ
implies f ( φ) = 0 almost surely with respect to p(xθ) (denoted by 'a.s. [pθ]') for all θ. An interpretation
of completeness for a sufficient statistic is that it makes the ancillary part of the data independent of φ
(Lehmann, 1981).
10
follows that
Var(θθ) = E[(θ − θ)2θ]
= E[(E[f (x)θ, φ] − θ)2θ]
= E[E[f (x) − θθ, φ]2θ]
≤ E[E[(f (x) − θ)2θ, φ]θ]
= E[(f (x) − θ)2θ]
= Var[f (x)θ],
where the inequality follows from Jensen's inequality. Particularly, if we take f (x) to
be an asymptotically efficient estimator (e.g., the MLE of p(xθ)), Var[f (x)θ] achieves
the lower bound, J −1
θ gives the asymptotic
efficiency of θ.
, which completes the proof of ii) because ρ2
θ
✷
From the interpretations and properties given in Lemmas 3, 4 and 5, ρ2
θ can be
used as a measure of decoding performance of q(xφ) when the true model is given by
p(xθ). We say that q(xφ) efficiently decodes θ if ρ2
θ > 0, θ is asymptotically
decodable with q(xφ), whereas if ρ2
θ = 0, θ is not decodable with q(xφ).
θ = 1. If ρ2
2.3 Results
By using ρ2
θ defined in Eq.(4), we now investigate the extent to which the decoders of
each scheme decode rate and temporal codes.
Theorem 6 In rate encoding, if the sample mean is a complete sufficient statistic for
θ = 1 with q(xφ) being the exponential
µ, the rate decoder efficiently decodes θ (i.e., ρ2
distribution).
Proof: Since µ(θ) is a one-to-one mapping, the sample mean is sufficient for θ. On
the other hand, the MLE of the rate parameter of the exponential distribution is given
by the sample mean. Therefore, the theorem follows from Lemma 5 ii).
✷
Theorem 7 Let q(xφ) be the MI model given by (2). Either in rate encoding or in
temporal encoding,
11
i) θ is efficiently decoded (i.e., ρ2
θ = 1) if G(x) is a complete sufficient statistic for
θ.
ii) θ is asymptotically decodable (i.e., ρ2
θ > 0) if ∂E[G(x)θ]
∂θ
6= 0.
Proof: From Eq.(2), the MLE of q(xφ) is given by φ = G(x)−1. Thus, i) follows
from Lemma 5 ii). For the proof of ii), we rewrite (4) as follows.
E[sp(x, θ)sq(x, φ)θ] = Z ∂ log p(xθ)
∂θ
sq(x, φ)p(xθ)dx
=
∂
∂θ Z sq(x, φ)p(xθ)dx
E[sq(x, φ)θ]
∂
E[G(x)θ],
∂θ
=
∂
∂θ
= −
where we used Eq.(2) to obtain the last equation. Inserting φ = φ(θ) into the above
equation leads to
E[sp(x, θ)sq(x, φ(θ))θ] = −
∂
∂θ
E[G(x)θ].
(16)
Through direct calculation, we also obtain
E[sq(x, φ(θ))2θ] = E{(G(x) − E[G(x)θ])2θ} ≡ Var[G(x)θ].
(17)
Substituting Eqs.(16) and (17) into Eq.(4), ρ2
θ is written as
θ = n ∂
ρ2
Therefore, ρ2
θ > 0 holds if ∂E[G(x)θ]
∂θ
∂θ E[G(x)θ]o2
JθVar[G(x)θ]
6= 0.
.
(18)
✷
The results and their consequences are summarized as follows.
1) In rate encoding, if the sample mean is a complete sufficient statistic for µ, the
rate decoder efficiently decodes the rate code.
2) If, on the other hand, the sample mean is not sufficient for µ in rate encoding, but
θ for the temporal decoder is larger than that
G(x) is chosen so that the value of ρ2
12
for the rate decoder, the temporal decoder can decode the rate code with greater
efficiency than the rate decoder.
3) In temporal encoding, if G(x) is chosen so that ∂E[G(x)θ]
6= 0, the temporal code
is asymptotically decodable with the temporal decoder. Particularly, if G(x) can
∂θ
be taken to be a complete sufficient statistic for θ, the temporal decoder decodes
the temporal code efficiently.
In the following, we will give three examples that illustrate the above consequences.
We first give an example illustrating consequence 2), where the rate decoder is not
efficient for decoding a rate code, and the temporal decoder achieves greater efficiency
than the rate decoder.
Example 8 Let p(xµ, κ) be a log-normal distribution:
(log x
1
p(xµ, κ) =
x√2πκ
exp(cid:20) −
2 )2
µ + κ
2κ
(cid:21).
(19)
See Levine (1991) for modeling the stochastic nature of ISIs with the log-normal distri-
bution. Suppose that the stimulus is encoded in µ, i.e., rate encoding. The sample mean
is not a sufficient statistic for µ of the distribution, which implies that the rate decoder
θ for the rate decoder is derived in Appendix A.1
does not decode efficiently. Indeed, ρ2
as
ρ2
θ =
κ
eκ − 1
.
(20)
θ → 0 if κ → ∞, as the distribution becomes more skewed and has a longer right-hand
ρ2
tail.
Instead of the rate decoder, consider using the temporal decoder with the MI model's
recovery function being
g(x) =
(αx/τ )α−1e−αx/τ
Γ(α, αx/τ )
,
(21)
where Γ(α, z) is the incomplete gamma function:
Γ(α, z) = Z ∞
z
tα−1e−tdt.
In Eq.(21), α(> 0) determines the shape of g(x) (i.e., g(x) ∼ xα−1 near x = 0), and τ
13
(a)
1.5
)
x
(
g
1
0.5
(b)
0.35
2
ρ
θ
0.3
0.25
0.2
0.15
0.1
0.05
α =0.8
α=1
α=3
α=10
0
0
τ
2τ
3τ
4τ
5τ
0
0
0.5
1
x
1.5
τ/µ
2
2.5
3
Figure 1:
(a) The shape of the recovery function (21) for α = 0.8, 1, 3 and 10. (b)
θ of temporal decoding as a function of τ /µ in Example 10. The value of the shape
ρ2
parameter of the gamma distribution was taken to be κ = 5. ρ2
θ reaches its maximum
when µ ≈ τ.
represents the correlation timescale between successive spikes. Figure 1(a) depicts the
shape of g(x) for several values of α. It is shown in Appendix A.2 that for each κ > 0,
θ ≈ 1 as closely as
possible by taking τ to be large enough and α to be small enough, because the sufficient
the temporal decoder with the recovery function (21) achieves ρ2
statistic G(x) for the parameter φ of the MI model approximates to log x, which is a
sufficient statistic for the mean parameter of the log-normal distribution.
✷
The next example illustrates consequence 1), i.e., a situation in which the sample
mean is sufficient for the mean parameter.
Example 9 Suppose that p(xµ, κ) is a gamma distribution with the mean µ and the
shape parameter κ:
p(xµ, κ) =
κκxκ−1e−κx/µ
µκΓ(κ)
,
(22)
where Γ(κ) is the gamma function. The gamma distribution has been used to describe
the stochastic nature of ISIs, and its information-theoretic properties have been stud-
ied (Ikeda & Manton (2009) and references therein). Also, suppose that the stimu-
lus is mapped onto µ (i.e., rate encoding). It is easy to see that the sample mean is
a complete sufficient statistic for µ, and thus the rate decoder efficiently decodes the
θ = 1), regardless of the value of κ. Note that the variance of the sample
mean achieves the Cram´er-Rao lower bound even with a finite sample size, because the
stimulus (ρ2
14
gamma distribution is an exponential family distribution (Schervish, 1995). Thus, nei-
ther the temporal decoder nor the gamma distribution (i.e., the true model) is necessary
for efficient decoding even with a finite sample size.
✷
The last example illustrates consequence 3).
Example 10 Consider that the true ISI distribution is given to be the gamma distribu-
tion (22), and that the stimulus is encoded in κ (i.e., temporal encoding). For temporal
decoding, let us take the recovery function of the MI model to be Eq.(21). From a direct
calculation (Appendix A.3), ρ2
θ is expressed as
ρ2
θ = (cid:8) ∂
∂κ E[log Γ(α, α µ
JκVar[log Γ(α, α µ
τ x)κ](cid:9)2
τ x)κ]
,
(23)
value of the parameters, (κ, µ/τ ). The value of α was taken so as to maximize ρ2
where E[·κ] and Var[·κ] are taken with respect to p(xµ = 1, κ). Note that ρ2
function of the dimensionless parameter, µ/τ. ρ2
θ is a
θ was numerically computed for each
θ for
θ as a function of τ /µ. It is seen from
θ takes its maximum near τ /µ ≈ 1, which indicates that the MI model
decodes best when the mean ISI of the true model, µ, matches the correlation timescale
each value of parameters. Figure 1(b) depicts ρ2
this figure that ρ2
of the MI model, τ.
3 Discussion
✷
Our main results are summarized as follows. First, the rate decoder efficiently decodes
rate codes if and only if the sample mean is a sufficient statistic for the mean parameter
of the true model. Second, the temporal decoder improves on the performance of the
rate decoder by a) decoding temporal codes that the rate decoder fails to read, and b)
achieving greater efficiency in decoding certain rate codes.
These results suggest that rate codes in stationary spike trains, which are defined
as the mapping from the stimulus to the mean firing rate, can further be divided into
two subcategories when the concept of sufficiency is taken into consideration: one is a
"strong" rate code, in which the sample mean is a sufficient statistic for decoding, and
the other is a "weak" rate code, in which the sample mean is not sufficient. We should
15
notice that spike count decoding matches the strong form of rate encoding, but not weak
form.
How can decoding results inform us whether or not rate coding is being used? In or-
der to answer this question in the context of neuronal data analysis, one may decode the
stimulus with rate and temporal decoders, and compare their decoding performances
(Jacobs et al., 2009). This procedure tells us whether or not the sample mean is suffi-
cient for decoding the stimulus: if the rate decoder performs as well as the temporal
decoder, then the sample mean is sufficient; if it does not, then the sample mean is not
sufficient. In terms of the original question of whether rate coding is being used, only
in the former case can we translate the decoding result into "strong" rate encoding; in
the latter case, we cannot conclude which scheme, "weak" rate encoding or temporal
encoding, is being used.
The key quantity in our theoretical analysis is the square correlation coefficient,
θ, which quantifies neural decoding performance.
ρ2
unnormalized quantity of ρ2
θ:
It is worth pointing out that the
J ∗
θ ≡ ρ2
θJθ =
E[sp(x, θ)sq(x, φ(θ))θ]2
E[sq(x, φ(θ))2θ]
,
θ has similar properties to Jθ; (i) J ∗
θ
can be regarded as a generalization of the Fisher information, Jθ, in the sense that J ∗
θ
−1 gives the
becomes Jθ if q(xφ) = p(xθ). J ∗
−1 gives that of p(xθ)
asymptotic variance of the MLE of q(xφ) (Lemma 3) as Jθ
(Schervish, 1995), and (ii) J ∗
θ appears in the leading term of the information, I ∗, of the
decoder with q(xφ) (Lemma 4), as Jθ does in the mutual information with the limit of
small input power (Kostal, 2010). As Jθ has been used to measure encoding accuracy
(for review, see Dayan & Abbott, 2001, chap. 3), J ∗
θ is used to measure the performance
of neural decoders.
It must be noted that our analysis is based on asymptotic theory, which assumes a
θ , give
the lower bounds of the variance of unbiased estimators, but generally do not corre-
large sample size. The inverse of the Fisher information and its generalization, J ∗
spond to the mean squared error of the estimators with a finite sample size, except for
special cases of exponential family distributions. Thus, the results based on asymptotic
analysis may not be justified for non-asymptotic cases. (Bethge et al. (2002) examined
16
this point in the context of population coding.) Especially, decoding using the "wrong"
model may severely compromise the accuracy of decoding in non-asymptotic cases.
θ provides correct results
in terms of minimum mean squared error when the asymptotic results are translated into
One therefore has to check carefully whether analysis using ρ2
non-asymptotic cases.
Our simple setting of stationary and renewal assumptions does not account for two
aspects of neuronal spikes that are relevant for neural coding. First, actual spike trains
exhibit nonstationarity due to both, the dynamics of the stimulus and the nature of the
neural encoding processes such as adaptation. Rate encoding for this case is generalized
to the scheme in which the stimulus is mapped onto a time-dependent firing rate, or, the
marginal intensity function. Then the question we would like to address is whether
reasonable estimates of the firing rate (e.g., based on spline models or histograms), are
asymptotically sufficient for decoding the stimulus, which may require more mathe-
matically careful treatment to be proven. Second, higher-order serial dependencies in
sequences of ISIs, for which the MI model (1) can not account, would certainly be rele-
vant for neural coding. Accordingly, temporal encoding is generalized to the scheme in
which the stimulus is mapped onto the higher-order serial dependencies. For temporal
decoding, the MI model can be generalized by taking the recovery function to depend
on the whole spiking history, rather than simply on the last spike. Taking into consid-
eration these two extensions, we suspect that our results summarized at the beginning
of the Discussion still hold. It would be interesting to examine the relation between en-
coding and decoding in a more realistic setting, for instance, with biophysically realistic
neuron models.
A Appendix: details of derivations
A.1 Derivation of equation (20)
Taking the parameter µ = µ(θ) and inserting G(x) = x into (18), ρ2
θ for the rate
becomes
θ = n ∂
ρ2
∂µE(xθ)o2
JµVar(xθ)
.
17
For the log-normal distribution (19), we have E(xθ) = µ, Var(xθ) = µ2(eκ − 1), and
Jµ = −E(cid:20) ∂2
∂µ2 log p(xµ, κ)(cid:12)(cid:12)(cid:12)(cid:12)
θ(cid:21) =
1
κµ2 .
Using these, we obtain Eq.(20).
A.2 Temporal decoding for the log-normal distribution
Here, we show that the temporal decoder with recovery function (21) can achieve ρ2
θ ≈
1 as closely as possible in Example 8. Taking the parameter µ = µ(θ) in (18), we have
θ = n ∂
ρ2
∂µ E[G(x)θ]o2
JµVar[G(x)θ]
,
where
G(x) =
= log Γ(α) −
for τ ≫ 1. Then,
τ
αx
α(cid:26) log Γ(α) − log Γ(cid:16)α,
τ (cid:17)(cid:27)
αα−1xα
Γ(α)τ α + O(τ −α−1),
∂E[G(x)θ]
∂µ
αα−1
Γ(α)τ α
∂E(xαθ)
∂µ
= −
+ O(τ −α−1).
A similar calculation leads to
Var[G(x)θ] = (cid:18) αα−1
Γ(α)τ α(cid:19)2
Var(xαθ) + O(τ −2α−1).
For the log-normal distribution (19), we also have Jµ = 1/(κµ2) and E(xmθ) =
µmeκ(m−1)m/2, m > 0. Thus, we obtain
ρ2
θ =
κα2
eκα2 − 1
+ O(τ −1).
Therefore, limτ →∞,α→0 ρ2
taking τ to be large enough and α to be small enough.
θ = 1, that is, we can achieve ρ2
θ ≈ 1 as closely as possible by
18
A.3 Derivation of equation (23)
Eq.(21) is rewritten as
g(x) = −
Then, we get
∂Γ(α,αx/τ )
τ
α
Γ(α, αx/τ )
∂x
τ
α
∂
∂x
= −
log Γ(α, αx/τ ).
G(x) = Z x
0
g(u)du =
τ
α
[log Γ(α) − log Γ(α, αx/τ )],
where we used Γ(α, 0) = Γ(α). Taking κ = κ(θ) in Eq.(18), we obtain
θ = n ∂
∂κ E[log Γ(α, αx/τ )θ](cid:9)2
JκVar[log Γ(α, αx/τ )θ]
Thus, the scaling property of the gamma distribution leads to Eq.(23).
∂κ E[G(x)θ]o2
JκVar[G(x)θ]
ρ2
= (cid:8) ∂
.
References
Aalen, O. (1978). Nonparametric inference for a family of counting processes. The
Annals of Statistics, 6, 701 -- 726.
Adrian, E. D. & Zetterman, Y. (1926). The impulse produced by sensory nerve endings:
Part II: The response of a single end organ. Journal of Physiology, 61, 151 -- 171.
Bethge, M., Rotermund, D. & Pawelzik, K. (2002). Optimal short-term population cod-
ing: when Fisher information fails. Neural Computation, 14, 2317 -- 2351.
Cover, T. & Thomas, J. (1991). Elements of Information Theory. New York: Winley.
Dayan, P. & Abbott, L. F. (2001). Theoretical Neuroscience. Cambridge: MIT Press.
DiMatteo, I., Genovese, C. R., & Kass, R. E. (2001). Bayesian curve-fitting with free-
knot splines. Biometrika, 88, 1055 -- 1071.
Ikeda, S. & Manton, J. H. (2009). Capacity of a single spiking neuron channel. Neural
Computation, 21, 1714 -- 1748.
19
Jacobs, A. L., Fridman, G., Douglas, R. M., Alam, N. M., Latham, P. E., Prusky, G. T.,
& Nirenberg, S. (2009). Ruling out and ruling in neural codes. Proceedings of the
National Academy of Sciences, 106, 5936 -- 5941.
Kass, R. E. & Ventura, V. (2001). A spike-train probability model. Neural Computation,
13, 1713 -- 1720.
Kostal, L. (2010). Information capacity in the weak-signal approximation. Physical
Review E, 82, 026115.
Kostal, L., Lansky, P., & Pokora. O. (2011). Variability measures of positive random
variables. PLoS, 6, e21998.
Latham, P. E., & Nirenberg, S. (2005). Synergy, redundancy, and independence in
population codes, revisited. Journal of Neuroscience, 25, 5195 -- 5206.
Lehmann, E. L. (1981). An interpretation of completeness and Basu's theorem. Journal
of the American Statistical Association, 76, 335 -- 340.
Levine, M. W. (1991). The distribution of the intervals between neural impulses in the
maintained discharges of retinal ganglion cells. Biological Cybernetics, 65, 459 --
467.
McCullagh, P., & Nelder, J. P. (1989). Generalized Linear Models, 2nd Edition, New
York: Chapman and Hall.
Merhav, N., Kaplan, G., Lapidoth, A., & Shamai Shitz, S. (1994). On information rates
for mismatched decoders.
IEEE Transactions on Information Theory, 40, 1953 --
1967.
Oizumi, M., Ishii, T., Ishibashi, K., Hosoya, T., & Okada, M. (2010). Mismatched
decoding in the brain. Journal of Neuroscience, 30, 4815 -- 4826.
Paninski, L. (2004). Maximum likelihood estimation of cascade point-process neural
encoding models. Network: Computation in Neural Systems, 15, 243 -- 262.
Paninski, L., Pillow, J., & Lewi, J. (2007). Statistical models for neural encoding,
decoding, and optimal stimulus design. Chapter in Computational Neuroscience:
20
Progress in Brain Research, eds. Cisek, P., Drew, T. & Kalaska, J., 493 -- 507. Ams-
terdam: Elsevier.
Paninski, L., Brown, E. N., Iyengar, S., & Kass, R. E. (2008). Statistical analysis
of neuronal data via integrate-and-fire models. Chapter in Stochastic Methods in
Neuroscience, eds. Laing, C. & Lord, G. Oxford: Oxford University Press.
Pillow, J. W., Paninski, L., Uzzell, V. J., Simoncelli, E. P., & Chichilnisky, E. J. (2005).
Prediction and decoding of retinal ganglion cell responses with a probabilistic spiking
model. Journal of Neuroscience, 23, 11003 -- 11013.
Rawlings, J. O., Pantula, S. G., & Dickey, D. A. (1998). Applied Regression Analysis:
A Research Tool, 2nd Edition. New York: Springer.
Rieke, F., Warland, D., de Ruyter van Steveninck, R., & Bialek, W. (1997). Spikes:
Exploring the Neural Code. Cambridge: MIT Press.
Schervish, M. J. (1995). Theory of Statistics. New York: Springer.
Theunissen, F. & Miller, J. P. (1995). Temporal encoding in nervous systems: a rigorous
definition. Journal of Computational Neuroscience, 2, 149 -- 162.
Troy, J., & Robson, J. (1992). Steady discharges of X and Y retinal ganglion cells of
cat under photopic illuminance. Visual Neuroscience, 9, 535 -- 553.
Truccolo, W., Eden, U. T., Fellows, M. R., Donoghue, J. P., & Brown, E. N. (2005). A
point process framework for relating neural spiking activity to spiking history, neural
ensemble and extrinsic covariate effects. Journal of Neurophysiology, 93, 1074 --
1089.
White, H. (1982). Maximum likelihood estimation of misspecified models. Economet-
rica, 50, 1 -- 25.
21
|
1811.04489 | 1 | 1811 | 2018-11-11T21:55:23 | Nonlinear analysis of EEG complexity in episode and remission phase of recurrent depression | [
"q-bio.NC"
] | Biomarkers of Major Depressive Disorder(MDD), its phases and forms have long been sought. Research indicates that the complexity measures of the cortical electrical activity (EEG) might be candidates for this role. To examine whether the complexity of EEG activity, measured by Higuchi fractal dimension (HFD) and sample entropy (SampEn), differs between healthy subjects, patients in remission and episode phase of the recurrent depression and whether the changes are differentially distributed between hemispheres and cortical regions. Resting state EEG with eyes closed was recorded from 26 patients suffering from recurrent depression and 20 age and sex-matched healthy control subjects. Artefact-free EEG epochs were analyzed by in-house developed programs running HFD and SampEn algorithms. Depressed patients had higher HFD and SampEn complexity compared to healthy subjects. Surprisingly, the complexity was even higher in patients who were in remission than in those in the episode. Altered complexity was present in the frontal and centro-parietal regions when compared to the control group. The complexity in frontal and parietal regions differed between the two phases of depressive disorder. SampEn manifested higher sensitivity than HFD in some cortical areas. Complexity measures of EEG distinguish between the three groups. Further studies are needed to establish whether these measures carry the potential to aid clinically relevant decisions about depression. | q-bio.NC | q-bio | Nonlinear analysis of EEG complexity in episode and remission phase of
recurrent depression
Cukic Milena PhD1, *, Stokic Miodrag PhD2,3, Radenkovic Slavoljub4, Ljubisavljevic
Milos MD PhD5, Simic Slobodan MD6, Danka Savic7 PhD
1Department of General Physiology and Biophysics, School of Biology, University of
Belgrade, Serbia
2Life Activities Advancement Center, Belgrade, Serbia
3Institute for Experimental Phonetics and Speech Pathology, Belgrade, Serbia
4TomTom, Amsterdam, Netherlands
5Department of Physiology, College of Medicine and Health Sciences, UAE University,
Al Ain, UAE
6Institute for Mental Health, Belgrade, Serbia
7 Vinca Institute, Laboratory of Theoretical and Condensed Matter Physics 020/2,
University of Belgrade, Belgrade, Serbia
Author note
Authors of this paper are Milena Čukić, PhD, Studentski trg 16, 11 000 Belgrade, Serbia,
tel. +381 60 0287704, e-mail: [email protected] ; Miodrag Stokić, PhD, Gospodar
Jovanova 35, 11 000 Belgrade, Serbia, tel. +381 11 3208 532, e-mail: [email protected] ,
Slavoljub Radenković, Oosterdoksstraat 114, 1011 DK Amsterdam, the Netherlands, tel.
+381621775149, e-mail: [email protected]; Slobodan Simić, M.D, Palmotićeva 37,
Belgrade, Serbia, tel.+381 11 330 7543, e-mail:[email protected] ; Miloš
1
Ljubisavljević, M.D,PhD, P. O. Box 17666, Al-Ain, United Arab Emirates, tel. +971 3 7137 707,
e-mail: [email protected] ; Danka Savić, PhD, Mike Petrovica Alasa 12-14, 11001 Belgrade,
Serbia, [email protected].
Corresponding author has been authorized by all other authors to act on their behalf in all
matters pertaining the publication of the manuscript. Order of the names has been agreed by all
authors.
* Correspondence concerning this article should be addressed to Milena Čukić, Ph.D.,
Studentskitrg 16, 11 000 Belgrade, Serbia, tel. +381 11 3208 532 (Current address:
KoninginWilhelminaplein 644, 1062 KS Amsterdam, The Netherlands, +31615178926), e-mail:
[email protected]
2
Abstract
Background: Biomarkers of Major Depressive Disorder(MDD), its phases and forms have long
been sought. Research indicates that the complexity measures of the cortical electrical activity
(EEG) might be candidates for this role.
Aims: To examine whether the complexity of EEG activity, measured by Higuchi's fractal
dimension (HFD) and sample entropy (SampEn), differs between healthy subjects, patients in
remission and episode phase of the recurrent depression and whether the changes are
differentially distributed between hemispheres and cortical regions.
Methods: Resting state EEG with eyes closed was recorded from 26 patients suffering from
recurrent depression and 20 age and sex-matched healthy control subjects. Artefact-free EEG
epochs were analyzed by in-house developed programs running HFD and SampEn algorithms.
Results: Depressed patients had higher HFD and SampEn complexity compared to healthy
subjects. Surprisingly, the complexity was even higher in patients who were in remission than in
those in the episode. Altered complexity was present in the frontal and centro-parietal regions
when compared to control group. The complexity in frontal and parietal regions differed between
the two phases of depressive disorder. SampEn manifested higher sensitivity than HFD in some
cortical areas.
Conclusions: Complexity measures of EEG distinguish between the three groups. Further studies
are needed to establish whether these measures carry a potential to aid clinically relevant
decisions about depression.
Keywords: Electroencephalogram, Higuchi's fractal dimension, sample entropy,
complexity, recurrent depression, episode, remission
3
Introduction
Major Depressive Disorder (MDD) is a serious mental illness associated with protracted
personal suffering, and significant social and functional impairment. It has become the leading
cause of ill health and disability worldwide, before heart diseases, arthritis and many forms of
cancer. Depression has a strong tendency to reoccur - a significant number of patients will suffer
from at least one more episode after the first one, reaching four episodes on average during the
lifetime [1]. In these patients, the risk of new episodes rises significantly with each subsequent
recurrence although the course of the disease can be unique [1]. Hence, the decision to stop the
therapy, or to initiate the maintenance therapy to prevent a relapse in patients with recurrent
depression who have achieved remission, often presents a significant clinical challenge [2].
Therefore, finding accurate and reliable biomarkers that can help differentiate MDD episode
from remission is of considerable importance. Also, there is a need to distinguish the (onset)
episode of bipolar disorder from a depressive episode in Major Depression [3] in order to treat it
appropriately.
Stress hormones, most of all cortisol, have long been the main candidate biomarkers for
depression, but the results of numerous studies are inconclusive [4]. It seems that combined
biomarkers are more promising than any single one [5], but there still are no biomarkers that are
sufficiently sensitive and specific [6].
Methods based on the theory of nonlinear dynamics are justly included in the quest for
biomarkers. They provide more accurate information than classical spectral analysis and have
already enabled insights into neural activity and connectivity in both healthy physiological
processes and pathological conditions. The first study that used nonlinear measures in depression
showed that EEG dynamics, measured by correlation dimension, in patients with Major
4
Depressive Disorder (MDD) is more predictable, i.e., less complex than in healthy subjects [7].
Contrary to this result, Lempel -- Ziv complexity measure of the EEG signal is higher in depressed
patients compared to control subjects, particularly in anterior brain regions [8]. Furthermore,
patients suffering multiple depressive episodes do not recover dynamics found in healthy
subjects as patients who had one depressive episode [9]. The analysis of EEG signal with
Approximate Entropy also showed higher values in normal controls compared to depression
patients [10]. Other studies utilizing Lempel-Ziv complexity and other complexity measures
found either no difference between MDD and control subjects [11], or increased EEG
complexity in patients with depression [8,11 -- 13]. Analyses of EEG complexity with Higuchi's
fractal dimension (HFD) also showed increased EEG complexity in patients with depression,
particularly in the beta and gamma sub-bands mostly in the frontal area [11,12]. Application of
wavelet-chaos methodology found significantly increased complexity in both parietal and frontal
regions in MDD patients, both in full-band activity and in beta and gamma EEG sub-bands [13].
De la Torre-Lugue and Xavier Bornas [14] in their review study concluded that 'EEG dynamics
for depressive patients appear more complex but may be more random than the dynamics of
healthy non-depressed individuals'.
In this study, we explore the use of different measures of complexity of brain activity,
estimated from the resting state electroencephalogram (EEG) records, in discriminating episode
from remission phase in patients with recurrent depression, as well as from healthy subjects.
Methods
Subjects
EEG data were recorded at the Institute for Mental Health in Belgrade, Serbia. The
participants were 26 patients (17 women and 9 men) suffering from recurrent depression, 25 to
5
68 years old (mean 32.40, SD 10.16). As a control, we used EEG recordings of 20 age-matched
(mean 30.14, SD 8.94) healthy controls (10 males, 10 females) with no history of any
neurological or psychiatric disorders, recorded at the Institute for Experimental Phonetic and
Speech Pathology in Belgrade, Serbia. All participants were right-handed, according to
Edinburgh Handedness Inventory. The participants were informed about the experimental
protocol and signed informed consent. The protocol was approved by the Ethical Committees of
the participating institutions. All participants with depression were on medications and under the
supervision of an experienced specialist in clinical psychiatry. Their diagnoses was made
according to ICD-10 scale. The study compared three groups: healthy controls (C), and
depressed patients (D) in an episode (E), and in remission (R).
Data acquisition
EEG was recorded in the resting state with 10/20 International system for electrode
placement, using NicoletOne Digital EEG Amplifier (VIAYSYS Healthcare Inc. NeuroCare
Group), with closed eyes without any stimulus (resting state EEG). EEGs were recorded from 19
electrodes in a monopolar montage (Electro-cap International Inc. Eaton, OH USA). The
sampling rate was 1 kHz. The resistance was kept less than five KOhm. A bandpass filter was
0.5-70 Hz. The same setup was used for the control group, using Nihon Kohden Inc. apparatus.
Since the protocol of recordings was the same, the fact that we used recordings from two
different EEG producers' apparata did not introduce the difference between the groups [15].
Each recording lasted for three minutes. Participants were instructed to reduce any
movement, staying in a comfortable sitting position with eyes closed. The EEG records of four
subjects were discarded from further analyses because of the high level of muscle activity or
blinking artifacts. Further, we used records from 22 patients and 20 healthy controls for this
6
study. Half of the patients were recorded while they were in an acute episode, while the other
half were in remission phase of the disease. Artifacts were carefully inspected and removed
manually from the records by two independent experts. From each artifact-free record, we chose
three epochs: one from the beginning, one from the middle and the last one close to the end of
the record. Each epoch comprises of 5000 samples. Positions of epochs in each person's EEG
recording were specified by the ordinal number of the first sample in that epoch. Therefore, for
each subject, epoch, and electrode, two nonlinear measures were calculated.
Data analysis
Initially, the classical spectral analysis was performed by constructing spectral power
maps (EEGLAB program [16]). Thereon, for all EEG epochs, Higuchi's fractal dimension
(HFD) and Sample entropy (SampEn) were calculated. Fractal and SampEn maps were
constructed on the whole spectrum, not dividing the signal into bands. The analysis was adopted
as it has been shown that the Fourier analysis is redundant to fractal analyses [17]. The fractal
dimension of EEG was calculated by using Higuchi's algorithm [18], demonstrated to be the
most appropriate for electrophysiological data [19]. This method provides a reasonable estimate
of the fractal dimension even if short signal segments are analyzed and it is computationally fast.
HFD was also chosen because it is widespread in the EEG literature facilitating comparison of
the results. We performed the Higuchi's algorithm [18], with the maximal length of an
epoch kmax= 8, shown to perform the best for this type of signals [20]. HFD of a time series is a
measure of its complexity and self-similarity in the time domain. HFD is not an integer, and the
value of fractal dimension (FD) of waveforms (e.g. EEG) can range between 1 and 2. Higher
self-similarity and complexity results in higher HFD [21]. Sample Entropy (SampEn) was
computed according to the procedure by Richman and Moorman [22]. SampEn estimates the
signal complexity by computing the conditional probability that two sequences of a given
7
length n, similar for m points, remain similar within tolerance r at the next data point (when self-
matches are not included). SampEn measures the irregularity of the data (the higher the values,
the less regular signal) that is related to signal complexity [23]. SampEn was calculated using
tolerance level of r = 0.15 times the standard deviation of the time series and m = 2, shown to be
optimal for EEG [24]. Both HFD and SampEn were calculated for each electrode for the duration
of signal (the epochs of artifact-free recorded EEG; three epochs from each recording), using the
in-house written algorithm in Java programming language. It should also be noted that
correlations with any medical data were not explored since the main aim of the study was to find
independent nonlinear markers based on analysis of the EEG signal, which could be utilized as
an additional tool in clinical practice.
Statistical analysis
Both HFD and SampEn values were used as an ensemble for analysis of variance
(ANOVA with post-hoc Bonferroni correction, SPSS Statistics version 20.0, SPSS Inc, USA).
The Kolmogorov Smirnov test showed that HFD and SampEn data were not normally
distributed. To obtain data with normal distribution and to include it in the analysis of the
difference in complexity in resting state EEG data, normalized values of SampEn and HFD
obtained from epochs of recorded EEG were calculated as log10 normalization in SPSS.
Normalized SampEn and HFD data were compared using ANOVA with factors state (Controls
vs. Depression, and Controls vs. Episode vs. Remissions) and position of the electrode (1 to 19).
For every electrode, ANOVA was repeated for each measure independently. Bonferroni
correction was used where appropriate. For all analyses, probability values p = 0.05 were
considered as statistically significant.
Principal Component analysis (PCA)
8
To reduce the dimensionality of the problem and decorrelate the measures (HFD and
SampEn calculated from the same epochs extracted from the raw EEG signal), we utilized PCA
[25] in order to obtain three principal components (PCs) corresponding to largest eigenvalues of
the sample covariance matrix. We defined percentage of the explained variance by first three
PCs as ratio between sums of variances of three PCs and original variables. Here we wanted to
demonstrate the possibility of classification of previously calculated nonlinear measures, by
utilizing only the first three components in order to see whether the data were separable. We
used Matlab 15b for this calculation (MathWorks, Masacushets, USA).
Results
Spectral power maps
The first level of analysis was to compare spectral power maps of low alpha (8-10Hz),
high alpha (10-12Hz), and beta (13-30Hz) bands between healthy controls (C) and patients in a
different phase of the disease (i.e. episode (E) or remission (R), Figure 1). Spectral power maps
in low alpha band showed an overall decrease in both E and R groups in posterior regions
(maximum at C3, C4, Cz, P3, P4, Pz, and T3) when compared to C group. In E and R group
there was an increase in low alpha spectral power in the right prefrontal region (Fp2) and lateral
right frontal region (F8) when compared to C group. There was a statistically significant effect of
group (C, E, R) on low alpha (8-10 Hz) spectral power at following regions: C3 - F(2,40)=8.748
p=0.021; C4 - F(2,40)=5.168, p=0.04; Cz - F(2,40)=6.381, p=0.03; P3 - F(2,40)=5.748, p=0.021;
P4 - F(2,40)=4.286, p=0.05; and Pz - F(2,40)=13.939, p<0.001. Spectral power maps in high
alpha band showed a decrease of high alpha spectral power in right prefrontal (Fp2), left
temporal (T3), and central (C3, Cz) regions in E and R groups when compared to C group. Also,
9
there was a statistically significant effect of group (C, E, R) on high alpha (10-12Hz) spectral
power at following regions: Fp2 - F(2,40)=9.211, p<0.01; C3 - F(2,40)=4.096, p=0.05; Cz -
F(2,40)=10.734, p<0.01; and T3 - F(2,40)=8.633, p=0.012.
Spectral power maps in beta band showed a decrease of beta (13-30 Hz) spectral power at
frontal (Fz, Fp1, Fp2, F3, F4, F7, F8), and central-temporal (C3, T3) regions in E and R groups
when compared to C group. In contrast, there was an increase of beta spectral power in posterior
regions (P3, P4, Pz, T5, T6, O1, O2) in E and R groups when compared to C group. In the E
group only there was a frontal (F3 > F4) and temporal-occipital (T5/O1 < T6/O2) asymmetry.
Significant effect of group (C, E, R) was found at following cortical regions: Fp1 -
F(2,40)=7.207, p=0.018; Fp2 - F(2,40)=8.107, p<0.01; F3 - F(2,40)=12.342, p<0.01; F4 -
F(2,40)=18.674, p<0.01; Fz - F(2,40)=7.997, p=0.023; F7 - F(2,40)=4.922, p=0.04; F8 -
F(2,40)=6.817, p=0.033; P3 - F(2,40)=7.002, p=0.018; P4 - F(2,40)=6.699, p=0.03; Pz -
F(2,40)=14.361, p<0.01; T5 - F(2,40)=13.207, p<0.01; T6 - F(2,40)=14.338, p<0.01; O1 -
F(2,40)=16.189, p<0.01; and O2 - F(2,40)=16.917, p<0.01.
10
Figure 1. Spectral power maps are showing the difference between healthy persons (C) and those in
the episode (E) and remission (R).
Higuchi's fractal dimension (HFD) and Sample entropy (SampEn)
Figure 2 summarizes the difference in HFD and SampEn values in all three groups
averaged across all electrodes. HFD values for healthy controls ranged from 1.0093 to 1.0530
(mean 1.0290), for patients in the episode from 1.01 to 1.3546 (mean 1.0876), and for those in
remission from 1.0133 to 1.483 (mean 1.1136). SampEn values for healthy controls were 0.0764
11
to 0.3959 (mean 0.1613), in the episode group 0.2006 to 0.7866 (mean 0.3992) and in remission
group 0.1983 to 0.861 (mean 0.4777).
ANOVA showed a significant effect of group on HFD (F(2,2334)=30.831, p<0.01) and
on SampEn values (F(2,2334)=38.863, p<0.01). The post-hoc test showed that both HFD and
SampEn values were lowest in the C group, followed by participants in the E group. Participants
in the R group had the highest HFD and SampEn values (HFDControl<HFDEpisode<HFDRemission,
p<0.01 for each comparison; SampEnControl<SampEnEpisode<SampEnRemission, p<0.01 for each
comparison).
Figure 2. Averaged values for HFD (left) and SampEn (right) values averaged across all 19 electrodes,
for all three groups (C- control, E-episode, and R-remission). Note the difference in scale for HFD and
SampEn. *** p<0.01.
The next level of analysis was to determine effect of a Group (C, E, R) on HFD and
SampEn values averaged across Regions of Interest -- ROIs (ROI 1: left frontal regions Fp1, F3,
F7; ROI 2: right frontal regions Fp2, F4, F8; ROI 3: Vertex regions Fz, Cz, Pz; ROI 4: left
temporal-parietal-central-occipital regions T3, T5, P3, C3, O1; and ROI 5: right temporal-
12
parietal-central-occipital regions T4, T6, P4, C4, O3). The data for ROI and group are shown in
Figure 3. ANOVA showed a significant effect of group on HFD values for the ROI 1:
F(2,366)=83.021, p<0.01, ROI 2: F(2,366)=85.400, p<0.01, ROI 3: F(2,366)=99.117, p<0.01,
ROI 4: F(2,612)=78.908, p<0.01, and ROI 5: F(2,612)=94.547, p<0.01. The post-hoc test
showed for each ROI that HFD is lowest in C group followed by E group while R group had the
highest HFD values. A significant effect of group on SampEn values (ANOVA) was found for
the ROI 1: F(2,366)=333.302, p<0.01, ROI 2: F(2,366)=308.381, p<0.01, ROI 3:
F(2,366)=446.666, p<0.01, ROI 4: F(2,612)=464.020, p<0.01, and ROI 5: F(2,612)=492.307,
p<0.01. The post-hoc test showed for each ROI that SampEn is lowest in C group followed by E
group while R group had the highest SampEn values.
Figure 3. The comparison of HFD (left) and SampEn (right) values averaged across each Region of
Interest (ROI). ROI1 included left frontal electrodes (Fp1, Fp3, F7). ROI2 included right frontal
electrodes (Fp2, F4, F8). ROI3 included vertex electrodes (Fz, Cz, Pz). ROI4 included left temporal-
parietal-central-occipital regions electrodes (T3, T5, P3, C3, O1) and ROI5 included right temporal-
parietal-central-occipital electrodes (T4, T6, P4, C4, O3). C-control, E-episode, R-remission. Note the
difference in scale for HFD and SampEn.
The final level of analysis was to determine the possible effect of a group on HFD and
SampEn values for each electrode. ANOVA showed a significant effect of Group (C, E, R) on
13
HFD values for each electrode (p<0.01). However, post-hoc test showed that significant
difference between each group was found for electrodes Fp1, Fp2, Fz, P3, P4, T5, T6, and Cz
(p<0.01). Post-hoc Bonferroni correction showed that for each electrode
HFDControl<HFDEpisode<HFDRemission, p<0.05 for each comparison. A similar result was found for
SampEn values. ANOVA showed a significant effect of Group (C, E, R) on SampEn values for
each electrode (p<0.01). Post-hoc showed that this difference is driven by significantly lower
SampEn values in C group when compared to E and R group. There was no significant
difference between E and R group for each electrode. However, the post-hoc test showed that
significant difference between each group was found for electrodes Fp1, Fp2, F4, F7, Fz, P3, P4,
O2, T5, T6, Cz, and Pz (p<0.01) and that for each electrode SampEnControl<SampEnEpisode
<SampEnRemission, p<0.05 for each comparison.
Two-way ANOVA found a statistically significant effect of factors Group (C, E, R) and
ROI (frontal left, frontal right, vertex, TCPO left, TCPO right), as well as the interaction (Group
x ROI) on HFD values (p < .001). The same analysis found a statistically significant effect of
factors Group (p < .001) and ROI (p < .01) on SampEn values. No significant effect of the
interaction was found.
14
Figure 4: Results of (ANOVA) comparison of calculated sample entropy for certain electrodes which
are particularly well discriminative: left, the values of SampEn for position Fp1 (fronto-parietal), and
right, the values of calculated SampEn for EEG recorded from position T3 (temporal). The significant
difference between those with depression and controls is particularly pronounced, and the difference
between Exacerbation and Remission is also significant.
Figure 4 shows a spatial representation of significant differences of values of SampEn
and HFD according to electrode position, i.e. Fractal and SampEn maps. The results show that
SampEn is more discriminative regarding the number of electrodes with the statistically
significant difference between C, E, and R.
Figure 5: The spatial representation of significant differences in HFD and SampEn. In orange,
electrodes from which SampEn values showed a significant difference in comparison to controls, and
in green electrodes where HFD had a significant difference. The left panel show the difference between
the whole group of patients compared to controls. Note that SampEn shows better discrimination than
HFD. The middle panel shows the comparison of al three groups (C vs. E vs. R). A significant
difference was found on 12 electrodes for SampEn, and on nine electrodes for HFD. Right panel shows
the comparison of E and R groups. SampEn values were significantly different on Fp1, Fp2, F7, F3, P3,
P4, T6, O2, Fz and Cz electrodes. HFD values were significantly different on Fz, Cz, P4, and T6
electrodes.
Figure 5-A shows that for both measures there is a significant difference between the
patient and control group for the majority of electrodes. However, SampEn showed to
discriminate R and E group (Figure 5-C) better. SampEn differed between R and E groups on
Fp1, Fp2, P3, P4, O2, F7, T6, Fz, Cz electrodes, while HFD was different on Fz, Cz, P4, T6
15
electrodes only. When all three phases are examined (Figure 5-B) SampEn shows a significant
difference on 12, and HFD on 8, out of 19 electrodes.
Principal Component Analysis (PCA)
With only the first three Principal Components (PCs) we want here to illustrate that those
calculated values are separable. Again, SampEn gave more clear separation of the data when
compared to HFD (on Figure 6, SampEn results are on the left and HFD on the right picture). In
this study we did not intend to deal with further classification, although a high accuracy could be
obtained for several machine learning algorithms.
Figure 6: Principal component analysis was used to show the separability of data. We used only first
three principal components. Blue diamonds are symbols for those who are diagnosed with depression
and are in remission; green stars are depressed patients who are in exacerbation and red squares are
representing control group. Panel left represents how separable are the values of SampEn calculated
from EEG of subjects from these three groups; panel on the right represents the values of calculated
Higuchi's fractal dimension which are obviously less separable data, but still distinctive.
Discussion
The results show that both Higuchi fractal dimension (HFD) and SampEn nonlinear
measures of resting state EEG signal discriminate between healthy controls and depressed
patients, the latter having higher complexity. These differences are widely distributed and
include frontal, midline (vertex), and temporal-parietal-occipital regions. Furthermore, the
16
complexity differs significantly between episode and remission, being higher in remission than
in the episode phase of the disease.
Although this last finding is counterintuitive (it would be expected that the remission
state is closer to the healthy one in every respect), it might be in line with the observation of
Willner et al. [26]: "It is evident that antidepressants do not normalize brain activity: mood and
behavior are restored to normal, but the antidepressant-treated brain is in a different state from
the non-depressed brain." According to the research on the first-episode depression, such brain is
initially different from the non-depressed one and the differences are structural [27], as well as
functional [28,29]. In terms of nonlinear dynamic systems, initial conditions of the depressed
brain system are different from the non-depressed. Even if the dynamics of information
processing were the same, these different initial conditions would direct the system to steady
states distinctive from normal. However, the dynamics is probably also different due to
compensatory mechanisms. For example, functional neuroimaging revealed reduced integrity of
the uncinate fasciculus and enhanced functional connectivity of anterior cingulate cortex and
medial temporal lobe in MDD [30]. The higher severity of depression, the more pronounced this
negative structure-function relation. The authors suggest that the increased functional
connectivity is a compensatory mechanism for decreased uncinate fasciculus integrity. Willner et
al. came to a similar conclusion that decreased hippocampal functioning in depression causes an
increase in the activity of the ventral "affective" system [26]. It is then easy to suppose that the
enrichment of fronto-limbic connectivity and reorganization of circuits is accompanied by
increment in complexity.
The EEG hallmark of depression is the presence of stable hemispheric asymmetry in the
alpha spectral band, although the differences in other spectral bands were also demonstrated
17
[31]. Interestingly, de Vinne et al [32] showed that frontal alpha asymmetry cannot be used as
biomarker. At present, there is no consensus about the direction of change. Our spectral analysis
shows that the power was decreased in alpha and high-alpha bands in majority of the cortical
regions, but increased in beta bands in posterior regions in patients. This may point towards the
presence of hyperactivity in posterior regions (alpha desynchronization) of the right hemisphere,
which is known to process the negative emotional content [33].
As said in the Introduction, the results concerning complexity of EEG in depressed
patients are discrepant. What may account for these opposing findings? One important aspect
pertains to methodological differences between the studies, related to signal acquisition (number
of EEG electrodes used, sampling frequency, pre-processing of raw signal, i.e., decomposition
on bands and filtering), as well as experimental design (probing the emotional content, using
different stimuli, performing cognitive task, etc.). The eyes-closed condition, unlike the eyes-
open condition, allows measurement of the resting state arousal without the influence of cortical
processing of the visual input in other bands on the complexity of brain dynamics. Also, it should
be noted that we did not divide the spectrum of the signal to standard bands, but observed the
changes in broadband. This is important as it has been shown that signal decomposition like
Fourier, Wavelet, or cosine transformation can impact the result of a subsequent nonlinear
analysis yielding erroneous results [34,35]. Other reasons may relate to inherent differences
between nonlinear algorithms that are based on different theoretical frameworks [36,37]. Our
results are in line with studies that also used Higuchi's fractal dimension [11,12], that was shown
to be more accurate than Katz's algorithm [12]. The difference in complexity values between
depressed and healthy subjects in our study were much larger than those reported in Bachman et
18
al. [11]. Another possible source of difference is choosing different values for k in Higuchi's FD
algorithm; Bachmann et al. [11] used 50 for their k value.
The results of Fractal and SampEn maps are in line with previous electrophysiological
studies demonstrating the presence of stable frontal asymmetry [38,39] in MDD. However, in
this study the signal was not divided to standard bands, hindering conclusion that current
findings are directly related to the alpha band asymmetry. The results point to elevated
complexity in frontal, central and right parieto-temporal regions. This is also in line with earlier
EEG studies [40], which reported similar topographical changes in distribution.
It should be noted that we used HFD and SampEn, two nonlinear measures able to detect
differential aspects of the signal under analysis. While HFD examines the complexity in the time
domain, SampEn can characterize the irregularity of a signal or its predictability. They both
showed higher complexity in patients with depression when compared to healthy control
subjects. The difference was much more pronounced when examined by SampEn suggesting
increased variability, or "irregularity" or unpredictability of the signal.
Conclusions
The idea of EEG-based classification of depression is not entirely novel. Our study
confirmed that it is possible to quantify the difference between depressed patients and controls
by employing two complexity measures - HFD and SampEn, on resting state EEG. Furthermore,
it showed, for the first time, that both measures could detect a statistically significant difference
between depressed patients who were in episode and remission. Whether these and other non-
linear measures may be used as potential clinical markers of disease stage or of the effectiveness
of various treatments in MDD remains to be confirmed on larger groups of patients.
19
Acknowledgments
We want to thank Goran Petrovic for the input on the graphical representation of SampEn
and HFD maps. This research was supported by the Ministry of Education, Science and
Technological Development of the Republic of Serbia under Projects 178027 and 32032, and
CMHS grant 31M201.
Author Contributions: M. C. - designed the study, analyzed the data and drafted the
manuscript; M. S. - recorded EEG data from healthy controls, performed statistical analyses and
revised the manuscript; S. S. - performed clinical assessment of patients, selected patients and
revised the manuscript; S.R. - wrote and tested both algorithms in Java for calculating HFD and
SampEn and revised the manuscript; M.Lj. -- critically interpreted the data and wrote the
manuscript; D.S.- critically interpreted the data and wrote the manuscript.
Competing Financial Interests Statement: The authors declare no competing financial
interests.
20
References
1.
2.
3.
4.
5.
Solomon DA, Keller MB, Leon AC, et al. Multiple Recurrences of Major Depressive
Disorder. Am. J. Psychiatry [Internet]. 157(2), 229 -- 233 (2000). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/10671391.
Teasdale JD, Segal Z V, Williams JMG, Ridgeway VA, Soulsby JM, Lau MA. Prevention
of relapse/recurrence in major depression by mindfulness-based cognitive therapy. J.
Consult. Clin. Psychol. [Internet]. 68(4), 615 -- 623 (2000). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/10965637.
Zhang K, Liu Z, Cao X, et al. Amplitude of low-frequency fluctuations in first-episode,
drug-naïve depressive patients: A 5-year retrospective study. PLoS One [Internet]. 12(4),
e0174564 (2017). Available from: http://www.ncbi.nlm.nih.gov/pubmed/28384269.
Juruena MF, Bocharova M, Agustini B, Young AH. Atypical depression and non-atypical
depression: Is HPA axis function a biomarker? A systematic review. J. Affect. Disord.
[Internet]. 233, 45 -- 67 (2018). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/29150144.
Strawbridge R, Young AH, Cleare AJ. Biomarkers for depression: recent insights, current
challenges and future prospects. Neuropsychiatr. Dis. Treat. [Internet]. 13, 1245 -- 1262
(2017). Available from: http://www.ncbi.nlm.nih.gov/pubmed/28546750.
6. Quevedo J, Yatham LN. Biomarkers in mood disorders: Are we there yet? J. Affect.
Disord. [Internet]. 233, 1 -- 2 (2018). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/29395259.
7. Nandrino JL, Pezard L, Martinerie J, et al. Decrease of complexity in EEG as a symptom
8.
of depression. Neuroreport [Internet]. 5(4), 528 -- 30 (1994). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/8003689.
Li Y, Tong S, Liu D, et al. Abnormal EEG complexity in patients with schizophrenia and
depression. Clin. Neurophysiol. [Internet]. 119(6), 1232 -- 1241 (2008). Available from:
http://dx.doi.org/10.1016/j.clinph.2008.01.104.
9. Arns M, Cerquera A, Gutiérrez RM, Hasselman F, Freund JA. Non-linear EEG analyses
predict non-response to rTMS treatment in major depressive disorder. Clin. Neurophysiol.
[Internet]. 125(7), 1392 -- 1399 (2014). Available from:
http://dx.doi.org/10.1016/j.clinph.2013.11.022.
10. Puthankattil SD, Joseph PK. Analysis of EEG Signals Using Wavelet Entropy and
Approximate Entropy : A Case Study on Depression Patients. Int. J. Medical, Heal.
Biomed. Bioeng. Pharm. Eng. 8(7), 420 -- 424 (2014).
11. Bachmann M, Lass J, Suhhova A, Hinrikus H. Spectral asymmetry and Higuchi's fractal
dimension measures of depression electroencephalogram. Comput. Math. Methods Med.
[Internet]. 2013, 1 -- 8 (2013). Available from: http://dx.doi.org/10.1155/2013/251638.
12. Ahmadlou M, Adeli H, Adeli A. Fractality analysis of frontal brain in major depressive
disorder. Int. J. Psychophysiol. [Internet]. 85(2), 206 -- 211 (2012). Available from:
http://dx.doi.org/10.1016/j.ijpsycho.2012.05.001.
13. Akar SA, Kara S, Agambayev S, Bilgic V. Nonlinear analysis of EEG in major depression
with fractal dimensions [Internet]. In: 2015 37th Annual International Conference of the
IEEE Engineering in Medicine and Biology Society (EMBC). IEEE, 7410 -- 7413 (2015)
[cited 2017 Aug 10]. Available from: http://ieeexplore.ieee.org/document/7320104/.
14. De La Torre-Luque A, Bornas X. Complexity and Irregularity in the Brain Oscillations of
21
Depressive Patients: A Systematic Review. Neuropsychiatry (London). 7(5), 466 -- 477
(2017).
15. Pivik RT, Broughton RJ, Coppola R, Davidson RJ, Fox N, Nuwer MR. Guidelines for the
recording and quantitative analysis of electroencephalographic activity in research
contexts. Psychophysiology [Internet]. 30(6), 547 -- 558 (1993). Available from:
http://dx.doi.org/10.1111/j.1469-8986.1993.tb02081.x.
16. Delorme A, Makeig S. EEGLAB: an open source toolbox for analysis of single-trial EEG
dynamics including independent component analysis. J. Neurosci. Methods [Internet].
134(1), 9 -- 21 (2004). Available from: http://dx.doi.org/10.1016/j.jneumeth.2003.10.009.
17. Kalauzi A, Bojić T, Vuckovic A. Modeling the relationship between Higuchi's fractal
dimension and Fourier spectra of physiological signals. Med. Biol. Eng. Comput.
[Internet]. 50(7), 689 -- 99 (2012). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/22588703.
18. Higuchi T. Approach to an irregular time series on the basis of the fractal theory. Phys. D
Nonlinear Phenom. 31(2), 277 -- 283 (1988).
19. Esteller R, Vachtsevanos G, Echauz J, Litt B. A comparison of waveform fractal
dimension algorithms. IEEE Trans. Circuits Syst. I Fundam. Theory Appl. [Internet].
48(2), 177 -- 183 (2001). Available from: http://dx.doi.org/10.1109/81.904882.
20. Spasic S, Kalauzi A, Grbic G, Martac L, Culic M. Fractal analysis of rat brain activity
after injury. Med. Biol. Eng. Comput. [Internet]. 43(3), 345 -- 8 (2005). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/16035222.
21. Eke A, Herman P, Kocsis L, Kozak LR. Fractal characterization of complexity in
temporal physiological signals. Physiol Meas. 23(0967 -- 3334 (Print)), R1-38 (2002).
22. Richman JS, Moorman JR. Physiological time-series analysis using approximate entropy
and sample entropy. Am J Physiol Hear. Circ Physiol [Internet]. 278(6), H2039-2049
(2000). Available from: http://ajpheart.physiology.org/content/278/6/H2039.long.
23. Pincus SM. Approximate entropy as a measure of irregularity for psychiatric serial
metrics. Bipolar Disord. [Internet]. 8(5p1), 430 -- 440 (2006). Available from:
http://dx.doi.org/10.1111/j.1399-5618.2006.00375.x.
24. Molina-Picó A, Cuesta-Frau D, Aboy M, Crespo C, Miró-Martínez P, Oltra-Crespo S.
Comparative study of approximate entropy and sample entropy robustness to spikes. Artif.
Intell. Med. [Internet]. 53(2), 97 -- 106 (2011). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/21835600.
Jolliffe IT. Principal Component Analysis and Factor Analysis [Internet]. Princ. Compon.
Anal. , 115 -- 128 (1986). Available from: http://dx.doi.org/10.1007/978-1-4757-1904-8_7.
25.
26. Willner P, Scheel-Krüger J, Belzung C. The neurobiology of depression and
antidepressant action. Neurosci. Biobehav. Rev. [Internet]. 37(10), 2331 -- 2371 (2013).
Available from: http://www.ncbi.nlm.nih.gov/pubmed/23261405.
27. Ramezani M, Abolmaesumi P, Tahmasebi A, et al. Fusion analysis of first episode
depression: where brain shape deformations meet local composition of tissue.
NeuroImage. Clin. [Internet]. 7, 114 -- 21 (2015). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/25610773.
28. Bluhm R, Williamson P, Lanius R, et al. Resting state default-mode network connectivity
in early depression using a seed region-of-interest analysis: Decreased connectivity with
caudate nucleus. Psychiatry Clin. Neurosci. [Internet]. 63(6), 754 -- 761 (2009). Available
from: http://doi.wiley.com/10.1111/j.1440-1819.2009.02030.x.
22
29. Zhang J, Wang J, Wu Q, et al. Disrupted Brain Connectivity Networks in Drug-Naive,
First-Episode Major Depressive Disorder. Biol. Psychiatry [Internet]. 70(4), 334 -- 342
(2011). Available from:
https://www.sciencedirect.com/science/article/pii/S0006322311005476?via%3Dihub.
30. de Kwaasteniet B, Ruhe E, Caan M, et al. Relation Between Structural and Functional
Connectivity in Major Depressive Disorder. Biol. Psychiatry [Internet]. 74(1), 40 -- 47
(2013). Available from: http://dx.doi.org/10.1016/j.biopsych.2012.12.024.
31. Gold C, Fachner J, Erkilla J, Erkkila J. Validity and reliability of electroencephalographic
frontal alpha asymmetry and frontal midline theta as biomarkers for depression. Scand. J.
Psychol. [Internet]. 54(2), 118 -- 126 (2012). Available from:
http://dx.doi.org/10.1111/sjop.12022.
32. van der Vinne N, Vollebregt MA, van Putten MJAM, Arns M. Frontal alpha asymmetry as
a diagnostic marker in depression: Fact or fiction? A meta-analysis. NeuroImage. Clin.
[Internet]. 16, 79 -- 87 (2017). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/28761811.
33. Coan JA, Allen JJB. Frontal EEG asymmetry as a moderator and mediator of emotion.
Biol. Psychol. [Internet]. 67(1 -- 2), 7 -- 50 (2004). Available from:
http://dx.doi.org/10.1016/j.biopsycho.2004.03.002.
34. Klonowski W. From conformons to human brains: an informal overview of nonlinear
dynamics and its applications in biomedicine. Nonlinear Biomed. Phys. [Internet]. 1(1), 5
(2007). Available from: http://dx.doi.org/10.1186/1753-4631-1-5.
35. Rabinovich MI, Varona P, Selverston AI, Abarbanel HDI. Dynamical principles in
neuroscience. Rev. Mod. Phys. [Internet]. 78(4), 1213 -- 1265 (2006). Available from:
http://dx.doi.org/10.1103/revmodphys.78.1213.
36. Goldberger AL, Peng C-K, Lipsitz LA. What is physiologic complexity and how does it
change with aging and disease? Neurobiol. Aging [Internet]. 23(1), 23 -- 6 (2002). Available
from: http://www.ncbi.nlm.nih.gov/pubmed/11755014.
37. Pincus SM, Goldberger AL. Physiological time-series analysis: what does regularity
quantify? Am. J. Physiol. [Internet]. 266(4 Pt 2), H1643-56 (1994). Available from:
http://www.ncbi.nlm.nih.gov/pubmed/8184944.
38. Allen JJB, Urry HL, Hitt SK, Coan JA. The stability of resting frontal
electroencephalographic asymmetry in depression. Psychophysiology [Internet]. 41(2),
269 -- 280 (2004). Available from: http://dx.doi.org/10.1111/j.1469-8986.2003.00149.x.
39. Davidson RJ. What does the prefrontal cortex "do" in affect: perspectives on frontal EEG
asymmetry research. Biol. Psychol. [Internet]. 67(1 -- 2), 219 -- 234 (2004). Available from:
http://dx.doi.org/10.1016/j.biopsycho.2004.03.008.
40. Haase L, Thom NJ, Shukla A, et al. Mindfulness-based training attenuates insula response
to an aversive interoceptive challenge. Soc. Cogn. Affect. Neurosci. [Internet]. 11(1), 182 --
190 (2014). Available from: http://dx.doi.org/10.1093/scan/nsu042.
23
|
1808.01503 | 1 | 1808 | 2018-08-04T16:17:44 | Dynamic self-organized error-correction of grid cells by border cells | [
"q-bio.NC"
] | Grid cells in the entorhinal cortex are believed to establish their regular, spatially correlated firing patterns by path integration of the animal's motion. Mechanisms for path integration, e.g. in attractor network models, predict stochastic drift of grid responses, which is not observed experimentally. We demonstrate a biologically plausible mechanism of dynamic self-organization by which border cells, which fire at environmental boundaries, can correct such drift in grid cells. In our model, experience-dependent Hebbian plasticity during exploration allows border cells to learn connectivity to grid cells. Border cells in this learned network reset the phase of drifting grids. This error-correction mechanism is robust to environmental shape and complexity, including enclosures with interior barriers, and makes distinctive predictions for environmental deformation experiments. Our work demonstrates how diverse cell types in the entorhinal cortex could interact dynamically and adaptively to achieve robust path integration. | q-bio.NC | q-bio |
Dynamic self-organized error-correction of grid cells by border cells
Eli Pollock,1, ∗ Niral Desai,2, † Xue-xin Wei,3, ‡ and Vijay Balasubramanian4, §
1MIT, Brain and Cognitive Sciences, 43 Vassar St, Cambridge, MA 02142, USA
2University of Texas at Austin, Department of Physics, 2515 Speedway, Austin, TX 78705, USA
3Dept. of Statistics, Center for Theoretical Neuroscience,
Columbia University, 3227 Broadway, New York, NY 10027, USA
4David Rittenhouse Laboratories, University of Pennsylvania, Philadelphia, PA 19004, USA
(Dated: August 7, 2018)
Grid cells in the entorhinal cortex are believed to establish their regular, spatially correlated
firing patterns by path integration of the animal's motion. Mechanisms for path integration, e.g.
in attractor network models, predict stochastic drift of grid responses, which is not observed ex-
perimentally. We demonstrate a biologically plausible mechanism of dynamic self-organization by
which border cells, which fire at environmental boundaries, can correct such drift in grid cells. In
our model, experience-dependent Hebbian plasticity during exploration allows border cells to learn
connectivity to grid cells. Border cells in this learned network reset the phase of drifting grids. This
error-correction mechanism is robust to environmental shape and complexity, including enclosures
with interior barriers, and makes distinctive predictions for environmental deformation experiments.
Our work demonstrates how diverse cell types in the entorhinal cortex could interact dynamically
and adaptively to achieve robust path integration.
To survive, animals must have a way of knowing where
they are within an environment. A key component of this
self-localization is path integration [1], namely the ability
to continuously integrate velocity to keep track of posi-
tion [2 -- 4]. What neural mechanisms underlie path inte-
gration? Decades of research suggest that several brain
regions may be involved, including hippocampus (which
contains place cells [5 -- 7]) and entorhinal cortex (which
contains grid cells [8, 9] along with other spatially cor-
related neural types). It is believed that the response of
grid cells may be generated through path integration of
an animal's trajectory through space.
In models of path integration, noise in neurons and
inaccuracies in the integrator typically lead to accumu-
lating errors in the location estimate [3, 4], sometimes
causing substantial drifts over time in the grid response
pattern [10]. However, such drifts are rarely seen exper-
imentally grid cell firing. This suggests that the brain
must contain mechanisms, perhaps involving environ-
mental landmarks [2 -- 4, 11 -- 20], to correct for drifts and
to allow a stable representation of an animal's location.
The boundaries of an environment are obvious land-
marks, and it has been proposed that they provide cues
for path integration [8, 11 -- 13, 18, 21, 22]. Experimen-
tally, it is known that environmental boundaries mod-
ulate the firing properties of place cells [23] and grid
cells [24, 25]. Recently, a new class of boundary-sensitive
neurons was discovered in the entorhinal cortex and
nearby regions [26 -- 30], and there is evidence that these
"border cells" may be involved in error correction of grid
cell responses [15, 22].
∗ [email protected]
† [email protected]
‡ [email protected]; Equal contribution
§ [email protected]; Equal contribution
The authors of
[22] proposed a model with a simple
hand-crafted pattern of connectivity between border cells
and grid cells to correct drifts of grid firing patterns in
a square environment. This pioneering effort, however,
did not address how border-to-grid connectivity might be
learned to provide error correction in environments of dif-
ferent, and more complex, shapes. This is an important
challenge because environments change, animals move to
new locations, and natural habitats often have complex
geometries with internal barriers and passageways.
Here, we develop a modeling framework in which con-
nections between border cells and grid cells are self-
organized to provide an error correction signal for grid
cells while the animal freely explores an environment. In
our model, grid responses are generated by attractor dy-
namics, similar to previous models [10, 31 -- 33], but are
modified by a learned interaction with border cells. We
show that the resulting border-grid network dramatically
reduces drift in the grid system. Our proposed mecha-
nism can correct grid drifts across environments with dif-
ferent shapes and complexities, for example when inter-
nal barriers are introduced, enabling a stable representa-
tion of self-location in ethologically relevant situations.
Our model also makes distinctive predictions for grid
drift after changes in the spatial environment. Together,
these results provide insights into the use of environmen-
tal boundaries, and landmarks in general, to achieve ro-
bust path integration. This paper expands work pre-
sented at the Computational and Systems Neuroscience
meeting (CoSyNe) in February 2017.
RESULTS
Grids drift linearly with time in an attractor model
We constructed a continuous attractor model of the
grid system following [10, 33] (Methods). In this model,
grid cells are subject to external excitation modulated
2
boundaries in a fixed allocentric direction [26, 27]. Thus,
if the environment was deformed or a new internal barrier
was introduced, the model border cells continued to fire
along segments of the new boundary with same absolute
orientation and over similar length scales (Methods).
Self-organization of border-to-grid connectivity
In our model, border cells develop connectivity with
grid cells via experience-dependent Hebbian plastic-
ity [34]. When an individual border cell fires, its synap-
tic weight with a synchronously firing grid cell increases.
Specifically, the connection strength between the ith pre-
synaptic neuron (border cell) and the jth post-synaptic
neuron (grid cell), Wij, is updated according to a learning
rule [34], W t
ij + γxiyj , where γ is a learning rate
and xi, yj = 1, 0 denote firing or silence of border cell i
and grid cell j respectively (Methods). Divisive normal-
ization is then applied to maintain the overall connectiv-
ity strength between border and grid populations [35],
i.e., W t
i,j. Such normalization is needed
for the stability of Hebbian learning (e.g., [35]).
i,j/(cid:80)
ij = W t−1
i,j → W t
j W t
The border-to-grid connectivity pattern stabilizes as
the animal gains experience in an environment, and
strong connections form between each border cell and
grid cells that fire at locations within its spatial firing
field (Fig. 2 insets). Grid cells whose spatial phases are
shifted by a grid lattice vector will have identical con-
nection strengths, leading to the connectivity patterns in
Fig. 2. Our model predicts that the spatial phase of a
grid cell will be correlated with its connection strength
with a border cell in a manner that depends on the spa-
tial boundary geometry and on the border field location.
These predictions can be tested experimentally by re-
constructing the functional connectivity of grid cells and
border cells via simultaneous recording of populations
using multi-electrode recording or calcium imaging.
Error-correction in simple environments
Intuitively, the mechanism described above will store
the relative phase of grid and border cells in the strength
of synapses connecting them. Thus, in a mature net-
work, border cell firing should on average reinstate the
phase of grid cells that may have drifted. To test this
we allowed grids to form with and without simultaneous
development of border-grid connectivity in a square envi-
ronment. For the model with connectivity, the corrective
input to each grid cell was a sum over the activity of bor-
der cells weighted by the border-grid synaptic strength
(Methods). We quantified the displacement of grid fir-
ing fields relative to their initial location in terms of a
mean squared grid drift (Methods). Learned border-grid
connectivity dramatically attenuated drift in the square
environment (Fig. 3a,b). Grid stabilization was apparent
in single cell firing patterns (compare blue and red points
in Fig. 3a vs. Fig. 1a), and also persisted in circular en-
vironments (Fig. 3c,d). Thus, grid drift can be corrected
Figure 1. Grid drifts in a continuous attractor network
model. a) Response of a single model neuron in a continuous
attractor network model for the grid system. At a given time
a grid cell responds if the animal is located near the vertices
of a triangular lattice in space (blue points, burgundy points,
or red points). However, as time passes the grid pattern shifts
in space (blue = early times, burgundy = intermediate times,
red = late times). Averaging over time can destroy the grid-
like appearance of the firing pattern. b) Drift, quantified as
mean squared shift between triangular grid patterns, increases
linearly in time (see Methods). Error bars are standard error
over 50 replicate simulations of animal trajectories.
by movement velocity and inhibit each other recurrently.
The network self-organizes into a grid pattern of activity
on the neural sheet. The pattern is translated on the neu-
ral sheet as the animal moves, with the consequence that
a given grid cell fires when the animal is physically lo-
cated in a triangular grid of locations in the environment
(Fig. 1a, blue points or red points). We found that the
firing pattern of individual grid cells drifts in space over
time (Fig. 1a, blue to red gradient). The average cross-
correlation of grid firing patterns at different times was
used to quantify the mean squared grid drift (Methods)
as the simulated animal explored the environment (tra-
jectories in Methods). We found that the mean squared
drift increased linearly with time, as expected for a two
dimensional random walk (Fig. 1b).
Modeling responses of border cells
We considered that the drift observed in Fig. 1 might
be corrected if the grid attractor network interacted ap-
propriately with landmarks, notably boundaries. To
this end, we constructed a model population of border
cells [26 -- 30] which are known to fire near environmen-
tal boundaries (Methods). Consistently with experiment
[26] we assumed that border cells fire over scales that
range between half to all of the sidelength of rectangular
boundaries (125 cm to 250 cm in our simulation; Fig. 2).
While we assumed that border cells have some bias to-
wards lying on only one wall in the environment, we also
allowed them to wrap around corners (Fig. 2), consistent
with the observation that border cells frequently have
a "dominant wall" but are not necessarily wall-specific
[26]. We also assumed that border cells map to new en-
vironmental configurations by firing in the proximity of
3
Figure 2. Border cell firing and connectivity. Four ex-
ample border cells and their firing fields are shown around
the boundary of the spatial enclosure (large square). Insets:
Heatmap indicates the strength of connections between indi-
cated border cell and grid cells organized by spatial phase on
the neural sheet. The periodic connectivity pattern occurs
because grid cells whose phase differs by a grid lattice vector
respond in the same physical locations.
Figure 3. Correction of grid drift by learned border input.
(a-b) Square environment. Top: firing pattern for a single
grid cell with error correction by border inputs, blue dots =
early times, red dots = late times. Bottom: Grid drift as
a function of time with (green) and without (black) border
cell inputs. Error bars = standard error over 50 replicate
simulations.
(e-f)
Similarly for a square with an internal barrier.
(c-d) Similarly for a circular enclosure.
by interaction with border cells in simple convex envi-
ronments where each border field is attached to a single,
compact region of the boundary.
Error correction in complex environments
Ethologically relevant environments can have complex,
non-convex shapes, sometimes with internal barriers. In
such spaces border cells may fire in multiple disjoint re-
gions. For example, in the presence of an internal barrier
(Fig. 3e), border cells can respond both to the barrier and
to walls parallel to the barrier [26, 27]. In such situations
grid cells with different phases can respond at the same
time as a given border cell with disjoint response fields.
This possibility complicates the grid phase re-setting by
border cell input that is required to correct drift. We
therefore tested our mechanism in an enclosure with an
internal barrier (Fig. 3e), with border cells assigned to
respond in the same allocentric direction at the barrier
and boundary walls (Methods), consistently with experi-
ment [26, 27]. Hebbian plasticity (Methods) caused bor-
der cells to form strong connections with concurrently
firing grid cells, so that cells with multiple border fields
synapsed with grid cells with different phases. Again,
grid drift was strongly attenuated by interaction with
border cells (Fig. 3f). Error-correction occurs robustly
so long as accumulated drift is small, because in this case
the correct grid cells will be primed to fire near the inter-
nal barrier and near the boundary, albeit a little out of
phase. Thus, border cell firing will simply reinstate the
proper phase in the grid cell, causing it to "snap back" to
the right spatial pattern. If the accumulated drift were
large, error correction could fail, but this is not the case
when there is regular contact with boundaries.
Error correction in changing environments
Real-world navigation occurs in a dynamic context
where environments change, and where objects, barri-
ers, and other animals enter or leave a space. To test
whether border cell input can stabilize grids during and
after a spatial deformation we first morphed a square
enclosure into a disc, while mapping border cells so that
they responded to boundaries in the same allocentric ori-
entations in both shapes (Fig. 4, top row). After the de-
formation, stability of the grid pattern was temporarily
disrupted and the rate of drift increased dramatically.
In fact the drift was higher than it would have been
without an error-correction mechanism. However, as the
animal continued to explore, ongoing plasticity reorga-
nized border-grid connections to be appropriate for the
new enclosure, once again stabilizing the grid. We also
tested two milder deformations -- the sudden insertion
of a narrow internal barrier (Fig. 4, middle row), and
compression of the square to a rectangle (Fig. 4, bot-
tom row). Again the grids were transiently disrupted
and then stabilized, but the effects were weaker than af-
ter the more dramatic square-to-disc deformation. These
results highlight the importance of ongoing learning and
self-organization in correcting errors and maintaining sta-
bility in path integration mechanisms operating in dy-
namically changing environments.
DISCUSSION
We have demonstrated that border cells that self-
organize connectivity with grid cells can correct accu-
mulating positional drift, even in complex geometries
that change in time. We expect that such interac-
tions with boundaries will also be able to correct path-
integration errors in other models of grid formation [36]
4
Figure 4. Correction of grid drift in changing environments. (Top Row, a-e) Square-Disc deformation. The square environment
in (a) is deformed to a disc in (b) at 1200s leading to disruption of the grid pattern which then stabilizes in (c) at late times.
Border firing fields in (d) map allocentrically between enclosures. Without border inputs grids drift similarly before and after
the deformation. Self-organized error correction from border cells leads to much lower drift before the deformation, a transient
increase in drift after the deformation, and re-stabilization as the changed environment is re-learned (error bars = standard
error over 50 replicates). (Middle Row, f-j) Insertion of an internal barrier. (Bottomg Row, k-o) Square-Rectangle deformation.
and boundary-related responses [16, 23, 27, 37]. Indeed,
such anchoring may be generally necessary for reliable
evidence accumulation with noisy neurons [38].
Many authors have noted that diverse sources of sen-
sory and non-sensory information including landmarks
and boundaries must be merged to maintain reliable spa-
tial representations during navigation [14, 17, 18, 39 -- 42].
In the grid system such external information must be
passed in without interrupting the internal dynamics, so
that grid cells can continue to perform path integration
while being informed by external cues. Our results show
that boundary information, at least, will be effectively
injected if border cells and grid cells that fire in phase
develop strong synapses. Perhaps this simple mechanism
provides a prototype for more general cue integration in
the spatial navigation systems of the brain.
Several authors have explored the learned use of land-
mark and sensory information to correct errors, e.g.,
in the head-direction system [14, 43, 44], and in spa-
tial representations maintained by robots [20]. One
study showed that learned connections between "land-
mark cells" and grid cells would lead to grid firing de-
formations consistent with experiment [45]. Meanwhile,
[46] explored how learned modifications of border-grid
synapses affect the error-correction scenario of [22]. We
add to this literature by showing how a particular kind
of landmark -- a boundary -- can provide cues for error
correction even in complex and changing environments.
Our model makes several testable predictions. First,
border-grid connection strengths will be correlated with
grid cell spatial phase in a boundary geometry and bor-
der field location dependent manner. Second, the con-
nectivity pattern will adapt to recent experience. Third,
grid firing patterns will become animal trajectory depen-
dent when an environment changes shape because grids
remain tethered to the now-deformed boundary. Very re-
cently, the authors of [47] presented direct evidence for
such trajectory dependence, and argued that averaging
over trajectories partially reproduces the apparent effect
of grid rescaling after environmental deformation [48, 49].
It is challenging to stabilize path integrators in com-
plex environments. Our paper proposes a simple solution
that might be realized in the brain, and will be useful in
the design of robots learning to navigate rich terrains.
ACKNOWLEDGMENTS
This work was partially supported by the Honda Re-
search Institute and NSF grant PHY-1734030 (VB), NSF
grant PHY-1607611 (Aspen Center for Physics), the
Penn Vagelos Molecular Life Sciences Program (EP, ND)
and the NSF NeuroNex Award DBI-1707398 (X-XW).
METHODS
Attractor network model for the grid cells
We implemented the grid cell attractor model in [10],
modified as in [33] to simplify neural connectivity. A
two-dimensional sheet of N × N neurons with periodic
boundary conditions follows an update equation
τ
dsi
dt
+ si = g
Mij sj + I + α vt cos(θt − θi) + Ci(t)
(cid:105)
+
(cid:104)(cid:88)
j
Here, si is the activity level of neuron i, g is a gain param-
eter, τ is the time constant, I is a constant external input,
α defines the velocity gain, vt and θt are the speed and
head direction at time t, and θi is the preferred direction
of neuron i. Preferred direction is chosen from the four
cardinal directions by tiling the neural sheet with 2 × 2
squares containing up, down, left, and right preferences
in a fixed pattern. The term Ci(t) is the corrective input
from the border cells to grid cell i at time t, and is defined
later. In simulations without border cells, Ci(t) = 0. The
notation(cid:2).(cid:3)
+ indicates a rectified linear function.
Connection strengths between neurons are determined
by their distance on the neural sheet. Neurons within
a disc of radius R have a constant inhibitory weight
M0, while neurons outside have zero weight. The disc
is slightly offset by a distance l in the neuron's pre-
ferred direction. All told the connection weights be-
tween neurons at (xi, yi) and (xj, yj) are given by Mij =
M0×Θ
,
where Θ is the Heaviside step function. For parameter
values, we used N = 32, τ = 10 ms, dt = 1 ms, g = 1,
I = 3, α = 2, M0 = −0.05, R = 13, and l=2.
R−(cid:112)(xi − xj − l cos θi)2 + (yi − yj − l sin θi)2
(cid:16)
(cid:17)
5
wall. The border field length was chosen uniformly from
the interval (L, L/2), where L is the wall length. The
average border cell thus took up 3/4 of a wall and was
defined to have a field width of 10 cm, consistently with
observation [26]. Border cells whose ends went beyond
the wall on which they were centered wrapped around
corners to other walls. Thus border cells have a bias
towards lying on a single "dominant wall", but are not
wall-specific. We used 16 border cells in all simulations.
To create border cells for circular environments, we
projected border fields from the square environment out
to the circumscribing circle. For rectangular environ-
ments, we rescaled border fields to occupy the same pro-
portion of space on walls that were compressed. Finally,
when introducing a barrier, we defined border cells so
that they responded in the same allocentric direction at
the barrier and the boundary walls. For example, if a
border cell fired when the animal was at the lower part of
the right wall, it also fired at the corresponding location
on the left face of a barrier inserted into the environment.
Learning border-to-grid connectivity
The update rule and divisive normalization for border-
grid connection weights Wij are given in the main text.
Grid and border cells spiked with Poisson statistics. In
each 1 ms timebin, border cell firing probability was 0.01
when the animal was inside the firing field and 0 oth-
erwise. For grid cells, we followed a hybrid of [10] and
[33] and defined firing probability in 1 ms timebins as
the synaptic activation (right hand side of the attractor
equation) times 0.118. The corrective input to the jth
grid cell was a weighted sum of border cell activities
(cid:88)
Cj = β
Wij xi
Navigation Simulation
Animal speed, head direction, and border interactions
were generated by simulating an animal's path.
Environments: The basic enclosure was a square with
side length 250 cm. Circular environments were chosen
to exactly circumscribe the square. The barrier environ-
ment had an internal wall of width 20 cm and length 125
cm, touching the southern wall at a location 30 cm to
the right of the midline of the square enclosure.
Paths: Trials ran for 2400 seconds of simulation time.
Animals moved at a constant speed of 1 m/s, starting
from the center of the environment in a random direc-
tion. Every 0.1 seconds of simulation time (100 time
steps), we chose a new direction by drawing from a Gaus-
sian distribution with a standard deviation of one radian.
Whenever the animal reached a boundary we randomly
reassigned the head direction for the next step so that
the trajectory remained within the enclosure.
Border cell assignment
In the square enclosure each border cell was centered
randomly along the middle 50% of a randomly chosen
i
with β = 200 and xi = 1, 0 if a border cell fires or not.
Quantifying grid cell drift
To quantify drift of grid response fields, we recorded
locations of the Poisson spikes of a single grid cell. Simu-
lations were broken into 12 windows of 200 seconds each
that were long enough to explore the environment, but
short enough that the grids remained stable.
In each
window, we created a two-dimensional histogram of spike
locations, with bin widths of 1 cm. We then computed
a cross-correlation matrix of time-adjacent histograms.
The drift between time windows was calculated as the
displacement from the origin of the most central peak
of the cross-correlogram. By adding drift vectors across
subsequent time windows, we quantified drift in single
trials. To generate drift plots (Fig. 1b, Fig. 3 bottom,
and Fig. 4 right, we averaged the squared magnitude of
drift across 50 trials in each time window.
6
[1] C. Darwin, Nature 7, 417 (1873).
[2] R. F. Wang and E. S. Spelke, Trends in cognitive sciences
6, 376 (2002).
[3] A. S. Etienne and K. J. Jeffery, Hippocampus 14, 180
(2004).
[4] B. L. McNaughton, F. P. Battaglia, O. Jensen, E. I.
Moser, and M.-B. Moser, Nature Reviews Neuroscience
7, 663 (2006).
[5] J. O'Keefe and J. Dostrovsky, Brain research (1971).
[6] J. O'Keefe, Experimental neurology 51, 78 (1976).
[7] J. O'keefe and L. Nadel, The hippocampus as a cognitive
map (Oxford: Clarendon Press, 1978).
[8] T. Hafting, M. Fyhn, S. Molden, M.-B. Moser, and E. I.
Moser, Nature 436, 801 (2005).
[9] M. Fyhn, S. Molden, M. P. Witter, E. I. Moser, and
M.-B. Moser, Science 305, 1258 (2004).
[10] Y. Burak and I. R. Fiete, PLoS computational biology 5,
[28] F. Savelli, D. Yoganarasimha, and J. J. Knierim, Hip-
pocampus 18, 1270 (2008).
[29] S. Stewart, A. Jeewajee, T. J. Wills, N. Burgess, and
C. Lever, Philosophical Transactions of the Royal Society
B: Biological Sciences 369, 20120514 (2014).
[30] L. Muessig, J. Hauser, T. J. Wills, and F. Cacucci, Neu-
ron 86, 1167 (2015).
[31] M. C. Fuhs and D. S. Touretzky, The Journal of Neuro-
science 26, 4266 (2006).
[32] T. Bonnevie, B. Dunn, M. Fyhn, T. Hafting, D. Derdik-
man, J. L. Kubie, Y. Roudi, E. I. Moser, and M.-B.
Moser, Nature neuroscience 16, 309 (2013).
[33] J. J. Couey, A. Witoelar, S.-J. Zhang, K. Zheng, J. Ye,
B. Dunn, R. Czajkowski, M.-B. Moser, E. I. Moser,
Y. Roudi, et al., Nature neuroscience 16, 318 (2013).
[34] D. O. Hebb, "The organization of behavior," (1949).
[35] K. D. Miller and D. J. MacKay, Neural Computation 6,
e1000291 (2009).
100 (1994).
[11] B. McNaughton, L. Chen,
and E. Markus, Cognitive
[36] M. E. Hasselmo, L. M. Giocomo, and E. A. Zilli, Hip-
Neuroscience, Journal of 3, 190 (1991).
pocampus 17, 1252 (2007).
[12] D. S. Touretzky and A. D. Redish, Hippocampus 6, 247
(1996).
[13] J. O'Keefe and N. Burgess, Hippocampus 15, 853 (2005).
[14] W. E. Skaggs, J. J. Knierim, H. S. Kudrimoti, and B. L.
McNaughton, in Advances in neural information process-
ing systems (1995) pp. 173 -- 180.
[37] C. Barry, C. Lever, R. Hayman, T. Hartley, S. Burton,
J. O'Keefe, K. Jeffery, and N. Burgess, Reviews in the
Neurosciences 17, 71 (2006).
[38] A. A. Faisal, L. P. Selen, and D. M. Wolpert, Nature
reviews neuroscience 9, 292 (2008).
[39] K. M. Gothard, W. E. Skaggs, and B. L. McNaughton,
[15] L. M. Giocomo, The Journal of physiology 594, 6501
Journal of Neuroscience 16, 8027 (1996).
(2016).
[16] C. Lever, N. Burgess, F. Cacucci, T. Hartley,
J. O'Keefe, Biological cybernetics 87, 356 (2002).
[40] J. A. Perez-Escobar, O. Kornienko, P. Latuske,
and
L. Kohler, and K. Allen, Elife 5 (2016).
[41] F. Raudies, J. R. Hinman, and M. E. Hasselmo, The
[17] L. F. Jacobs and F. Schenk, Psychological review 110,
Journal of physiology 594, 6513 (2016).
285 (2003).
[42] F. Raudies and M. E. Hasselmo, PLoS computational
[18] A. Cheung, D. Ball, M. Milford, G. Wyeth, and J. Wiles,
biology 11, e1004596 (2015).
PLoS computational biology 8, e1002651 (2012).
[19] F. Savelli, J. Luck, and J. J. Knierim, Elife 6 (2017).
[20] M. Mulas, N. Waniek, and J. Conradt, Frontiers in com-
putational neuroscience 10, 13 (2016).
[43] K. Zhang, Journal of Neuroscience 16, 2112 (1996).
[44] J. Knierim, H. Kudrimoti, W. Skaggs,
and B. Mc-
Naughton, Pergamon studies in neuroscience 13, 343
(1996).
[21] J. Krupic, M. Bauza, S. Burton, and J. O'keefe, The
[45] S. Ocko, K. Hardcastle, L. M. Giocomo, and S. Ganguli,
Journal of physiology 594, 6489 (2016).
bioRxiv , 326793 (2018).
[22] K. Hardcastle, S. Ganguli, and L. M. Giocomo, Neuron
86, 827 (2015).
[23] J. O'keefe and N. Burgess, Nature 381, 425 (1996).
[24] T. Stensola, H. Stensola, M.-B. Moser, and E. I. Moser,
[46] D. Santos-Pata, R. Zucca, S. C. Low, and P. F. Ver-
schure, Frontiers in computational neuroscience 11, 65
(2017).
[47] A. T. Keinath, R. A. Epstein, and V. Balasubramanian,
Nature 518, 207 (2015).
bioRxiv , 174367 (2017).
[25] J. Krupic, M. Bauza, S. Burton, C. Barry, and J. OKeefe,
[48] C. Barry, R. Hayman, N. Burgess, and K. J. Jeffery,
Nature 518, 232 (2015).
Nature neuroscience 10, 682 (2007).
[26] T. Solstad, C. N. Boccara, E. Kropff, M.-B. Moser, and
[49] H. Stensola, T. Stensola, T. Solstad, K. Frøland, M.-B.
E. I. Moser, Science 322, 1865 (2008).
Moser, and E. I. Moser, Nature 492, 72 (2012).
[27] C. Lever, S. Burton, A. Jeewajee, J. O'Keefe,
and
N. Burgess, The journal of neuroscience 29, 9771 (2009).
|
1603.02007 | 3 | 1603 | 2017-08-29T12:55:07 | Fluctuation analysis in nonstationary conditions: single Ca channel current in cortical pyramidal neurons | [
"q-bio.NC"
] | Fluctuation analysis is a method which allows measurement of the single channel current of ion channels even when it is too small to be resolved directly with the patch clamp technique. This is the case for voltage-gated Ca2+ channels (VGCCs). They are present in all mammalian central neurons, controlling presynaptic release of transmitter, postsynaptic signaling and synaptic integration. The amplitudes of their single channel currents in a physiological concentration of extracellular Ca2+, however, are small and not well determined. But measurement of this quantity is essential for estimating numbers of functional VGCCs in the membrane and the size of channel-associated Ca2+ signaling domains, and for understanding the stochastic nature of Ca2+ signaling. Here, we recorded the VGCC current in nucleated patches from layer 5 pyramidal neurons in rat neocortex, in physiological external Ca2+ (1-2 mM). The ensemble-averaging of current responses required for conventional fluctuation analysis proved impractical because of the rapid rundown of VGCC currents. We therefore developed a more robust method, using mean current fitting of individual current responses and band-pass filtering. Furthermore, voltage ramp stimulation proved useful. We validated the accuracy of the method by analyzing simulated data. At an external Ca2+ concentration of 1 mM, and a membrane potential of -20 mV, we found that the average single channel current amplitude was about 0.04 pA, increasing to 0.065 pA at 2 mM external Ca2+, and 0.12 pA at 5 mM. The relaxation time constant of the fluctuations was in the range 0.2-0.8 ms. The results are relevant to understanding the stochastic properties of dendritic Ca2+ spikes in neocortical layer 5 pyramidal neurons. With the reported method, single channel current amplitude of native VGCCs can be resolved accurately despite conditions of unstable rundown. | q-bio.NC | q-bio | Fluctuation analysis in nonstationary conditions:
single Ca channel current in cortical pyramidal neurons
C. Scheppach, H. P. C. Robinson
Running title: fluctuation analysis: Ca channel current
Abstract
Fluctuation analysis is a method which allows measurement of the single channel current of
ion channels even when it is too small to be resolved directly with the patch clamp technique.
This is the case for voltage-gated calcium channels (VGCCs). They are present in all
mammalian central neurons, controlling presynaptic release of transmitter, postsynaptic
signaling and synaptic integration. The amplitudes of their single channel currents in a
physiological concentration of extracellular calcium, however, are small and not well
determined. But measurement of this quantity is essential for estimating numbers of
functional VGCCs in the membrane and the size of channel-associated calcium signaling
domains, and for understanding the stochastic nature of calcium signaling. Here, we recorded
the voltage-gated calcium channel current in nucleated patches from layer 5 pyramidal
neurons in rat neocortex, in physiological external calcium (1-2 mM). The ensemble-
averaging of current responses required for conventional fluctuation analysis proved
impractical because of the rapid rundown of calcium channel currents. We therefore
developed a more robust method, using mean current fitting of individual current responses
and band-pass filtering. Furthermore, voltage ramp stimulation proved useful. We validated
the accuracy of the method by analyzing simulated data. At an external calcium concentration
of 1 mM, and a membrane potential of −20 mV, we found that the average single channel
current amplitude was about 0.04 pA, increasing to 0.065 pA at 2 mM external calcium, and
0.12 pA at 5 mM. The relaxation time constant of the fluctuations was in the range 0.2–0.8 ms.
The results are relevant to understanding the stochastic properties of dendritic Ca2+ spikes in
neocortical layer 5 pyramidal neurons. With the reported method, single channel current
amplitude of native voltage-gated calcium channels can be resolved accurately despite
conditions of unstable rundown.
Introduction
When an ion channel opens, a characteristic fixed current flows through it, which can be
directly measured with the patch clamp method (1). Even when the single-channel current is
too small to be resolved with this technique, it can still be inferred by fluctuation analysis,
where the current through many channels is measured and the fluctuations arising from the
opening and closing of individual channels are analyzed (2–5).
Voltage-gated calcium channels (VGCC) (6–8) are central to many physiological processes,
and are present in most neuronal membranes, where they have important roles in signaling.
The apical dendrites of neocortical layer 5 (L5) pyramidal neurons (9, 10) exhibit action
potentials carried by Ca2+ (11). These dendritic Ca2+ spikes provide an amplification
mechanism for distal synaptic inputs, triggering bursts of sodium action potentials at the soma
and hence constitute a fundamental mechanism of synaptic integration in these neurons (12,
13). Similar dendritic calcium spikes have been found in hippocampal CA1 pyramidal cells
(14), and Purkinje neurons (15). At presynaptic terminals, calcium influx through voltage-
gated calcium channels triggers the vesicular release of neurotransmitter (16).
The amplitude of currents through native single Ca2+ channels at physiological extracellular
Ca2+ concentrations (1-2 mM) (17) is a key parameter. Knowing it is essential to be able to
relate the density of functional channels to the size of whole-cell calcium currents, to
characterize the stochasticity of calcium signaling due to random opening and closing of
single ion channels (18–20), and to understand the formation of calcium nanodomains, highly
localized and concentrated plumes of calcium on the cytoplasmic side of open calcium
channels, allowing specific signaling (21), which depends on the rate of Ca2+ entry through
individual channels.
recording under sophisticated
However, physiological single CaV channel current has not been extensively investigated,
with only a few estimates in the literature (22, 23, 21, 24). In chick dorsal root ganglion
neurons, values between 0.2 and 0.33 pA were estimated for different VGCC subtypes (21),
at a membrane potential of −65 mV and an extracellular calcium concentration of 2 mM, by
direct single-channel
low-noise conditions, and by
extrapolating from measurements at more elevated calcium concentrations. Such a small
amplitude makes it extremely challenging to resolve and estimate accurately. Single calcium
channels are therefore studied almost exclusively in high extracellular barium solutions (e.g.
(25)), since barium permeates most VGCC subtypes at higher rates than does calcium, giving
much larger, detectable channel current amplitudes close to 1 pA in size.
To better understand the stochastic properties of Ca2+ spikes in neocortical L5 pyramidal
neurons, we wanted to measure the single channel current of the native Ca2+ channels in these
cells, in physiological [Ca2+]o, for which fluctuation analysis appeared a suitable approach.
However, because of rapid run-down of the channel current (a common problem in the study
of Ca2+ channels, see (26, 27)), fluctuation analysis cannot be applied in its usual form,
relying on stable conditions to obtain mean current and current variance by ensemble-
averaging across several recording sweeps. We therefore developed a modified fluctuation
analysis protocol which extracts information from individual sweeps, based on mean current
fitting and band-pass filtering. Furthermore, we used voltage ramp stimuli rather than the
more customary voltage steps. The method was validated by modeling. We obtain a mean
single Ca2+ channel current of 0.065 pA at −20 mV and 2 mM extracellular calcium, which is
about one order of magnitude below the resolution limit of standard patch-clamp single-
channel recordings.
Methods
Tissue preparation
Animal procedures adhered to U.K. Home Office legislation and University of Cambridge
guidelines. Acute brain slices were obtained from Wistar rats aged 6-14 days. Animals were
killed by dislocation of the neck, and parasagittal brain slices were prepared as detailed in
(28). The brain was cut along the midline and one hemisphere was glued by the medial
surface to a tilted block (12°) on top of the horizontal slicer stage, slightly raising the ventral
part of the brain. 300 µm thick slices were obtained with a vibratome (VT1200S by Leica,
Milton Keynes, U.K.), with the dorsal cortical surface facing the horizontal blade. Slices were
incubated at 34°C for 30 min and then kept at room temperature.
Solutions, drugs, chemicals
Extracellular solution (used for slicing and perfusion of slices) had the following
composition: 125 mM NaCl, 2.5 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 1.25 mM NaH2PO4,
25 mM NaHCO3, 25 mM glucose, bubbled with carbogen gas (95% O2, 5% CO2), pH 7.4 (c.f.
(29) p. 200, (30, 28)). Intracellular Cs-based pipette solution had the following composition:
90 mM Cs-methanesulfonate, 30 mM CsCl, 10 mM BAPTA-Na4 (Ca2+ chelator), 11 mM
HEPES buffer, to which was added 4 mM ATP (ATP Mg salt obtained from Sigma,
Gillingham, U.K.; free Mg: about 1 mM), 0.3 mM GTP (GTP Na salt hydrate obtained from
Sigma; 1.2 mM Na+), and 10 mM creatine phosphate (CrP) Na2 (Na CrP dibasic tetrahydrate
obtained from Sigma), to mitigate run-down of Ca2+ channels (26), and pH was adjusted to
7.3 with NaOH and HCl (about 4.5 mM NaOH for HEPES, about 8 mM NaOH for ATP). The
liquid junction potential of 12.2 mV was corrected for. HEPES-buffered extracellular solution
consisted of: 145 mM NaCl, 2.5 mM KCl, 2 mM CaCl2, 1 mM MgCl2, 10 mM HEPES, pH
adjusted to 7.4 with about 4.2 mM NaOH. Tetrodotoxin (TTX, Na channel blocker) was
obtained from Alomone Labs (Jerusalem, Israel) and Tocris (Bristol, U.K.).
Recording
Slices were viewed with an Olympus (Southend-on-Sea, U.K.) BX50WI fixed stage upright
microscope and a x60 objective, or a x10 objective for larger-scale overview, with infrared
(IR) or visible light differential interference contrast (DIC) optics and a Lumenera (Ottawa,
Canada) Infinity 3M CCD camera. To assist orientation in the slice, the position in the slicing
plane was monitored with an optical position encoder (Renishaw, Wotton-under-Edge, U.K.),
displaying x and y coordinates at micron resolution. Pipette pressure was monitored by a
pressure meter. Voltage clamp data were collected with a MultiClamp 700B (Axon
Instruments, Molecular Devices, Wokingham, U.K.) amplifier, with a feedback resistor of 50
GΩ and a 10 kHz 4-pole Bessel filter, with waveform generation and sampling at 16-bit
resolution at a frequency of 50 kHz, using a National Instruments (Newbury, U.K.) analog
interface and custom software written in C++ and Matlab (Mathworks, Cambridge, U.K.).
During the experiment, perfusion with extracellular solution was maintained in the recording
chamber with a perfusion line and a suction line. All experiments were performed at room
temperature (20-23°C).
Pyramidal neurons in cortical layer V were visually identified. In order to minimize
variability, cells from the same cortical location were consistently targeted. Slices were
always obtained from the right hemisphere. In consecutive parasagittal slices, the orientation
of pyramidal cell dendrites changes from slice to slice, and slices in which the dendrites run
roughly parallel to the slice surface were selected, which came from about half-way between
the midline and the lateral brain surface. The central part of the cortex in the rostral-caudal
direction was targeted, corresponding to the sensory hindlimb region (31).
Patch pipettes were pulled from borosilicate glass capillaries (Harvard Apparatus, Cambridge,
U.K.; 1.5 mm outer diameter, 0.86 mm inner diameter, filamented glass) and fire-polished.
Tip resistances (with the Cs-based pipette solution) were between 5.5 and 6.5 MΩ. Nucleated
patch recordings were performed as described in (32). A whole-cell recording was obtained,
and a negative pressure of about −200 mbar was applied to the pipette, aspirating the cell
nucleus to the pipette tip. Maintaining the negative pressure, the pipette was slowly and
carefully retracted, such that the nucleus was pulled out of the cell and acted as a place holder
around which the cell membrane resealed. The nucleated patch was pulled out of the slice and
then raised well above it. Finally, the negative pipette pressure was reduced to about −50
mbar. The process from initial contact of the membrane with the pipette tip to the first
recorded sweep took about 4 to 7 minutes.
Ca2+ channel currents were recorded during depolarizing ramps, from a holding potential of
−70 mV rising to 0 mV (slope 0.6 mV/ms, stimulus repeated every 3 s), and also during
depolarizing voltage steps (holding potential −70 mV, 100 ms long steps to test potentials
cycling through −60, −40, −30, −20, −10 and 0 mV, stimulus repeated every 3 s).
Data analysis, modeling and statistics
Analysis was carried out using custom code in MATLAB and R. Leak subtraction for step
responses (Fig. 1A) was carried out by subtraction of the smoothed, scaled response to a small
step (holding potential +10 mV) below the activation range of the calcium current. All current
traces were digitally filtered with a low-pass Gaussian filter ((29) p. 485) at a −3dB corner
frequency fc = 1 kHz. To isolate current fluctuations, voltage-ramp-evoked currents were
fitted with a piecewise polynomial, consisting of a flat segment and two linear segments
linked by cubic splines (see Supporting Material), giving a smooth, continuous-derivative
heuristic function whose eight parameters were fitted by minimizing squared deviation from
the data, using a Nelder-Mead simplex algorithm (Matlab function fminsearch), which was
taken as the estimate of the mean current 𝜇!. Following subtraction of this function, the
band-pass filtered fluctuations (𝜎!!) was computed over brief (5 ms) time bins, and the single
channel current i estimated from the relationship 𝑖≈𝜎!! 𝜇! (see Results).
fluctuating current responses were passed through a high-pass filter, by subtracting from the
signal its low-pass-filtered version (Gaussian filter with fc = 50 Hz, resulting in a high-pass
filter with corner frequency 94 Hz), to reject extraneous low-frequency noise and error
introduced at low frequencies by the mean-fitting procedure. The variance of the resulting
Weighted straight-line fits with fit parameter estimates, errors and correlations were obtained
with the statistics package R.
Results
Calcium channel currents in nucleated patches
To perform fluctuation analysis effectively, one needs a large membrane containing as many
channels as possible, but which can be voltage-clamped reliably. Normal whole-cell voltage-
clamp recording is impractical in pyramidal neurons because the large apical dendrite leads to
space clamp problems. Therefore, we used the nucleated patch technique, which combines a
large available membrane area with excellent space clamp properties.
Voltage-gated Ca2+ channel currents were recorded in nucleated patches from neocortical L5
pyramidal neurons, using a cesium-based intracellular pipette solution to block potassium
channel currents (26), 1 µM TTX in the Ringer solution to block sodium channel currents,
and 2 mM extracellular calcium. Figure 1A shows an example family of leak-subtracted
currents in response to step depolarizations. The current is activated around −30 mV and has a
more pronounced inactivating component as the membrane is depolarized further. The small
amplitude of these currents at physiological calcium concentration means that the signal-to-
noise ratio is also small. Responses to ramp depolarization (Fig. 1B) also clearly show the
onset and voltage-dependence of the current. Application of 200 µM cadmium completely
abolished the current (Fig. 1C, n = 11 cells), allowing its unambiguous identification as a
voltage-gated calcium channel current. As expected for calcium currents (26, 27), despite
inclusion of ATP and GTP in the pipette, there was often a relatively fast rundown over the
first 5-10 minutes following establishment of the recording (Fig. 1B).
FIGURE 1 Voltage-gated Ca2+ channel currents in nucleated patches from L5 pyramidal
neurons. A) Leak-subtracted inward voltage-gated Ca2+ channel current during a family of
voltage steps from −70 mV to −40, −30, −20, −10, 0 mV. B) Three current responses during
voltage ramps from a holding potential of −70 mV to 0 mV. During the initial part of the
ramp, there is a linear leak current (indicated by grey lines; grey vertical lines indicate the
range of the straight-line fit, from 5 ms to 40 ms after ramp onset). From about −40 mV, an
inward current develops. (Ramp stimulus was repeated every 3 s. The three responses shown
are each 21 s apart. Dashed horizontal line: zero current.) C) The inward current is eliminated
18 s after addition of extracellular Cd2+ (200 µM).
Fluctuation analysis in slowly nonstationary conditions
Single calcium channel openings at physiological calcium concentrations are effectively
impossible to resolve directly, because the combination of their small amplitude and short
open duration means that their signal falls below the noise level of the patch-clamp technique.
With the fluctuation analysis method, it is possible to infer the single channel current
nevertheless, by measuring the current through many channels and quantifying its fluctuations
arising from the opening and closing of individual channels (2–5). Assuming N identical,
independent channels with open probability p leads to the following relationship of total
current variance σI
where i is the single channel current (review: (29) p. 81, (6) p. 384). If p is small, then
Fluctuation analysis of voltage-gated channels is usually done by measuring channel currents
during steps to a constant voltage (4, 5). Current mean and variance are obtained by ensemble
averaging a number of consecutive sweeps. Computing a suitable ensemble average relies on
complete stability of conditions and number of available channels over a considerable number
(at least ≈20) of consecutive sweeps. However, this is not possible in the present case of
𝜎!!=𝑖𝜇!−!!!! ,
𝜎!!≈𝑖𝜇!.
2 to the ensemble mean level of current µI:
(1)
(2)
native voltage-gated calcium channel currents. The channel current runs down visibly within
one minute (Fig. 1B). In addition, the resistance and reversal potential of the linear leak
current are not perfectly constant from sweep to sweep (Fig. 1B). While these instabilities
constitute only very low-frequency variability and occur despite a good noise level at medium
frequencies, they thwart the customary approach of obtaining current mean and variance from
multiple successive ensemble trials.
We therefore aimed to extract information from individual recording sweeps. For that, we
obtained the mean current not by averaging several successive sweeps, but by fitting a smooth
curve to the current trace. The approach is similar to that used for quantally-varying
nonstationary synaptic currents (33, 34).
Secondly, we found it advantageous to use not voltage steps, but voltage ramp stimuli, which
allows easy leak subtraction by fitting a straight line to the initial part of the ramp response,
produces a smooth mean channel current time course that can be well fitted, and naturally
samples the full voltage range where the channels are active.
Fluctuation analysis of voltage ramp responses
The leak current was obtained from a straight line fit to the initial part of the ramp response
from 5 ms to 40 ms after the start of the ramp (Fig. 1B), corresponding to a potential range of
−67 to −46 mV, below the activation range of the calcium current. After subtraction of the
linear leak current, to obtain the mean current, each ramp response was fitted by least-squares
(see Methods) with a piecewise polynomial function (Fig 2A, see Methods). This heuristic
function gave very good fits to current traces.
FIGURE 2 Isolation of current fluctuations. A) A heuristic piecewise polynomial function
with continuous derivative (see methods) was fitted by least squares to individual leak-
subtracted ramp responses (black trace). There were no systematic deviations of fits (grey
curve) from individual current traces. Grey dots indicate the range of the fitted data. B)
Current variance (black crosses) was computed over 5 ms time bins through the current
fluctuations (central black trace, obtained by subtracting from the current trace (panel A) the
fitted mean current, and high-pass filtering). Black curve in the lower part indicates the mean
current fit.
We then applied high-pass filtering to the mean-subtracted responses. This step increased the
robustness of the measurement, by confining the signal to a bandwidth (94 Hz - 1 kHz)
containing a substantial fraction of the channel noise power, while rejecting potentially large
fluctuations of baseline at low frequencies. Secondly, it dealt with the problem that fitting the
mean current to each individual ramp response leads to an underestimation of the current
variance (see also Discussion). Since we subtract from the data not the underlying true mean
current, but the fitted mean current, the obtained fluctuation trace is contaminated by their
difference. This difference contains only low frequencies and is hence strongly attenuated by
the high-pass filter. Fig. 2B shows an example of how the variance of the filtered fluctuations
Iresidual(t), computed as the mean of Iresidual(t)2 over a time bin, evolves during an individual
current response to a voltage ramp.
The drawback of high-pass filtering the fluctuations is that it also removes some of the
fluctuation power due to channel gating and therefore the variance must be corrected for this.
To that end, we investigated the spectral composition of the fluctuations, so that the effect of
filtering could be quantified. We found that the autocorrelation function of the fluctuations
was well-fitted by a single exponential relaxation (Lorentzian frequency component) passed
through the same filtering (see Supplemental Material, Fig. S2A) with noise time constant τ in
the range 0.2-0.8 ms (detailed data: Supplemental Material, Fig. S2B). The combined effect
of high- and low-pass filtering cuts a substantial amount of the power of channel gating
fluctuations (see Supplemental Material, Fig. S2C), resulting in a correction factor γ by which
is the experimentally
the current variance is reduced, i.e. 𝛾=𝜎filtered
measured, filtered current variance and 𝜎! is the full variance of the unfiltered channel
!
/𝜎!, where 𝜎filtered
!
revealing that the single channel current has an approximately linear current-voltage
relationship.
current fluctuations. γ is a function of τ and can be calculated by integrating the filtered
Lorentzian power spectral density over frequency, and dividing by the total variance (see
Supplemental Material, Fig. S2D). γ is actually rather insensitive to τ, remaining between 0.5
and 0.574 over the measured range of τ. Thus the measured variance is converted to true total
variance by dividing by γ.
Using ramps means that the voltage is not constant, but the single channel current i depends
on voltage. Therefore, fluctuation analysis has to be performed in each of the ten time bins
separately. In the end, the results can be re-combined in a single-channel current-voltage plot.
Applying this to actual data (Fig. 4) using one patch as an example, the baseline noise
was measured in many repeated trials over the 20-40 ms time bin (Fig. 4A),
in bins 1 to 10 (Fig. 4B,C), and plotted against the mean current
variance 𝜎baseline
!
as well as the variance 𝜎exptl!
in the bins. An error-weighted least-squares fit of the linear relationship 𝜎exptl! =𝜎filtered
+
!
=𝛾𝜎!+𝜎baseline
𝜎baseline
!
!
points). The fitted slopes 𝛾𝑖 are plotted as a function of membrane potential in Fig. 4D,
was computed. (Since baseline points are
calculated from 20 ms segments, they receive four times the weight of the 5 ms test bin
=𝛾𝑖𝜇!+𝜎baseline
!
FIGURE 4 Fluctuation measurements in one patch. A) Scatter plot of current variance versus
mean current, measured in the baseline bin. B) Same for bin 1 of the ramp (grey points)
relative to the baseline (black, same as in A, for comparison). Black line: straight line fit
through the baseline and test bin data. The resulting slope estimate and its error are given. C)
Same for bin 10. D) Plot of the obtained slope values, with error bars, corresponding to 𝑖∙𝛾,
Dividing by 𝛾 to give the final estimates of single channel current, we plot data for 7 patches
against bin voltage (i: single channel current; γ: correction factor). Straight line indicates
error-weighted linear fit, predicting a highly positive reversal potential as expected for
calcium channels. - Extracellular Ca2+ conc.: 2 mM.
at three different calcium concentrations in Fig. 5, as a function of membrane potential. Two
patches at 1 mM, two patches at 2 mM, and three at 5 mM were analysed, and straight line
fits to the single channel current-voltage relationships were calculated for the pooled data at
each concentration. Single channel current is reduced with depolarization, with a shallow
slope, implying a highly positive reversal potential, as expected, but the relationship may be
better approximated by a Goldman-Hodgkin-Katz current equation than an ohmic linear
relationship. However, given the limited voltage range of the data, nonlinear extrapolation is
not justified. Table 1 summarizes the parameters of the fits. The single channel current is
highly sensitive to the extracellular calcium concentration, as expected (Fig. 5). At 1 and 2
mM, in the physiological range, the single channel currents at −40 mV, just above activation
are 0.045 and 0.07 pA respectively. At the peak of the inward current, around −10 mV, the
corresponding values are 0.04 pA (1 mM) and 0.065 pA (2 mM). This means that
approximately 200 channels are open at the maximum of the ramp response in somatic
nucleated patches.
FIGURE 5 Summary of single calcium channel current amplitude, at different membrane
potentials and calcium concentrations. Data for 1 mM [Ca2+]o in green (2 patches), 2 mM in
blue (2 patches) and 5 mM in red (3 patches). Data from Fig. 4D are shown as blue filled
circles. Different patches indicated by different symbols. Straight line fits of the pooled data
at each calcium concentration are shown. There is relatively weak voltage-dependence, but a
large increase in single channel current amplitude with increasing extracellular calcium
concentration.
TABLE 1 Parameter values for the single Ca2+ channel current voltage
relationship at different Ca2+ concentrations
Single Ca2+ channel current-voltage curve (without correction): 𝐼=𝑚∙𝑈+𝑐
𝑟(𝑐,𝑚)
Corrected current-voltage curve: I = m·U + c
Voltage range: −45 mV to −18 mV
[Ca2+]o
1 mM
2 mM
5 mM
[Ca2+]o
1 mM
2 mM
5 mM
At the extracellular calcium concentrations [Ca2+]o of 1 mM, 2 mM and 5 mM, the slope m
and intercept c of the fitted current (I) voltage (U) curves (see Fig. 5) are given together with
𝑚 [pA/mV]
0.00009 ± 0.00020
0.00012 ± 0.00016
0.00041 ± 0.00019
m [pA/mV]
0.00017 ± 0.00037
0.00022 ± 0.00030
0.00076 ± 0.00036
c [pA]
−0.0203 ± 0.0052
−0.0336 ± 0.0042
−0.0573 ± 0.0050
c [pA]
−0.038 ± 0.010
−0.0625 ± 0.0089
−0.107 ± 0.012
0.97
0.97
0.97
r(c,m)
0.93
0.83
0.67
parameter errors and the correlation r(c,m) between c and m. The top set of values (m, c)
correspond to the uncorrected single-channel current, as obtained from straight line fits to data
like shown in Fig. 4D. The bottom set of values (m, c) take into account the correction factor
γ (see text).
Effect of run-down on standard fluctuation analysis
Here, we provide a quantitative criterion to determine when run-down is so strong that a
method like the one suggested above becomes necessary, and when run-down is only mild
such that standard fluctuation analysis would work in principle. (However, the problem of
low-frequency noise in recordings is a separate one and may favor the approach of high-pass
filtering even when run-down is manageable.) Assume N recordings of the current Ij (j=1...N),
e.g. N=20 to enable sufficiently accurate ensemble averages, and a linear run-down of the
number of available channels between successive trials, such that the expectation value of the
current µj decreases by Δµ = µj−µj+1. If one computes the mean of the Ij as if there was no run-
down, the expectation value of the result is µ with 𝜇!=𝜇+(!!!! −𝑗)∆µ. The variances of the
𝑠!=
expectation value of s2, 𝑠! ≈𝑖𝜇+!!!"∆𝜇!. The expectation value of the resulting measured
single channel current is 𝑖measured ≈ !!! . If one wants the systematic relative error due to
2=iµj, where i is the single channel current (for the case of small channel open
Ij are σj
probability). If one now proceeds to estimate the variance from the data by computing
, again as if there was no run-down, the estimate will be too large, because
the residuals (Ij−µ) are taken with respect to the wrong mean. A brief calculation yields the
(!!!!)!!!!
!!!!
neglecting run-down to be below 5%, one arrives at the criterion
Applied to our dataset, in 2 patches, the channel current stabilized such that only about 10%
of the data would have been unusable for standard fluctuation analysis, in 3 patches, about
50% of the data would have been unusable, and in 2 patches, run-down was so strong that
standard fluctuation analysis would have been impossible. Note that it is usually the first
couple of traces where run-down is fastest, but these are particularly valuable because the
channel current is largest, therefore having the best signal to noise ratio, and providing data at
large channel numbers.
(3)
=!!!"∆!!!"<5%
!measured!!
!
Comparison with difference trace method
Another method for dealing with slow drifts in the experimental conditions has been
suggested, based on forming differences of consecutive current traces (35, 36). If traces Ij(t)
(j=1,2,...) are recorded, one can analyze 𝐼!,!(𝑡):=(𝐼!(𝑡)−𝐼!(𝑡))/ 2, I2,3(t) etc. instead of
I1(t), I2(t) etc. Ij,j+1 have the same noise spectrum as Ij, but any slow drift which is
approximately stable between adjacent sweeps will be eliminated. This trick should deal well
with normal rundown (unless the number of available channels changes very rapidly or
irregularly). However, the method can be affected by low-frequency instabilities that are
faster than the time between current traces (like baseline wobble or variations in the leak
conductance) just like standard fluctuation analysis.
For comparison, we analyzed the above dataset (seven patches) with the difference trace
method. As for the individual traces method, the ramp responses were divided into bins in
order to achieve conditions of constant voltage, and the final results were collected in current-
voltage plots (Fig. 6A). Here, the data are not high-pass filtered, but only low-pass filtered
(Gaussian filter, cut-off frequency fc), and the result depends on fc, since the filter attenuates
high frequency components of the channel fluctuation spectrum. We therefore performed the
analysis for fc=0.5, 1, 2, 5 and 10 kHz and extrapolated the results to infinite fc, using as a fit
function 𝑖Ca(𝑓!)=𝑖Ca!⋅!!tan!!(𝑓!/𝑓!), which is the expected fc-dependence of iCa for a brick-
wall filter (fN: half-power corner frequency of the Lorentzian channel fluctuation spectrum)
(35).
The results for the above example patch (cf. Fig. 4) are shown in Fig. 6A. For comparison, the
results from the same patch when using the individual traces analysis are shown below (Fig.
6B). At higher membrane voltages, the two methods yield similar results, but at lower
voltages, the difference traces method breaks down below about −35 mV, whereas the
individual traces methods remains stable up to about −45 mV. The same picture emerged
when analyzing the other patches of the dataset. The breakdown points of the two methods
were quantified as a deviation of the obtained single channel current from the fit line of 50%
or more, and the difference traces method systematically broke down earlier than the
individual traces method.
FIGURE 6 Comparison with difference traces method. A) Single calcium channel current at
a range of membrane potentials, for one patch, obtained by analysis with the difference traces
method. Solid line: straight line fit. B) Results for the same patch when analyzed with the
individual traces method (cf. Fig. 4D).
This difference is related to the accuracy with which the slope can be determined in mean-
variance scatter plots as shown in Fig. 4B,C. Indeed, in both methods, and for all seven
patches, the breakdown points as defined above coincided with the points at which the
relative error of the slope estimation from the straight line fit in the scatter plots became 50%
or more. (For the difference traces method, the results at fc=1 kHz were investigated.)
The reason for this discrepancy in performance is that the high-pass filtering of the individual
traces method improves the signal-to-noise ratio by attenuating the baseline noise while
leaving the signal strength relatively unaffected. When estimating the slope in mean-variance
scatter plots, the key problem is the scatter of the baseline noise variance, which is also
present in the test bins (cf. Fig. 4A,B,C). We determined the scatter of the baseline variance
for the two methods, which was reduced by factors around 6 for the individual traces method
compared to the difference traces method. E.g. in the patch analyzed above, the baseline noise
was 0.20±0.13 pA2 for the difference traces method, but only 0.115±0.020 pA2 for the
individual traces method. In contrast, the signal (channel noise variance) is only attenuated by
a factor of less than 2 (cf. correction factor γ). Therefore, especially at lower voltages, where
above was modeled. A Ca2+ channel model C⇌O⇌D (C: closed, O: open, D: deactivated
Simulated data
To test the proposed method, we analyzed simulated data. The situation of the example patch
the signal (total current through the Ca2+ channels) is smaller, the difference traces method
performs worse than the individual traces method.
state) was adjusted to fit the experimental current activation time course during the voltage
ramp (details: see Supplemental Material), rundown was implemented as a decline of the
number of available channels, matching the experimental decay of the Ca2+ channel current,
and the synthetic channel current traces were generated by simulating the stochastic opening
and closing of the channels (Gillespie algorithm, see (18, 37, 38)) and using the
experimentally determined single channel current. Care was taken to model the structure of
the background noise as accurately as possible. Two noise mechanisms were identified, one
which was low-frequency, i.e. long-range correlated at τ≈300 ms and was well captured by
assuming a fluctuation of the leak conductance δg(t), leading to a voltage-dependent current
noise δg(t)⋅U(t), and one short-range (details: see Supplemental Material).
The simulated noisy current traces were analyzed with the individual traces method. Fig. 7B
shows that the method faithfully reproduces the correct single channel current. For
comparison, the simulated data were also analyzed with the difference traces method (Fig.
7A). As for the real data (cf. Fig. 6A,B), both methods yield similar results at higher
membrane voltages, but the individual traces method is more accurate (results closer to the
real single channel current) and remains stable at lower voltages, where the difference traces
method breaks down. Therefore, the positive slope of the single channel current-voltage curve
was correctly resolved by the individual traces method (0.00037±0.00015 pA/mV; correct
value: 0.00022 pA/mV), but not by the difference traces method (−0.00043±0.00052 pA/mV).
As above, we determined the breakdown points of the two methods (−35 mV and < −45 mV),
which coincided with the points where the relative error of the slope estimation in the mean-
variance scatter plots became 50% or higher. The baseline noise levels were 0.20±0.11 pA2
and 0.110±0.017 pA2, in agreement with the results from the real data.
FIGURE 7 Analysis of simulated data. A) Analysis with the difference traces method. Solid
line: straight line fit to the data. Dashed line: true single channel current. B) Analysis with the
individual traces method.
We also used the simulated data to assess the extent of underestimation of fluctuations
resulting from fitting the mean current to the data. The current variance was underestimated
by up to 8% (bin 10), but after application of the high-pass filter, the underestimation was
reduced to below 0.3%.
in neocortical L5 pyramidal neurons,
Discussion
In this study, we aimed to measure the single-channel current of native voltage-gated Ca2+
in physiological extracellular Ca2+
channels
concentrations. Conventional fluctuation analysis cannot be applied, because of the rapid run-
down of Ca2+ channel currents. We therefore developed a modified fluctuation analysis
protocol which still works under such unstable conditions. Our results provide an essential
step towards understanding how many channels participate in physiological calcium currents,
how variable these currents are, and how the calcium signaling complex of the channel
microenvironment operates. In particular, the results will contribute to a better understanding
of stochastic properties of the dendritic Ca2+ spike in neocortical L5 pyramidal neurons.
Measuring native Ca2+ channels in neocortical L5 pyramidal neurons
Voltage-gated Ca2+ channels consist of one principal subunit α1, which interacts with
auxiliary subunits (α2δ, β) (7). Channel properties are further modified by accessory proteins
and lipid environment and are subject to extensive modulation in a cell-type-specific manner.
It is hence important to assess them in native channels in situ, rather than in simplified α1
expression systems.
We recorded calcium currents in nucleated patches and encountered current rundown. One
approach to counteracting rundown is the perforated patch method ((29) p. 7, p. 45ff), where
wash-out of natural intracellular components is slowed. However for nucleated patches, this
cannot be applied since a conventional whole-cell configuration needs to be established to be
able to aspirate the nucleus (see Methods).
The nucleated patch method obtains somatic membrane (although a small fraction of the
membrane may also be derived from the proximal apical dendrite). It should be noted that the
method is mechanically more invasive than normal whole-cell measurements, such that some
disruption e.g. of the cytoskeleton is possible, which is known to potentially influence
channel kinetics (39, 40). Even an alteration of the single channel current cannot be excluded,
although we are not aware of studies showing such an effect.
Measuring single Ca2+ channel current amplitude using fluctuation analysis
Accurate estimation of the physiological unitary amplitude of voltage-gated calcium channels
has proved a difficult problem. The concentration of the permeant ion is low (1-1.5 mM),
while even in near-isotonic (100-150 mM) barium or calcium, single channel conductance is
small (7-25 pS, depending on subtype, see (7)). Under the assumption of a linear relationship
of single channel conductance to the permeant ion concentration, values in the region of
0.007-0.025 pA might be expected with a driving force of 100 mV and at physiological Ca2+
concentration. With mean channel open times in the millisecond or sub-millisecond range, a
bandwidth of at least 1 kHz is required to adequately identify channel openings, at which
even in optimal patch-clamp recording conditions (41) one has to expect at least 20 fA rms
noise. Thus direct resolution of the single physiological calcium channel amplitude is, at best,
at the extreme limit of what is achievable with the patch-clamp technique. However, this
approach has been attempted, by prolonging channel openings pharmacologically, which
allows a more aggressive noise reduction by low-pass filtering (23), or by using low-noise
quartz patch pipettes (21, 24). Nevertheless, the accuracy of direct single-channel current
measurements remains hard to assess (see discussion below, in Comparison with previous
estimates).
Another possible approach, as we have adopted here, is to analyze the statistics of current
fluctuations in a population containing a large number of channels. Indeed, this was used in
an early landmark study on single voltage-gated calcium channels (22). Although fluctuation
analysis can in principle measure much lower unitary channel amplitudes, it is important to
highlight its limitations. In general, one must assume that all channel openings are identical in
amplitude – if not, the single channel current estimate will be an average for the different
subtypes of channel openings, weighted according to their prevalence, although this may still
be useful information. Fluctuation analysis also assumes independence of channels in the
population – however this is usually a valid assumption. In conventional fluctuation analysis,
the expectation current is estimated as the ensemble mean current, by averaging many
identical trials. However, this approach has the major drawback of requiring a long period of
stationarity, to accumulate at least 20 to 30 responses under identical conditions. We found
that this was not realizable in this preparation, and may be a questionable strategy to apply to
calcium channels under any circumstances, owing to the progressive rundown to which they
are prone, even when ATP and GTP are provided in the pipette solution (26, 27). Therefore,
we adapted this method in several ways. First, we took the approach of fitting individual
responses to a smooth function which was sufficiently flexible to follow the form of
macroscopic currents accurately without over-fitting, i.e. following the time course of
fluctuations. This resembles previous fitting methods applied to synaptic currents (33, 34). In
both cases, fitting a function by least-squares can result in a slight underestimation of the
variance – i.e. the fitted function is closer overall, in a least-squares sense, to the response
than is the underlying expectation current. However, this kind of error is reduced by using a
reasonably extended trial, essentially allowing time for a high number of fluctuations around
the mean within each trial. In the present situation, we found in simulations that the
underestimation was up to 8%. Applying a high-pass filter to the fluctuation trace dealt with
this problem in a controlled way, reducing the error to below 0.3%. Secondly, the high-pass
filter rejected extraneous low-frequency artifacts, which we found to be a useful enhancement
to the robustness of the method. We showed how to compensate for this step, taking into
account the temporal correlation of the fluctuations (Fig. 3A). We verified the proposed
method by simulation, showing that it faithfully recovers the single-channel current.
Finally, in the present experimental situation, ramp stimuli had a number of advantages over
voltage steps. Firstly, leak subtraction is easily done in individual sweeps by fitting a straight
line to the initial part of the ramp, such that the sampled baseline and leak are maximally
close to the measurement range. Also, there is no capacitative transient to worry about, only a
step capacitative current, most of which is cancelled by the amplifier circuitry, and any
residual current is eliminated by the leak subtraction. For voltage steps, one would have to
work with additional steps to hyperpolarized and/or slightly depolarized voltages and subtract
the linearly scaled response from the test traces to correct for leak and residual uncancelled
capacitative currents. Secondly, a ramp stimulus leads to a smooth, gradual time course of the
mean current response, which can be well fitted with a heuristic function. Step responses are
more peaked and therefore more prone to overfitting. The overfitting-correction with a high-
pass filter only works well if the mean current contains only low frequencies (cf. Results
section Fluctuation analysis of voltage ramp responses), which is not the case for a brief peak.
Thirdly, with ramps, the full voltage range where channels are activated is automatically and
efficiently sampled with one generic stimulus, whereas for steps, one would have to pick
specific test voltages to resolve the voltage dependence of the single-channel current.
However, in other situations, particularly for fast inactivating channels, and when a wide
range of sampled popen values is desired (e.g. Na+ channels), step stimuli could be a better
choice, and the presented method can also be applied there, although the results may be less
reliable in places where the mean current changes rapidly.
The method we developed here for calcium channels should be useful in the study of other
ion channel types, particularly in situations where channels run down quickly and baseline
and linear leak are not perfectly stable. We provided a quantitative criterion when run-down is
too strong so that standard fluctuation analysis becomes systematically flawed. Moderate run-
down can also be eliminated by analyzing the differences of consecutive traces, as has been
suggested before (35, 36). However, the high-pass filtering of the presented method has the
additional advantage of attenuating other low-frequency noise. In our dataset, this gave rise to
more accurate results, and better stability of the method at lower signal strength (low channel
current at voltages where only few channels are activated), compared to the difference traces
method.
We assumed that channel open probability p is small, such that the linear approximation of Eq.
2 is valid. The maximal error from this approximation is given by p, i.e. if p is 10%, one
underestimates the single channel current by up to 10%. (If data at smaller p are also available,
the error decreases.) For Ca2+ channels, at the voltages of this study (below −18 mV), the
channel open probability has been measured to be small, below 0.2 even at the top end of the
voltage range (25), such that an error above 10% arising from this source is unlikely. If an ion
channel type has larger open probabilities (as would for example be the case for Na channels,
see (4, 5)), the method presented here is not valid, because the number of channels N, which
decreases due to run-down, does not disappear from the mean-variance equation. However, it
could be extended to the use of the full quadratic relationship (Eq. 1) for channel types for
which this proves necessary, or experimental conditions could be adjusted such that one is
again in the low popen regime.
Ca2+ channel subtypes in neocortical L5 pyramidal neurons
Of which subtype are the Ca2+ channels encountered in neocortical L5 pyramidal neurons?
Eleven subtypes of the principal channel-forming α1 subunit are known, distinguished by
electrophysiological and pharmacological properties and grouped in three gene families (7).
The first family CaV1, encodes L-type channels, while CaV2 genes encode N-type, splice
variants P- and Q-type, and R-type channels, and the CaV3 genes encode T-type channels.
CaV1 and CaV2 channels require higher depolarisations to activate than CaV3 channels and are
therefore called high-voltage activated (HVA) calcium channels, while CaV3 channels are
called low-voltage activated (LVA).
In (27), the pharmacology of calcium currents in nucleated patches from neocortical L5
pyramidal neurons was examined and no T-type channels were found, but all other subtypes,
L-, N-, R-, P- and Q-types were inferred from pharmacology. In (42), Ca2+ channels in
dendrites and somata of hippocampal pyramidal neurons were studied. In dendrites, this study
found only occasional L-type channels, but predominantly a class of HVA medium
conductance channels which was tentatively identified with R-type, and LVA T-type channels
were frequently encountered. In somata, a larger density of L-type channels and fewer R-type
channels were found.
Our own pharmacological data (see Supporting Material), while preliminary, suggest that the
bulk of the Ca2+ channel current we measured in somatic nucleated patches was due to R-type
and/or Q-type channels, with a small contribution from L-type channels. The activation
threshold of our currents was much more depolarized than typical for LVA T-type channels.
Moreover, the single exponential component of the autocorrelation of fluctuations (Fig. 3A)
suggests a reasonably homogeneous population of channels being activated under our
conditions. (For more detailed discussion, see Supporting Material.). However, in general, the
values that we have measured reflect the overall average current amplitude of the subtypes
present in the cell which are activated by the particular voltage stimulus used, weighted in
proportion to their contribution to the total current.
Comparison with previous estimates of single Ca2+ channel current
The calcium concentration in the cerebrospinal fluid of rats is 1.0 to 1.5 mM ((17), see also
note in (21)), while a standard value for artificial cerebrospinal fluid recipes that is often used
in brain slice experiments (e.g. (11, 30)) is 2 mM. As for the single channel current in
extracellular barium iBa, iCa depends on the calcium channel subtype, but there is no uniform
ratio between iBa and iCa. T-type channels produce about the same current in barium and in
calcium, whereas L-type and N-type channel currents are larger in barium (43). Hence the
recognized hierarchy of single channel currents in barium, CaV1 (L-type) > CaV2 (N,P,Q,R-
type) > CaV3 (T-type), is not valid in calcium. In (21), iCa for L-, N- and T-type channels was
measured and the current hierarchy N-type > L-type > T-type was obtained, with the N-type
current 38% larger than the L-type current, and the T-type current 16% smaller than the L-
type current.
iCa does not depend linearly on the extracellular calcium concentration [Ca2+]o or on the
membrane voltage. iCa saturates at high [Ca2+]o and the dependence can be described with the
Hill equation (see discussion in (21)). The voltage-dependence of iCa follows the nonlinear
Goldman-Hodgkin-Katz current equation ((6) p. 445ff). Therefore, we restrict our comparison
to iCa values at the same extracellular calcium concentration and membrane voltages, rather
than comparing slope conductances. In (22), calcium channels in bovine adrenal chromaffin
cells were studied. In these cells, half of the calcium current is thought to flow through P/Q-
type channels, and the other half through N- and also L-type channels (reviewed in (36)). At a
membrane potential of −12 mV, fluctuation analysis yielded a single channel current of 0.03
pA in 1 mM Ca2+ and 0.09 pA in 5 mM Ca2+, which is in good agreement with our results
(0.04 pA in 1 mM Ca2+, 0.11 pA in 5 mM). In (23), L-type calcium channels were studied in
arterial smooth muscle and channel openings were prolonged with BayK8644, which resulted
in channel open times of 10 ms and longer, allowing filtering at 500 Hz. In cell-attached
recordings in 2 mM [Ca2+], 0.10 pA at −20 mV and 0.20 pA at −40 mV were obtained.
Compared with our almost constant value of 0.07 pA in the range between −20 and −40 mV,
however, this current-voltage relationship is much steeper. In fact, extrapolating the values
beyond −20 mV predicts a reversal potential of about 0 mV, although ECa should be highly
positive. Taking into account that our currents are probably predominantly due to CaV2
channels and should hence be about 40% larger than L-type currents under otherwise equal
conditions, the values in (23) seem comparatively large. In (21), a direct single-channel
recording approach was also used, with quartz pipettes employed to reduce noise, and like in
(23), larger values were obtained than the ones reported here. For N-type channels in 2 mM
Ca2+, 0.21 pA at −25 mV was reported. Compared with our value of 0.068 pA, this is about a
factor of 3 larger. Part of the discrepancy may be due to different channel subtypes present in
layer 5 pyramidal neurons. For example if R-type channels predominate, they might have a
smaller iCa than N-type channels, despite both being members of the CaV2 family. L-type
channels may also contribute, which in (21) are reported to have a slightly smaller iCa.
Thus, single channel current estimates obtained by direct single-channel recordings appear to
yield larger values than fluctuation analysis. When trying to judge by eye the size of a single
channel current step at a low signal-to-noise ratio, close to the resolution limit, there is a
danger of overestimating the single-channel current, if only a small subset of events are large
enough to detect, or if insignificant fluctuations are mistaken for channel openings (see
discussion in (19)). In (22), single-channel currents in 95 mM Ba obtained by single-channel
recordings were compared with fluctuation analysis results, yielding about 40% larger values
in direct single-channel recordings. As a possible reason, a loss of noise power due to filtering
was discussed, which would reduce the single-channel current estimate from fluctuation
analysis. However, we fully corrected for the effects of filtering, so this should not be an issue.
Also discussed was the possibility of heterogeneity in the channel population, i.e. that small
channel openings can be missed by direct single-channel recordings but are picked up by
fluctuation analysis.
The relaxation time constants τ we obtained from autocorrelations were in the range 0.2-0.8
ms. This is consistent with mean channel open times between 0.4 and 1 ms reported in (42),
obtained by single-channel recordings in high extracellular barium.
Functional implications
In computational models of ion channels, it is usual to consider the limit of many ion channels,
leading to deterministic differential equations for neuronal dynamics (45). Increasingly,
though, it is appreciated that the stochastic current fluctuations due to the opening and closing
of individual ion channels are not negligible, for example they can lead to spontaneous action
potentials and have a large effect on action potential timing (18). In (20), a detailed
computational study of stochastic calcium spikes in cerebellar Purkinje cell dendrites was
performed, and it was found that large variability in calcium spike bursts, also seen
experimentally, is produced, in a model, by the interaction of stochastic calcium influx with
downstream calcium-dependent conductances, via stochastic intracellular calcium transport
and buffering. A single channel permeability of 2.5 x 10-5 µm3 ms-1 was assumed for P-type
channels, resulting in a single channel current of about 0.035 pA (at body temperature, 2 mM
extracellular calcium and −50 mV membrane voltage), which is close to our values for 1 mM
extracellular calcium.
The extreme sensitivity of calcium nanodomain concentrations to the calcium influx (46)
means that downstream biochemical signaling is also critically dependent on the exact value
of iCa. For example, stochastic calcium influx is predicted to have a major effect on CaMKII
activation in dendritic spines (47) and therefore potentially on long-term potentiation.
Although some of this calcium enters through NMDA receptors, dendritic calcium channels
will also contribute, and therefore their physiological unitary properties, as well as those of
NMDA receptors, are relevant in this regard.
To our knowledge, there have been no studies which directly recorded single voltage-gated
calcium channels in physiological calcium concentration in the apical dendrite of layer 5
pyramidal neurons. Based on current-clamp and ionic substitution and blocker experiments, it
has been suggested that there may be a "hot spot" of Ca2+ channel density at the dendritic
Ca2+ spike initiation zone (12, 48). It is conceivable that the subtype composition of dendritic
calcium channels is different to that in the soma, but in the absence of further evidence, our
data give the best available indication of the physiological single-channel current amplitude in
the dendrite as well as the soma.
Conclusion
By adapting and refining a fluctuation analysis approach, we have been able to provide an
improved estimate of the single channel current iCa of Ca2+ channels in physiological levels of
extracellular Ca2+, a quantity which has rarely been measured, despite its importance for
stochastic single-channel effects or calcium nanodomains. We further measured the timescale
of autocorrelation of calcium channel gating. These results should assist in building more
realistic computational models to understand calcium-dependent signaling in neurons.
Author contributions
C.S. and H.P.C.R. designed the experiments. C.S. performed the experiments and analyzed
the data. C.S. and H.P.C.R. performed the simulations and wrote the manuscript.
Acknowledgments
C.S. acknowledges funding by the German Academic Exchange Service (DAAD), the
Biotechnology and Biological Sciences Research Council (BBSRC), the Cambridge European
Trust (CET) and the Research Innovation Fund of the University of Freiburg.
Supporting citations
References (49, 50) appear in the Supporting Material.
Neher, E., and B. Sakmann. 1976. Single-channel currents recorded from membrane
Stevens, C.F. 1972. Inferences about membrane properties from electrical noise
Neher, E., and C.F. Stevens. 1977. Conductance fluctuations and ionic pores in
Sigworth, F.J. 1980. The variance of sodium current fluctuations at the node of
Tsien, R.W., and C.F. Barrett. 2005. A Brief History of Calcium Channel Discovery.
Sigworth, F.J. 1977. Sodium channels in nerve apparently have two conductance states.
References
1.
of denervated frog muscle fibres. Nature. 260: 799–802.
2.
measurements. Biophys J. 12: 1028–1047.
3.
membranes. Annu Rev Biophys Bioeng. 6: 345–381.
4.
Nature. 270: 265–267.
5.
Ranvier. J Physiol. 307: 97–129.
Hille, B. 2001. Ion Channels of Excitable Membranes. 3rd ed. Sinauer.
6.
7.
Catterall, W.A., E. Perez-Reyes, T.P. Snutch, and J. Striessnig. 2005. International
Union of Pharmacology. XLVIII. Nomenclature and structure-function relationships of
voltage-gated calcium channels. Pharmacol. Rev. 57: 411–425.
8.
In: Voltage-Gated Calcium Channels. Springer US. pp. 27–47.
9.
Nat. Rev. Neurosci. 9: 206–221.
Stuart, G.J., N. Spruston, and M. Häusser. 2008. Dendrites. 2nd ed. OUP.
10.
Schiller, J., Y. Schiller, G. Stuart, and B. Sakmann. 1997. Calcium action potentials
11.
restricted to distal apical dendrites of rat neocortical pyramidal neurons. J Physiol. 505: 605–
616.
12. Larkum, M.E., T. Nevian, M. Sandler, A. Polsky, and J. Schiller. 2009. Synaptic
integration in tuft dendrites of layer 5 pyramidal neurons: a new unifying principle. Science.
325: 756–760.
13. Major, G., M.E. Larkum, and J. Schiller. 2013. Active properties of neocortical
pyramidal neuron dendrites. Annu Rev Neurosci. 36: 1–24.
14. Golding, N.L., H. Jung, T. Mickus, and N. Spruston. 1999. Dendritic Calcium Spike
Initiation and Repolarization Are Controlled by Distinct Potassium Channel Subtypes in CA1
Pyramidal Neurons. J. Neurosci. 19: 8789–8798.
15. Llinas, R., and M. Sugimori. 1980. Electrophysiological properties of in vitro Purkinje
cell dendrites in mammalian cerebellar slices. J Physiol. 305: 197–213.
16.
Spruston, N. 2008. Pyramidal neurons: dendritic structure and synaptic integration.
Südhof, T.C. 2013. Neurotransmitter Release: The Last Millisecond in the Life of a
Jones, H.C., and R.F. Keep. 1988. Brain fluid calcium concentration and response to
Fenwick, E.M., A. Marty, and E. Neher. 1982. Sodium and calcium channels in
Synaptic Vesicle. Neuron. 80: 675–690.
17.
acute hypercalcaemia during development in the rat. J Physiol. 402: 579–593.
18. Chow, C.C., and J.A. White. 1996. Spontaneous action potentials due to channel
fluctuations. Biophys J. 71: 3013–3021.
19. Cannon, R.C., C. O'Donnell, and M.F. Nolan. 2010. Stochastic ion channel gating in
dendritic neurons: morphology dependence and probabilistic synaptic activation of dendritic
spikes. PLoS Comput Biol. 6: e1000886,1-18.
20. Anwar, H., I. Hepburn, H. Nedelescu, W. Chen, and E. De Schutter. 2013. Stochastic
calcium mechanisms cause dendritic calcium spike variability. J Neurosci. 33: 15848–15867.
21. Weber, A.M., F.K. Wong, A.R. Tufford, L.C. Schlichter, V. Matveev, and E.F.
Stanley. 2010. N-type Ca2+ channels carry the largest current: implications for nanodomains
and transmitter release. Nat Neurosci. 13: 1348–1350.
22.
bovine chromaffin cells. J Physiol. 331: 599–635.
23. Gollasch, M., J. Hescheler, J.M. Quayle, J.B. Patlak, and M.T. Nelson. 1992. Single
calcium channel currents of arterial smooth muscle at physiological calcium concentrations.
Am J Physiol. 263: C948–952.
24. Church, P.J., and E.F. Stanley. 1996. Single L-type calcium channel conductance with
physiological levels of calcium in chick ciliary ganglion neurons. J Physiol. 496 (Pt 1): 59–68.
25.
Fox, A.P., M.C. Nowycky, and R.W. Tsien. 1987. Single-channel recordings of three
types of calcium channels in chick sensory neurones. J Physiol. 394: 173–200.
26. Bean, B.P. 1992. Whole-cell recording of calcium channel currents. Meth Enzym.
207: 181–193.
27. Almog, M., and A. Korngreen. 2009. Characterization of voltage-gated Ca(2+)
conductances in layer 5 neocortical pyramidal neurons from rats. PLoS ONE. 4: e4841.
28. Davie, J.T., M.H.P. Kole, J.J. Letzkus, E.A. Rancz, N. Spruston, G.J. Stuart, and M.
Häusser. 2006. Dendritic patch-clamp recording. Nat. Protoc. 1: 1235–1247.
29.
30.
into neocortical pyramidal dendrites. Nature. 367: 69–72.
31.
6th edn. Academic Press/Elsevier.
32.
Sather, W., S. Dieudonne, J.F. MacDonald, and P. Ascher. 1992. Activation and
desensitization of N-methyl-D-aspartate receptors in nucleated outside-out patches from
mouse neurones. J Physiol. 450: 643–672.
33. Robinson, H.P., Y. Sahara, and N. Kawai. 1991. Nonstationary fluctuation analysis
and direct resolution of single channel currents at postsynaptic sites. Biophys. J. 59: 295–304.
34. Traynelis, S.F., R. Angus Silver, and S.G. Cull-Candy. 1993. Estimated conductance
of glutamate receptor channels activated during EPSCs at the cerebellar mossy fiber-granule
cell synapse. Neuron. 11: 279–289.
35. Heinemann, S.H., and F. Conti. 1992. Nonstationary noise analysis and application to
patch clamp recordings. Methods Enzymol. 207: 131–148.
36. Conti, F., B. Neumcke, W. Nonner, and R. Stämpfli. 1980. Conductance fluctuations
from the inactivation process of sodium channels in myelinated nerve fibres. J. Physiol. 308:
217–239.
37. Kispersky, T., and J.A. White. 2008. Stochastic models of ion channel gating -
Scholarpedia.
http://www.scholarpedia.org/article/Stochastic_models_of_ion_channel_gating. .
38. Gillespie, D.T. 1977. Exact stochastic simulation of coupled chemical reactions. J.
1995. Single-Channel Recording. 2nd ed. Plenum Press.
Stuart, G.J., and B. Sakmann. 1994. Active propagation of somatic action potentials
Paxinos, G., and C. Watson. 2009. The Rat Brain in Stereotaxic Coordinates. compact
Phys. Chem. 81: 2340–2361.
39. Undrovinas, A.I., G.S. Shander, and J.C. Makielski. 1995. Cytoskeleton modulates
gating of voltage-dependent sodium channel in heart. Am. J. Physiol. 269: H203-214.
40. Bennett, P.B. 2000. Anchors Aweigh!: Ion Channels, Cytoskeletal Proteins, and
Cellular Excitability. Circ. Res. 86: 367–368.
41. Levis, R.A., and J.L. Rae. 1993. The use of quartz patch pipettes for low noise single
channel recording. Biophys J. 65: 1666–1677.
42. Magee, J.C., and D. Johnston. 1995. Characterization of single voltage-gated Na+ and
Ca2+ channels in apical dendrites of rat CA1 pyramidal neurons. J Physiol. 487: 67–90.
43.
Fox, A.P., M.C. Nowycky, and R.W. Tsien. 1987. Kinetic and pharmacological
properties distinguishing three types of calcium currents in chick sensory neurones. J Physiol.
394: 149–172.
44. Garcia, A.G., A.M. Garcia-De-Diego, L. Gandia, R. Borges, and J. Garcia-Sancho.
2006. Calcium signaling and exocytosis in adrenal chromaffin cells. Physiol Rev. 86: 1093–
1131.
45. Koch, C. 1999. Biophysics of Computation. OUP.
46. Tadross, M.R., R.W. Tsien, and D.T. Yue. 2013. Ca2+ channel nanodomains boost
local Ca2+ amplitude. Proc. Natl. Acad. Sci. 110: 15794–15799.
47. Zeng, S., and W.R. Holmes. 2010. The Effect of Noise on CaMKII Activation in a
Dendritic Spine During LTP Induction. J. Neurophysiol. 103: 1798–1808.
48.
Perez-Garci, E., M.E. Larkum, and T. Nevian. 2013. Inhibition of dendritic Ca2+
spikes by GABAB receptors in cortical pyramidal neurons is mediated by a direct Gi/o-βγ-
subunit interaction with Cav1 channels. J Physiol. 591: 1599–1612.
49. Nowycky, M.C., A.P. Fox, and R.W. Tsien. 1985. Three types of neuronal calcium
channel with different calcium agonist sensitivity. Nature. 316: 440–443.
50.
IUPHAR/BPS. 2015. Guide to Pharmacology: Voltage-gated calcium channels. .
Supporting material to:
"Fluctuation analysis in nonstationary conditions:
single Ca channel current in cortical pyramidal neurons"
Biophysical Journal, 2017
C. Scheppach, H. P. C. Robinson
This supplementary document contains more detailed information on the fitting function to
voltage-ramp evoked calcium currents, the filter correction, pharmacological results on the
observed Ca2+ channel currents and details on the modeling of Ca2+ channels and the
background noise.
Fitting of current responses to voltage ramp stimuli
Current responses to voltage ramp stimuli were fitted with a piecewise polynomial, consisting
of a flat segment and two linear segments linked by cubic splines (Fig. S1), giving a smooth
heuristic function whose eight parameters were fitted by minimizing squared deviation from
the data, which was taken as the estimate of the mean current 𝜇! for fluctuation analysis.
FIGURE S1 Fit function for voltage-ramp-evoked calcium currents. A heuristic piecewise
polynomial function with continuous derivative was fitted by least squares to individual leak-
subtracted ramp responses (see main text, Fig. 2A). The function consisted of an initial flat
segment and two linear sections joined by cubic splines. The transition times t1…4, gradients
ma,b and intercepts Ia,b were all free parameters.
Correction for effects of filtering
FIGURE S2 Correction for effects of filtering. A) Channel noise is dominated by a single
exponential relaxation. Black trace shows the autocorrelation of the signal computed in one
time bin (bin 10), with baseline autocorrelation subtracted. The red curve shows the least-
squares fit to a model of exponentially-relaxing fluctuations (single Lorentzian spectral
component). Undershoot is caused by the high-pass filter. B) Mean channel open/burst time τ,
the relaxation time constant of fluctuations, extracted by this procedure is plotted for multiple
patches and time bins, as a function of membrane potential. Green is 1 mM [Ca2+], blue is 2
mM and red is 5 mM. Filled circles with error bars are mean and standard error of the mean
obtained from about five adjacent data points. τ values are concentrated in the range 0.2-0.8
ms. C) Illustration of the effect of the bandpass filtering in the frequency domain. L(f)
indicates the Lorentzian distributed power of the underlying current fluctuations (exponential
noise of τ = 0.8 ms, therefore Lorentzian half-power corner frequency = 200 Hz; PSD: power
spectral density). The combined effect on noise power of the low- and high-pass filter
characteristics (A(f) and B(f) respectively) is shown as the curve 𝐴(𝑓)!∙𝐵(𝑓)! . Channel
noise power which passes through filtering is shown in red: 𝐿(𝑓)∙𝐴(𝑓)!∙𝐵(𝑓)!. D) The
ratio of the integral of the noise power density passed by the filter to the total noise power
gives the correction factor γ, whose dependence on the noise relaxation time constant τ is
shown as the solid curve, the combined effect of the low-pass filtering (dashed curve) and
high-pass filtering (dotted curve). γ is quite insensitive to τ over the experimentally-relevant
range.
mean of 𝐼(𝑡!)∙𝐼(𝑡!) with 𝑡!−𝑡!=𝑡 and at least one of t1 and t2 inside the bin. Hence the
Methods
The autocorrelation R(t) of a current trace I(t) for a time bin (Fig. S2A) was computed as the
autocorrelation could also be estimated for t larger than the bin width of 5 ms.
Pharmacology of observed Ca2+ channel currents
Methods
Solutions, drugs, chemicals
Nifedipine (L-type Ca2+ channel blocker) was obtained from Sigma (Gillingham, U.K.). A 40
mM stock solution in ethanol was produced on the day of the experiment. (S)-(−)-Bay K 8644
(L-type activator) was obtained from Tocris (Bristol, U.K.). A 10 mM stock solution in
ethanol was prepared, kept in the fridge and used within two days. ω-Conotoxin GVIA (N-
type blocker) and ω-agatoxin IVA (P-type blocker) were obtained from Tocris. 100 µl of each
drug dissolved in HEPES-buffered extracellular solution was applied with a pipette into the
recording chamber, and the bath solution was gently stirred by aspiration with the pipette
several times. Since the recording chamber solution volume was around 3 ml, the
concentration of the drug was diluted by a factor of about 30.
Recording
Ca2+ channel currents were recorded during depolarizing ramps, e.g. from a holding potential
of −70 mV rising to 0 mV (slope 0.6 mV/ms, stimulus repeated every 3 s), and also during
depolarizing voltage steps (e.g. holding potential −70 mV, 100 ms long steps to test potentials
cycling through −80 mV, −60 mV, −30 mV, −20 mV, −10 mV and 0 mV, stimulus repeated
every 3 s). In nifedipine block experiments: in 6 of 8 patches, ramp stimulus with a holding
potential of −70 mV, prepulse at −90 mV for 200 ms, rise to 0 mV, stimulus repeated every 5
s. In one patch, holding potential of −60 mV, no prepulse, depolarization to 0 mV, repeated
every 2 s. In one patch, holding potential of −70 mV, no prepulse, depolarization to 0 mV,
repeated every 2 s.
Data analysis
Leak subtraction for step responses (Fig. S2 E) was carried out by subtraction of scaled
ensemble-averaged responses to small steps (holding potential +/− 10 mV) below the
activation range of the calcium current.
Results: Pharmacology of Ca2+ channels in nucleated patches from neocortical L5
pyramidal neurons
Ca2+ channel currents were recorded during depolarizing ramps, and also during depolarizing
voltage steps. Application of 200 µM cadmium completely blocked the current (see main text,
Fig. 1C; n = 11 cells). The block could not be reversed by perfusing the bath again with Cd-
free solution (n=2). Partial block was observed at 20 µM Cd2+ (n=2), which was reversible
(n=1). 2 µM Cd2+ did not block the current (n=1).
In most cases (5 of 8 patches), nifedipine (L-type Ca2+ channel blocker) produced no clear
blocking effect (Fig. S2 A), but a partial block was noted in 3 patches (Fig. S2 D). Generally,
a small partial block was difficult to differentiate from the natural time course of rundown.
Therefore, the Ca2+ channel current was plotted against time (Fig. S1 A-D), and if addition of
the blocker resulted in a drop of the current below what would be expected from the
extrapolated rundown beforehand, the data were interpreted as showing a block. Nifedipine
was used at several concentrations: 3 µM (2 patches, partial block in 1), 30 µM (4 patches,
partial block in 1), 50 µM (1 patch, partial block) and 70 µM (1 patch, no block). Bay K8644
(1 µM), an L-type channel agonist, had no discernible effect on the ramp-activated inward
current (n=2), but boosted the current during depolarizing voltage steps (n=2, Fig. S2 E). We
conclude that some L-type channels are present in somata of L5 pyramidal neurons, but they
do not account for the major component of the calcium current. ω-Conotoxin GVIA (N-type
Ca2+ channel blocker) at 1 µM generally did not block the Ca2+ channel current (Fig. S2 B,
n=7, only in one patch possible slight block). ω-Agatoxin IVA (P-type Ca2+ channel blocker,
30 nM) also had no blocking effect (Fig. S2 C, n=4).
In conclusion, the full block by Cd2+ shows that the inward current observed in ramps is
carried by voltage-gated calcium channels. The lack of block by ω-conotoxin GVIA and ω-
agatoxin IVA indicates that N and P type channels are not involved, while T-type channels
would have a much lower activation threshold (1), suggesting that the ramp-evoked current
corresponds mostly to R-type (CaV2.3), or possibly Q-type (CaV2.1) channels, with a small
contribution from L-type channels.
FIGURE S2 Pharmacology of Ca2+ channels. In the top row, three example patches are
shown in which peak current was insensitive to A) nifedipine (30 µM), B) ω-conotoxin GVIA
(CTX, 1 µM) and C) ω-agatoxin IVA (ATX, 30 nM). D) Patch in which Ca2+ channel current
was partially blocked by nifedipine (3 µM). Red bars indicate wash-in period during which
currents were not measured, black bars indicate the presence of blocker. Note the quite typical
time course of rundown of the calcium channel current in panel A. E) Patch in which BayK (1
µM) boosted the current during depolarizing voltage steps.
Discussion: Ca2+ channel subtypes in neocortical L5 pyramidal neurons
We concluded that the current which we studied in voltage ramp responses in nucleated
patches was likely to be due to high-voltage-activated R- and/or Q-type channels, with a small
contribution from L-type channels. Cadmium (200 µM) totally blocked the current.
Nifedipine (3-70 µM), an L-type blocker, only resulted in a partial block in some patches, and
Bay K8644 (1 µM), an L-type agonist, only boosted the current in some patches. The toxin
blockers ω-conotoxin GVIA (1 µM), which fully blocks N-type channels (2), or ω-agatoxin
IVA (30 nM), which is reported to fully block P-type channels, did not affect the current. We
did not test concentrations of ω-agatoxin IVA which would be required to block Q-type
channels (Kd = 100-200 nM, see (3)). The activation threshold of the current was much more
depolarized than typical for T-type channels. Our results differ somewhat from those of a
previous study (4), in which the pharmacology of calcium currents in nucleated patches from
neocortical L5 pyramidal neurons was also examined, but using barium as the permeant ion.
There, also no T-type channels were found, but the presence of all other subtypes, L-, N-, R-,
P- and Q-types was reported: nifedipine, ω-conotoxin GVIA, ω-agatoxin IVA, as well as ω-
conotoxin MVIIC (blocks Q-, N- and P-type) and SNX-482 (may block R-type) all blocked
fractions of the current. In (5), Ca2+ channels were studied in dendrites and somata of
hippocampal pyramidal neurons, again with barium as the permeant ion. In dendrites,
consistent with the results reported here, only occasional L-type channels were found, but
predominantly a class of HVA medium conductance channels which was tentatively
identified with R-type, based on the resistance of the channels to ω-conotoxin MVIIC.
However, in somata, ω-conotoxin MVIIC was found to have a significant blocking effect, and
LVA T-type channels were frequently encountered. The reasons for these discrepancies are
not clear, and may well partly be due to our use of physiological calcium as the permeant ion,
the different cell type studied in (5) and the difficulty of performing clean Ca2+ channel
pharmacology due to rapid Ca2+ channel rundown and the lack of highly subtype-specific
Ca2+ channel blockers.
Modeling of Ca2+ channels and background noise
Ca2+ channel model
We used a model with one closed, one open and one deactivated state:
C 𝑘!(𝑈)⇌𝑘!(𝑈) O 𝑘!!⇌𝑘!! D
+ and kd
(1)
kc(U)=B⋅exp(−β⋅U)
kd
kd
+=0.015 kHz
−=0.0015 kHz
− were chosen such that the deactivation time constant is 70 ms, and the current
kd
deactivates to 10% of the peak value (cf. Fig 1A of main text):
For kc(U), an exponential voltage dependence was used:
The parameters were chosen such that the mean channel open time is 0.4 ms at −30 mV and
0.8 ms at 0 mV (cf. Fig. 3B of main text):
ko(U) was chosen such that the experimental current time course in response to the ramp was
reproduced. To achieve this, a sigmoidal voltage dependence was necessary:
Furthermore, an open probability of 1% was imposed at the maximum of channel activation
within the analyzed data range (90 ms after onset of the ramp). This value is arbitrary, but was
chosen such that the linear approximation of the relationship between current mean and
variance in fluctuation analysis is valid (see main text). The parameters were obtained by a
least squares fit to the experimental current time course:
A=0.0215 kHz
K=0.00983
α=0.167 mV−1
!!!⋅exp(!!")
B=1.25 ms−1
β=0.023 mV−1
𝑘!(𝑈)=
!
Background noise
To characterize the background noise present in our recordings, we plotted autocovariances
R(t,t+Δt)=<I(t)⋅I(t+Δt)> of the data, centered at different times t. On longer timescales, the
autocovariances decayed exponentially with a time constant of about 300 ms, but in the
course of the voltage ramp, the autocovariance values became progressively smaller than
expected. This can be explained by assuming a leak current Ileak(t)=g(t)⋅U(t) which reverses at
0 mV and has a fluctuating conductance g(t)=g0+δg(t). The noisy δg(t) traces with
exponential autocovariance were generated by convolving white Gaussian noise with an
exponential function.
To capture the remaining short-range (i.e. high frequency) noise of the data, the power
spectral density PSD(f) (f: frequency) was obtained from the baseline bin (width: 20 ms).
Then, the expected PSDl-r for the above long-range noise was computed (which was
concentrated almost exclusively in the zero-frequency component), and subtracted from the
experimental PSD, to obtain PSDs-r of the remaining short-range noise. The noise traces were
then generated by convolving white Gaussian noise with the Fourier transform of PSDs-r(𝑓).
Nowycky, M.C., A.P. Fox, and R.W. Tsien. 1985. Three types of neuronal calcium
Supporting references
1.
channel with different calcium agonist sensitivity. Nature. 316: 440–443.
2.
IUPHAR/BPS. 2015. Guide to Pharmacology: Voltage-gated calcium channels. .
3.
Catterall, W.A., E. Perez-Reyes, T.P. Snutch, and J. Striessnig. 2005. International
Union of Pharmacology. XLVIII. Nomenclature and structure-function relationships of
voltage-gated calcium channels. Pharmacol Rev. 57: 411–425.
4.
conductances in layer 5 neocortical pyramidal neurons from rats. PLoS ONE. 4: e4841.
5. Magee, J.C., and D. Johnston. 1995. Characterization of single voltage-gated Na+ and
Ca2+ channels in apical dendrites of rat CA1 pyramidal neurons. J Physiol. 487: 67–90.
Almog, M., and A. Korngreen. 2009. Characterization of voltage-gated Ca(2+)
|
1509.00171 | 2 | 1509 | 2016-03-20T19:33:00 | Topological schemas of cognitive maps and spatial learning in the hippocampus | [
"q-bio.NC"
] | Spatial navigation in mammals is based on building a mental representation of their environment---a cognitive map. However, both the nature of this cognitive map and its underpinning in neural structures and activity remains vague. A key difficulty is that these maps are collective, emergent phenomena that cannot be reduced to a simple combination of inputs provided by individual neurons. In this paper we suggest computational frameworks for integrating the spiking signals of individual cells into a spatial map, which we call schemas. We provide examples of four schemas defined by different types of topological relations that may be neurophysiologically encoded in the brain and demonstrate that each schema provides its own large-scale characteristics of the environment---the schema integrals. Moreover, we find that, in all cases, these integrals are learned at a rate which is faster than the rate of complete training of neural networks. Thus, the proposed schema framework differentiates between the cognitive aspect of spatial learning and the physiological aspect at the neural network level. | q-bio.NC | q-bio | Topological schemas of cognitive maps and spatial learning
A. Babichev1, S. Cheng2 and Yu. Dabaghian1∗
1Department of Neurology Pediatrics, Jan and Dan Duncan Neurological Research Institute,
Baylor College of Medicine, Houston, TX 77030 USA
Department of Computational and Applied Mathematics, Rice University, Houston, TX 77005 USA,
2Mercator Research Group "Structure of Memory" and Department of Psychology,
Ruhr-University Bochum, Universitaetsstrasse 150, 44801, Bochum, Germany
E-mail: [email protected], [email protected], [email protected]∗
(Dated: August 5, 2018)
Spatial navigation in mammals is based on building a mental representation of their environment --
a cognitive map. However, both the nature of this cognitive map and its underpinning in neural
structures and activity remains vague. A key difficulty is that these maps are collective, emergent
phenomena that cannot be reduced to a simple combination of inputs provided by individual neurons.
In this paper we suggest computational frameworks for integrating the spiking signals of individual
cells into a spatial map, which we call schemas. We provide examples of four schemas defined by
different types of topological relations that may be neurophysiologically encoded in the brain and
demonstrate that each schema provides its own large-scale characteristics of the environment -- the
schema integrals. Moreover, we find that, in all cases, these integrals are learned at a rate which is
faster than the rate of complete training of neural networks. Thus, the proposed schema framework
differentiates between the cognitive aspect of spatial learning and the physiological aspect at the
neural network level.
6
1
0
2
r
a
M
0
2
]
.
C
N
o
i
b
-
q
[
2
v
1
7
1
0
0
.
9
0
5
1
:
v
i
X
r
a
2
I.
INTRODUCTION
In the 1940's, Tolman proposed that animals build an internal representation -- a cognitive map -- of
their environment and that this map allows the animal to perform space-dependent tasks such as navigat-
ing paths, finding shortcuts, and remembering the location of their nest or food source [1]. Three decades
later, O'Keefe and Dostrovsky discovered pyramidal neurons in the hippocampus, named place cells, that
become active only in a particular region of the environment -- their respective place fields [2] (Figure 1A).
The striking spatial selectivity of these place cells led O'Keefe and Nadel [3] to suggest that they form a
neuronal basis of Tolman's cognitive map, thus providing this abstract concept with a concrete neurophysio-
logical basis. In the ensuing decades, it was realized that there are many brain regions involved in cognitive
mapping of the environment [4], yet there is still no consensus on either the physiological mechanisms of
this phenomenon or the theoretical principles that explain them [5]. Overall, it is believed that individual
cells encode elements of the cognitive map, much like contributing pieces to a jigsaw puzzle. However, this
analogy is not direct: the spiking activity of each separate neuron has no intrinsic spatial or geometrical
properties -- these properties appear only at the population level, emerging from the synchronous spiking
activity of large neuronal ensembles [6, 7]. The mechanism of this phenomenon remains unknown, i.e.,
there exists a disconnect between the level of individual neurons from which the preponderance of neuro-
physiological data is acquired and the level of neuronal ensembles where the large-scale representations of
space are believed to emerge [8].
In a recently proposed a model of spatial learning [9, 10], we attempted to bridge this gulf by combining
recent experimental results pointing out the topological nature of the hippocampal map [11 -- 17] and methods
of Algebraic Topology. This model allowed demonstrating that place cell activity can encode an accurate
topological map of the environment and estimating the time needed to accumulate the required connectivity
information. Further analyses of the model suggested to us that it is indicative of a more general theoretical
framework that may lead to a systematic understanding of how spiking activity of neurons can be integrated
to produce large-scale characteristics of space. In this paper, we outline the general principle and provide
four specific models, which we call schemas, of integrating the activity of simulated neurons into a coherent
representation of the explored environment. For each schema we find that a large-scale spatial map is
produced within a short, biologically plausible period, which could be used to estimate the spatial learning
time in different environments.
II. THEORETICAL FRAMEWORK
A. The model
A schema model of a cognitive map contains the following three key components:
i. An abstract schema S(R, PS ) represents the spatial information contained in the map at any given
time. It consists of a set of formal regions R = {r1, r2, ..., rN} and a set of relationships, PS = {ρ1, ρ2, ..., ρM},
that express how these regions combine. We presume that each region ri in the schema can be related to
any other region r j through a chain of relationships with intermediate regions ρα(ri, rk), ρβ(rk, rl, rm),...,
ργ(rn, r j). A specific selection of the relationships included in PS determines the type of spatial information
encoded in the schema and the global arrangement of the encoded regions, which is crucial both for the
properties of the resulting map as well as for the information encoded in it.
ii. The neural implementation, NS , is a neural network that encodes the schema S. For the sake of
simplicity, we model NS using a basic, two layer, feed-forward neural network inspired by cell assembly
theory [18], which consists of a layer of cells that represent regions of space and another layer of readout
neurons that represent the relationships between these regions (Figure 1B). When a cell ci fires a spike, we
say that the region ri is "active"; otherwise it is "latent" [19]. When a readout neuron fires a spike, we
3
FIG. 1: Physiological components of the schema. A. The simulated trajectory in a 1 × 1 m environment (left) and
200 randomly scattered place fields (clusters of colored dots) produced by place cells with a mean firing rate f = 12
Hz and a mean place field size s = 20 cm. B. Schematic representation of three overlapping assemblies of place cells
(shown by black dots) that project synaptically onto their respective readout neurons (blue pentagons). The active
place cells (black dots with red centers) of the ignited cell assembly (in the middle of the figure) produce spike trains
that drive the spiking activity of the readout neuron (blue pentagon with the red center).
say that the corresponding relationship is "instantiated." Thus, by construction, the relationships between
regions are represented via temporal relationships between the spike trains and by the parameters of synaptic
connections between the two layers in NS .
iii. The spatial map and the representing space. The goal of introducing schemas is to model the
assemblage of the cognitive maps from the cells' spiking activity. However, in absence of a mechanism ex-
plaining how spatial representations emerge from the spike trains, this task remains undefined. Statements
such as "a given place cell's activity encodes a region" or "the coactivity of a set of place cells represents a
spatial overlap between the encoded regions" require an interpretation. In the analysis of electrophysiolog-
ical data, this interpretation is acquired by mapping the neuronal activity into an auxiliary, external space X
which is selected according to the experimenters' best judgment. For example, constructing the place fields
by ascribing Cartesian x− y coordinates to the place cells' spikes and identifying the areas where the spikes
cluster is one attempt to map the unobservable formal regions encoded in the cognitive map into observable
regions of the spatial environment [20]. In the following discussion, we will refer to this algorithm as to
standard place field mapping. Spaces that have been used to interpret the activity of place cells and other
cells include Euclidean domains in one [21, 22] and in three dimensions [24, 25]; circles [26], tori [27],
spheres [28], and even Klein bottles [29]. To capture this aspect of cognitive map analysis, we define a
spatial mapping from the schema S to a representing space X,
f : S → X
(1)
in which the formal regions of S are mapped into the "concrete" regions of X, xk = f (rk). We will refer to
xk as the X-representations of the formal region rk and to the resulting layout of the representing regions in
X as the spatial map of the schema, MX(S).
Although the representing regions are selected to reproduce the relationships between the formal re-
gions as well as possible, the resulting map does not always capture the structure of the original schema:
some relationships may be lost in the mapping or the mapping may produce relationships between the rep-
resenting regions that are not encoded in S. For example, the place field maps are believed to reflect an
animal's cognitive map's structure but their faithfulness has not been established or even addressed in the
neurophysiological literature. In the case when the set of relationships between the regions xk (PX) matches
the schematic relationships exactly, so that PX = PS , the mapping will be referred to as faithful. The corre-
sponding spatial map may then be viewed as a model of S, i.e., the structure of S can be deduced from the
layout of the representing regions.
4
Thus, each specific schema model includes these three components -- the abstract schema S, its neuronal
network implementation NS and the spatial mapping (1) into a representing space X. For brevity, we will
refer to this triad as to "schema," when no ambiguities can arise.
B. Spatial learning
A key property of our approach, crucial for modeling spatial learning, is that schemas are dynamic
objects. As an animal explores a novel environment, new regions become represented by the activity of
place cells and new relationships are inferred from the spike trains' temporal patterns [9, 10]. According
to the standard approach of neural network theory, the process of learning a schema may be viewed as
the process of training the readout neurons to represent the set PS by detecting repetitive patterns in the
incoming spike trains. From this perspective, a schema is learned after its network is trained, i.e., after the
readout neurons stop adopting their spiking responses to the patterns of the incoming spike trains.
On the other hand, from a cognitive perspective, the purpose of spatial learning is to acquire qualitative,
large-scale characteristics of the environment, which enable spatial planning, spatial navigation and spatial
reasoning, such as path connectivity, shortcuts and obstacles, geometric and topological properties, global
symmetries and so forth. Such large-scale characteristics of the environment that are captured through the
relationships of a given schema will be referred to as schema integrals, IS. Below we demonstrate that
the minimal time Tmin required for the schema's integrals to emerge is typically shorter than the time, TN,
required to train all readout neurons, i.e., large-scale information can be extracted from a partially trained
network. Thus, the schema approach captures two complementary aspects of spatial learning: physiolog-
ical learning -- the process of forming and training the cell assembly network and schematic or cognitive
learning -- the emergence of information about the global structure of space, expressed as the corresponding
production of schema integrals.
C. Topological Schemas
What aspect of space is represented in the hippocampal map? The answer to this question depends on the
information captured by the readout neurons in the hippocampal cell assemblies. Since correlating neuronal
spiking with geometrical properties of the representing space sometimes produces useful interpretations of
electrophysiological data, most authors assume that the spiking patterns of place cells encode geometric
properties of space [30, 31]. For example, it has been shown that combining the spiking activity of a
relatively small number of place cells with the information about the sizes, shapes, and locations of their
respective place fields allows a reconstruction of the animal's trajectory in a typical experimental enclosure
on a moment-to-moment basis [32, 33].
However, the read-out neurons have access only to the place cells' spikes, and not to their respective
place fields. Obtaining the shape and size of any given place field, which is nothing but a cumulative spatial
histogram of spikes used for illustrational purposes, requires accumulating a substantial number of spikes
from the corresponding place cell. Yet the spike trains produced during the activity period of a given place
cell are short -- typically hundreds of milliseconds in duration -- and highly variable, not only because of
the animal's movements, but also because of the intrinsic stochasticity of neuronal spiking [34]. Thus, the
spike trains of place cells contain little information about a place field, such as its shape, location and other
computationally expensive parameters. Furthermore, recent experimental studies point out that these spike
trains do not provide the geometric information on the synaptic integration timescale of seconds or fractions
of a second [17, 22, 23].
Since the temporal pattern of place cell firing is the only information available to downstream neurons, a
physiologically adequate class of schemas of the hippocampal map may be constructed based on capturing
qualitative, topological relationships between regions, e.g., overlap, adjacency, ordering and containment,
5
FIG. 2: Graph schema. A. A schematic illustration of the spike trains produced by seven place cells whose coactivity
is indicated by the dashed rectangles connecting the spike trains. B. The corresponding seven place fields traversed
by the animal's trajectory (dashed line). C. The corresponding graph schema, the seven vertexes of which correspond
to seven formal regions encoded by the active place cells. The edges mark the relationships encoded in the schema,
e.g., the connection (r4, r5) is in the schema, but (r5, r1) is not.
from the temporal relationships between the spike trains [16, 35]. The resulting maps will then produce a
topological representation of space rather than a geometrical one [36, 37], in which the relative arrangement
of the locations is more important than mapping the precise positions. Topological schemas have several
advantages over the more precise geometric schemas, e.g., higher stability (e.g., faithfulness of a topological
map is not destroyed under continuous deformations of the representing space) and lower computational
cost, which may make them biologically more viable (see Discussion).
There remain many possibilities in which to read out qualitative information about the spike trains
and thus there are many topological schemas. In this perspective, a particular readout mechanism, which
responds to specific patterns of place cell coactivity, defines the type of spatial information encoded in the
schema. The following discussion presents four different topological schemas based on different qualitative
relations between regions and the rate at which these schemas are acquired.
III. RESULTS
A. Graph Schema G
The simplest topological schema is based on binary connections:
its set of relationships consists of
pairs of connected regions, PG = {(ri, r j), (ri, rk), (rm, rn), ...}. Such schema can be viewed as a graph,
G, whose vertices are linked if the corresponding regions are related according to PG (Figure 2). The
corresponding neuronal implementation is produced by training the pair-coactivity detector readout neurons
to respond to nearly simultaneous spiking of their respective pair of presynaptic cells [38, 39]. In other
words, physiological learning of a graph schema G amounts to detecting pairs of cells that exhibit frequent
coactivity [42].
We modeled this process by simulating place cell spiking activity induced by a rat's movements across
a place field map in a small environment (Figure 1B). To simplify the analyses, we assume that as soon
as the coactivity occurs, the corresponding connection is immediately "learned," i.e., incorporated into the
schema. As a result, at every moment of time t, the connectivity matrix of the graph is defined by the
coactivity observed prior to that moment. Thus, Ci j = 1 if cells ci and c j cofired before t and Ci j = 0
otherwise. Figure 3A shows that the number of links in the graph, which is the number of recruited pair-
coincidence detectors, grows as the schema is learned and saturates at ca. TN = 5 mins, i.e., after this time
new incoming spike trains do not produce new connections in NG.
The saturation of the schema could be a trivial result if the graph becomes fully connected or remains
6
FIG. 3: Spatial learning based on the graph schema. A. The number of links in the graph schema as a function of
time, computed for a simulated ensemble of 200 place cells with randomly scattered place fields (see Methods). The
blue line represents the mean and the red lines show the error margins. B. Graph schema entropy (blue) and place
field map entropy (red) as a function of time. The green line shows the mutual information between the map and the
schema. Both entropies and the mutual information saturate at the time when the number of links saturates. C. The
probability of establishing a maximal length connection in the graph schema stabilizes in about 2.2 minutes, when
only 50% of connections appeared.
mostly empty. A simple characteristic capturing the efficiency of G, which generalizes to other schemas in
a natural way is its entropy [40, 41]. This is the specific entropy of the readout neurons,
HG = −pc log2(pc) − pd log2(pd),
where pc and pd = 1 − pc are the fractions of the connected and disconnected vertex pairs in the graph.
For a fully discrete (pc = 0) or a fully connected graph (pc = 1) the entropy vanishes and maximal entropy
HG = 1 is achieved for pc = 1/2 (in which case the absence of a link is as informative as its presence).
Figure 3B demonstrates that for the place cell ensemble used in our simulations (see Methods), the entropy
of the graph schema asymptotically approaches a maximal value of about HG ≈ 0.8 in about five minutes, a
value implying that the schema network NG is neither underloaded nor oversaturated.
To quantify the correspondence between the schema G and its place field map MX(G), we calculated
the entropy HX, of the occurrences of place field pairwise overlaps across time and compared HX to HG.
Figure 3B demonstrates that both entropies remain close throughout the entire learning period, indicating
that the complexity of the place field layout remains similar to the complexity of the encoded relationships
at all times. In addition to this correspondence we also computed the mutual information (MI, see Methods)
between the place field overlap and place cell coactivities, which also grows with the rat's navigational
experience (Figure 3B). Thus, we have convergent lines of evidence indicating efficient spatial learning
captured by the graph schema.
As a cognitive map model, the graph schema G provides a stratum for implementing graph-theoretical
navigation algorithms, that is, for establishing paths connecting spatial locations [43, 44]. Its integrals IG
are the global characteristics of the region-to-region connectivity graph, e.g., its partitioning, the colorabil-
ity of its vertexes and edges [45], its planarity, and the existence of a path between two given vertexes. As
an example of such large-scale characteristics we identified the shortest paths connecting pairs of the most
distant vertexes and computed the time required to establish these connections. The results shown in Fig-
ure 3C demonstrate that the animal establishes connections between the most distant locations in the graph
in about Tmin = 2.2 minutes, a time when only about 50% of the readout neurons are trained. Similarly,
emergence of the information required to establish existence of a circuit of the graph which traverses each
edge exactly once, called an Eulerian path, takes about 2.2 minutes, while the correct number of cliques in
G, which are sets of pairwise connected vertices, can be deduced within two minutes. These observations
suggest that the emergence of schema integrals before the network is trained may be a general phenomenon.
7
FIG. 4: Simplicial schema. A schematic representation of the connections between the vertexes associated with
the seven place cells and their spatial map shown on Figure 1B. B. The existence of the two-dimensional simplexes,
corresponding to higher order coactivity relationships, permits the links to be deformed. This is illustrated for paths
γ1 and γ2: the transformation can be visualized by slipping one path across the two-dimensional facets, thus demon-
strating the topological equivalence between paths γ1 and γ2. C. The nerve of the map shown on Figure 1B matches
the simplicial schema.
B. Higher-order Overlap (Simplicial) Schema T
A topological schema may be based on representing not only binary, but also ternary, quaternary and
other higher-order connectivity relations between spatial domains. For example, a schema may represent
the overlaps between regions, including triple, quadruple, etc., overlaps. The key property of the overlap
relation is that if k regions, r1, r2, ..., rk, have a common intersection, then so does any subcollection of
them. The simplest mathematical object that is closed under the operation of taking non-empty subsets
is an abstract simplex, which can be viewed as a list of k elements [46]. Hence, a (k + 1)-order overlap
relationship ρ(r0, r1, ..., rk) may be represented by a k-dimensional simplex σ = [r1, r2, ..., rk]. A set of
overlap relationships therefore forms an abstract simplicial complex, T , and we will hence refer to a higher
order overlap schema as to simplicial schema.
Under the standard mapping of the place cell spiking activity into the environment, the simplicial
schemas' relationships, PT , represent the overlaps between the place fields. For example, the place field
map shown in Figure 2B can be faithfully encoded by a simplicial schema with four 3d order relationships
P3 = {(r6, r1, r7), (r7, r1, r2), (r1, r2, r4), (r2, r3, r4)} and an additional binary relation (r4, r5), as shown in Fig-
ure 4. The neuronal marker of these overlaps is the spiking coactivity: if the animal enters a location where
several place fields overlap, their respective place cells produce (with a certain probability) temporally over-
lapping spike trains. Hence the neural network implementation of a simplicial schema, NT , should be built
to detect the coactivity events, using coincidence detector readout neurons (which, in fact, corresponds to
the current view on the hippocampal cell assembly network organization [18, 47 -- 49]).
Physiological learning of a simplicial schema hence amounts to training the readout neurons to detect
place cell coactivities. Our learning algorithm (see Methods) ensures that, at every stage of learning, only
the highest order relationships are kept while the redundant lower-order relationships are eliminated. For
example, pairwise connections between three neurons become redundant after a triple coactivity between
them is detected, at which point the three pair-detector readout neurons can be replaced with a single triple-
coincidence detector. Numerical simulations demonstrate that, as the rat explores the environment, the more
probable, lower-order coactivity events are captured first and the less probable higher order coactivities
accumulate more slowly (Figure 5A). Moreover, although rapid changes of the readout neurons' order stops
after five or six minutes, slow regroupings continue during the entire navigation period, T = 25 minutes.
Thus, unlike the pairwise overlaps in G that can be instantly identified, the orders of the readout neurons
cannot be deduced from a single coactivity event. In this sense, the orders of the readout neurons are integral
characteristics of place cells' spiking activity, and therefore may be viewed as integrals of the simplicial
8
FIG. 5: Learning the simplicial schema. A. The development of the population of readout neurons as a function of
time. The typical order of co-activity is about 20, the highest order is 33. The population of the high order readout
neurons (order above 11) increases through the entire duration of the experiment (over 20 minutes). Other populations
reach a stable plateau (e.g., orders 9 and 10) or reach a maximum and then drop (e.g., order under 8). The decrease of
the number of the low order relationships after an initial increase indicates the elimination of redundant information.
B. The development of the overlap relationships between the standard representing regions as a function of time. The
typical overlap order is about 20, the highest order is 35. As in the case with the readout neurons, the number of
high-order overlaps (order above 17) saturates on the rise, unlike the lower order overlaps. C. The entropy of the
dimensions of registered simplexes (red) is similar to the entropy of the orders of the overlaps of the concrete regions
xk (blue). The mutual information between these two variables is computed along the trajectory (green).
schema.
There exists an additional important set of T -integrals, which capture the topology of the representing
space. This property of the simplicial schemas can be illustrated using the Cech theorem, which states that
the pattern of overlaps between regions U1, U2, ..., Un, covering a topological space X = ∪iUi, encodes
homological invariants of X, provided that every intersection Ui ∩ U j ∩ ...∩ Uk is contractible [50, 51]. The
proof is based on building the "nerve" of the covering -- a simplicial complex, the d-dimensional simplexes
of which correspond to the (d + 1)-fold overlaps between covering regions, and showing that it is topologi-
cally equivalent to X (Figure 4C). This theorem implies that the spatial map of a sufficiently rich simplicial
schema may encode the topology of the space navigated by the rat, and suggests that if this map is faithful,
i.e., if the nerve of the spatial map matches the schema's relationship set PT exactly, then the schema also
captures the large-scale topological representation of the space.
To study the correspondence between the simplicial schema and its map, we compared the schema's
entropy HT , defined by the probabilities for a readout neuron to become a kth-order co-activity detector,
to the entropy HX of the place field map MX(T ), defined via the probabilities of producing a kth-order
overlap between the place fields (Figure 5B). As shown on Figure 5C, both entropies closely follow one
another: they both grow initially and reach similar asymptotes in approximately four minutes. However,
the mutual information between these two series of events decreases with time. The reason for this effect
lies in the idealized nature of the representing regions xk, built as convex hulls of the spike clusters in the
two-dimensional environment (for other place field construction algorithms see [53, 54]). The xk's crisp
boundaries produce high-order overlaps, which are not captured by the place cell coactivity and hence by
the schema -- compare the orders of the readout neurons on Figure 5A with the orders of overlap between
the corresponding representing regions xk in Figure 5B.
This result can be viewed from several perspectives. First, it illustrates that the scope of reliable in-
formation that can be drawn from the spatial map is limited: only sufficiently robust, qualitative aspects
of the place field map, such as low dimensional overlaps, can be trusted. Second, the regions ri that were
originally introduced as "formal," that is, devoid of intrinsic properties, should fundamentally be viewed as
"fuzzy" and not as Euclidean domains with crisp boundaries [55].
Direct computations show that the coactivity complexes do, in fact, capture the topology of the repre-
9
FIG. 6: Topological loops in simplicial schema. A sequence of the place cell combinations ignited along a path γ
(black line) corresponds to a sequence of simplexes -- a simplicial path Γ that represents γ in T . B. The dynamics
of the total number of zero-dimensional loops (red) and one-dimensional loops (blue). Unlike the growing number
of links in the graph schema (Figure 2A), the number of topological loops decreases with time. Eventually, a single
loop survives. The margins of error are shown above and below each graph by a pair of pink and light-blue lines,
respectively. C. The barcodes -- timelines of the one-dimensional loops in the simplicial complex. The topological
noise vanishes after ca. 4 minutes, which is the schematic estimate of the cognitive learning time.
senting space, provided that the place cells' spiking parameters fall into the biological domain [9, 10, 47],
and hence that simplicial schemas provide a framework for representing topological information. For ex-
ample, cell assemblies ignited along the physical paths traversed by the animal correspond to sequences of
coactivity simplexes -- the simplicial paths that represent the physical paths in T (Figure 6A). The structure
of the simplicial paths allows establishing topological (in)equivalences between navigational paths, e.g.,
topologically equivalent simplicial paths represent physical paths that can be deformed into one another, a
non-contractible simplicial path corresponds to a class of the physical paths that enclose inaccessible or yet
unexplored parts of the environment. As a result, the simplicial schema produces a qualitative description
of navigational routes: while the total number of paths grows exponentially, the number of topologically
distinct loops, which represent topologically distinct paths is small (Figure 6B).
However, this information does not emerge immediately: as the animal begins to navigate a new environ-
ment, most topological loops reflect transient connections. As the spiking information accumulates, these
"spurious" loops disappear and only the loops that correspond to the physical signatures of the environment
persist (Figure 6C). With methods drawn from persistent homology theory [56, 57] one can determine the
minimal period Tmin required for removing the spurious loops, which provides a theoretical estimate of the
time required to learn the environment [9, 10]. Figure 6C demonstrates that in our test map, after Tmin = 4
minutes most topological loops have vanished and only the loops that correspond to the physical holes in
the environment survive. Thus, as in the case of the graph schema, the topological connectivity of the cog-
nitive map is captured by the simplicial schema before the underlying neuronal network is fully trained,
Tmin < TN.
C. Mereological Schema M
Although a sufficiently rich simplicial schema can capture the topological invariants of the representing
space X as its integrals, it does not capture all the qualitative topological aspects of the connectivity between
regions. For example, the identical simplicial schema (represented by a tetrahedron) can faithfully represent
the overlap relationships in the two maps shown in Figure 7, because both maps contain the same set of
regions R = {r1, r2, r3, r4} and one fourth-order overlap relation PR = {(r1, r2, r3, r4)}, as well as all their
consequent ternary and binary sub-relations. However, these maps are topologically different, since they
cannot be transformed from one into another by a continuous deformation of the plane R2. The obstruction
10
FIG. 7: Mereological schema. In A and B, the overlap pattern does not capture the cover relationship. Four regions,
x1, x2, x3 and x4 form a quadruple overlap in both cases. However, in map shown in A, the region x4 is contained in
the union of x1, x2 and x3. In the map shown in B, x4 is not covered. C. The cover and the overlap relationships in a
mereological schema corresponding to the map of Figure 1A. The covering regions are connected by red links (e.g.,
r6, r1, and r2) and the red arrows point to the covered region (e.g., r7). D. A neuronal implementation of the covering
relationship includes three inhibitory neurons (magenta) that provide inhibitory input into the readout neuron (purple).
If each inhibitory input of an active interneuron exceeds the excitatory input of the driving cell c7, the readout neuron
can spike only if the activity of the cell c7 is not accompanied by the inputs from any of the cells c1, c2 or c6. As
a result, the readout neuron will remain silent as long as the activity of the cells c1, c2 and c6 temporally covers the
activity of cell c7.
to such deformation is that the region x4 on Figure 7A is contained in the union of the regions x1, x2
and x3, i.e., x4 ⊂ (x1 ∪ x2 ∪ x3), and no containment relationships exist between any combinations of the
regions on Figure 7B. Neither a graph schema G nor a simplicial schema T can capture this difference;
what is required is the additional covering relation, (x1, x2, ..., xl) (cid:74) (y1, y2, ..., yk), (x(cid:48)s are covered by y(cid:48)s),
in terms of which the map on Figure 7A is described by the relationship r4 (cid:74) (r1, r2, r3), whereas the regions
shown on Figure 7B produce no containment relation. The cover relation produces a new -- mereological --
schema M, in which the information is encoded in terms of topological containment (Figure 7C). The
intuition behind neuronal implementation of the formal cover relation is the following. If the activity of
one ensemble of place cells, U = {c1, c2, ..., ck}, outlasts, or covers in time, the activity of cells in another
ensemble V = {d1, d2, ..., dl}, then the region XU, representing the U-ensemble, contains the region XV
representing the V-ensemble:
XV ⊂ XU if V (cid:74) U.
From this perspective, the set of covering cells provides contextual information about the covered cells, i.e.,
the cover relation combines the basic formal regions into more complex, composite regions.
The cover relationship can be implemented, e.g., by a combination of the excitatory and inhibitory
neurons shown on Figure 7D. In such a cell assembly, the readout neuron signals a violation of the cover
relationship, i.e., the latter is represented by an absence of the readout neurons' spiking activity up to the
moment t. Hence, in contrast with simplicial schemas, where readout neurons learn to detect ever higher-
order coactivities, a readout neuron in a mereological schema M learns to detect ever larger groups of cells
that together inhibit its activity (Figure 7D).
Similarly to the overlap orders in T , the cover relationships, as a rule, cannot be deduced from a single
coactivity event. Thus, these relationships represent integral information that can be viewed as the M-
integrals which characterize the large-scale topology of a space. We are currently unaware of additional
mereological algorithms that would allow large-scale computations of the environment's global topological
characteristics, similar to computing the homological invariants in a simplicial schema. Nevertheless, a
mereological schema encodes an important type of topological information, which may be used in physio-
logical neural networks to represent spatial maps.
In general, covering relationships can be established between arbitrary (including multiply connected)
regions. As a result, the number of possible combinations of covered and covering regions dramatically
11
FIG. 8: Learning the mereological schema. A. Time development of the covering relationships: a pair of covering
place cells, U = {ci, c j}, and a covered cell, V = {dk}. Each line corresponds to a specific choice of U and V. Each
line begins as soon as the covering relationship is detected and stops as soon as it is violated, that is, as soon as the
readout neuron shown in Figure 7D would fire. Note that the majority of relationships are short-lived: a large number
of spurious relationships are detected at the beginning of exploration. After about seven minutes the majority of them
disappear, similar to the behavior of topological loops computed in the simplicial schema shown in Figure 6B. This
diagram shows about 1% of the detected pairs, selected at random. B. The number of detected cover relationships
between pairs of place cells and a single place cell as a function of time.
increases. Even if the covered region V = {d1, d2, ..., dl} is spatially "bundled" (e.g., if each pair di and d j is
coactive at some moment of time, so that V forms a connectivity clique) the selection of possible covering
regions remains very large. Therefore, in order to test the development of cover relationships in time, we
opted to limit our study to neuron pairs covering an individual neuron (k = 2 and l = 1).
The results of simulations show that the time required to learn second-order covering relationships in M
is comparable to the learning times in the graph schema G (Figure 8). As spatial exploration begins, a large
number of transient covering relationships is produced because of insufficient spiking data. With accumu-
lating spike trains most cover relationships become violated, so that the number of surviving relationships
quickly drops. As the animal completes its first turn around a central hole of the environment (Figure 1B),
a new set of (mostly transient) relationships is injected into the schema which produces the peak shown
in Figure 8B. Subsequently, the number of cover relations steadily diminishes to about 200 pairs, which
corresponds to a saturated schema. This result reflects qualitative behavior of higher order covering re-
lationships, though a full implementation of the algorithm for the higher-order covering combinations (k,
l > 1) is computationally substantially more complex.
D. Complex Relations and the RCC Schema R
Qualitative Space Representations (QSR) are discrete, region-based versions of the conventional
point-set theoretical geometries and topologies [58] used to formalize "intuitive" qualitative spatial
reasoning [59, 67, 75], and thus are particularly important for modeling cognitive representations of
space [62 -- 65].
Important examples of QSRs are the Region Connection Calculi (RCC) -- formal log-
ical theories based on a family of binary topological relations between regions [66]. For example,
the most basic RCC theory, RCC5, which applies to the case of regions with fuzzy boundaries, is
built using the five relations shown in Figure 9A: disconnect (DR), partial overlap (PO), proper part
and its inverse (PP and PPi), and equality (EQ) [67].
the arrange-
ment of regions shown on Figure 2B is described by the following set of RCC5 relationships: P =
{PO12, PO14, PO16, PO17, PPi23, PO24, PO27, PP32, PP34, PPi43, PPi45, PO46, PP54, PO67; DR for all other
pairs} (Figure 9B). More elaborate RCC calculi can capture tangencies [66], convexity [67], qualitative di-
In terms of these relations,
12
FIG. 9: Illustration of RCC5 schema. A. RCC5 relationships: five logically possible pairwise relations: "x is
discrete from y" (denoted as DR), "x partially overlaps with y" (PO), "x is a proper part of y" (PP), "y is a proper
part of x" (PPi), and "x is identical to y" (EQ). B. An RCC5 schema, R5, of the spatial map from Figure 2B. For
convenience, the DR connections are shown with gray dashed lines. The structure of the rest of the relations produces
a graph similar to the one shown in Figure 2C. The black lines indicate PO connections. Cyan and blue arrows show
the PP and PPi connections, respectively. C. A U-track having two dead ends and a W-track having three dead ends
and a junction, j, marked in red. Every time the rat visits the junction point it must choose between the left and the
right turn, indicated by the red and blue trajectories, respectively. D. Topological relationships between regions on a
U- and a W-track that allow capturing the tracks qualitative geometries. The endpoints, e1, e2 and e3 are regions that
overlap with only one other region. The midpoints, m1, m2 and m3, overlap with two regions and the junction overlaps
with three regions.
rections [68], and distances [69] as well as complex hierarchies of all these relationships [70]. As a result,
RCC methods can capture not only standard topological signatures of spaces, such as loops and holes [71],
but also more subtle qualitative features, such as branching points, linear sections, and dead ends. These
qualitative features produce fundamental differences in spatial reasoning required for navigating the corre-
sponding environments. For example, the junction point on the W-tracks, which are often used in behavioral
experiments (Figure 9C), forces an animal to choose between a right or left turn, which is reflected in the
place cell code [72, 73]. The RCC5 theory allows capturing such features, e.g., distinguishing between
the U- and W-tracks, which, from the perspective of algebraic topology, are but contractible manifolds
(Figure 9D).
To model spatial learning based on a specific RCC approach, one can construct an RCC schema, in
which the readout neurons are trained to recognize the appropriate set of binary relationships. However, an
important aspect of RCC constructions is that the set of relationships that can be simultaneously imposed
on a set of regions is restricted [74, 75]. For example, if x and y partially overlap and y is a proper subset
of z, then z and x must have a non-null intersection and z cannot be a subset of x. Therefore, we define an
RCC schema R as a schema with a set of consistent RCC relations between regions.
To model the process of physiological learning in the RCC5 (R5) schema, we trained five types of
readout neurons to recognize the five RCC5 relationships, starting from the initial DR relationship. This
however requires more complex algorithms than in G and T schemas: while the partial temporal overlap can
always be interpreted as partial spatial overlap, other temporal relationships cannot be uniquely assigned
to a spatial RCC5 relation (Figure 10A). For example, passing through two partially overlapping regions
13
FIG. 10: Temporal vs. spatial relationships. A. Temporal relationships between the spike trains (o overlap, s
separation, d during, id inverse d and e equal [114]) and the corresponding spatial relationships. The relationships
DRxy, PPxy, PPixy and EQxy between the regions can be imitated by partial overlap, depending on the shape of the
trajectory, which shows that these relations cannot be directly deduced from the spike train structure. B. The transitions
between the RCC5 relationships, showing the immediate conceptual neighborhood (continuity table) structure of
RCC5. These are the possible sequences of gradual transformation of the RCC5 relationships. For example, if at
some moment of time two regions, x and y, were disconnected (DRxy) then this relationship cannot instantly jump to
a containment relationship (PPxy or PPixy) without going through, at least instantaneously, the partially overlapping
(POxy) relationship.
along a particular trajectory can generate a temporal disconnect, a temporal cover, or a temporal equality
relationship between two spike trains which can be mistaken for spike trains produced by a DR, PP/PPi,
or EQ relationship, respectively. Because of this ambiguity, the spiking activity of the presynaptic cells
in the cell assemblies produced during individual runs through a pair of place fields can invoke different
interpretations of the spatial relationships. Thus, learning a R5 schema rests on encoding, at each moment,
the best guesses for the relationships between pairs of regions and then updating them based on the available
spiking history and the qualitative analogue of continuity constraints, as shown in Figure 10B.
In our simulations, the relationships evolved rapidly and saturated within about TN ≈ 4 minutes from
the onset of the exploration (Figure 11A). Figure 11B shows that at the beginning of the exploration, the
number of inconsistencies between independently trained readout neurons is high. Subsequently, their
number quickly diminishes as the information about coactivity is acquired. An increase in the number of
PP relationships in Figure 11A produces a splash of inconsistencies occurring at about 3 minutes, which is
at the time when the animal completes its first turn around the central hole. This phenomenon has the same
origin as the splash of transient cover relationships occurring in the mereological schema M (Figure 8).
As the statistical information about place cell coactivity accumulates, a stable set of RCC5 relationship
emerges. The schema's specific entropy, defined using the probabilities of observing all five relationships,
saturates about the same time, TN ≈ 4 minutes. The entropy of the RCC5 relationships between the
representing regions in the map MX(R) remains similar to the schema entropy during the entire course
of learning, reaching the asymptotic value of H ≈ 0.84 (Figure 11C). Moreover, the mutual information
between the map and the schema increases with the acquisition of information in a way similar to the case
of graph schema G but unlike the case of simplicial schema T (cf. Figures 11C, 3C and 5E). Once again,
this data indicates that spatial maps built on regions with diffuse boundaries may better reflect the nature
of the encoded regions. In the meantime, the integrals of the R5 schema, i.e., the combinations of RCC5
relationships that represent the junction and the endpoints on the W track, emerge from neuronal spiking in
under Tmin ≈ 2 minutes -- much sooner than the readout neurons in R5 network are trained.
14
FIG. 11: Learning the RCC5 schema. A. Evolution of RCC5 relationships: at the beginning of the exploration of
the new space, the regions are mostly disconnected and the partial overlap of relationships is accumulated over time.
Similarly to the graph schema, the number of all types of relationships saturates in about five minutes. B. The number
of inconsistencies in the randomly initialized relational network is higher at the beginning of exploration and decays
to a low, steady level by the time the number of relationships stabilize. C. The entropy of the relations encoded in the
schema (red), the entropy of the RCC5 relationships in the map (blue), and the mutual information (green) between
them saturate at a similar time scale.
IV. DISCUSSION
We have presented a framework for integrating the place cells' spiking information into a global map of
space, implemented via simple cell assembly neural networks, wired to encode spatial relationships. The
approach is motivated by the experimental results [11 -- 17, 78 -- 83] and by the theoretical models proposed
in [9, 10] and in [42 -- 44]. From the perspective of the current approach these models are particular imple-
mentations of the simplicial and the graph schema, respectively; the mereological and the RCC schemas
are new -- to our knowledge, such models have not yet been considered.
In the following we outline several important aspects of this framework and provide a general context
for the model.
4.1 Emergence of the memory map in schemas. There is a clear parallel between a coherent repre-
sentation of space emerging from the integrated inputs of many individual neurons and a continuous state
of matter (e.g., a solid or a liquid) emerging from the collective dynamics of molecules. From a descriptive
point of view, the common element in both phenomena is that neither macro-system can be reduced to a
trivial aggregation of the properties of its elementary constituents. Even when the properties at both the
microscopic and macroscopic levels are well understood, it can be difficult to correlate the properties at one
level with those at the other. For example, the measurement of the macroscopic properties of a liquid does
not allow one to determine its molecular structure. Conversely, a detailed description of the properties of a
water molecule does not explain directly key phenomena of the physics of water. In physics, the solution
to this problem historically proceeded from a simplified, phenomenological models, which bridged the gap
between the microscopic and macroscopic levels of matter. In a similar way, the present discussion offers a
testbed model with which to bridge the gap between place cells and the large-scale spatial map.
4.2 Topological spatial maps. Topological maps have several biological advantages over geometric (or
topographic [36]) maps, which follow from the qualitative nature of topological relationships [16]. First,
natural environments are dynamic, so that it is often impossible for an animal to know when and how its
navigational task may change. Hence acquiring a qualitative map based on the invariants of the space of an
environment, may be biologically more effective than spending time on producing a computationally costly
precise answer from mutable relationships between dynamic cues.
One implication of this hypothesis is that in morphing environments the place fields will retain the pat-
tern of topological connectivity and may adjust their shapes in order to compensate for the deformation of
the representing space. This hypothesis is supported by experiments which demonstrate that place fields
15
maintain their relative configurations in morphing environments [76 -- 83] and that place cell coactivity pat-
tern in an animal traversing remains invariant over a significant range of geometric transformations [17, 22].
If the map is Cartesian, i.e., based on precise coordinates, distances, angles and so forth, such deformation
can be achieved by redrawing the place fields at each stage of the deformation, via some complex path in-
tegration mechanism [5, 11, 12, 63, 84 -- 86]. From the topological perspective, the observed deformation of
the place fields is simply a result of projecting the same stable neuronal map into a morphing environment,
which does not require extra computations and hence may be biologically more plausible.
4.3. Schemas constrain the generation of intrinsic sequences in the hippocampus. Place cells
become active in temporal sequences that either match with or are inverse to the spatial ordering of their
place fields during the active, resting, or sleep states. Initially, temporal sequences were observed after
or during the recording of the place fields, leading various authors to suggest that the observed temporal
sequences were a replay of sequences imprinted by sensory inputs [96, 97]. More recent experiments have
observed temporal sequences that corresponded to trajectories along which the animal had never traveled
[98]. Furthermore, experiments have revealed that temporal sequences observed before the animal entered
an environment for the first time were predictive of the place field sequence measured later [99, 100].
These observations strongly suggest that temporal sequences are not merely replays of previously imprinted
sequences [23, 98]. The better interpretation is that sequences are drawn from a pool of sequences that are
intrinsic to the hippocampal network and this network structure gives rise to the location of place fields
[101 -- 103]. The CRISP (for Context Representation, Intrinsic Sequences, and Pattern completion) theory
goes further to argue that the intrinsic sequences in the hippocampus are crucial for the storage of episodic
memories [101]. However, this theory does not explain the origin or properties of such sequences.
In the schema framework, all neural activity produced in the hippocampus has to be consistent with its
schema. For example, in the graph schema, spontaneously replayed sequences of neural activity would have
to be consistent with the connectivity of the graph. In other words, a cell ci may fire a spike after cell c j
only if the relationship ρi j = (ri, r j), or a chain of intermediate relationships ρii1, ρi1i2, ..., ρin−1in= j, is present
in PR. Other schemas impose different constraints on which sequences can be produced, and the elements
in the sequences may be ensembles of place cells, rather than single cells. In other words, schemas serve as
"topological templates" off which sequences are generated.
Physiologically, this implies that the hippocampal network that implements a particular schema can
produce sequences with specific "grammar" which may not have been directly imprinted or previously
produced by the network. In fact, such offline state sequences of place cell activations, which the animal
had never experienced, were recently observed in the experiments [99, 100]. Moreover, these sequences
were consistent with the topology of the spatial environment [15]. Thus, schemas can explain the intrinsic
sequences postulated by CRISP theory as well as in preplay and replay. This intimate relationship between
spontaneous sequences and schemas may be exploited in future investigations in order to infer the schema
based on recordings of sequences or to predict the properties of intrinsic sequences from a given schema.
4.4. Spatial vs. non-spatial memories. The hippocampus has been suggested to encode both spatial and
nonspatial memories [87 -- 91]. For example, it plays a key role in the ability to remember visual, odor, action
and memory sequences, and to put a specific memory episode into the context of preceding and succeeding
events, as well as the ability to produce complete memory sequence from a single structured input [92 -- 95].
The topological view on the hippocampal spatial representations [9, 17, 35] provides a common framework
for understanding both spatial and nonspatial memory functions as manifestations of a single mechanism,
which simply produces a topological arrangement of memory elements, irrespective of the nature of their
content. According to this view, there is no principal difference between the internalized topological map
of spatial locations and a topological map of memory sequences in the mnemonic domain.
4.5. Connections to experiment. Given the place cells' spiking parameters and a hypothesis about how
the downstream neurons might process place cell (co)activity, a schematic computation can be used to assess
the effectiveness of the corresponding spatial learning mechanism: how much time will be required to map
a space, how many integrals can such mapping produce, how quickly these integrals will emerge and how
16
stable they will be. This scope of computations suggests a possibility for experimental verifications of the
proposed framework. For example, a decline in spatial learning caused by neurodegenerative diseases (e.g.,
in Alzheimer's rat models), by aging or by psychoactive substances is assessed in behavioral experiments in
terms of the extra times required to learn various memory tasks. On the other hand, such cognitive changes
are associated with changes in the place cell spiking parameters [104 -- 109]. It may therefore be possible to
compare the downturn of spatial memory observed in topological learning tasks [11, 12] with the increase
of the learning time(s) estimated via a particular schema model for the same change in spiking variables.
Another alternative was suggested to us by our recent studies of hippocampal mapping of 3D spaces in
bats, using two types of simplicial schemas. The results suggest that in the 3D case, the readout neurons
in the place cell assemblies should operate by integrating synaptic inputs over working memory periods,
rather than detecting coactivities on synaptic plasticity timescale [110]. Of course, until these predictions
are proved or disproved experimentally, their value is discussable; meanwhile, the schema approach allows
theoretical reasoning and generates predictions about hippocampal neurophysiology.
V. METHODS
Place cells. Spiking is produced by the rat's movement through the environment covered by the place
fields (Figure 1A-B). The Poisson rate of the firing of place cells is a function of the animal's position r(t)
at time t,
λi(r) = fie
− (r−ri)2
2s2
i
,
where fi is the maximal firing rate of cell ci, si defines the width of its firing field centered at ri [20]. In
an N-cell ensemble, the parameters fi, and si, i = 1, ..., N are modeled as random variables drawn from
stationary unimodal distributions characterized by their respective means ( f and s) and standard deviations
(see Figure 1 and Methods in [9]). For the computations we used an ensemble with N = 200 neurons, with
mean firing rate f = 12 Hz and the mean place field size s = 20 cm. Larger ensembles typically affect the
numerical values of the computed quantities, but not the essence of the phenomena described in the paper.
This spiking is modulated by theta-oscillations, which are a subcortical EEG cycle in the hippocampus with
a frequency of ∼ 8 Hz (for details see [10]).
Learning Algorithm. The physiological processes responsible for emergence of cell assemblies with
readout neurons trained to integrate presynaptic inputs and to produce a particular response that "actualizes"
the information encoded by the place cell coactivity are complex and multifaceted [18]. For example, the
readout neurons that encode place field overlap must identify a group of place cells and learn to respond
to the coactivity of this specific group. However, what matters for our study, are the qualitative results of
this process: the number of readout neurons , the order of the coactivity detected by these neurons, how
this order grows in a typical cell assembly during the learning process and so forth. Therefore, we set aside
a neural network simulation of schema learning and employ the following schematic, phenomenological
algorithm:
If a relationship ρ of an appropriate type is detected, then:
1. if ρ is already listed in PR, ignore;
2. else if ρ can be inferred from the known relationships, ignore;
3. if ρ provides nontrivial information, then add ρ to PR.
4. if the known relationships can be inferred from ρ, then remove the redundant relationships.
5. continue
17
Steps 2 and 3 ensure that only the highest order relationships are kept in the schema, eliminating re-
dundant, lower-order relationships. At the beginning, every state of the readout neurons can be empty and
trained as the simulated animal explores a novel environment, or these states can be randomly initialized
and then relearned. The transitions between the readout neuron types may be regarded as a rudimentary,
schematic model of the synaptic plasticity mechanisms. In novel environments, place fields stabilize in
about four minutes [111], even though cognitive learning of the environment may take days or even weeks
[112]. This implies that the readout neurons can be trained using constant spiking parameters fi and si.
Temporal relationships between the spike trains and the physiological mechanisms underlying the
downstream neurons' readout process are in general very complex. For the sake of simplicity, we consider
only the rate-based representation of neuronal activity [113], which allows for a variety of possibilities
for encoding relationships. Such relationships may entail that the firing rates of the pre- and postsynaptic
neurons may be required to fall within a particular interval of values and the period of activity of a neuron
ci may be required to precede, to follow, or to overlap with the activity of a neuron c j by a certain minimal,
maximal or fixed amount of time [114]. The present analysis works from the three mutually exclusive
logical possibilities for the activity of any two neurons ci and c j :
1. there is an empty intersection of activity, i.e., the two cells are active at different times;
2. there is a non-null intersection of activity, i.e., their activities overlap;
3. the activity of one cell is a proper subset of the other cell, i.e., the activity of one cell occurs entirely
within the timespan of the activity of the other cell.
The time window for defining the co-activity of two or more cells is two θ-periods [10, 115].
Schema entropy and mutual information. For each relationship ρk of the schema we computed its nor-
malized frequency of appearance pk and evaluated the resulting specific entropy,
H = −Σk pk log2 pk.
The specific entropy for the corresponding spatial map was evaluated by identifying the relationships ρk(cid:48) that
obtain between the corresponding representing regions and computing their appearance probabilities pk(cid:48).
Following the trajectory of the animal (Figure 1B), we could also detect the joint probability of appearance
pk,k(cid:48) of a given pair of relationships, both in the schema as well as in the map (ρk, ρk(cid:48)), and then compute
their mutual information between the map and the schema,
MI = −ΣkΣk(cid:48) pk,k(cid:48) log2
pk,k(cid:48)
pk pk(cid:48) .
The computational software used for topological analysis is JPlex, an open-source package implement-
ing Persistent Homology Theory methods developed by the Computational Topology group at Stanford
University [116].
VI. ACKNOWLEDGMENTS
We thank Robert Phenix, Vicky Brandt and Loren Frank for helpful comments. The work was supported
in part by Houston Bioinformatics Endowment Fund (A.B. and Y.D.), the W. M. Keck Foundation grant for
pioneering research (A.B. and Y.D.) and by the NSF 1422438 grant (A.B. and Y.D.), and by the German
Research Foundation (Deutsche Forschungsgemeinschaft, DFG): SFB 874, project B2 (S.C.), a grant from
the Stiftung Mercator (S.C.).
VII. REFERENCES
18
[1] Tolman, E. (1948). Cognitive maps in rats and men, Psychol. Rev., 55: 189-208.
[2] Best, P., White, A., Minai, A. (2001). Spatial processing in the brain: the activity of hippocampal place cells.
[3] O'Keefe, J., Nadel, L. (1978). The hippocampus as a cognitive map, New York: Clarendon Press; Oxford
Annual. Rev. Neurosci. 24: 459-486.
University Press. xiv, 570 pp.
[4] Redish A., Touretzky D. (1997). Cognitive maps beyond the hippocampus, Hippocampus, 7: pp. 15-35.
[5] McNaughton B., Battaglia F., Jensen O., Moser E., Moser M. (2006). Path integration and the neural basis of
the 'cognitive map', Nat. Rev. Neurosci. 7: 663-678.
[6] Eichenbaum, H. (1999). Conscious awareness, memory and the hippocampus, Nat. Neurosci., 2: 775-776.
[7] Pouget, A., Dayan, P., Zemel, R. (2000). Information processing with population codes, Nat. Rev. Neurosci. 1:
125-132.
[8] Harnad, S. (1994). Why and how we are not zombies, Journal of Consciousness Studies 1: 164-167.
[9] Dabaghian, Y., Mmoli, F., Frank, L., Carlsson, G. (2012). A Topological Paradigm for Hippocampal Spatial
Map Formation Using Persistent Homology, PLoS Comput. Biol. 8: e1002581.
[10] Arai, M., Brandt, V., Dabaghian, Y. (2014). The Effects of Theta Precession on Spatial Learning and Simpli-
cial Complex Dynamics in a Topological Model of the Hippocampal Spatial Map, PLoS Comput. Biol. 10:
e1003651.
[11] Alvernhe, A., Sargolini, F., Poucet, B. (2012). Rats build and update topological representations through ex-
[12] Poucet, B., Herrmann, T. (2001). Exploratory patterns of rats on a complex maze provide evidence for topo-
[13] Alvernhe, A., Save, E., Poucet, B. (2011). Local remapping of place cell firing in the Tolman detour task, The
ploration, Anim. Cogn. 15: 359-368.
logical coding, Behav. Processes 53: 155-162.
European Journal of Neuroscience, 33:16961705.
[14] Shapiro, M., Tanila, H., Eichenbaum, H. (1997). Cues that hippocampal place cells encode: dynamic and
hierarchical representation of local and distal stimuli, Hippocampus 7: 624-642.
[15] Wu, X., Foster, D. (2014). Hippocampal Replay Captures the Unique Topological Structure of a Novel Envi-
ronment, The Journal of Neuroscience, 34: 6459-6469.
[16] Chen, Z., Kloosterman, F., Brown, E., Wilson, M. (2012). Uncovering spatial topology represented by rat
hippocampal population neuronal codes, J. Comput. Neurosci., 33: 227-255.
[17] Dabaghian, Y., Brandt, V., Frank, L. (2014). Reconceiving the hippocampal map as a topological template,
eLife 3:03476.
[18] Buzsaki, G. (2010). Neural syntax: cell assemblies, synapsembles, and readers, Neuron 68: 362-385.
[19] Russell, B. (1921) The analysis of mind, London, New York,: G. Allen & Unwin ltd.; The Macmillan company.
[20] Barbieri, R., Frank, L., Nguyen, D., Quirk, M., Solo, V., Wilson, M., Brown, E. (2004). Dynamic analyses of
information encoding in neural ensembles, Neural Comput. 16: 277-307.
[21] Frank, L., Stanley, G., Brown, E. (2004). Hippocampal plasticity across multiple days of exposure to novel
environments, J. Neurosci. 24: 7681-7689.
[22] Diba, K., Buzsaki, G. (2008). Hippocampal network dynamics constrain the time lag between pyramidal cells
across modified environments, J. Neurosci., 28: 13448-13456.
[23] Cheng, J., Ji, D. (2013). Rigid firing sequences undermine spatial memory codes in a neurodegenerative mouse
[24] Hayman, R., Verriotis, M., Jovalekic, A., Fenton, A., Jeffery, K. (2011). Anisotropic encoding of three-
dimensional space by place cells and grid cells, Nat. Neurosci., 14: 1182-1188.
[25] Yartsev, M., Ulanovsky, N. (2013). Representation of Three-Dimensional Space in the Hippocampus of Flying
[26] Taube, J. (2011). Head direction cell firing properties and behavioural performance in 3-D space, J. Physiol.
model, eLife 2:e00647.
Bats, Science 340: 367-372.
589: 835-841.
[27] Finkelstein, A., Derdikman, D., Rubin, A., Foerster, J., Las, L., Ulanovsky, N. (2015). Three-dimensional
head-direction coding in the bat brain, Nature 517: 159-164.
19
[28] Chen, A., DeAngelis, G., Angelaki, D. (2011). Representation of Vestibular and Visual Cues to Self-Motion in
Ventral Intraparietal Cortex, J. Neurosci. 31: 12036-12052.
[29] Swindale, N. (1996). Visual cortex: Looking into a Klein bottle, Current Biol. 6: 776-779.
[30] Barry, C., Burgess, N. (2007). Learning in a geometric model of place cell firing, Hippocampus, 17: 786-800.
[31] O'Keefe, J., Burgess, N. (1996). Geometric determinants of the place fields of hippocampal neurons, Nature
381: 425-428.
[32] Guger, C., Gener, T., Pennartz, C., Brotons-Mas, J., Edlinger, G., Bermudez i Badia, S., Verschure, P., Schaffel-
hofer, S., Sanchez-Vives, M. (2011). Real-time Position Reconstruction with Hippocampal Place Cells, Front.
Neurosci. 5:85.
[33] Brown, E., Frank, L., Tang, D., Quirk, M., Wilson, M. (1998). A statistical paradigm for neural spike train de-
coding applied to position prediction from ensemble firing patterns of rat hippocampal place cells, J. Neurosci.,
18: 7411-7425.
[34] Fenton, A., Muller, R., (1998). Place cell discharge is extremely variable during individual passes of the rat
through the firing field, Proc. Natl. Acad. Sci. 95: 3182-3187.
[35] Dabaghian, Y., Cohn, A., Frank, L., (2011). Topological Coding of the Hippocampus. In: Igelnik B, editor,
Computational Modeling and Simulation of Intellect: Current State and Future Perspectives, BMI Research
Inc., USA. pp. 293-320
[36] Chen, Z., Gomperts, S., Yamamoto, J., Wilson, M. (2014). Neural representation of spatial topology in the
[37] Stella, F., Cerasti, E., Treves, A. (2013). Unveiling the metric structure of internal representations of space,
rodent hippocampus, Neural Comput. 26: 1-39.
Front. Neural Circuits 7:81.
[38] Katz, Y., Kath, W., Spruston, N., Hasselmo, M. (2007). Coincidence detection of place and temporal context
in a network model of spiking hippocampal neurons, PLoS Comput. Biol. 3: e234.
[39] Brette, R. (2012). Computing with Neural Synchrony, PLoS Comput. Biol., 8: e1002561.
[40] Dehmer, M., Mowshowitz, A. (2011). A history of graph entropy measures, Info. Sci. 181: 57-78.
[41] Mowshowitz, A. (1968). Entropy and the complexity of graphs: I. An index of the relative complexity of a
graph, Bulletin of Mathematical Biophysics 30, pp. 175-204.
[42] Muller, R., Stead, M., Pach, J. (1996). The hippocampus as a cognitive graph, J. Gen. Physiol., 107: 663-694.
[43] Trullier, O., Meyer, J-A. (2000). Animat navigation using a cognitive graph, Biol. Cybern., 83: 271-285.
[44] Chrastil, R., Warren, W. (2014). From Cognitive Maps to Cognitive Graphs, PLoS One 9: e112544.
[45] Berge, C. (1982). The theory of graphs and its applications, Westport, Conn.: Greenwood Press.
[46] Aleksandrov, P. (1965). Elementary concepts of topology, New York: F. Ungar Pub. Co. 63 pp.
[47] Babichev, A., Memoli, F., Ji, D., Dabaghian, Y. (2015). Combinatorics of Place Cell Coactivity and Hippocam-
pal Maps, (arXiv:1509.01677).
hippocampus, Nature 424: 552-556.
[48] Harris, K., Csicsvari, J., Hirase, H., Dragoi, G., Buzsaki, G. (2003). Organization of cell assemblies in the
[49] Harris, K. (2005). Neural signatures of cell assembly organization, Nat. Rev. Neurosci., 6: 399-407.
[50] Hatcher, A. (2002). Algebraic topology, Cambridge; New York: Cambridge University Press.
[51] Dubrovin, B., Fomenko, A., Novikov, S. (1992). Modern geometry -- methods and applications. New York:
Springer-Verlag.
[52] Curto, C., Itskov, V. (2008) .Cell groups reveal structure of stimulus space, PLoS Comput. Biol., 4: e1000205.
[53] Maurer, P., Cowen, S., Burke, S., Barnes, C., McNaughton, B. (2006). Organization of hippocampal cell
assemblies based on theta phase precession, Hippocampus 16: 785-794.
[54] Muller, R., Kubie, J., Ranck, J. (1987). Spatial firing patterns of hippocampal complex-spike cells in a fixed
environment, J. Neurosci. 7: 1935-1950.
[55] Liu, Y-M., Luo, M-K. (1997). Fuzzy topology, River Edge, NJ: World Scientific Pub. x, 353 pp.
[56] Zomorodian, A., Carlsson, G. (2005). Computing persistent homology, Discrete & Computational Geometry
33: 249 -- 274.
45: 61-75.
[57] Ghrist, R. (2008). Barcodes: The persistent topology of data, Bulletin of the American Mathematical Society
[58] Hazarika, S., Cohn, A. (2001). Qualitative Spatio-Temporal Continuity, Proceedings of the International
Conference on Spatial Information Theory: Foundations of GIS, Springer-Verlag., New York, pp. 92-107.
doi:10.1007/3-540-45424-17
[59] Gotts, N., Gooday, J., Cohn, A. (1996). A Connection Based Approach to Common-Sense Topological De-
scription and Reasoning The Monist 79: 51-75.
20
[60] Renz, J., Rauh, R., Knauff, M. (2000). Towards Cognitive Adequacy of Topological Spatial Relations, In:
Freksa C, Habel C, Brauer W, Wender K, editors, Spatial Cognition II Springer Berlin Heidelberg. pp. 184-
197.
[61] Cohn, A., Hazarika, S. (2001). Qualitative Spatial Representation and Reasoning: An Overview, Fundam. Inf.,
46: 1-29.
[62] Knauff, M., Rauh, R., Renz, J. (1997). A Cognitive Assessment of Topological Spatial Relations: Results from
an Empirical Investigation. Proceedings of the International Conference on Spatial Information Theory: A
Theoretical Basis for GIS, Springer-Verlag. pp. 193-206.
[63] Goodrich-Hunsaker, N., Howard, B., Hunsaker, M., Kesner, R. (2008). Human topological task adapted for
rats: Spatial information processes of the parietal cortex, Neurobiol. Learn Mem., 90: 389-394.
[64] Wallgrun, J. (2010). Qualitative Spatial Reasoning for Topological Map Learning, Spatial Cognition & Com-
putation, 10: 207-246.
[65] Zeithamova, D., Schlichting, M., Preston, A. (2012). The hippocampus and inferential reasoning: Building
memories to navigate future decisions, Front. Hum Neurosci., 6.
[66] Cui, Z., Cohn, A., Randell, D. (1993). Qualitative and Topological Relationships in Spatial Databases, Proceed-
ings of the Third International Symposium on Advances in Spatial Databases Springer-Verlag. pp. 296-315.
[67] Cohn, A., Bennett, B., Gooday, J., Gotts, N. (1997). Qualitative Spatial Representation and Reasoning with the
Region Connection Calculus. Geoinformatica, 1: 275-316.
[68] Li, S., Cohn, A. (2012). Reasoning with topological and directional spatial information, Computational Intel-
ligence 28: 579-616.
[69] Gerevini, A., Renz, J. (1998). Combining Topological and Qualitative Size Constraints for Spatial Reason-
ing, Proceedings of the 4th International Conference on Principles and Practice of Constraint Programming,
Springer-Verlag. pp. 220-234.
[70] Lehmann, F., Cohn, A. (1994). The EGG/YOLK reliability hierarchy: semantic data integration using sorts
with prototypes, Proceedings of the third international conference on Information and knowledge management,
Gaithersburg, Maryland, United States: ACM. pp. 272-279.
[71] Gotts, N. (1994). Defining a 'doughnut' made difficult. In: C. Eschenbach C. Habel and B. Smith (Eds.),
Topological Foundations of Cognitive Science, University of Hamburg, Hamburg, pp. 105-130.
[72] Frank, L., Brown, E., Wilson, M. (2000). Trajectory encoding in the hippocampus and entorhinal cortex,
Neuron 27: 169-178.
[73] Huang, Y., Brandon, M., Griffin, A., Hasselmo, M., Eden, U. (2009). Decoding movement trajectories through
a T-maze using point process filters applied to place field data from rat hippocampal region CA1, Neural
Comput., 21: 3305-3334.
[74] Bennett, B. (1998). Determining Consistency of Topological Relations, Constraints 3: 213-225.
[75] Renz, J. (2002). Qualitative spatial reasoning with topological information, Berlin ; New York: Springer.
[76] Muller, R., Kubie, J. (1987). The effects of changes in the environment on the spatial firing of hippocampal
complex-spike cells, J. Neurosci. 7: 1951-1968.
[77] Colgin, L., Leutgeb, S., Jezek, K., Leutgeb, J., Moser, E., McNaughton, B., Moser, MB. (2010). Attractor-map
versus autoassociation based attractor dynamics in the hippocampal network, J. Neurophys., 104: 35-50.
[78] Lever, C., Wills, T., Cacucci, F., Burgess, N., O'Keefe, J. (2002). Long-term plasticity in hippocampal place-
cell representation of environmental geometry, Nature 416: 90-94.
[79] Leutgeb, J., Leutgeb, S., Treves, A., Meyer, R., Barnes, C., McNaughton, L., Moser, MB., Moser, E. (2005).
Progressive transformation of hippocampal neuronal representations in "morphed" environments, Neuron 48:
345-358.
[80] Wills, J., Lever, C., Cacucci, F., Burgess, N., O'Keefe, J. (2005). Attractor dynamics in the hippocampal
representation of the local environment. Science 308: 873-876.
[81] Touretzky, D., Weisman, W., Fuhs, M., Skaggs, W., Fenton, A., Muller, R. (2005). Deforming the hippocampal
map, Hippocampus 15: 41-55.
[82] Gothard, K., Skaggs, W., McNaughton, B. (1996). Dynamics of mismatch correction in the hippocampal en-
semble code for space: interaction between path integration and environmental cues, J. Neurosci., 16: 8027-
8040.
[83] Gothard, K., Skaggs, W., Moore, K., McNaughton, B. (1996). Binding of hippocampal CA1 neural activity to
multiple reference frames in a landmark-based navigation task. J. Neurosci. 16: 823-835.
[84] Etienne, A., Jeffery, K. (2004). Path integration in mammals, Hippocampus 14: 180-192.
[85] McNaughton, B., Barnes, C., Gerrard, J., Gothard, K., Jung, M., Knierim, J., Kudrimoti, H., Qin, Y., Sk-
21
aggs, W., Suster, M., Weaver, K. (1996). Deciphering the hippocampal polyglot: the hippocampus as a path
integration system, J. Exp. Biol. 199: 173-185.
[86] Poucet, B. (1993). Spatial cognitive maps in animals: new hypotheses on their structure and neural mecha-
nisms, Psychol. Rev., 100: 163-182.
[87] Eichenbaum, H. (2000). A cortical-hippocampal system for declarative memory, Nat. Rev. Neurosci. 1: 41-50.
[88] Konkel, A., Cohen, N. (2009). Relational memory and the hippocampus: Representations and methods, Front.
Neurosci. 3(2): 166174.
[89] Shrager, Y., Bayley, P., Bontempi, B., Hopkins, R., Squire, L. (2007). Spatial memory and the human hip-
pocampus, Proc. Natl. Acad. Sci. 104: 2961-2966.
[90] Soei, E., Koch, B., Schwarz, M., Daum, I. (2008). Involvement of the human thalamus in relational and non-
relational memory, European J. Neurosci., 28: 2533-2541.
[91] Eichenbaum, H., Dudchenko, P., Wood, E., Shapiro, M., Tanila, H. (1999). The hippocampus, memory, and
place cells: is it spatial memory or a memory space? Neuron 23: 209-226.
[92] Wood, E., Dudchenko, P., Eichenbaum, H. (1999). The global record of memory in hippocampal neuronal
[93] Fortin, N., Agster, K., Eichenbaum, H. (2002). Critical role of the hippocampus in memory for sequences of
[94] Fortin, N., Wright, S., Eichenbaum, H. (2004). Recollection-like memory retrieval in rats is dependent on the
[95] Sauvage, M., Fortin, N., Owens, C., Yonelinas, A., Eichenbaum, H. (2008). Recognition memory: opposite
effects of hippocampal damage on recollection and familiarity, Nat. Neurosci. 11: 16-18.
[96] Foster, D., Wilson, M. (2006). Reverse replay of behavioural sequences in hippocampal place cells during the
awake state, Nature 440: 680-683.
[97] Louie, K., Wilson, M. (2001). Temporally Structured Replay of Awake Hippocampal Ensemble Activity during
Rapid Eye Movement Sleep, Neuron 29: 145-156.
[98] Gupta, A., van der Meer M., Touretzky, D., Redish, A. (2010). Hippocampal Replay Is Not a Simple Function
[99] Dragoi, G., Tonegawa, S. (2011). Preplay of future place cell sequences by hippocampal cellular assemblies,
activity, Nature 397: 613-616.
events, Nat. Neurosci. 5: 458-462.
hippocampus, Nature 431: 188-191.
of Experience, Neuron 65: 695-705.
Nature 469: 397-401.
Acad. Sci., 110(22):9100-5.
[100] Dragoi, G., Tonegawa, S. (2013). Distinct preplay of multiple novel spatial experiences in the rat Proc. Natl.
[101] Cheng, S. (2013). The CRISP theory of hippocampal function in episodic memory, Front. Neural Circuits 7:88.
[102] Buhry, L., Azizi, A., Cheng, S. (2011). Reactivation, Replay, and Preplay: How It Might All Fit Together.
[103] Azizi, A., Wiskott, L., Cheng, S. (2013). A computational model for preplay in the hippocampus, Front.
Neural Plast. vol. 2011, Article ID 203462.
Comput. Neurosci. 7:161.
[104] Cacucci, F., Yi, M., Wills, T., Chapman, P., O'Keefe, J. (2008). Place cell firing correlates with memory deficits
and amyloid plaque burden in Tg2576 Alzheimer mouse model, Proc. Natl. Acad. Sci. 105: 7863-7868.
[105] Robbe, D., Montgomery, S., Thome, A., Rueda-Orozco, P., McNaughton, B., G. Buzsaki. (2006). Cannabinoids
reveal importance of spike timing coordination in hippocampal function, Nat. Neurosci., 9: 1526-1533.
[106] Silvers, J., Tokunaga, S., Berry, R., White, A., Matthews, D. (2003). Impairments in spatial learning and
memory: ethanol, allopregnanolone, and the hippocampus, Brain. Res. Rev., 43: 275-284.
[107] Gerrard, J., Kudrimoti, H., McNaughton, B., Barnes, C. (2001). Reactivation of hippocampal ensemble activity
patterns in the aging rat, Behav. Neurosci., 115: 1180-1192.
[108] Robitsek, R., Fortin, N., Koh, M., Gallagher, M., Eichenbaum, H. (2008). Cognitive aging: a common decline
of episodic recollection and spatial memory in rats, J. Neurosci., 28: 8945-8954.
[109] Wilson, I., Ikonen, S., Gureviciene, I., McMahan, R., Gallagher, M., Eichenbaum, H., Tanila, H. (2004).
Cognitive aging and the hippocampus: how old rats represent new environments, J. Neurosci. 24: 3870-3878.
[110] Hoffman, K., Babichev, A., Dabaghian, Y. (2016). Topological mapping of space in bat hippocampus,
(arXiv:1601.04253).
[111] Brown, N., Nguyen, D., Frank, L., Wilson, M., Solo, V. (2001). An analysis of neural receptive field plasticity
by point process adaptive filtering. Proc. Natl. Acad. Sci. 98: 12261-12266.
[112] Frank, L., Brown, E., Stanley, G. (2006). Hippocampal and cortical place cell plasticity:
implications for
episodic memory, Hippocampus 16: 775-784.
[113] Ahmed, O., Mehta, M. (2009). The hippocampal rate code: anatomy, physiology and theory, Trends Neurosci.
[114] Ligozat, G. (2013). Allen's Calculus. Qualitative Spatial and Temporal Reasoning, John Wiley & Sons, Inc.
32: 329-338.
Hoboken, NJ., pp. 1-28.
[115] Mizuseki, K., Sirota, A., Pastalkova, E., Buzsaki, G. (2009). Theta oscillations provide temporal windows for
local circuit computation in the entorhinal-hippocampal loop, Neuron 64: 267-280.
[116] (JPlex freeware). (2011). Computational Topology group, Stanford University.
22
|
1604.01359 | 3 | 1604 | 2016-07-04T00:17:05 | Physical limits to magnetogenetics | [
"q-bio.NC",
"q-bio.BM"
] | This is an analysis of how magnetic fields affect biological molecules and cells. It was prompted by a series of prominent reports regarding magnetism in biological systems. The first claims to have identified a protein complex that acts like a compass needle to guide magnetic orientation in animals (Qin et al., 2016). Two other articles report magnetic control of membrane conductance by attaching ferritin to an ion channel protein and then tugging the ferritin or heating it with a magnetic field (Stanley et al., 2015; Wheeler et al., 2016). Here I argue that these claims conflict with basic laws of physics. The discrepancies are large: from 5 to 10 log units. If the reported phenomena do in fact occur, they must have causes entirely different from the ones proposed by the authors. The paramagnetic nature of protein complexes is found to seriously limit their utility for engineering magnetically sensitive cells. | q-bio.NC | q-bio | Physical limits to magnetogenetics
Division of Biology and Biological Engineering
California Institute of Technology
Markus Meister
Pasadena, CA 91125
[email protected]
Abstract
This is an analysis of how magnetic fields affect biological molecules and cells. It was prompted
by a series of prominent reports regarding magnetism in biological systems. The first claims to
have identified a protein complex that acts like a compass needle to guide magnetic orientation
in animals (Qin et al., 2016). Two other articles report magnetic control of membrane
conductance by attaching ferritin to an ion channel protein and then tugging the ferritin or
heating it with a magnetic field (Stanley et al., 2015; Wheeler et al., 2016). Here I argue that
these claims conflict with basic laws of physics. The discrepancies are large: from 5 to 10 log
units. If the reported phenomena do in fact occur, they must have causes entirely different from
the ones proposed by the authors. The paramagnetic nature of protein complexes is found to
seriously limit their utility for engineering magnetically sensitive cells.
Introduction
There has been renewed interest recently in the effects of magnetic fields on biological cells. On
the one hand we have the old puzzle of magnetosensation: How do organisms sense the Earth's
magnetic field for the purpose of navigation? The biophysical basis for this ability is for the most
part unresolved. On the other hand lies the promise of "magnetogenetics": the dream of making
neurons and other cells responsive to magnetic fields for the purpose of controlling their activity
with ease. The
for
magnetosensation can point the way to effectively engineering magnetogenetics.
The physical laws by which magnetic fields act on matter are taught to science students in
college (Feynman et al., 1963). Obviously those principles impose some constraints on what
biological mechanisms
and
magnetogenetics. A recent spate of high-profile articles has put forward audacious proposals in
this domain without any attempt at such reality checks. My goal here is to offer some
calculations as a supplement to those articles, which makes them appear in a rather different
light. These arguments should also help in evaluating future hypotheses and in engineering new
molecular tools.
A molecular biocompass?
Generally speaking, magnetic fields interact only weakly with biological matter. The reason
magnetic fields are used for whole-body medical imaging, and why they have such appeal for
magnetogenetics, is that they penetrate through tissues essentially undisturbed. The other side of
this coin is that evolution had to develop rather special mechanisms to sense a magnetic field at
all, especially one as weak as the Earth's field.
linked, because uncovering Nature's method
two are closely
are plausible
candidates,
for both magnetosensation
1
This mechanism is well understood in just one case: that of magnetotactic bacteria (Bazylinski
and Frankel, 2004). These organisms are found commonly in ponds, and they prefer to live in the
muck at the bottom rather than in open water. When the muck gets stirred up they need to return
to the bottom, and they accomplish this by following the magnetic field lines down. For that
purpose, the bacterium synthesizes ferrimagnetic crystals of magnetite and arranges them in a
chain within the cell. This gives the bacterium a permanent magnetic moment, and allows it to
act like a small compass needle. The cell's long axis aligns with the magnetic field and flagella
in the back of the cell propel it along the field lines. It has been suggested that magnetosensation
in animals similarly relies on a magnetite mechanism, for example by coupling the movement of
a small magnetic crystal to a membrane channel (Kirschvink et al., 2001). A competing proposal
for magnetosensation suggests that the magnetic field acts on single molecules in certain
biochemical reactions (Ritz et al., 2010). In this so-called "radical pair mechanism" the products
of an electron transfer reaction depend on the equilibrium between singlet and triplet states of a
reaction intermediate, and this equilibrium can be biased by an applied magnetic field. These two
hypotheses and their respective predictions for magnetosensation have been reviewed
extensively (Johnsen and Lohmann, 2005; Kirschvink et al., 2010).
On this background, a recent article by Qin et al (2016) introduces a new proposal. As for
magnetotactic bacteria, the principle is that of a compass needle that aligns with the magnetic
field, but here the needle consists of a single macromolecule. This putative magnetic receptor
protein was isolated from the fruit fly and forms a rod-shaped multimeric complex that includes
40 iron atoms. The authors imaged individual complexes by electron microscopy on a sample
grid. They claim (1) that each such rod has an intrinsic magnetic moment, and (2) that this
moment is large enough to align the rods with the earth's magnetic field: "about 45% of the
isolated rod-like protein particles oriented with their long axis roughly parallel to the
geomagnetic field". We will see that neither claim is plausible based on first principles:
1. The protein complex has a permanent dipole moment. The smallest iron particles known to
have a permanent magnetic moment at room temperature are single-domain crystals of magnetite
(Fe3O4), about 30 nm in size (Dunlop, 1972). Those contain about 1 million iron atoms, closely
packed to produce high exchange interaction, which serves to coordinate their individual
magnetic moments (Feynman et al., 1963, Ch 37). The protein complex described by Qin et al
(2016) contains only 40 Fe atoms, and those are spread out over a generous 24 nm. There is no
known mechanism by which these would form a magnetic domain and thus give the complex a
permanent magnetic moment. Despite intense interest in making single-molecule magnets, their
blocking temperatures (when the magnetic spins become locked to the molecular axes) are still
below 14 degrees Kelvin (Demir et al., 2015). So the amount of iron in this putative molecular
compass seems too small by about 5 log units.
2. Individual complexes align with the earth's field. Let us suppose generously that the 40 Fe
atoms could in fact conspire – by a magical mechanism unknown to science – to align their
individual spins perfectly, and to make a single molecule with a permanent magnetic moment at
room temperature. How well would this miniature compass needle align with the earth's
magnetic field? This is a competition between the magnetic force that aligns the particle and
thermal forces that randomize its orientation. What is that balance?
2
An atom with n unpaired electrons has an effective magnetic moment of
where 1
µeff = n n + 2
(
)µB ,
µB = Bohr magneton = 9 ×10−24 J
T
,
(1)
(2)
For iron atoms, n is at most 5, and a complex of 40 aligned Fe atoms would therefore have a
magnetic moment of at best
m = 40 × 5 5 + 2
(
)µB = 2 ×10−21 J
T
.
The interaction energy of that moment with the earth's field (about 50 µT) is at most
mBEarth = 1×10−25 J .
Meanwhile the thermal energy per degree of freedom is
The ratio between those is
kT = 4 ×10−21J .
mBEarth
kT
= 2 ×10−5 .
(3)
(4)
(5)
(6)
That is the degree of alignment one would expect for the protein complex. Instead, the authors
claim an alignment of 0.45. Again, this claim exceeds by about 5 log units the prediction from
basic physics, even allowing for a magical alignment of the 40 Fe spins. Clearly the reported
observations must arise from some entirely different cause, probably unrelated to magnetic
fields.
An ion channel gated by magnetic force?
With the goal of controlling the activity of neurons, Wheeler et al (2016) reported the design of a
molecular system intended to couple magnetic fields to ionic current across the cell membrane.
Their single-component protein consists of a putative mechano-sensitive cation channel
(TRPV4) fused on the intracellular face to two subunits of ferritin. The hope was that "the
paramagnetic protein would enable magnetic torque to tug open the channel to depolarize cells".
Indeed, the report includes experimental results from several preparations suggesting that neural
activity can be modulated by static magnetic fields.2 What could be the underlying biophysical
mechanism?
Ferritin is a large protein complex with 24 subunits that forms a spherical shell about 12 nm in
diameter. Wheeler et al (2016) suppose optimistically that the two subunits of ferritin attached to
1 In the spirit of order-of-magnitude calculations, I will use single-digit precision for all
quantities.
2 There is a similar claim in Stanley et al (2015), but the evidence is scant and hard to interpret:
only 18 of ~2000 cells "responded" (their Supplementary Figure 10).
3
the channel protein are able to nucleate an entire 24-subunit ferritin complex. The hollow core of
this particle can be filled with iron in the form of a ferric hydroxide (Arosio et al., 2009). At
room temperature ferritin has no permanent magnetization: it is strictly paramagnetic or
superparamagnetic (Papaefthymiou, 2010). Unlike the magnetite particles in magnetotactic
bacteria, the iron core of ferritin is too small (~5 nm) to sustain a permanent dipole moment
(blocking temperature ~40 K). Instead the direction of the Fe spins in the core fluctuates
thermally. An external magnetic field biases these fluctuations, producing a magnetic moment m
proportional to the field B of
where ξ is the magnetizability of a ferritin particle. This quantity can be derived from bulk
measurements of ferritin magnetic susceptibility (see Methods) at
m =ξB ,
(7)
ξ= 2.4 ×10−22 J
T2 .
(8)
I will consider four scenarios by which such a ferritin particle might be manipulated with an
external magnetic field. In the first, the magnetic field has a gradient, and the particle is pulled in
the direction of higher field strength. In the second, the force arises from interactions among
neighboring ferritins through their induced magnetic moments. In the third, the magnetic field
exerts a torque assuming that the ferritin core is anisotropic, with a preferred axis of
magnetization. Finally, the collective pull of many ferritins on the cell membrane may induce a
stress that opens stretch-activated channels.
Fig. 1: A TRPV4 channel (pink) inserted in the membrane with a ferritin complex (green) attached on the
cytoplasmic side, approximately to scale. The magnetic field B induces a moment m in the ferritin core,
leading to a force F or a torque N on the ferritin particle, and resulting forces tugging on the channel. See text
for details.
1. A magnetic field gradient pulls on ferritin (Fig 1a). Paramagnetic particles experience a force
that is proportional to the magnetic field gradient and the induced magnetic moment (Feynman et
al., 1963, Ch 35). In the experiments of Wheeler et al (2016) the field strength was ~0.05 T and
the field gradient ~6.6 T/m (their Supplementary Figure 2). What is the resulting force on a
ferritin particle?
The interaction energy between the moment and the magnetic field is
4
U = −
1
2
mB ,
(9)
where the factor of 1 2 arises because the moment m is in turn induced by the field (Jackson,
1998, Ch 5.16). The force produced by the field gradient is the spatial derivative of that energy,
namely
F1 = −
d
dx
U = ξB dB
dx = 2 ×10−22 × 0.05 × 7 N = 7 ×10−23N .
(10)
This would be the force exerted by one ferritin complex on its linkage under the reported
experimental conditions.
How does this compare to the force needed to open an ion channel? That has been measured
directly for the force-sensitive channels in auditory hair cells (Howard and Hudspeth, 1988), and
amounts to about 2 ×10−13 N . So this mechanism for pulling on ferritin seems at least 9 log units
too weak to provide an explanation.
2. Two ferritins pull on each other (Fig 1b). As proposed by Davila et al (Davila et al., 2003),
neighboring paramagnetic particles linked to the cell membrane could tug on each other by the
interaction between their magnetic moments, rather than by each being drawn into a magnetic
field gradient. If the field is oriented parallel to the cell membrane, then nearby ferritins will have
induced magnetic moments that are collinear and thus attract each other. If the field is
perpendicular to the membrane their magnetic moments will repel (Fig 1b). These dipole-dipole
interactions decline very rapidly with distance. For example, in the attractive configuration the
force between two dipoles of equal magnetic moment m at distance d is given by
where
F =
3µ0
2π
m2
d4 ,
(11)
is the vacuum permeability. The strongest interaction will be between two ferritins that are
nearly touching, so that d = 2R = 12 nm . In that situation one estimates that
(12)
µ0 = 4π×10−7 N
A2
F2 =
3µ0
2π
ξB(
2R(
)2
)4 = 3×10−21 N .
(13)
Unfortunately we are again left with an exceedingly tiny force, about 8 log units weaker than the
gating force of the hair cell channel.
What if the mechano-sensitive channel used in this study is simply much more sensitive to tiny
forces than the channel in auditory hair cells? An absolute limit to sensitivity is given by thermal
fluctuations. Whatever molecular linkage the ferritin is pulling on, it needs to provide at least kT
of energy to that degree of freedom to make any difference over thermal motions. Because of the
steep distance dependence, the force between ferritins drops dramatically if they move just one
5
radius apart. The free energy gained by that motion compared to the thermal energy is
approximately
F2R
kT =
3×10−21 × 6 ×10−9
4 ×10−21
= 4 ×10−9 ,
(14)
again 8 log units too small to have any noticeable effect.
3. The magnetic field exerts a torque on the ferritin (Fig 1c). Although at room temperature
ferritin has no permanent magnetic moment, its induced moment may exhibit some anisotropy.
In general this means that the iron core is more easily magnetized in the "easy" direction than
orthogonal to it. For example, this may result from an asymmetric shape of the core. While the
exact value of that anisotropy is unknown, we can generously suppose it to be infinite, so the
ferritin particle has magnetizability ξ in one direction and zero in the orthogonal directions.
Thus the induced magnetic moment may point at an angle relative to the field (Fig 1c), resulting
in a torque on the ferritin particle that could tug on the linkage with the channel protein.
However, the magnitude of such effects is again dwarfed by thermal fluctuations: The interaction
energy between the moment and a magnetic field pointing along the easy axis is
1
2
mB = −
1
2ξB2 = −3×10−25 J
(15)
U! = −
and zero with the field orthogonal. This free energy difference is about 4 log units smaller than
the thermal energy. Following the same logic as for Qin et al's compass needle, the magnetic
field can bias the alignment of the ferritins by only an amount of 10-4. Another way to express
this is that any torque exerted by the ferritin on its ion channel linkage will be 10,000 times
smaller than the thermal fluctuations in that same degree of freedom.
4. Many ferritins exert a stress that gates mechanoreceptors in the membrane (Fig 1d). Perhaps
the magnetic responses are unrelated to the specific linkage between ferritin and a channel
protein. Instead one could imagine that a large number of ferritins exert a collective tug on the
cell membrane, deforming it and opening some stress-activated channels in the process. The
membrane stress required to gate mechanoreceptors has been measured directly by producing a
laminar water flow over the surface of a cell: For TRPV4 channels it amounts to ~20 dyne/cm2
(Soffe et al., 2015); for Piezo1 channels ~50 dyne/cm2 (Ranade et al., 2014). Suppose now that
the membrane is decorated with ferritins attached by some linkage, and instead of viscous flow
tugging on the surface one applies a magnetic field gradient to pull on those ferritins with force
F1 (Eqn 10). The density of ferritins one would need to generate the required membrane stress is
20 dyn/cm2
7 ×10−23 N
= 3×1010 ferritins
µm2
(16)
Unfortunately, even if the membrane is close-packed with ferritin spheres, one could fit at most
104 on a square micron. So this hypothetical mechanism produces membrane stress at least 6 log
units too weak to open any channels.
6
An ion channel gated by magnetic heating?
For a different mode of activating membrane channels, Stanley et al (2015) combined the
expression of ferritin protein with that of the temperature-sensitive membrane channel TRPV1.
The hope was that a high-frequency magnetic field could be used to heat the iron core of ferritin,
leading to a local temperature increase sufficient to open the TRPV1 channels, allowing cations
to flow into the cell. Stanley et al (2015) compared three different options for interaction
between the ion channels in the plasma membrane and the ferritin protein: In one case the ferritin
was expressed in the cytoplasm, in another it was targeted to the membrane by a myristoyl tail,
and in the third it was tethered directly to the channel protein by a camelid antibody linkage. The
direct one-to-one linkage between ferritin and ion channel worked best for generating Ca influx
via high-frequency magnetic fields, leading the authors to concluded that "Because temperature
decays as the inverse distance from the particle surface, heat transfer is likely to be most efficient
for this construct, suggesting that heat transfer from the particle could be limiting the efficiency
of the other constructs." Here I consider whether heat transfer from the ferritin particle is a likely
source of thermal activation for the TRPV1 channel at all.
Magnetic heating of nanoparticles is indeed a very active area of research. A sample biomedical
application would be to steer the nanoparticles within the body to certain cells, and then kill
those selectively by magnetic heating (Babincova et al., 2000). Typical nanoparticles of interest
are made of magnetite or maghemite, sometimes doped with other metals, and measure some
tens of nanometers in size (Hergt et al., 2006). A typical heating apparatus for small preparations
– like in the experiments of Stanley et al (2015) – consists of an electric coil with a few
windings, several centimeters in diameter, that carries a large oscillating current. The magnetic
fields generated inside the coil are on the order of tens of kA/m at frequencies of several 100
kHz 3.
Owing to the small size of the nanoparticles, the physics of heating are quite different from the
processes in our kitchen. A microwave oven heats water primarily by flipping molecular dipoles
in an oscillating electric field. And an induction stove works by inducing electric eddy currents
in the pot's bottom with an oscillating magnetic field. Neither of these electric effects plays any
role for nanoparticles. Instead the heat is generated purely by magnetic forces (Hergt et al.,
2006). Part of this comes from reorienting the magnetization of the material at high frequency,
which is opposed by internal relaxation processes, causing dissipation and heat. For larger
nanoparticles, the oscillating magnetic field may also make the particle move, with resulting
dissipation from external friction in the surrounding medium. These physical processes have
been modeled in great detail, and there is a large experimental literature to determine the heating
rates that can be accomplished with different kinds of nanoparticles. A figure of merit is the
"specific loss power (SLP)", namely the heating power that can be generated per unit mass of the
3 The literature sometimes refers misleadingly to heating by "radio waves" or "electromagnetic
radiation". At these frequencies the wavelength of a radio wave is about a kilometer, so of no
practical relevance to the experiments. There is no radiation involved in the interaction between
the solenoid coil and the nanoparticle.
7
magnetic material (see Methods). What sort of heating rate would we expect for the ferritin
particles used by Stanley et al (2015)?
Given the ease with which magnetic heating can be measured, the long-standing interest in
ferritin for medical engineering (Babincova et al., 2000), and the extensive research on its
magnetic properties (Papaefthymiou, 2010), it is surprisingly difficult to find any published
evidence for magnetic heating in ferritin. One report on the subject concludes simply that there is
none: ferritin shells reconstituted with a magnetite core produced no measurable magnetic
heating (SLP < 0.1 W/g), whereas doping the iron with varying amounts of cobalt did produce
some modest heating rates (Fantechi et al., 2014). Why is native ferritin such a poor heater? Both
theory and experiment show that the efficiency of heating magnetic nanoparticles depends
strongly on the particle size, and plummets steeply below 10 nm (Fortin et al., 2007;
Purushotham and Ramanujan, 2010). Magnetite particles smaller than 8 nm are not considered
useful for magnetic hyperthermia (Fantechi et al., 2015). The iron core of ferritin measures only
5-6 nm in diameter. Furthermore, the ferric hydroxide material in native ferritin has much lower
magnetic susceptibility than magnetite (~8-fold, Zborowski et al., 1996). Finally, the large
organic shell around the magnetic core may inhibit the motions that would allow for frictional
heating.
So, based on the literature, the most likely heating rate for ferritin is zero. Obviously that casts
doubt on the claims of Stanley et al (2015) that they activated ion channels through heating
ferritin. For the sake of keeping the argument alive, and to evaluate potential future
developments, let us instead suppose that ferritin could be engineered to produce a specific
heating rate of
P = 30 W
g of metal
(17)
This is the highest value obtained by filling the ferrite shell with cobalt-doped magnetite
(Fantechi et al., 2014) and thus a generous estimate of what might be accomplished by future
engineering of ferritin complexes inside cells. Assuming this specific heating rate, a single
ferritin particle with 2400 iron atoms generates heat at a rate of
This heat flux will produce a temperature gradient in the surrounding medium (Fig 2). As
Stanley et al (2015) state, the temperature indeed decays as the inverse distance r
from the
particle (Feynman et al., 1963, Ch 12), namely
Q
4πC
Q = 7 ×10−18 W
T r( ) =
(18)
1
r
(19)
where
C = 0.61 W
m ⋅K
(20)
is the thermal conductivity of water. Right at the surface of the ferritin sphere the temperature
increase is highest, namely
8
) = 1.5 ×10−10 K
Tferritin = T 6 nm(
This is a very tiny increase. Activation of a TRPV1 channel requires about 5 K of increase
relative to body temperature (Cao et al., 2013). So the temperature increase expected, even from
a futuristic optimized ferritin, is more than 10 orders of magnitude too small. The contention by
Stanley et al (2015) – that the proximity of ferritin to the channel allows for effective heating –
seems to lack a scientific basis.
(21)
12 nm
e
r
u
t
a
r
e
p
m
e
T
0
Distance (Radii)
5
Fig. 2: The steady-state temperature
profile around a heated sphere in an
infinite bath varies inversely with the
distance from the center of the sphere.
The same argument applies to a ferritin
sphere heated from its magnetic core
(top) and a spherical cell with a large
number of heated ferritins on its surface
(bottom).
10 µm
Perhaps the many other ferritins expressed on the same cell, though they are at greater distance,
might contribute to heating the local environment. Suppose one can express Nferritins = 10,000
TRPV1-ferritin complexes on the surface of a spherical cell with rcell = 5 µm radius. That is
about 10-fold the natural expression level in sensory neurons. One can treat the heat production
of those 10,000 ferritins as distributed evenly over the surface of the cell. Then the temperature
gradient outside the cell again follows a 1/r profile (Fig 2). At the surface of the cell the resulting
temperature increase will be
Tcell =
Q Nferritins
4πC rcell
= 1.7 ×10−9 K
(22)
Unfortunately this is still too low by 9 orders of magnitude. So one cannot achieve activation of
single neurons this way, which is of course a central goal of genetically expressed activators.
Suppose now that one expresses this number of ferritin-TRPV1 complexes on every neuron in
the brain. Would that perhaps be sufficient to heat the entire organ? At that density, the heating
rate per unit mass of brain will be
Pbrain =
Q Nferritins
4
3π r3
cell
1
ρbrain
= 1.2 ×10−4 W
g ,
(23)
where ρbrain = 1.03 g cm3 is the specific density. For comparison, the resting metabolic rate of
brain tissue is ~1.2 ×10−2 W g , and the resulting heat is carried away and regulated by the
processes that keep the organ's temperature stable. Heating of ferritin throughout the entire brain
would therefore contribute only a 1% increase to the heat already being generated from basal
9
activity: this will not overwhelm the homothermic regulation mechanisms sufficiently to open
TRPV1 channels.
In summary, it seems very unlikely that the effects reported in Stanley et al (2015) have anything
to do with heating ferritin. The available evidence says that native ferritin produces no
measurable magnetic heating at all. Even if we ignore that and assume a generous heating rate,
namely the largest reported using a custom metal alloy for the ferritin core, the resulting effects
are too small to matter by astronomical factors of 1010 (single-channel activation) and 109 (for
single-neuron activation).
Discussion
The calculations presented here evaluate the mechanisms that might underlie recent observations
on a molecular compass (Qin et al., 2016) and neural activation with static magnetic fields
(Wheeler et al., 2016) or high-frequency magnetic fields (Stanley et al., 2015). By all accounts,
none of the biophysical schemes proposed in these articles is even remotely plausible, and a few
additional proposals were eliminated along the way. The forces or torques or temperatures they
produce are too small by many orders of magnitude for the desired effects on molecular
orientation or on membrane channels. If the phenomena occurred as described, they must rely on
some entirely different mechanism. Barring dramatic new discoveries about the structure of
biological matter, the proposed routes to magnetogenetics, based on either pulling or heating a
ferritin/channel complex with magnetic fields, have no chance of success.
One does have to ask why none of these reports attempted a back-of-the-envelope estimate to
bolster their scientific claims. Neither, it seems, did the reviewers perform such a calculation, nor
the authors of three pieces that heralded the new achievements (Leibiger and Berggren, 2015;
Lewis, 2016; Lohmann, 2016). Why is it important to do so? First of all, claims that violate the
known laws of physics often turn out to be wrong (Maddox et al., 1988). There is, of course,
always a small chance of discovering new physics, but only if one understands what the old
physics predicts and recognizes the discrepancy. For example, if Qin and colleagues really
discovered a room-temperature molecular magnet – with a permanent magnetic moment
100,000-fold larger than the sum of all its iron spins – they will undoubtedly earn a Nobel Prize,
followed by financial bliss from the application of those molecular magnets to data storage
technology.
More importantly though, calculations are most useful when done ahead of time, to guide the
design of experiments. For example, Stanley et al (2015) and Wheeler et al (2016) evoke an
image in which the magnetic field pulls on the ferritin particles. This is possible only if the
magnetic field has a strong gradient (Fig 1a, Eqn 10). None of their experiments on animals were
designed to produce a strong gradient, nor do the articles report what it was. It is in fact possible
to pull on cells that express lots of ferritin, and this has been exploited for magnetic separation
(Owen and Lindsay, 1983). It requires very high magnetic fields, and separation columns with a
meshwork of fine steel fibers that produce strong gradients on a microscopic scale. Inserting
such a wire mesh into the brain would of course negate the goal of non-invasive control.
Two other hypothetical mechanisms for the ferritin effects require a strong field but no gradient.
This would be of great experimental value, because a homogeneous magnetic field could then
10
deliver the same control signal throughout an extended volume, like the brain of a mouse.
Among these, the dipole interaction between ferritins (Fig 1b, Eqn 14) offers little hope. Even
with a 100-fold larger field (5 T), these forces are still 4 log units too small to open a channel.
That field strength represents a practical limit: Small movements of the animal, or switching of
the field, will cause inductive eddy currents that activate the brain non-specifically, a
phenomenon experienced also by MRI subjects (Schenck et al., 1992).
On the other hand, exploiting anisotropy of the ferritin particle (Fig 1c, Eqn 15) may be within
range of utility. A 100-fold larger field could produce torque comparable with the thermal
energy, which when applied to thousands of channels might have a noticeable effect on
membrane currents. To enhance the shape anisotropy of the magnetic particles, perhaps one
could engineer the ferritin shell into an elongated shape. More fundamentally, it is clear that the
weak effects computed here are a consequence of ferritin's paramagnetism. A particle with a
permanent magnetic moment, such as the magnetosomes made by bacteria (Bazylinski and
Frankel, 2004), could exert much larger forces, torques, and temperatures (Hergt et al., 2006),
and may offer a physically realistic route to magnetogenetics.
With an eye towards such future developments, it is unfortunate that these three pseudo-
inventions were published, especially in high-glamour journals, because that discourages further
innovation. Now that the prize for magnetogenetics has seemingly been claimed, what motivates
a young scientist to focus on solving the problem for real? There is an important function here
for post-publication peer review, to make up for pre-publication failures and to reopen the
claimed intellectual space for future pioneers.
11
Methods
Magnetizability of native ferritin
Central to the arguments about magnetogenetics is the proportionality factor ξ between the
magnetic moment m of a single ferritin molecule and the magnetic field B ,
(24)
m =ξB .
Experimental measurements are usually performed on bulk samples of ferritin and report the
magnetic susceptibility χ, defined by
where M is the magnetization of the material, namely the magnetic moment per unit volume,
and
M = χH = χB µ0 ,
(25)
µ0 = 4π×10−7 N
A2
is the vacuum permeability. Therefore
ξ= χ
ρµ0
,
(26)
(27)
where ρ is the number of ferritin particles per unit volume. In practice, we will see that the
reported measurements of magnetization are more often normalized by the iron content of the
sample or by the mass, rather than by volume. Then the choice of ρ must be adjusted
accordingly.
• Michaelis et al (1943) report the susceptibility χFe of ferritin at 5.9 ×10−3 CGS units per mole
of iron in the preparation. Therefore we must divide by the number of ferritins per mole of iron,
ρFe . The authors report iron loading of maximally 23% w/w, which amounts to 2400 Fe atoms
per ferritin, and so
ρFe =
NA
2400
,
(28)
where NA is Avogadro's number. Furthermore, note that one CGS unit of molar susceptibility is
equivalent to 4π×10−6 SI units. Therefore the magnetizability of one ferritin is
T2 = 2.35 ×10−22 J
T2 .
2400 × 5.9 ×10−3 × 4π×10−6
6.02 ×1023 × 4π×10−7
ξMic = χFe
ρFeµ0
(29)
=
J
As a sanity check for all the conversions, we can use the authors' statement that the susceptibility
followed the Curie Law with an equivalent moment per iron atom of
(30)
µeff = 3.78 µB.
12
From this one derives
ξMic =
2
Nµeff
3kT =
2400 × 3.78 × 9.27 ×10−24
(
3× 4.11×10−21
)2
J
T2 = 2.37 ×10−22 J
T2
(31)
in close agreement with Eqn 23.
• Schoffa et al (1965) again report the susceptibility χFe referred to the iron content with a value
of 6.05 ×10−3 CGS units per mole of iron. Assuming again an iron loading of 2400 Fe per
ferritin, this results in
ξSch = 2.41×10−22 J
T2 .
(32)
• Jandacka et al (2015) report a susceptibility per unit mass χmass in SI units of
2.5 ×10−4 Am2 gT. Therefore we must divide by the number of ferritins per unit mass, ρmass .
At 2400 Fe per particle, one ferritin weighs ~580 kD, so that
and
ρmass =
NA
5.8 ×105 g
,
ξJan = χmass
ρmassµ0
=
2.5 ×10−4 × 5.8 ×105
6.02 ×1023
J
T2 = 2.41×10−22 J
T2 .
(33)
(34)
Given that these three measurements span the better part of a century using three different
instruments, the agreement is remarkable. I will use the value
Magnetic heating of nanoparticles
ξ= 2.4 ×10−22 J
T2 .
(35)
Table 1 summarizes some published measurements on magnetic heating of small nanoparticles
with diameter below 10 nm. The loss power per unit mass (SLP) depends on the apparatus used
for heating, in particular the SLP varies proportionally to the frequency of the alternating
magnetic field, and to the square of the field strength (Hergt et al., 2004). The table therefore
corrects all the SLP numbers to the conditions used by Stanley et al (2015): field strength
H = B µ0 = 25.5kA m and frequency f = 465 kHz .
13
Reference
Fortin et al
(2006)
Fortin et al
(2006)
Fortin et al
(2006)
Fantechi et
al (2015)
Hergt et al
(2004)
Fantechi et
al (2014)
Fantechi et
al (2014)
Fe2O3
Fe2O3
Fe2O3
Fe3O4
Fe2O3
ferritin with Fe3O4
ferritin with
Co0.15Fe2.85O4
8
8
7
6
5.3
6.7
24.8
700
24.8
700
24.8
700
12
15
183
410
4
14
37
6.5
15
2.8
10
26
75
49
12.4
183
<0.01
<0.1
per mass of only the metal ions
Material
d
[nm]
H
[kA/m]
f
[kHz]
SLP
[W/g]
SLP corr
[W/g]
Notes
6.8
12.4
183
2.81
30
per mass of only the metal ions
Table 1: Published measurements of specific loss power (SLP) for various magnetic particles of diameter d,
taken at a magnetic field strength H and frequency f. The values in the column "SLP corr" are corrected for the
field and frequency used by Stanley et al (2015).
14
References
Arosio, P., Ingrassia, R., and Cavadini, P. (2009). Ferritins: a family of molecules for iron
storage, antioxidation and more. Biochim Biophys Acta 1790, 589-599.
Babincova, M., Leszczynska, D., Sourivong, P., and Babinec, P. (2000). Selective treatment of
neoplastic cells using ferritin-mediated electromagnetic hyperthermia. Medical Hypotheses
54, 177-179.
Bazylinski, D. A., and Frankel, R. B. (2004). Magnetosome formation in prokaryotes. Nat Rev
Microbiol 2, 217-230.
Cao, E., Cordero-Morales, J. F., Liu, B., Qin, F., and Julius, D. (2013). TRPV1 channels are
intrinsically heat sensitive and negatively regulated by phosphoinositide lipids. Neuron 77,
667-679.
Davila, A. F., Fleissner, G., Winklhofer, M., and Petersen, N. (2003). A new model for a
magnetoreceptor in homing pigeons based on interacting clusters of superparamagnetic
magnetite. Physics and Chemistry of the Earth 28, 647-652.
Demir, S., Jeon, I.-R., Long, J., R., and Harris, T. D. (2015). Radical ligand-containing single-
molecule magnets. Coordination Chemistry Reviews 289, 149-176.
Dunlop, D. J. (1972). Magnetite: Behavior near the Single-Domain Threshold. Science 176, 41-
43.
Fantechi, E., Innocenti, C., Zanardelli, M., Fittipaldi, M., Falvo, E., Carbo, M., Shullani, V., Di
Cesare Mannelli, L., Ghelardini, C., Ferretti, A. M., Ponti, A., Sangregorio, C., and Ceci, P.
(2014). A smart platform for hyperthermia application in cancer treatment: cobalt-doped
ferrite nanoparticles mineralized in human ferritin cages. ACS Nano 8, 4705-4719.
Fantechi, E., Innocenti, C., Albino, M., Lottini, E., and Sangregorio, C. (2015). Influence of
cobalt doping on the hyperthermic efficiency of magnetite nanoparticles. Journal of
Magnetism and Magnetic Materials 380, 365-371.
Feynman, R. P., Leighton, R. B., and Sands, M. L. (1963). The Feynman Lectures on Physics,
Vol. 2 (Addison-Wesley). http://www.feynmanlectures.info/.
Fortin, J.-P., Wilhelm, C., Servais, J., Menager, C., Bacri, J.-C., and Gazeau, F. (2007). Size-
sorted anionic iron oxide nanomagnets as colloidal mediators for magnetic hyperthermia. J
Am Chem Soc 129, 2628-2635.
Hergt, R., Hiergeist, R., Zeisberger, M., Glockl, G., Weitschies, W., Ramirez, P., Hilger, I., and
Kaiser, W. A. (2004). Enhancement of AC-losses of magnetic nanoparticles for heating
applications. Journal of Magnetism and Magnetic Materials 280, 358-368.
Hergt, R., Dutz, S., Mueller, R., and Zeisberger, M. (2006). Magnetic particle hyperthermia:
nanoparticle magnetism and materials development for cancer therapy. Journal of Physics-
Condensed Matter 18, S2919-S2934.
Howard, J., and Hudspeth, A. J. (1988). Compliance of the hair bundle associated with gating of
mechanoelectrical transduction channels in the bullfrog's saccular hair cell. Neuron 1, 189-
199.
Jackson, J. D. (1998). Classical Electrodynamics, Chapter 5.16 (Hoboken, NJ: Wiley).
Jandacka, P., Burda, H., and Pistora, J. (2015). Magnetically induced behaviour of ferritin
corpuscles in avian ears: can cuticulosomes function as magnetosomes? J R Soc Interface 12,
20141087.
15
Johnsen, S., and Lohmann, K. J. (2005). The physics and neurobiology of magnetoreception. Nat
Rev Neurosci 6, 703-712.
Kirschvink, J. L., Walker, M. M., and Diebel, C. E. (2001). Magnetite-based magnetoreception.
Curr Opin Neurobiol 11, 462-467.
Kirschvink, J. L., Winklhofer, M., and Walker, M. M. (2010). Biophysics of magnetic
orientation: strengthening the interface between theory and experimental design. J R Soc
Interface 7 Suppl 2, S179-91.
Leibiger, I. B., and Berggren, P. O. (2015). Regulation of glucose homeostasis using
radiogenetics and magnetogenetics in mice. Nat Med 21, 14-16.
Lewis, S. (2016). Techniques: Magnetic manipulation. Nat Rev Neurosci 17, 262-263.
Lohmann, K. J. (2016). Protein complexes: A candidate magnetoreceptor. Nat Mater 15, 136-
138.
Maddox, J., Randi, J., and Stewart, W. W. (1988). "High-dilution" experiments a delusion.
Nature 334, 287-291.
Michaelis, L., Coryell, C. D., and Granick, S. (1943). Ferritin III. The magnetic properties of
ferritin and some other colloidal ferric compounds. J Biol Chem 148, 463-480.
Owen, C. S., and Lindsay, J. G. (1983). Ferritin as a label for high-gradient magnetic separation.
Biophys J 42, 145-150.
Papaefthymiou, G., C. (2010). The Mössbauer and magnetic properties of ferritin cores. Biochim
Biophys Acta 1800, 886-897.
Purushotham, S., and Ramanujan, R. V. (2010). Modeling the performance of magnetic
nanoparticles in multimodal cancer therapy. Journal of Applied Physics 107, 114701.
Qin, S., Yin, H., Yang, C., Dou, Y., Liu, Z., Zhang, P., Yu, H., Huang, Y., Feng, J., Hao, J., Hao,
J., Deng, L., Yan, X., Dong, X., Zhao, Z., Jiang, T., Wang, H. W., Luo, S. J., and Xie, C.
(2016). A magnetic protein biocompass. Nat Mater 15, 217-226.
Ranade, S. S., Qiu, Z., Woo, S. H., Hur, S. S., Murthy, S. E., Cahalan, S. M., Xu, J., Mathur, J.,
Bandell, M., Coste, B., Li, Y. S., Chien, S., and Patapoutian, A. (2014). Piezo1, a
mechanically activated ion channel, is required for vascular development in mice. Proc Natl
Acad Sci U S A 111, 10347-10352.
Ritz, T., Ahmad, M., Mouritsen, H., Wiltschko, R., and Wiltschko, W. (2010). Photoreceptor-
based magnetoreception: optimal design of receptor molecules, cells, and neuronal
processing. J R Soc Interface 7 Suppl 2, S135-46.
Schenck, J. F., Dumoulin, C. L., Redington, R. W., Kressel, H. Y., Elliott, R. T., and McDougall,
I. L. (1992). Human exposure to 4.0-Tesla magnetic fields in a whole-body scanner. Med
Phys 19, 1089-1098.
Schoffa, G. (1965). Der Antiferromagnetismus des Ferritins bei Messungen der magnetischen
im Temperaturbereich von 4,2 bis 300 degrees K. Zeitschrift für
Suszeptibilitat
Naturforschung B 20, 167-172.
Soffe, R., Baratchi, S., Tang, S. Y., Nasabi, M., McIntyre, P., Mitchell, A., and Khoshmanesh, K.
(2015). Analysing calcium signalling of cells under high shear flows using discontinuous
dielectrophoresis. Sci Rep 5, 11973.
16
Stanley, S. A., Sauer, J., Kane, R. S., Dordick, J. S., and Friedman, J. M. (2015). Remote
regulation of glucose homeostasis in mice using genetically encoded nanoparticles. Nat Med
21, 92-98.
Wheeler, M. A., Smith, C. J., Ottolini, M., Barker, B. S., Purohit, A. M., Grippo, R. M.,
Gaykema, R. P., Spano, A. J., Beenhakker, M. P., Kucenas, S., Patel, M. K., Deppmann, C.
D., and Guler, A. D. (2016). Genetically targeted magnetic control of the nervous system.
Nat Neurosci 19, 756-761.
Zborowski, M., Fuh, C. B., Green, R., Baldwin, N. J., Reddy, S., Douglas, T., Mann, S., and
Chalmers, J. J. (1996). Immunomagnetic isolation of magnetoferritin-labeled cells in a
modified ferrograph. Cytometry 24, 251-259.
Acknowledgements: The author thanks Bill Bialek, Justin Bois, Stephen J. Royle, and an
anonymous contributor on PubPeer for helpful comments on an earlier version.
Competing interests: No competing interests were disclosed.
Grant information: The author declared that no grants were involved in supporting this work.
17
|
1711.06967 | 4 | 1711 | 2017-11-24T22:30:51 | Neural correlates of flow using auditory evoked potential suppression | [
"q-bio.NC"
] | "Flow" is a hyper-engaged state of consciousness most commonly described in athletics, popularly termed "being in the zone." Quantitative research into flow has been hampered by the disruptive nature of gathering subjective reports. Here we show that a passive probe (suppression of Auditory Evoked Potential in EEG) that allowed our participants to remain engaged in a first-person shooting game while we continually tracked the depth of their immersion corresponded with the participants' subjective experiences, and with their objective performance levels. Comparing this time-varying record of flow against the overall EEG record, we identified neural correlates of flow in the anterior cingulate cortex and the temporal pole. These areas displayed increased beta band activity, mutual connectivity, and feedback connectivity with primary motor cortex. These results corroborate the notion that the flow state is an objective and quantifiable state of consciousness, which we identify and characterize across subjective, behavioral and neural measures. | q-bio.NC | q-bio | Neural correlates of flow using auditory evoked potential suppression
Kyongsik Yun1,2,3*, Saeran Doh4*, Elisa Carrus5, Daw-An Wu1,2, Shinsuke Shimojo1,2,6
1Computation and Neural Systems, California Institute of Technology, Pasadena, CA 91125,
USA
2Division of Biology, California Institute of Technology, Pasadena, CA 91125, USA
3BBB Technologies Inc., Seoul, South Korea
4Department of Food Business Management, School of Food, Agricultural and Environmental
Sciences, Miyagi University, Miyagi, Japan
5Division of Psychology, School of Applied Sciences, London South Bank University,
London UK
6Japan Science and Technology Agency, Saitama, Japan
*These authors contributed equally to this work.
Correspondence should be addressed to K.Y. [email protected]
1
Abstract
"Flow" is a hyper-engaged state of consciousness most commonly described in athletics,
popularly termed "being in the zone." Quantitative research into flow has been hampered by
the disruptive nature of gathering subjective reports. Here we show that a passive probe
(suppression of Auditory Evoked Potential in EEG) that allowed our participants to remain
engaged in a first-person shooting game while we continually tracked the depth of their
immersion corresponded with the participants' subjective experiences, and with their
objective performance levels. Comparing this time-varying record of flow against the overall
EEG record, we identified neural correlates of flow in the anterior cingulate cortex and the
temporal pole. These areas displayed increased beta band activity, mutual connectivity, and
feedback connectivity with primary motor cortex. These results corroborate the notion that
the flow state is an objective and quantifiable state of consciousness, which we identify and
characterize across subjective, behavioral and neural measures.
Keywords: flow; auditory evoked potential; EEG; anterior cingulate cortex; temporal pole
2
Introduction
"Flow" is a mental state of full immersion into an activity in which one is intensively
engaged, accompanied with a feeling of extreme concentration, full control, achievement and
pleasure in the activity 1. In popular culture, this state is often described in professional sports
as "being in the zone." According to Csikszentmihalyi (1990), flow represents the perfect
experience of control of emotion and cognition during performance and learning. Verbal
protocols he gathered from experts skilled in various activities, mainly via interview, indicate
that the flow state can be experienced during online gaming as well as in various other
activities, such as sports, education, singing, dancing, climbing, and even surgery.
Flow research has received wide interest in the field of marketing, mainly through
the study of Computer Mediated Environments (CME), as a process of optimal experience
during internet use2. It has been found that flow can create compelling consumer online
experiences and can be fun and pleasurable; however, besides marketing research on flow,
very little is known with regards to quantifiable measures of this experience, such as the
neural dynamics underpinning this state.
Investigating the neural correlates of flow requires that the timecourse of neural data
be matched to information about the states of flow during that time period. One challenge to
gathering this information is that flow requires extended, uninterrupted engagement in an
activity. Thus, investigations generally use post-experiment questionnaires to gather data
that accompany flow experiences, this means that flow data lack timing information-they
are usually blanket ratings of the entire performance period.
Several previous studies have gathered neural data from participants experiencing
flow states while performing tasks 3,4, but they all possess this same limitation. Because they
do not track the dynamics of flow in a way that can be aligned to the neural data, they are
3
limited to contrasting data from different blocks of play, in which they have manipulated task
conditions to either increase or decrease the likelihood of entering the flow state. These
studies have provided many candidate regions and neural signatures for the flow state, but the
methodology cannot dissociate between internal flow states and the external conditions that
they set up to induce flow.
A study by Klasen et al. (2011) created a timecourse to align with the neural data by
recording the participants' gameplay. They analyzed the game videos for events and content
that would support or inhibit flow, and aligned them with the fMRI data. By finding brain
areas that correlated with each content class and then calculating their overlap, they identified
a cerebellar-somatosensory cortex network as a likely substrate for flow. The use of fine-scale
timecourse in this study was a clear advantage, but here too the analysis actually correlated
the brain activity to external game conditions rather than to data about internal states. Thus it
is unclear whether this somato-motor brain network reflects the higher levels of action in the
content-defined gameplay epochs, or the (inferred) higher frequency of flow experience
during those epochs.
Here, we isolate the internal phenomenon of flow by correlating neural data directly
with timecourses of internal states. These timecourses are based on both subjective and
objective measures. Participants provided subjective data via a guided parametric version of
the retropective Think Aloud design 5. While they reviewed videos of their gameplay, they
rated their experience of flow in each time segment. Objective data regarding internal states
was obtained by a novel EEG probing method, which we validate and then apply here.
The EEG-based probe design is based on "telepresence" as a signature feature of the
flow experience during online gaming. This refers to the gameplayer's tendency to feel like
they "exist" in the game world, which leads to a neglect of sensory stimulation from the real
4
world. The player becomes more sensitive to game-relevant sensory stimuli in the virtual
word, and less sensitive to game-irrelevant, real-world stimuli 6. Accordingly, we hypothesize
that when subjects are in the flow state, they will tend to neglect sounds that aren't strictly
relevant to the game. Thus, we measure the AEPs (Auditory Evoked Potentials) elicited by
game-irrelevant, random beeping sounds7 inserted throughout gameplay. Attention-based
modulation of AEP amplitude is an extremely robust, and well-established effect in the
evoked-related potential literature8. For example, a previous study showed that the AEPs
elicited by a faint clicking sound were prominent when subjects counted them, but then
became suppressed when subjects read a book intently9. Thus, we expect the AEP to be
suppressed during immersion in the flow state. A recent study applied an auditory oddball as
a secondary task to objectively quantify flow while playing video games 10. Our AEP
suppression has advantages in that it is completely passive and does not explicitly interfere
with the subjects' cognitive processes during the flow experience.
It must be noted, however, that the AEP suppression alone is not sufficient to define
the flow state during game play, because it directly captures only some of the aspects of flow,
such as intense, focused concentration, and perhaps telepresence indirectly. Subjective
behavioral reporting on the experience of flow is also neither sufficient nor reliable in
defining flow. It was for this reason that we decided to define the flow state by a correlation
of objective and subjective measures, assessing flow state by comparing the AEP amplitude
suppression with the respective post-gameplay ratings of flow for each participant.
The present study aims: 1) to validate AEP suppression as a measure of the flow state,
thereby providing a passive probe that can be used to periodically measure the participant's
flow state without disrupting their engagement in the task. 2) to use the measured timecourse
of flow to localize its neural correlates, and to characterize functional connectivity between
5
those brain regions. By doing this, we would like to characterize the dynamics of flow in a
quantitative manner, and link it to objective neural substrates.
6
Methods
Participants
Twenty-nine healthy subjects participated in this study (24 males, 5 females, age: 23.5±3.4
years (mean±SD)). We recruited subjects on the basis of previous experience with video
games in general (average 13.6±10.1 hours/week, range 7~49 hours/week) with ads posted on
the main campus of the California Institute of Technology. All participants provided written
informed consent after receiving a detailed explanation of the experimental procedures. The
Institutional Review Board of California Institute of Technology approved all experimental
procedures and this study was carried out in accordance with the approved guidelines.
Participants were excluded if they had a history of neurological disorder such as seizure,
stroke, head injury, or a substance use disorder other than caffeine or nicotine. We obtained a
set of self-reported flow questionnaires 11 at the end of the experiment.
Task
The participants played a FPS video game (Call of Duty : Modern Warfare 2, Activision).
During pre-experiment game training, we adjusted the difficulty of the game depending on
the skill of each participant so that it was right above the player's skills. During the
experimental session, participants played the game for an hour, consisting of a 30min low
challenge and a 30min high challenge game scenario (Figure 1). During the game play,
randomly distributed beep sounds (inter-beep interval=1~120sec: duration=200ms; 40dB)
were presented via speakers to the participants, in order to measure auditory evoked
potentials.
7
Behavioral analysis
While participants were playing, we recorded the game play using Fraps, a realtime video
capture software (www.fraps.com). This recording was used in the review session following
game play. Here, participants were asked to review their own game play by replaying a video
of their own game session and responding whether they experienced flow or not for each 5
min time period of the video. The 5 min time period was chosen based on the previous
studies of neurofeedback and subjective behavioral ratings 12,13. Player performance (i.e.,
number of kills – deaths) was tracked across the same time bins of game play. Flow
experience and performance distribution histograms were compared using Pearson's
correlation. For additional skill-based analysis, fifteen highest performing participants were
grouped as "experts" and the others as "beginners."
Insert figure 1.
Electrophysiological recording and analysis
We recorded EEG activity from 128 scalp electrodes (EGI System 200; Electrical Geodesics,
Eugene, OR.) during gameplay. Electrode impedance was kept under 40 kΩ for all
recordings14. Electrode nets were covered with a shower cap to prevent electrodes from
drying so that we could maintain low electrode impedance. Vertical and horizontal ocular
movements were also recorded. The EEG was continuously recorded at a 500 Hz sampling
frequency and filtered (high pass 0.1Hz, low pass 200Hz, notch filter 60Hz and 120Hz). The
EEGs were segmented from −500ms to 1000ms relative to the onset of each beep. Ocular
artifact reduction was performed using ICA component rejection in EEGLAB 15.
8
Following pre-processing, we computed the evoked activity aligned to each task-
irrelevant beep sound, and this was done to evaluate epochs with and without AEP
suppression. Consequently, we separated each epoch as flow and non-flow groups based on
the participants' ratings as well as AEP suppression (Table 1) and computed each epoch's
source localization and partial directed coherence values to define the network of regions
involved and effective connectivity within it. The time window used for source localization
and partial directed coherence was 1000ms following the beep.
Time-frequency analysis
The epoched data were analyzed by means of a event-related spectral perturbation (ERSP) 16
(window length, 250 ms; step, 25 ms; window overlap, 90%). This was done to select the
epochs that showed AEP suppression. We removed the baseline of 500 ms preceding the
beep onset with duration of 500 ms from time frequency ERSP charts. We used a
nonparametric permutation test with 5000 randomizations for comparisons between the
activation and baseline, corrected for multiple comparisons 17.The corrected threshold was set
to p < 0.05, and the time-frequency representation only shows statistically significant results.
sLORETA
Standardized low resolution brain electromagnetic tomography (sLORETA) 18 was used for
source localization. Thousands of synchronized postsynaptic potentials from pyramidal
neurons of the cortex produce scalp EEG activity 19. sLORETA computes the three
dimensional localization of these activities. The subject-specific 3D coordinates of the 128
electrode positions were estimated (Geodesic Photogrammetry System; Electrical Geodesics,
Eugene, OR.) and applied to a digitized MRI version of the Talairach Atlas (McConnell Brain
9
Imaging Centre, Montréal Neurological Institute, McGill University). These Talairach
coordinates were then used to compute the sLORETA transformation matrix for each
participant. Following the transformation to an average reference, the EEG activity of the
flow and non-flow epochs selected from the above time frequency analysis was used to
calculate cross spectra in sLORETA for each participant. We used delta (1~4Hz), theta
(4~8Hz), alpha (8~12Hz), and beta (12~30Hz) frequency bands for the following analyses.
Using the sLORETA transformation matrix, the cross spectra of each participant and
frequency band were then transformed into sLORETA files. These files included the 3D
cortical distribution of the electrical neuronal generators for each participant. The computed
sLORETA image displayed the cortical neuronal oscillators in 6239 voxels, with a spatial
resolution of 5 mm 20. We used a nonparametric permutation test with 5000 randomizations
for comparisons between the flow and non-flow17. The threshold was set to p < 0.01.
Partial directed coherence
Partial directed coherence (PDC) is a form of frequency-domain Granger-causality, which
quantifies the direction of information transfer between brain regions 21,22. The EEG data
from ROIs defined by source localization were windowed in 500-sample long intervals (i.e.,
1000ms in length). The PDC values were evaluated in the delta (1~4Hz), theta (4~8Hz),
alpha (8~12Hz), and beta (12~30Hz) ranges. Model order (i.e., time delay of the
autoregressive parameters) was set to 15, based on previous studies for sufficient frequency
resolution 23,24. We used the nonparametric bootstrap approach for further statistical analyses
25,26. For details of this method, see Snijders and Borgatti 26. The threshold was set to p <
0.001, considering Bonferroni correction for multiple comparisions (3 ROIs, 3X3=9
comparisons). Each arrow pointing from the region i to its target region j represents a
10
significant PDC. SPSS (Windows version 15.0; SPSS, Inc., Chicago, IL) was used for the
statistical analyses.
11
Results
Mean duration of flow was 8.31±3.61 min (13.9% of total time). 53.0% of the flow
experience occurred between the 25 and 45 minute marks of the total one-hour game session.
A significant positive correlation was found between the overall occurrence of the flow
experience and the performance distribution throughout the game play (Pearson's correlation,
R=0.68, p=0.014) (Figure 2). These results indicate the close relationship between subjective
flow experience and performance, suggesting that the higher the experience of flow, the
better the gaming performance.
Insert Figure 2.
Next we tested AEPs as a neural marker for the flow state. Since we did not observe
the conventional clear AEP waveform, due to insufficient number of trials and background
noise from the game play, we analysed the ERSPs of the signal at low frequencies (1-30Hz).
In EEG analysis, the classical ERP model is limited to study complex brain dynamics because
the trial-by-trial frequency components tend to be averaged out by this method. ERSP has
been adapted for this reason27-30. Given the low signal-to-noise ratio, we used ERSPs, which
represent the power of oscillations of the ERP. This was done because evoked potentials
predominantly contain low-frequency information, and by doing this we would still be able to
evaluate the suppression of the auditory evoked potential in the absence of a clear AEP.
Following established protocols, ERSPs were calculated for each tone at the averaged
channels Fz and Pz8,31. We compared the EEG ERSP of the post-stimulus window to the
baseline (500~0ms before the beep onset) and separated epochs with significant post-stimulus
activation from those showing deactivation at averaged. Out of the epochs showing
12
suppressed signal, we categorised them as flow segments only if they were also rated as flow
by participants. The self-reported non-flow trials showed larger evoked potential activation
than the flow trials (t(1861)=3.84, p=0.0001) (Figure 3 and Figure S1). We also found a
significant correlation between suppressed evoked potential and self-reported experience of
flow (Table S1. Chi-square test; Χ2(1)=97.0, p<0.001).
Insert Figure 3.
Following the analysis of the auditory evoked potential, we source localized the
epochs corresponding to the flow state and those corresponding to the non-flow state at time
window of 0~1000ms after the stimulus. This was done to estimate the possible sources of
the activity that gives rise to flow experience, and therefore to find the brain areas associated
with flow. Figure 4 depicts localization results, which showed that the anterior cingulate
cortex (ACC: MNI x=-6, y=25, z=21, Figure 4a); and temporal pole (TP: x=-55, y=10, z=-25,
Figure 4b) were significantly activated in the flow state as compared to the non-flow state
only in the beta frequency range. We also found significant beta frequency oscillations in the
primary motor cortex (precentral gyrus (M1); x=15, y=-20, z=70; Figure 4c) in the beginners
compared to experts.
Insert Figure 4.
To examine the link between these three brain regions and the experience of flow, we
computed correlations between the subjective behavioral flow ratings and beta activity . We
found significant positive correlations between the flow ratings and activation in ACC
13
(Pearson's correlation; R2=0.22, P=0.017) and TP (R2=0.26, P=0.009) (Figure 5A and B) and
a significant negative correlation with activation in M1 (R2=0.32, P=0.003) (Figure 5C).
These correlation results therefore suggest that ACC and TP are the core areas associated with
the flow state.
Insert Figure 5.
To further understand the network properties of the neural correlates of flow, we analyzed
effective connectivity using PDC across the same three regions of interest. We found that the
PDC connectivity increased during the flow compared to non-flow state. More specifically,
the top-down connectivity increased (ACC->M1, TP->M1) and bottom-up connectivity
decreased (M1->TP) during the flow state (nonparametric jackknife procedure, p < 0.001)
compared to non-flow state (Figure 6).
Insert Figure 6.
14
Discussion
The current study aimed to generate the mental state of "flow" in the laboratory and to
quantify and to characterize it using a novel combination of objective and subjective data,
namely the suppression of AEP, and video-assisted self-reports. By correlating the timecourse
of flow with the EEG data, we localized the potential neural correlates of flow and
characterized effective connectivity between these brain areas. Specifically, the ACC and TP
emerged as important hubs associated with the experience of flow, showing increased beta
frequency power and increased effective connectivity patterns during the flow state compared
to non-flow.
There are several psychological criteria for flow 32, including (1) intense and focused
concentration, (2) a loss of self-consciousness, and (3) a feeling of control and ease of
performance. For example, when a game player experiences flow, his/her attention is
completely focused on the game character and the opponent, and he/she ignores any task-
irrelevant stimuli. The actions he/she performs are perceived as if they were his/her own,
rather than in the virtual world. We therefore exploited this notion by evaluating modulation
of the AEP amplitude in flow and non-flow. We found significantly suppressed neural activity
following the beep when in flow, compared to non-flow in the oscillatory evoked activity,
which represents the power of oscillations of the ERP. This result suggests that when
participants experienced flow, their brains shut off task-irrelevant stimuli in the literal sense
(even at the sensory level). Through the use of the AEP, we were able to quantify the flow
state without disrupting the participant's engagement with the game.
Our source localization results are consistent with the flow characteristics mentioned
above. Activation of the ACC known to be involved in the processing attention and focus 33,34,
which is in line with the first criterion of flow. The TP, an empathy-related brain region, was
15
also activated and we interpret this in light of the second characteristic of flow, a loss of self-
consciousness and telepresence. The TP has been known to be crucial in the process of
distinguishing self vs. other as well as in perspective-taking in social affective processes35,36.
Lastly, our current results show a pattern of negative correlation between motor activity and
the flow experience. We also found that the motor activity decreased in experts compared to
beginners. These results are consistent with previous studies, which have found that motor
activity decreases when motor behavior is processed more efficiently37,38, and that experts
develop a focused and efficient organization of neural networks39. This accords with the third
characteristic of flow, which is a feeling of control and ease of performance.
Our effective connectivity results showed a top-down connectivity from the ACC and
the TP to M1 in the flow state. Enhanced top-down processing has been known to be
consistent with heightened performance and concentration40,41. However, our top-down
connectivity results are different from the typical top-down executive control represented by
explicit (conscious) cognitive process of the prefrontal cortex. Rather, ACC and TP
performed as implicit (unconscious) cognitive and social processes, including extreme focus,
empathy, and telepresence. In summary, we argue that the ACC and the TP are the two main
regions that may provide the neural basis of the flow state, or "being in the zone".
Our study took several measures to maximize the likelihood of inducing a deep flow
state. First, we used a first-person-shooter video game known to be very addictive42, and our
participants were all experienced players, many of whom routinely devoted large amounts of
time to gameplay. This allowed us to recreate a state of flow in the laboratory in players who
had spontaneously experienced flow in the past. The laboratory setting was comfortable, and
provided a typical computer game playing environment. This contrasts with an MRI
environment, where players are prone, immobilized and in a confined space. Flow is less
16
likely to occur if the environment is unfamiliar and/or uncomfortable43. Most importantly, we
allowed the participants to play the game continuously for an hour. According to our results,
87.5% of the participants require at least 25min to get into the flow state (Figure 2). This
suggests that the relative strength of flow experiences in previous neuroscientific studies may
be limited, as individual blocks of their experiments range from 3-12 minutes.
Lastly, we
implemented an objective electrophysiological evaluation (AEP
suppression) to test whether participants were in flow and cross-validated these results
through subjective behavioral flow ratings. This technique allowed us to track the strength of
flow and build a timeline of the flow experience. Previous studies have been limited in their
characterization of flow: they tended to quantify objective external features such as game
performance (killing vs. being killed, active fighting vs. boring scenes in the case of shooting
games; easy vs. hard levels in general). Although there is a positive correlation between such
external criteria and the flow experience, using them as substitutes for direct measures of
flow causes clear confounds in the analysis of neural correlates. We argue that our cross-
validation method (across neural, behavioral and subjective measures) captures the actual
emergence of flow in the closest possible way to the original definition and the
phenomenology of the state.
One limitation of our study is related to our use of EEG, in that it cannot capture the
activity of deep brain structures, such as amygdala and midbrain reward circuits. These
regions have been known to be closely related with empathy, attention, motivation, and
reward, which are highly probable to be correlated with flow experience44. Future studies
using different techniques are necessary to pursue such targets. For example a combination of
EEG and fMRI could be used to simultaneously measure the AEPs and image deep brain
areas.
17
We should also mention the limitation that we did not resolve a clear AEP waveform,
probably because our experimental setting was ill-suited to extracting classical ERPs through
trial averaging. Rather than having hundreds of regularly repeated trials, we had a relatively
small number of beeps randomly inserted into the dynamic and irregular auditory game
environment. However, anlaysis in the time-frequency domain revealed clear patterns of
evoked oscillatory activity, showing large positive components of the AEP phase-locked to
the beep onset and suppression of those components during the flow state.
Techniques of detection and quantification of the flow state can contribute not only
to neuroscience of the altered state of consciousness, but also to game content design. The
behavioral characteristics of the flow experience, including its mean duration, temporal
location, and the relationship with gaming performance, would be very useful for effective,
immersive game design as well as marketing research on internet shopping, etc. Degrees of
the feeling of being in the zone have been known to be highly associated with amount of
information search in game players and purchasing intention among internet shoppers45. Thus,
whether our technique can be generalized to other situations of flow, particularly in the
domain of marketing, would be an intriguing question to address in future. Thus all together,
the flow is a psychological and neurophysiological phenomenon which can be uniquely
defined, and feasibly assessed by a combination of subjective and objective techniques as
demonstrated in the current study. Whether or not the neural correlates and the neural
measure are applicable to the flow state in other kinds of context, such as sports, gambling,
singing, dancing, internet surfing/shopping, etc. would be an obvious next step of future
research.
18
References
1
2
3
4
Csikszentmihalyi, M. The domain of creativity. Theories of creativity 4, 61-91 (1990).
Hoffman, D. L. & Novak, T. P. Marketing in hypermedia computer-mediated
environments: conceptual foundations. The Journal of Marketing, 50-68 (1996).
Plotnikov, A. et al. in Advanced Learning Technologies (ICALT), 2012 IEEE 12th
International Conference on. 688-689 (IEEE).
Berta, R., Bellotti, F., De Gloria, A., Pranantha, D. & Schatten, C.
Electroencephalogram and physiological signal analysis for assessing flow in games.
Computational Intelligence and AI in Games, IEEE Transactions on 5, 164-175
(2013).
5
Van Someren, M. W., Barnard, Y. F. & Sandberg, J. A. The think aloud method: A
practical guide to modelling cognitive processes. Vol. 2 (Academic Press London,
1994).
6
7
8
9
Cowley, B., Charles, D., Black, M. & Hickey, R. Toward an understanding of flow in
video games. Computers in Entertainment (CIE) 6, 20 (2008).
Allison, B. Z. & Polich, J. Workload assessment of computer gaming using a single-
stimulus event-related potential paradigm. Biol. Psychol. 77, 277-283 (2008).
Picton, T. & Hillyard, S. Human auditory evoked potentials. II: Effects of attention.
Electroencephalogr. Clin. Neurophysiol. 36, 191-200 (1974).
Picton, T., Hillyard, S., Krausz, H. & Galambos, R. Human auditory evoked potentials:
I. Evaluation of components. Electroencephalography & Clinical Neurophysiology;
Electroencephalography & Clinical Neurophysiology (1974).
10
Nuñez Castellar, E. P., Antons, J.-N., Marinazzo, D. & van Looy, J. Being in the zone:
19
Using behavioral and EEG recordings for the indirect assessment of flow. PeerJ
Preprints 4, e2482v2481 (2016).
11
Csikszentmihalyi, M. Toward a psychology of optimal experience. Review of
personality and social psychology 2, 13-36 (1982).
12
Prinsloo, G. E., Derman, W. E., Lambert, M. I. & Rauch, H. L. The effect of a single
session of short duration biofeedback-induced deep breathing on measures of heart
rate variability during laboratory-induced cognitive stress: A pilot study. Appl.
Psychophysiol. Biofeedback 38, 81-90 (2013).
13
Potteiger, J. A., Schroeder, J. M. & Goff, K. L. Influence of music on ratings of
perceived exertion during 20 minutes of moderate intensity exercise. Percept. Mot.
Skills 91, 848-854 (2000).
14
Ferree, T. C., Luu, P., Russell, G. S. & Tucker, D. M. Scalp electrode impedance,
infection risk, and EEG data quality. Clin. Neurophysiol. 112, 536-544 (2001).
15
Delorme, A. & Makeig, S. EEGLAB: an open source toolbox for analysis of single-
trial EEG dynamics including independent component analysis. J. Neurosci. Methods
134, 9-21 (2004).
16
Pfurtscheller, G. & Lopes da Silva, F. H. Event-related EEG/MEG synchronization
and desynchronization: basic principles. Clin. Neurophysiol. 110, 1842-1857 (1999).
17
Nichols, T. E. & Holmes, A. P. Nonparametric permutation tests for functional
neuroimaging: a primer with examples. Hum. Brain Mapp. 15, 1-25 (2002).
18
Pascual-Marqui, R. D., Esslen, M., Kochi, K. & Lehmann, D. Functional imaging
with low-resolution brain electromagnetic tomography (LORETA): a review. Methods
Find. Exp. Clin. Pharmacol. 24, 91-95 (2002).
19 Martin, J. H. The collective electrical behavior of cortical neurons:
the
20
electroencephalogram and the mechanisms of epilepsy. Principles of Neural Science.
4th ed. New York: McGraw-Hill, Health Professions Division, 777-791 (2000).
20
Pascual-Marqui, R. D. Standardized low resolution brain electromagnetic tomography
(sLORETA): technical details. Methods Find. Exp. Clin. Pharmacol. 24, 5-12 (2002).
21
Baccal, L. A. & Sameshima, K. Partial directed coherence: a new concept in neural
structure determination. Biol. Cybern. 84, 463-474 (2001).
22
Baccal, L. A. & Sameshima, K. Comments on 'Is partial coherence a viable technique
for identifying generators of neural oscillations?'. Biol. Cybern. 95, 135-141 (2006).
23
Brovelli, A. et al. Beta oscillations in a large-scale sensorimotor cortical network:
Directional influences revealed by Granger causality. Proceedings of the National
Academy of Sciences 101, 9849-9854 (2004).
24
Supp, G. G., Schlogl, A., Trujillo-Barreto, N., Muller, M. M. & Gruber, T. Directed
cortical information flow during human object recognition: analyzing induced EEG
gamma-band responses in brain's source space. PLoS One 2 (2007).
25
Efron, B. Nonparametric estimates of standard error: the jackknife, the bootstrap and
other methods. Biometrika 68, 589-599 (1981).
26
Snijders, T. A. B. & Borgatti, S. P. Non-parametric standard errors and tests for
network statistics. Connections 22, 61-70 (1999).
27 Makeig, S. Auditory event-related dynamics of the EEG spectrum and effects of
exposure to tones. Electroencephalogr. Clin. Neurophysiol. 86, 283-293 (1993).
28
Grandchamp, R. & Delorme, A. Single-trial normalization for event-related spectral
decomposition reduces sensitivity to noisy trials. Front Psychol 2, 236 (2011).
29
Rossi, A., Parada, F. J., Kolchinsky, A. & Puce, A. Neural correlates of apparent
motion perception of impoverished facial stimuli: a comparison of ERP and ERSP
21
activity. Neuroimage 98, 442-459 (2014).
30
Engell, A. D., Huettel, S. & McCarthy, G. The fMRI BOLD signal tracks
electrophysiological spectral perturbations, not event-related potentials. Neuroimage
59, 2600-2606 (2012).
31
Hegerl, U. & Frodl-Bauch, T. Dipole source analysis of P300 component of the
auditory evoked potential: a methodological advance? Psychiatry Research:
Neuroimaging 74, 109-118 (1997).
32
Nakamura, J. & Csikszentmihalyi, M. Flow theory and research. Oxford handbook of
positive psychology, 195-206 (2009).
33
Cohen, R., Kaplan, R., Moser, D., Jenkins, M. & Wilkinson, H. Impairments of
attention after cingulotomy. Neurology 53, 819-819 (1999).
34
Lane, R. D. et al. Neural correlates of levels of emotional awareness: evidence of an
interaction between emotion and attention in the anterior cingulate cortex. J. Cogn.
Neurosci. 10, 525-535 (1998).
35
Ruby, P. & Decety, J. How would you feel versus how do you think she would feel? A
neuroimaging study of perspective-taking with social emotions. J. Cogn. Neurosci. 16,
988-999 (2004).
36
Yun, K., Watanabe, K. & Shimojo, S. Interpersonal body and neural synchronization
as a marker of
implicit
social
interaction. Scientific Reports 2, 959,
doi:doi:10.1038/srep00959 (2012).
37 Mazaheri, A., Nieuwenhuis, I. L. C., van Dijk, H. & Jensen, O. Prestimulus alpha and
mu activity predicts failure to inhibit motor responses. Hum. Brain Mapp. 30, 1791-
1800 (2009).
38
Henson, R. & Rugg, M. Neural response suppression, haemodynamic repetition
22
effects, and behavioural priming. Neuropsychologia 41, 263-270 (2003).
39 Milton, J., Solodkin, A., Hluštík, P. & Small, S. L. The mind of expert motor
performance is cool and focused. Neuroimage 35, 804-813 (2007).
40
Engel, A. K., Fries, P. & Singer, W. Dynamic predictions: oscillations and synchrony
in top-down processing. Nature Reviews Neuroscience 2, 704-716 (2001).
41
Pessoa, L., Kastner, S. & Ungerleider, L. G. Neuroimaging studies of attention: from
modulation of sensory processing to top-down control. The Journal of neuroscience
23, 3990-3998 (2003).
42 Montag, C. et al. Internet addiction and personality in first-person-shooter video
gamers. Journal of Media Psychology: Theories, Methods, and Applications 23, 163
(2011).
43
Csikszentmihalyi, M. Finding flow: The psychology of engagement with everyday life.
(Basic Books, 1998).
44
Klasen, M., Weber, R., Kircher, T. T. J., Mathiak, K. A. & Mathiak, K. Neural
contributions to flow experience during video game playing. Social cognitive and
affective neuroscience 7, 485-495 (2012).
45
Doh, S. Flow Construct: Its Mediating Roles in an On-line Search Model. 11th
International Marketing Trends Conference proceedings (2012).
23
Acknowledgments
This work was supported by JST-ERATO, JST-CREST, Tamagawa-Caltech gCOE programs,
Grant-in-Aid for Scientific Research (Kakenhi) in Japan, and Basic Science Research
Program through the National Research Foundation of Korea (NRF) funded by the Ministry
of Education, Science and Technology (2013R1A6A3A03020772).
Author Contributions
K.Y., S.D., E.C., and S.S. designed and performed the experiments. K.Y. analyzed the data.
K.Y., S.D., E.C., and S.S. wrote the manuscript.
Additional information
Competing financial interests: The authors declare no competing financial interests.
24
Figure Legends
Figure 1. Experimental procedure.
Participants played a first-person shooter video game for an hour, consisting of 30min low
challenge and 30min high challenge. Randomly distributed beep sounds (inter-beep
interval=1~120sec) were presented to participants via speakers during the game.
Figure 2. Histograms of (A) the flow distribution and (B) the performance distribution across
time of game play. Performance was defined as the mean number of "kills" minus "deaths"
across all participants (R=0.68, p=0.014).
Figure 3. Time-frequency analyses of auditory evoked activity during (A) flow and (B) non-
flow states. The non-flow epochs showed significant evoked potential activation and the flow
trials showed significant evoked potential suppression compared to the baseline (500~0ms
before the beep onset) (average of channels Fz and Pz; nonparametric permutation test,
p<0.05). The vertical line represents the onset of the beep.
Figure 4. (A) Source localization showing the contrast of flow state minus non-flow state for
(A) the anterior cingulate cortex (ACC) (x=-6, y=25, z=21), and (B) temporal pole (TP) (x=-
55, y=10, z=-25). (C) the flow minus non-flow contrast for beginners minus experts in the
precentral gyrus (M1) (x=15, y=-20, z=70) (all: non-parametric permutation test, red: p<0.05,
yellow: p<0.01).
Figure 5. Correlation between the behavioral flow ratings and source localized EEG activity.
For the (A) temporal pole (TP) (R2=0.26, P=0.009), (B) anterior cingulate cortex (ACC)
25
(R2=0.22, P=0.017), and (C) primary motor cortex (M1) (R2=0.32, P=0.003).
Figure 6. Effective connectivity using PDC between regions of interest, including anterior
cingulate cortex (ACC), temporal pole (TP), and primary motor (M1) cortex. These regions
were selected based on a priori source localized activity in Figure 4. Red arrows indicate the
contrast for PDC connectivity of flow with non-flow, and the blue arrows indicate the
contrast of non-flow with flow.
26
Figure 1
27
Figure 2
28
Figure 3
29
Figure 4
30
Figure 5
31
Figure 6
32
Supplementary Information
Neural correlates of flow using auditory evoked potential suppression
Kyongsik Yun1,2,3*, Saeran Doh4*, Elisa Carrus5, Daw-An Wu1,2, Shinsuke Shimojo1,2,6
1Computation and Neural Systems, California Institute of Technology, Pasadena, CA 91125, USA
2Division of Biology, California Institute of Technology, Pasadena, CA 91125, USA
3BBB Technologies Inc., Seoul, South Korea
4Department of Food Business Management, School of Food, Agricultural and Environmental Sciences, Miyagi
University, Miyagi, Japan
5Division of Psychology, School of Applied Sciences, London South Bank University, London UK
6Japan Science and Technology Agency, Saitama, Japan
*These authors contributed equally to this work.
Correspondence should be addressed to K.Y. [email protected]
S1 Table. Percentage of epochs with and without suppressed AEP rated as flow and non-flow
Flow rating
Non-flow rating
AEP suppression
No AEP suppression
26.1%
14.6%
18.7%
40.6%
Epochs rated as flow showed significantly larger number of suppressed AEP compared with
the AEP of epochs rated as non-flow (chi-square test; Χ2(1)=97.0, p<0.001).
Table S1. Venn diagram of the percentage distribution of all the trials
S1 Figure. Peak auditory evoked potential (0~500ms) in non-flow and flow state. The non-
flow trials showed larger evoked potential activation than the flow trials (t(1861)=3.84,
p=0.0001).
|
1907.13004 | 2 | 1907 | 2019-12-27T15:41:37 | Visual illusions via neural dynamics: Wilson-Cowan-type models and the efficient representation principle | [
"q-bio.NC"
] | In this work we have aimed to reproduce supra-threshold perception phenomena, specifically visual illusions, with Wilson-Cowan-type models of neuronal dynamics. We have found that it is indeed possible to do so, but that the ability to replicate visual illusions is related to how well the neural activity equations comply with the efficient representation principle. Our first contribution is to show that the Wilson-Cowan equations can reproduce a number of brightness and orientation-dependent illusions, and that the latter type of illusions require that the neuronal dynamics equations consider explicitly the orientation, as expected. Then, we formally prove that there can't be an energy functional that the Wilson-Cowan equations are minimizing, but that a slight modification makes them variational and yields a model that is consistent with the efficient representation principle. Finally, we show that this new model provides a better reproduction of visual illusions than the original Wilson-Cowan formulation. | q-bio.NC | q-bio |
Visual illusions via neural dynamics: Wilson-Cowan-type models
and the efficient representation principle
DTIC, Universitat Pompeu Fabra, Barcelona, Spain
Marcelo Bertalm´ıo
[email protected]
Luca Calatroni
Universit´e Cote d'Azur, CNRS, INRIA,
Laboratoire d'Informatique, Signaux et Syst`emes de Sophia Antipolis, France
[email protected]
Valentina Franceschi
IMO, Universit´e Paris-Sud, Orsay, France
[email protected]
Benedetta Franceschiello
FAA, LINE, Radiology, CHUV, Lausanne, Switzerland
[email protected]
Alexander Gomez-Villa
DTIC, Universitat Pompeu Fabra, Barcelona, Spain
[email protected]
Dario Prandi
Universit´e Paris-Saclay, CNRS, CentraleSuplec,
Laboratoire des signaux et syst`emes, Gif-sur-Yvette, France
[email protected]
Abstract
We reproduce supra-threshold perception phenomena, specifically visual illusions, by Wilson-
Cowan-type models of neuronal dynamics. Our findings show that the ability to replicate the
illusions considered is related to how well the neural activity equations comply with the efficient
representation principle. Our first contribution consists in showing that the Wilson-Cowan
(WC) equations can reproduce a number of brightness and orientation-dependent illusions.
Then, we formally prove that there can't be an energy functional that the Wilson-Cowan dy-
namics are minimizing. This leads us to consider an alternative, variational modelling which
has been previously employed for local histogram equalization (LHE) tasks. In order to adapt
our model to the architecture of V1, we perform an extension that has an explicit dependence
on local image orientation. Finally, we report several numerical experiments showing that LHE
provides a better reproduction of visual illusions than the original WC formulation and that its
cortical extension is capable to reproduce also complex orientation-dependent illusions.
New & Noteworthy: We show that the Wilson-Cowan equations can reproduce a number
of brightness and orientation-dependent illusions. Then, we formally prove that there can't be
an energy functional that the Wilson-Cowan equations are minimizing, making them sub-optimal
with respect to the efficient representation principle. We thus propose a slight modification that is
consistent with such principle and show that this provides a better reproduction of visual illusions
than the original Wilson-Cowan formulation. We also consider the cortical extension of both
models in order to deal with more complex orientation-dependent illusions.
1
1 Introduction
The goal of this work is to point out the intimate connections existing between three popular
approaches in vision science: the Wilson-Cowan equations, the study of visual brightness illusions,
and the efficient representation theory.
As other articles in this special issue make abundantly clear, Wilson-Cowan equations have a
long and successful story of modelling cortical low-level dynamics [Cowan et al., 2016]. Nonetheless,
the study of psychophysics by Wilson-Cowan equations ([Adini et al., 1997, Herzog et al., 2003,
Bertalm´ıo and Cowan, 2009, Ernst et al., 2016, Bertalm´ıo et al., 2017, Wilson, 2003, Wilson,
2007, Wilson, 2017]) is a topic that hasn't been addressed much in neuroscience, and we are not
aware of publications in which Wilson-Cowan equations are used for predicting brightness illusions.
In this work, we aim to fill this gap.
The study of visual illusions has always been key in the vision science community, as the
mismatches between reality and perception provide insights that can be very useful to develop
new models of visual perception [Kingdom, 2011] or of neural activity [Eagleman, 1959, Murray
and Herrmann, 2013], and also to validate the existing ones. It is commonly accepted that visual
illusions arise due to neurobiological constraints [Purves et al., 2008] that modify the underpinned
mechanisms of the visual system.
The efficient representation principle, introduced by Attneave [Attneave, 1954] and Barlow
[Barlow et al., 1961], states that neural responses aim to overcome these neurobiological constraints
and to optimize the limited biological resources by being tailored to the statistics of the images that
the individual typically encounters, so that visual information can be encoded in the most efficient
way. This principle is a general strategy observed across mammalian, amphibian and insect species
[Smirnakis et al., 1997] and is embodied by neural processing according to abundant experimental
evidence [Fairhall et al., 2001, Mante et al., 2005, Benucci et al., 2013].
Our work aims at pulling together the three approaches just mentioned, providing a more unified
framework to understand vision mechanisms. First, we show that the Wilson-Cowan equations are
able to qualitatively reproduce a number of visual illusions. Secondly, we formally prove that
Wilson-Cowan equations (with constant input) are not variational, in the sense that they are not
minimizing any energy functional. Next, we detail how a simple modification turning the Wilson-
Cowan equations variational yields a local histogram equalisation method that is consistent with
the efficient representation principle. We finally show how this new formulation provides a better
reproduction of visual illusions than the Wilson-Cowan model.
We remark that our model has to be intended as a proof of concept, whose objective is the
reproduction of perceptual phenomena at a macroscopic level with no quantitative assessment on
analogous psychophysical data. There are in fact very important limitations for doing that, since
such comparison would require both a perfect knowledge of how behavioural data were collected,
and a tuning of the model parameters to match with the observed perception. Nonetheless, we
believe that the numerical evidence of our experiments and our theoretical considerations can be
used for future research studies comparing our computational results with the ones corresponding
to experiments coming from psychophysics.
2 Materials and methods
2.1 Visual illusions
Computational models able to reproduce visual illusions represent very effective methods to test
new hypotheses and generate new insights, both for neuroscience and applied disciplines such as
image processing. Illusions can be classified according to the main feature detection mechanisms
involved during the visual process [Shapiro and Todorovic, 2016]. In this contribution we consid-
ered two main groups of visual illusions to assess the efficacy of our model in reconstructing the
perceptual process: brightness illusions and orientation-dependent illusions.
2
2.1.1 Brightness illusions
Brightness illusions are a class of phenomena where image regions with the same gray level are
perceived as having different brightness, depending on the shapes, arrangement and gray level of
the surrounding elements. Fig. 1 shows the nine brightness illusions we have chosen to perform
tests on in this paper. They are all very popular and at the same time they represent a diverse
set, as we can see from the following descriptions.
White's illusion:
identical [White, 1979] (Fig. 1(a)).
the left gray rectangle appears darker than the right one, while both are
Simultaneous brightness contrast:
while both are identical [Brucke, 1865] (Fig. 1(b)).
the left gray square appears lighter than the right one,
Checkerboard illusion:
in the seventh column, while both are identical [DeValois and DeValois, 1990] (Fig. 1(c)).
the mid-gray square in the fifth column appears darker than the one
Chevreul illusion:
a pattern of homogeneous bands of increasing intensity from left to right
is presented. However, the bands in the image are perceived as inhomogeneous, i.e. darker and
brighter lines appear at the borders between adjacent bands [Ratliff, 1965] (Fig. 1(d)).
Chevreul cancellation: when the order of the bands is reversed, now decreasing in intensity
from left to right, the effect is cancelled [Geier and Hud´ak, 2011] (Fig. 1(e)).
Dungeon illusion:
two gray rectangles are perceived as darker or lighter depending on the gray
intensities of both the background and the grid, see [Bressan, 2001]. The left rectangle is perceived
as darker than the one on the right (Fig. 1(f)).
Grating induction:
the background grating (which can be tuned to different orientations) in-
duces the appearance of a counter-phase grating in the homogeneous gray horizontal bar [McCourt,
1982] (Fig. 1(g)).
Hong-Shevell illusion:
right, while both are identical [Hong and Shevell, 2004] (Fig. 1(h)).
the mid-gray half-ring on the left appears darker than the one on the
Luminance illusion:
four identical dots over a background where intensity increases from left
to right, and the dots on the left are perceived being lighter than the ones on the right [Kitaoka,
2006] (Fig. 1(i)).
2.1.2 Orientation-dependent illusions
We also consider orientation-dependent illusions, where the perceptual phenomenon (e.g. in terms
of brightness or contrast) is affected by the orientation of the image elements.
Poggendorff illusion. The Poggendorff illusion, presented in the modified version considered in
this work in Fig. 2(a), is a very well known geometrical optical illusion in which the presence of a
central surface induces a misalignment of the background lines. This illusion depends both on the
orientation of the background lines and the width of the central surface [Weintraub and Krantz,
1971], as the more the angle is close to π/2 the less is the bias, but in this example the perceived
bias is also dependent on the brightness contrast between central surface and background lines.
3
(a) White
(b) Brightness con-
trast
(c) Checkerboard
(d) Chevreul
(e) Chevreul cancel-
lation
(f) Dungeon
(g) Grating induc-
tion
(h) Hong-Shevell
(i) Luminance
Figure 1: From left to right, top to bottom: White's illusion, Brightness contrast, the Checkerboard
illusion, the Chevreul illusion, Chevreul cancellation, the Dungeon illusion, the Grating induction,
the Hong-Shevell illusion and the Luminance illusion.
Tilt illusion. The Tilt illusion is a phenomenon where the perceived orientation of a test line
or grating is altered by the presence of surrounding lines or a grating with a different orientation.
In our case we consider the effect that the orientation of a surround grating pattern has on the
perceived contrast of a grating pattern in the center: the inner circles in Figs. 2(b) and 2(c) are
identical but the latter is perceived as having more contrast than the former.
(a) Poggendorff illusion.
(b) Tilt illusion, same θ.
(c) Tilt illusion, different θ.
Figure 2: From left to right: a modified version of the Poggendorff illusion based on Grating In-
duction, a modified Tilt illusion with concentric circles having the same orientation and a modified
Tilt illusion with concentric circles having different orientations.
2.2 Wilson-Cowan-type models for contrast perception
In this section we introduce four different evolution equations derived from the Wilson-Cowan
formulation, that will be studied in this paper. We recall that, denoting by a(x, t) the state of a
population of neurons with spatial coordinates x ∈ R2 at time t > 0, the Wilson-Cowan equations
4
proposed in [Wilson and Cowan, 1972, Wilson and Cowan, 1973] can be written1 as
(cid:90)
ω(x(cid:107)y)σ(a(y, t)) dy + h(x),
(2.1)
a(x, t) = −βa(x, t) + ν
∂
∂t
R2
where β > 0 and ν ∈ R are fixed parameters, σ : R → R is a non-linear sigmoid saturation function,
the kernel ω(x(cid:107)y) models interactions at two different spatial locations x and y (we will assume
that the integral of ω is normalised to 1) and h is the input signal.
2.2.1 Wilson-Cowan equations do not fulfill any variational principle
Over the last thirty years, the use of variational methods in imaging has become increasingly
popular as a regularisation strategy for solving general ill-posed imaging problems in the form
find u s.t.
f = T(u).
(2.2)
Here, f represents a given degraded image and T a (possibly non-linear) operator describing the
degradation (e.g. noise, blur, under-sampling, etc.)
Due to the lack of fundamental properties such as existence, uniqueness and stability of the
solution of the problem (2.2), the idea of regularisation consists of incorporating a priori informa-
tion on the desired image u(cid:63) and on its closeness to the data f by means of suitable variational
terms. This gives rise, in particular, to variational methods where one looks for an approximation
u(cid:63) of the real solution u by solving
u(cid:63) = arg min E(u),
(2.3)
where E is the energy functional combining regularisation and data fit, depending also on the given
image f . A popular way to solve the variational problem consists in finding u(cid:63) as the steady-state
solution of the evolution equation given by the gradient descent of the energy functional
u = −∇E(u),
∂
∂t
ut=0 = f,
(2.4)
under appropriate conditions on the boundary of the image domain.
In the context of vision science, evolution equations have been originally used as a tool to
describe the physical transmission, diffusion and interaction phenomena of stimuli in the visual
cortex [Beurle and Matthews, 1956, Wilson and Cowan, 1972, Wilson and Cowan, 1973]. Varia-
tional methods are the main tool of ecological approaches, that pose the efficient coding problem
[Olshausen and Field, 2000] as an optimisation problem to be solved with evolution equations
that minimise an energy functional [Atick, 1992] involving natural image statistics and biological
constraints. The resulting solution is optimal because it has minimal redundancy.
However, we must remark that, while considering the gradient descent of an energy functional
gives always an evolution equation, the reverse is not true: not every evolution equation is minimis-
ing an energy functional. In fact, this is the case for the Wilson-Cowan equations, which do not
fulfil any variational principle, as we prove in Appendix A. As a consequence, they are sub-optimal
in reducing the redundancy.
We remark that it is possible to define an energy that decreases along trajectories of (2.1), as
done in [French, 2004]. This ensures in particular that even though the evolution is not variational,
its steady states (i.e., solutions of (2.1) that are constant in time) can indeed be obtained as critical
points of this energy.
1In [Wilson and Cowan, 1972] the sigmoid function is applied outside of the integral term and not only on
the activity a(y, t) as in (2.1). This corresponds to an "activity-based" model of neuron activation, while (2.1)
corresponds to a "voltage-based" one. See [Faugeras, 2009], where the two models are shown to be equivalent.
5
2.2.2 A modification of the Wilson-Cowan equations complying with efficient repre-
sentation
Remarkably, the efficient representation principle has correctly predicted a number of neural pro-
cessing aspects and phenomena like the photoreceptor response performing histogram equalisation,
the dominant features of the receptive fields of retinal ganglion cells (lateral inhibition, the switch
from bandpass to lowpass filtering when the illumination decreases, and, remarkably, colour op-
ponency, with photoreceptor signals being highly correlated but color opponent signals having
quite low correlation), or the receptive fields of cortical cells having a Gabor function form [Atick,
1992, Daugman, 1985, Olshausen and Field, 2000]. Efficient representation is the only framework
able to predict the functional properties of neurons from a simple principle, and given how simple
the assumptions are it's really surprising that this approach works so well [Meister and Berry,
1999].
In [Bertalm´ıo and Cowan, 2009] it is shown how a slight modification of the Wilson-Cowan
formulation leads to a variational model, as we now present. Assuming that the activity signal a
is in the range [0, 1], we can re-write equation (2.1) in terms of a sigmoid σ shifted by 1
2 (which we
take as the average signal value) and inverted in sign, thus getting:
(cid:18)
(cid:19)
a(x, t) = −βa(x, t) − ν
∂
∂t
R2
ω(x(cid:107)y)σ
a(y, t) − 1
2
dy + h(x).
(2.5)
(cid:90)
(cid:90)
Note that this is just a re-writing of equation (2.1), so it is still not associated to any variational
method. However, if we now assume σ to be odd and replace the 1
2 term by a(x, t), we obtain
a(x, t) = −βa(x, t) + ν
∂
∂t
R2
ω(x(cid:107)y)σ(a(x, t) − a(y, t)) dy + h(x),
(2.6)
and this equation is now a gradient descent equation, as it does fulfil a variational principle.
Furthermore, under the proper choice of parameters β, ν and input signal h, this evolution
equation performs local histogram equalisation (LHE) [Bertalm´ıo et al., 2007]. This is key for
our purposes, since, as Atick points out [Atick, 1992], one of the main types of redundancy or
inefficiency in an information system like the visual system happens when some neural response
levels are used more frequently than others, and for this type of redundancy the optimal code is
the one that performs histogram equalisation.
It is therefore expected that the modification of the Wilson-Cowan equations in (2.6), which
better complies with the efficient representation principle, should be more effective in reducing
redundancy than the original Wilson-Cowan model of equation (2.1).
2.2.3 Accounting for orientation
Models (2.1) and (2.6) ignore orientation and as such they are not well-suited to explain a number
of visual phenomena. For this reason, following [Bertalm´ıo et al., 2019a], we extend them to a
third dimension, representing local image orientation, as follows. We let La : Q× [0, π) → R be the
cortical activation in V1 associated with the signal a, so that La(x, θ) encodes the response of the
neuron with spatial preference x and orientation preference θ to a. Mathematically, such activation
is obtained via a suitable convolution with the receptive profiles of V1 neurons, as explained in
Appendix B, see also [Duits and Franken, 2010, Petitot, 2017, Prandi and Gauthier, 2017, Citti
and Sarti, 2006, Sarti and Citti, 2015]. Then, denoting by A(x, θ, t) the cortical response at time t
for any t > 0, the natural extension of equations (2.1) and (2.6) to the orientation dependent case
is given by the two models:
(cid:16)
(cid:90) π
(cid:90)
(cid:90)
ω(x, θ(cid:107)y, φ)σ(cid:0)A(x, θ, t)−A(y, φ, t)(cid:1) dy dφ + Lh(x, θ), (2.8)
ω(x, θ(cid:107)y, φ)σ
dy dφ + Lh(x, θ),
A(y, φ, t)
(cid:17)
(2.7)
0
Q
(cid:90) π
A(x, θ, t) = −βA(x, θ, t) + ν
∂
∂t
A(x, θ, t) = −βA(x, θ, t) + ν
∂
∂t
0
Q
6
where Lh(x, θ) denotes the cortical activation in V1 corresponding to the visual input h at spatial
location x and orientation preference θ. We remark that these models describe the dynamic
behaviour of activations in the 3D space of positions and orientation. As explained in Appendix B,
once a stationary solution is found, the two-dimensional perceived image can be found by simply
applying the formula
A(x, θ) dθ.
(2.9)
(cid:90) π
0
1
π
a(x) =
2.2.4 Models under consideration
We summarise here the four models we are going to test in the following sections. The orientation-
independent WC and LHE models are:
a(x, t) = −(1 + λ)a(x, t) +
a(x, t) = −(1 + λ)a(x, t) +
∂
∂t
∂
∂t
1
2M
1
2M
ω(x, y)σ (a(y, t)) dy + λf0(x) + µ(x)
(WC-2D)
ω(x, y)σ (a(x, t) − a(y, t)) dy + λf0(x) + µ(x),
(LHE-2D)
(cid:90)
(cid:90)
Q
Q
which relate to (2.1) and (2.6) by simply choosing parameters as β = 1 + λ and ν = 1/2M where
M > 0 is a normalisation constant, and input signal h(x) = λf0(x)+µ(x), where λ > 0, f0(x) is the
local intensity at x ∈ Q of given image f0 and µ(x) denotes a local average of the initial stimulus
f0 around x (a choice motivated by the averaging behaviour of cells in the magnocellular pathway
[?] and already considered in similar models e.g. [Bertalm´ıo et al., 2007, Bertalm´ıo, 2014]).
The orientation-dependent WC and LHE models can be similarly written as:
A(x, θ, t) = − (1 + λ)A(x, θ, t) +
∂
∂t
1
2M
A(x, θ, t) = − (1 + λ)A(x, θ, t) +
+ λLf0(x, θ) + Lµ(x, θ),
1
2M
∂
∂t
+ λLf0(x, θ) + Lµ(x, θ),
ω(x, θy, φ)σ (A(y, φ, t)) dy dφ
ω(x, θy, φ)σ (A(x, θ, t) − A(y, φ, t)) dy dφ
(WC-3D)
(LHE-3D)
(cid:90) π
(cid:90) π
0
(cid:90)
(cid:90)
Q
0
Q
which can analogously be related to (2.7) and (2.8) by choosing the very same parameters as
above and by now taking as cortical activation in V1 corresponding to h the quantity Lh(x, θ) =
λLf0(x, θ) + Lµ(x, θ).
2.2.5 Numerical implementation
All four relevant equations (WC-2D), (LHE-2D), (WC-3D), and (LHE-3D) are numerically imple-
mented via a forward Euler time-discretisation, as presented in [Bertalm´ıo et al., 2007]. For a given
image a, the cortical activation La is recovered via standard wavelet transform methods, as pre-
sented in [Bertalm´ıo et al., 2019a] (see also [Duits and Franken, 2010]). The codes, written in Julia
[Bezanson et al., 2017], are available at the following link: http://www.github.com/dprn/WCvsLHE.
All the considered images are of size 200 × 200 pixels, and take values in the interval [.15, .85]
in order to avoid out-of-range issues. We always consider K = 30 discretised orientations, as
done in [Boscain et al., 2018] for instance. As presented in Appendix B, the receptive profiles
associated to the discretised orientations selected are obtained via cake wavelets [Bekkers et al.,
2014], for which the frequency band bw is set to bw= 5. The interaction kernel is taken to be a 2D
or 3D Gaussian with standard deviation σω, the local mean average µ is obtained via Gaussian
filtering with standard deviation σµ.
In our experiments we used the following two piece-wise
linear functions as sigmoids:
(cid:19)
(cid:18)
x − 1
2
,
(2.10)
σ(ρ) := min{1, max{αρ,−1}},
σ(ρ) := −σ
7
with α = 5, see Figure 3. Note that σ, which will be used for LHE models, is odd and centered in
zero while σ, which will be used for WC models, is shifted in 1/2 and shows a reversed behaviour.
This in fact corresponds to a change of sign in the integral terms of LHE models w.r.t. the WC
ones, as discussed in Section 2.2.2.
(a) σ and the line y = x
(b) σ and the line y = −x + 1
2
Figure 3: Sigmoid functions in the form (2.10), with α = 5, as considered in our experiments.
Finally, the evolution stops when the L2 relative distance between two successive iterations
is smaller than a tolerance τ = 10−2, which identifies convergence of the iterates to a stationary
state.
3 Results
In this section, we present the results obtained by applying the four models described above to
the visual illusions described in Section 2.1. Our objective is to understand the capability of these
models to replicate the visual illusions under consideration. That is, we are interested in whether
the output produced by the models qualitatively agrees with the human perception of the phenom-
ena. We stress that our study is purely qualitative as it has to be intended as a proof of concept
showing how Wilson-Cowan-type dynamics can be effectively used to replicate the perceptual ef-
fects due to the observation of visual illusions. We do not address here the match with empirical
data since those depend on several experimental conditions for which a correspondence with the
model parameters is not clear. A dedicated study on experiments motivated by psychophysics,
addressing the validation of our models and, possibly, allowing for the creation of ground-truth
references for a quantitative assessment is left for future research.
Due to the lack of a universal metric adapted to the task of assessing the replication of visual
illusions, we will evaluate replication or lack thereof by presenting relevant line profiles, i.e., plots of
brightness levels along a single row (line), of images produced by the four models in consideration (a
common tool used by several brightness/lightness/color models before [Blakeslee et al., 2016, Otazu
et al., 2008]). These lines are chosen as to cross a section of the image called target: A gray region
in the image (or set of regions in the case of the Chevreul illusion), where the brightness illusion
appears.
In all the results shown in this section, the original visual stimulus profile is represented as a
blue dashed line. The line profiles of the output models are represented as solid red (LHE-2D),
green (WC-2D), magenta (LHE-3D), and cyan (WC-3D) lines.
The parameters appearing in the models have been chosen independently for each illusion and
each model, in order to obtain the best possible replication of the visual illusion. Here, by best-
replication we mean that the extracted line-profiles correctly mimic the perceptual outcome from
a qualitative point of view.The chosen parameters are presented in Table 1.
8
WC-2D
LHE-2D
WC-3D
LHE-3D
Illusion
White
Brightness
Checkerboard
Chevreul
Chevreul canc.
Dungeon
Gratings
Hong-Shevell
Luminance
Poggendorff
Tilt
σµ σω
20
10
10
2
10
70
5
2
2
2
10
6
2
6
20
5
6
2
λ M σµ σω
50
.7
10
.7
.7
70
10
.7
3
.9
40
.7
.7
6
.5
.7
6
.7
1.4
1.4
1.4
1
1
1.4
1
1
1
10
2
10
2
5
5
2
5
2
λ M σµ σω
30
.7
10
.7
.7
70
40
.7
20
.9
50
.7
.7
6
30
.7
6
.7
20
2
10
2
2
2
2
10
2
1
1
1
1
1
1
1
1
1
λ M σµ σω
50
.7
10
.7
.7
70
7
.5
3
.5
50
.7
.7
6
30
.7
6
.7
10
20
1.4
1.4
1.4
1
1.4
1.4
1
1
1
2
2
10
5
5
5
2
10
2
3
15
λ M
1
.7
1
.7
.7
1
1
.7
1
.9
1
.7
.7
1
1
.7
1
.7
.5
1
1
.7
Table 1: Parameters used in the tests.
3.1 Orientation-independent brightness illusions
Table 1 summarises the replication results obtained for the illusions described in Section 2.1:
if the model replicates the illusion we indicate in the table the used parameters, otherwise a
cross () denotes no replication, i.e. the failure of the model to reproduce computational results
corresponding to the visual perception of the considered illusion.
White's illusion. The chosen line profile for the plots in Fig. 4 corresponds to the central
horizontal line of the image, which crosses both gray patches. As both plots show, all four models
correctly predict the left target to be darker than the right one.
Figure 4: Predicted brightness in White illusion
Simultaneous brightness contrast. The plots in Fig. 5 show the line profiles of the central
horizontal line of the image, which crosses the two gray squares. We see that our four models
replicate this illusion (left square lighter than the right square). In both the 2D and the 3D case,
we observe that LHE methods result in an enhanced contrast effect w.r.t. WC methods.
Checkerboard illusion. The chosen line profiles for this illusion are the two horizontal lines
crossing, respectively, the left gray target and the right one. In Fig. 6, we chose to plot the first
half of the line profile corresponding to the left target and the second half of the one corresponding
to the right target. The profiles of all the four models show replication of this illusion, by which
the left target is perceived darker than the right one.
9
01002000.10.50.9Visual stimulusLHE-2DWC-2DLHE-3DWC-3DPredicted brightness 2D models0100200Predicted brightness 3D modelsvisualstimulus Figure 5: Predicted brightness in simultaneous brightness contrast
Figure 6: Predicted brightness in Checkerboard illusion
Chevreul illusion. Fig. 7 presents the line profiles for the central horizontal line.All four models
correctly replicate the perceived changes within each band.
Figure 7: Predicted brightness in Chevreul illusion
.
Chevreul cancellation. The line profiles for the central horizontal line are presented in Fig. 8.
In this case all models are able to correctly replicate the effect, although in the case of (WC-2D)
and (LHE-3D) the perceptual response is not perfect, due to the presence of some oscillations.
We also remark that the correct replication of this illusion is extremely sensitive to the chosen
parameters.
10
0.10.50.901002000100200Visual stimulusLHE-2DWC-2DLHE-3DWC-3DvisualstimulusPredicted brightness 2D modelsPredicted brightness 3D models0.10.50.94010016040100160visualstimulusVisual stimulusLHE-2DWC-2DLHE-3DWC-3DPredicted brightness 2D modelsPredicted brightness 3D models0.30.50.7Visual stimulusLHE-2DWC-2DLHE-3DWC-3D1001604010016040visualstimulusPredicted brightness 2D modelsPredicted brightness 3D models Figure 8: Predicted brightness in Chevreul cancellation
Dungeon illusion. Profiles of the central section (3 middle squares) of each target are shown
in Fig. 9. The first part of the plot (left to right) represents the profile of the rectangle on black
background. The second plot shows the target on white background. As these profiles show,
our four proposed models replicate human perception (first target is predicted as darker than
the second). Nevertheless, the assimilation effect (target intensity goes towards surrounding) is
stronger in the 3D models.
Figure 9: Predicted brightness in Dungeon illusion
Grating induction.
In Fig. 10 the continuous and dashed blue lines respectively show the profile
of the grating and of the central horizontal line row of the visual stimulus. Then, the line profiles
of the central horizontal line of the outputs have been plotted. We observe that for both 2D and
3D models a counter-phase grating appears in the middle row, which successfully coincides with
human perception. Notice that LHE methods have a higher amplitude in both cases.
Hong-Shevell illusion. Fig. 11 shows the line profiles of the central horizontal line around the
target (gray ring) neighbourhood rings in the first half of the image. As in the case of the Dungeon
illusion, we present in the first half of the plot (left to right) the output of the first stimulus (light
background) and in the second half the output of the second (dark background). We see how our
four proposed models replicate the assimilation effect. Hence, the gray ring in the first image is
predicted as brighter than the gray ring in the second visual stimulus.
11
40100160visualstimulusPredicted brightness 2D modelsPredicted brightness 3D modelsVisual stimulusLHE-2DWC-2DLHE-3DWC-3D0.30.70.540100160508050805080Visual stimulusLHE-2DWC-2DLHE-3DWC-3D0.20.50.95080Predicted brightness 2D modelsPredicted brightness 3D modelsvisualstimulus Figure 10: Predicted brightness in grating induction
Figure 11: Predicted brightness in Hong-Shevell illusion
Luminance illusion. Horizontal profiles crossing top left and right targets (gray circles) are
depicted in Fig. 12. For each target our four models reconstruct the left target as brighter than
the right one. Hence, all models correctly predict this contrast effect. In this case, LHE presents
a higher contrast response in both responses (2D and 3D).
Figure 12: Predicted brightness in luminance gradient illusion
We observe that in all the considered brightness illusions both the 3D methods present neighbourhood-
dependent oscillations.
12
0.10.50.9Predictedbrightness 2D modelsPredicted brightness 3D models01002000100200visualstimulusVisual stimulusLHE-2DWC-2DLHE-3DWC-3DVisual stimulusgrating stripevisualstimulus60306030Visual stimulusLHE-2DWC-2DLHE-3DWC-3DPredicted brightness 2D modelsPredicted brightness 3Dmodels0.10.50.9603060300.20.50.8Predicted brightness 2D modelsPredicted brightness 3D models20205015018050150180visualstimulusVisual stimulusLHE-2DWC-2DLHE-3DWC-3D 3.2 Orientation-dependent illusions
Poggendorff illusion. The output images and a zoom of the target gray middle area are pre-
sented in Fig. 13.
In this case (WC-2D), (WC-3D), and (LHE-2D) are not able to completely
replicate the illusion, since induced white lines on the gray area are not connected. On the other
hand, (LHE-3D) successfully replicates the perceptual completion over the gray middle stripe.
Figure 13: Zoom of the predicted completion for Poggendorff illusion
Tilt illusion.
In Fig. 14 we present line profiles, for both visual stimuli, for a diagonal line
starting at the bottom left corner of the image and ending at the top right one. In order to be able
to correctly compare the two images, the line profile of the second image (from top to bottom) has
been extracted after flipping the outer circle along the vertical axis, so that the responses to both
stimulus have the same background. Although there is a noticeable effect, such as a reduction in
contrast for the (WC-2D), the difference between the responses to the two stimuli is very mild for
all models with the exception of (LHE-3D).
The fact that indeed this model is replicating the effect can be better appreciated looking at
Fig. 15, which shows a composite of the inner circle for the responses to the two visual stimuli
of the two orientation-dependent models. It is then evident that the (LHE-3D) model yields a
stronger result than the (WC-3D) one. In fact, the former shows increased visibility (measured
here as the contrast) for the half of the circle corresponding to the second stimulus than the one
corresponding to the first stimulus. On the other hand, in the case of the (WC-3D) model (or of
2D models, not depicted here), the circle shows no difference among its two halves. This justifies
our claim that the (LHE-3D) model can increase the visibility of the inner circle (replicate the
illusion) based on the orientation of the outer circle.
Figure 14: Predicted brightness in Tilt illusion
13
LHE-2DWC-2DLHE-3DWC-3D5025050250Visual stimulusLHE-2DWC-2DLHE-3DWC-3D010.55025050250visualstimulusPredicted brightness2D modelsPredicted brightness 3D models Figure 15: Detail in predicted brightness in Tilt illusion
4 Discussion
The results presented in the previous section show that the four models are able to reproduce several
brightness illusions. Concerning orientation-dependent illusions we observe that, as expected, 2D
models cannot reproduce them, while the only 3D model that correctly reproduces the perceptual
outcome is the (LHE-3D). However we stress that determining replication or lack thereof in the
Tilt illusion is subtle, as the observed effects are very mild.
As already mentioned, the parameters of the presented results are chosen independently from
one illusion to the other in order to qualitatively optimise the perceptual replication in terms
of suitable line profiles. Empirical observations show that the value of the model parameters
involved are indeed related with the size of the target and the spatial frequency of the background.
Nevertheless, if one settles for milder replications, it would be possible to choose more uniform
parameters. For instance, this happens for the (WC-3D) model in the Chevreul and Chevreul
cancellation illusions, which can be reproduced simultaneously with parameters σµ = 3 and σω =
30, although with less striking results.
Regarding the 3D models, we want to point out that we have chosen to use K = 30 orientations
whereas this number commonly takes values in the 12-18 range in the literature (e.g. [Scholl et al.,
2013, Chariker et al., 2016, Pattadkal et al., 2018]). Our selection of 30 orientations is motivated
by some preliminary tests (which we are not presenting here) showing that a coarser orientation
discretisation seems not to be sufficient to reproduce most of the orientation-dependent illusions.
As future research we will test whether or not a different selection of parameters allows to reproduce
those illusions with less orientations, but we should also mention that some works in the literature
actually use a high number of orientations in cortical models (e.g. 64 orientations in [Teich and
Qian, 2010]).
Finally, we notice that the output of 3D models often shows oscillations. For some illusions
(white and dungeon), the (WC-3D) model produces more oscillatory solutions than (LHE-3D),
and for others (Chevreul brightness, grating induction, and luminace gradient), the (LHE-3D)
have stronger oscillations than (WC-3D). The relation between the model parameters and possible
dependence of the target surrounding is a matter of future research.
14
5 Conclusions and future work
We consider Wilson-Cowan-type models describing neuronal dynamics and apply them to the study
of replication of brightness visual illusions.
We show that Wilson-Cowan equations are able to replicate a number of brightness illusions and
that their variational modification, accounting for changes in the local contrast and performing local
histogram equalisation, outperforms them. We consider also extensions of both models accounting
for explicit local orientation dependence, in agreement with the architecture of V1. Although
in the case of pure brightness illusions we found no real advantage in considering models taking
into account orientations, these turned out to be necessary for the replication of two exemplary
orientation-dependent illusions, which only the 3D LHE variational model is able to reproduce.
In order to understand and fully exploit the potential of the orientation-dependent LHE model,
further research should be done. In particular, a more accurate modelling reflecting the actual
structure of V1 should be addressed. This concerns first the lift operation, where the cake wavelet
should be replaced by the more physiologically plausible Gabor filters, as well as the interaction
weight ω which could be taken to be the anisotropic heat kernel of [Citti and Sarti, 2006, Sarti and
Citti, 2015, Duits and Franken, 2010]. The design of appropriate psychophysics experiments testing
the visual illusions considered in this work and their match with our models' outputs is clearly a
further important research direction, which would turn our qualitative study into a quantitative
one. The problem of matching computational models of perception with psychophysical data is in
fact not trivial, but necessary to provide insights about how visual perception works and to identify
the computational parameters able to reproduce the perceptual bias induced by these phenomena.
Acknowledgements and Grants
M. B. would like to thank the organizers of the conference to celebrate Jack Cowan's 50 years at the
University of Chicago for their kind invitation to attend that meeting, which served as inspiration
for this work, and also acknowledges the support of the European Unions Horizon 2020 research and
innovation programme under grant agreement number 761544 (project HDR4EU) and under grant
agreement number 780470 (project SAUCE), and of the Spanish government and FEDER Fund,
grant ref. PGC2018-099651-B-I00 (MCIU/AEI/FEDER, UE). L. C., V. F. and D. P. acknowledge
the support of a public grant overseen by the French National Research Agency (ANR) as part
of the Investissement d'avenir program, through the iCODE project funded by the IDEX Paris-
Saclay, ANR-11-IDEX-0003-02 and of the research project LiftME funded by INS2I, CNRS. L. C.
and V. F. acknowledge the support provided by the Fondation Math´ematique Jacques Hadamard.
V. F. acknowledges the support received from the European Union's Horizon 2020 research and
innovation programme under the Marie Sk(cid:32)lodowska-Curie grant No 794592. V. F. and D. P. also
acknowledge the support of ANR-15-CE40-0018 project SRGI - Sub-Riemannian Geometry and
Interactions. B. F. acknowledges the support of the Fondation Asile des Aveugles.
Disclosures
All authors declare that there is no commercial relationship relevant to the subject matter of
presentation.
Author contributions
All authors equally contributed to this work.
15
A Non-variational nature of Wilson-Cowan equation
In this section we show that, for non-trivial choices of weight and sigmoid functions, Wilson-Cowan
equations do not admit a variational formulation.
equations, with constant input. Namely, for a : R → Rn, we consider
For the sake of simplicity, we will consider only a finite dimensional variant of Wilson-Cowan
d
dt
a(t) = −µa(t) + W σ(a(t)) + h.
(A.1)
Here, h ∈ Rn is the input, µ > 0 is a parameter, σ ∈ C1(R) is any function (we denoted σ(v) =
(σ(vi))i for v ∈ Rn), and W ∈ Rn×n is a symmetric interaction kernel. For a proof in the infinite-
dimensional setting we refer to [Bertalm´ıo et al., 2019b]
associated with a functional J : Rn → R, i.e.,
Equation (A.1) admits a variational formulation if it can be written as the steepest descent
We have the following.
a(t) = −∇J(a(t)).
d
dt
(A.2)
Theorem. The Wilson-Cowan equation (A.1) admits the variational formulation (A.2) only if
either W is a diagonal matrix, or σ is an affine function, i.e., σ(x) = αx + β for some α, β ∈ R.
Proof. Writing (A.1) and (A.2) componentwise, we find the following relation for J:
W(cid:96),kσ(v(cid:96)) − hi,
v = (v1, . . . , vn) ∈ Rn,
i = 1, . . . , n.
By differentiating again the above, and letting δij denote the Kroenecker delta symbol, we have
W(cid:96),kσ(cid:48)(v(cid:96))δj(cid:96) = µδij − Wijσ(cid:48)(vj),
i, j = 1, . . . , n.
(A.3)
Namely, Hess J(v) = (µδij − Wijσ(cid:48)(vj))ij. Assume that W is not a diagonal matrix. Then, since
both the Hessian matrix and W are symmetric, by choosing i (cid:54)= j such that Wij (cid:54)= 0 we get
k
σ(cid:48)(vi) = σ(cid:48)(vj)
∀v ∈ Rn.
(A.4)
This clearly implies that σ(cid:48) is constant, thus showing that σ must be an affine function.
We observe that the above reasoning does not apply to the LHE algorithm. Indeed, the discrete
Wi(cid:96)σ(cid:0)ai(t) − a(cid:96)(t)(cid:1) + h.
(cid:88)
(cid:96)
form of the latter is
d
dt
a(t) = −µa(t) +
Then, the corresponding variational equation (for µ = 0 and h = 0) is
v ∈ Rn.
Wi(cid:96)σ(vi − v(cid:96)),
∂iJ(v) = −(cid:88)
(cid:96)(cid:54)=i
(LHE)
(A.5)
∂iJ(v) = µvi −(cid:88)
∂ijJ(v) = µδij −(cid:88)
k
This yields
∂jiJ(v) = Wijσ(cid:48)(vi − vj),
for v ∈ Rn,
i (cid:54)= j.
(A.6)
This does not contradict the symmetry of the Hessian, as σ was chosen to be odd an thus σ(cid:48) is
even. Indeed, we know by [Bertalm´ıo and Cowan, 2009] that we can let
Wk(cid:96)Σ(vk − v(cid:96)),
(A.7)
J(v) :=
where Σ is such that Σ(cid:48) = σ.
(cid:88)
k,(cid:96)
16
B Encoding orientation-dependence via cortical-inspired models
Orientation dependence of the visual stimulus is encoded via cortical inspired techniques, following
e.g., [Citti and Sarti, 2006, Duits and Franken, 2010, Petitot, 2017, Prandi and Gauthier, 2017, Bohi
et al., 2017]. The main idea at the base of these works goes back to the 1959 paper [Hubel and
Wiesel, 1959] by Hubel and Wiesel (Nobel prize in 1981) who discovered the so-called hypercolumn
functional architecture of the visual cortex V1. Following [Hubel and Wiesel, 1959], each neuron ξ
in V1 detects couples (x, θ) where x ∈ R2 is a retinal position and θ is a direction at x. Orientation
preferences θ are then organised in hypercolumns over the retinal position x, see [Petitot, 2017,
Section 2].
Let Q ⊂ R2 be the visual plane. To a visual stimulus f : Q → [0, 1] is associated a cortical
activation Lf : Q × [0, π) → R such that Lf (ξ) encodes the response of the neuron ξ = (x, θ).
Letting ψξ ∈ L2(R2) be the receptive profile (RP) of the neuron ξ, such response is assumed to be
given by
Lf (ξ) = (cid:104)ψξ, f(cid:105)L2(R2) =
ψξ(x)f (x) dx.
(B.1)
(cid:90)
Q
Motivated by neuro-phyisiological evidence, we assume that RPs of different neurons are "de-
ducible" one from the other via a linear transformation. As detailed in [Duits and Franken,
2010, Prandi and Gauthier, 2017], see also [Bertalm´ıo et al., 2019a, Section 3.1], this amounts to
the fact that the linear operator L : L2(Q) → L2(Q× [0, π)) is a continuous wavelet transform (also
called invertible orientation score transform). That is, there exists a mother wavelet Ψ ∈ L2(R2)
such that Lf (x, θ) =(cid:2)f ∗ (Ψ∗ ◦ R−θ)(cid:3)(x). Here, f ∗ g denotes the standard convolution on L2(R2)
and R−θ is the counter-clock-wise rotation of angle θ. Notice that, although images are functions
of L2(R2) with values in [0, 1], it is in general not true that Lf (x, θ) ∈ [0, 1].
Concerning the choice of the mother wavelet, we remark that neuro-physiological evidence
suggests that a good fit for the RPs is given by Gabor filters, whose Fourier transform is the
product of a Gaussian with an oriented plane wave [Daugman, 1985]. However, these filters are
quite challenging to invert, and are parametrised on a bigger space than M, which takes into
account also the frequency of the plane wave and not only its orientation. For this reason, in
this work we instead considered cake wavelets, introduced in [Duits et al., 2007, Bekkers et al.,
2014]. These are obtained via a mother wavelet Ψcake whose support in the Fourier domain is
concentrated on a fixed slice, depending on the number of orientations one aims to consider in
the numerical implementation. For the sake of integrability, the Fourier transform of this mother
wavelet is then smoothly cut off via a low-pass filtering, see [Bekkers et al., 2014, Section 2.3] for
details. Observe, however, that, since we are considering orientations on [0, π) and not directions
on [0, 2π), we choose a non-oriented version of the mother wavelet, given by ψcake(ω) + ψcake(eiπω),
in the notations of [Bekkers et al., 2014].
An important feature of cake wavelets is that, in order to recover the original stimulus from
its cortical activation, it suffices to simply "project" the cortical activations along hypercolumns.
This yields
f (x) :=
Lf (x, θ) dθ.
(B.2)
This justify the assumption, implicit in equation (2.9), that the projection of a cortical activation
F (not necessarily given by a visual stimulus) to the visual plane is given by
(cid:90) π
1
π
0
(cid:90) π
0
P F (x) =
1
π
References
F (x, θ) dθ.
(B.3)
[Adini et al., 1997] Adini, Y., Sagi, D., and Tsodyks, M. (1997). Excitatory -- inhibitory network in
the visual cortex: Psychophysical evidence. Proceedings of the National Academy of Sciences,
94(19):10426 -- 10431.
17
[Atick, 1992] Atick, J. J. (1992). Could information theory provide an ecological theory of sensory
processing? Network: Computation in Neural Systems, 3(2):213 -- 251.
[Attneave, 1954] Attneave, F. (1954). Some informational aspects of visual perception. Psycho-
logical review, 61(3):183.
[Barlow et al., 1961] Barlow, H. B. et al. (1961). Possible principles underlying the transformation
of sensory messages. Sensory communication, 1:217 -- 234.
[Bekkers et al., 2014] Bekkers, E., Duits, R., Berendschot, T., and ter Haar Romeny, B. (2014). A
multi-orientation analysis approach to retinal vessel tracking. JMIV, 49(3):583 -- 610.
[Benucci et al., 2013] Benucci, A., Saleem, A. B., and Carandini, M. (2013). Adaptation maintains
population homeostasis in primary visual cortex. Nature neuroscience, 16(6):724.
[Bertalm´ıo, 2014] Bertalm´ıo, M. (2014). From image processing to computational neuroscience: a
neural model based on histogram equalization. Front. Comput. Neurosc., 8:71.
[Bertalm´ıo et al., 2019a] Bertalm´ıo, M., Calatroni, L., Franceschi, V., Franceschiello, B., and
Prandi, D. (2019a). A cortical-inspired model for orientation-dependent contrast perception:
A link with Wilson-Cowan equations. In Lellmann, J., Burger, M., and Modersitzki, J., edi-
tors, Scale Space and Variational Methods in Computer Vision, pages 472 -- 484, Cham. Springer
International Publishing.
[Bertalm´ıo et al., 2019b] Bertalm´ıo, M., Calatroni, L., Franceschi, V., Franceschiello, B., and
Prandi, D. (2019b). Cortical-inspired Wilson-Cowan-type equations for orientation-dependent
contrast perception modelling. arXiv preprint: https://arxiv.org/abs/1910.06808.
[Bertalm´ıo et al., 2007] Bertalm´ıo, M., Caselles, V., Provenzi, E., and Rizzi, A. (2007). Perceptual
color correction through variational techniques. IEEE T. Image Process., 16(4):1058 -- 1072.
[Bertalm´ıo and Cowan, 2009] Bertalm´ıo, M. and Cowan, J. D. (2009). Implementing the retinex
algorithm with wilson -- cowan equations. Journal of Physiology-Paris, 103(1-2):69 -- 72.
[Bertalm´ıo and Cowan, 2009] Bertalm´ıo, M. and Cowan, J. D. (2009). Implementing the retinex
algorithm with Wilson-Cowan equations. J. Physiol. Paris, 103(1):69 -- 72.
[Bertalm´ıo et al., 2017] Bertalm´ıo, M., Cyriac, P., Batard, T., Martinez-Garcia, M., and Malo, J.
(2017). The wilson-cowan model describes contrast response and subjective distortion. J Vision,
17(10):657 -- 657.
[Beurle and Matthews, 1956] Beurle, R. L. and Matthews, B. H. C. (1956). Properties of a mass of
cells capable of regenerating pulses. Philosophical Transactions of the Royal Society of London.
Series B, Biological Sciences, 240(669):55 -- 94.
[Bezanson et al., 2017] Bezanson, J., Edelman, A., Karpinski, S., and Shah, V. B. (2017). Julia:
A fresh approach to numerical computing. SIAM review, 59(1):65 -- 98.
[Blakeslee et al., 2016] Blakeslee, B., Cope, D., and McCourt, M. E. (2016). The oriented difference
of gaussians (ODOG) model of brightness perception: Overview and executable Mathematica
notebooks. Behav. Res. Methods, 48(1):306 -- 312.
[Bohi et al., 2017] Bohi, A., Prandi, D., Guis, V., Bouchara, F., and Gauthier, J.-P. (2017). Fourier
descriptors based on the structure of the human primary visual cortex with applications to object
recognition. Journal of Mathematical Imaging and Vision, 57(1):117 -- 133.
[Boscain et al., 2018] Boscain, U. V., Chertovskih, R., Gauthier, J.-P., Prandi, D., and Remi-
zov, A. (2018). Highly corrupted image inpainting through hypoelliptic diffusion. Journal of
Mathematical Imaging and Vision, 60(8):1231 -- 1245.
18
[Bressan, 2001] Bressan, P. (2001). Explaining lightness illusions. Perception, 30(9):1031 -- 1046.
[Brucke, 1865] Brucke, E. (1865). uber erganzungs und contrasfarben. Wiener Sitzungsber, 51.
[Chariker et al., 2016] Chariker, L., Shapley, R., and Young, L.-S. (2016). Orientation selectiv-
ity from very sparse lgn inputs in a comprehensive model of macaque v1 cortex. Journal of
Neuroscience, 36(49):12368 -- 12384.
[Citti and Sarti, 2006] Citti, G. and Sarti, A. (2006). A cortical based model of perceptual com-
pletion in the roto-translation space. JMIV, 24(3):307 -- 326.
[Cowan et al., 2016] Cowan, J. D., Neuman, J., and van Drongelen, W. (2016). Wilson -- cowan
equations for neocortical dynamics. The Journal of Mathematical Neuroscience, 6(1):1.
[Daugman, 1985] Daugman, J. G. (1985). Uncertainty relation for resolution in space, spatial
frequency, and orientation optimized by two-dimensional visual cortical filters. J. Opt. Soc. Am.
A, 2(7):1160 -- 1169.
[DeValois and DeValois, 1990] DeValois, R. L. and DeValois, K. K. (1990). Spatial vision. Oxford
university press.
[Duits et al., 2007] Duits, R., Felsberg, M., Granlund, G., and ter Haar Romeny, B. (2007). Image
analysis and reconstruction using a wavelet transform constructed from a reducible representa-
tion of the euclidean motion group. International Journal of Computer Vision, 72(1):79 -- 102.
[Duits and Franken, 2010] Duits, R. and Franken, E. (2010). Left-invariant parabolic evolutions
on SE(2) and contour enhancement via invertible orientation scores. Part I: linear left-invariant
diffusion equations on SE(2). Quart. Appl. Math., 68(2):255 -- 292.
[Eagleman, 1959] Eagleman, D. M. (1959). Visual illusions and neurobiology. Nature reviews
Neuroscience, (2):920926 (2001).
[Ernst et al., 2016] Ernst, U. A., Schiffer, A., Persike, M., and Meinhardt, G. (2016). Contextual
interactions in grating plaid configurations are explained by natural image statistics and neural
modeling. Frontiers in systems neuroscience, 10:78.
[Fairhall et al., 2001] Fairhall, A. L., Lewen, G. D., Bialek, W., and van Steveninck, R. R. d. R.
(2001). Efficiency and ambiguity in an adaptive neural code. Nature, 412(6849):787.
[Faugeras, 2009] Faugeras, O. (2009). A constructive mean-field analysis of multi population neu-
ral networks with random synaptic weights and stochastic inputs. Frontiers in Computational
Neuroscience, 3.
[French, 2004] French, D. (2004). Identification of a free energy functional in an integro-differential
equation model for neuronal network activity. Applied Mathematics Letters, 17(9):1047 -- 1051.
[Geier and Hud´ak, 2011] Geier, J. and Hud´ak, M. (2011). Changing the chevreul illusion by a
background luminance ramp: lateral inhibition fails at its traditional stronghold-a psychophys-
ical refutation. PloS One, 6(10):e26062.
[Herzog et al., 2003] Herzog, M. H., Ernst, U. A., Etzold, A., and Eurich, C. W. (2003). Local
interactions in neural networks explain global effects in gestalt processing and masking. Neural
Computation, 15(9):2091 -- 2113.
[Hong and Shevell, 2004] Hong, S. W. and Shevell, S. K. (2004). Brightness contrast and assimi-
lation from patterned inducing backgrounds. Vision Research, 44(1):35 -- 43.
[Hubel and Wiesel, 1959] Hubel, D. and Wiesel, T. (1959). Receptive fields of single neurones in
the cat's striate cortex. The Journal of physiology, 148(3):574591.
19
[Kingdom, 2011] Kingdom, F. A. (2011). Lightness, brightness and transparency: A quarter cen-
tury of new ideas, captivating demonstrations and unrelenting controversy. Vision Research,
51(7):652 -- 673.
[Kitaoka, 2006] Kitaoka, A. (2006). Adelsons checker-shadow illusion-like gradation lightness illu-
sion. http://www.psy.ritsumei.ac.jp/~akitaoka/gilchrist2006mytalke.html. Accessed:
2018-11-03.
[Mante et al., 2005] Mante, V., Frazor, R. A., Bonin, V., Geisler, W. S., and Carandini, M. (2005).
Independence of luminance and contrast in natural scenes and in the early visual system. Nature
neuroscience, 8(12):1690.
[McCourt, 1982] McCourt, M. E. (1982). A spatial frequency dependent grating-induction effect.
Vision Research, 22(1):119 -- 134.
[Meister and Berry, 1999] Meister, M. and Berry, M. J. (1999). The neural code of the retina.
Neuron, 22(3):435 -- 450.
[Murray and Herrmann, 2013] Murray, M. M. and Herrmann, C. (2013).
Illusory contours: a
window onto the neurophysiology of constructing perception. Trends Cogn Sci., (17(9)):471 -- 81.
[Olshausen and Field, 2000] Olshausen, B. A. and Field, D. J. (2000). Vision and the coding of
natural images: The human brain may hold the secrets to the best image-compression algorithms.
AmSci, 88(3):238 -- 245.
[Otazu et al., 2008] Otazu, X., Vanrell, M., and Parraga, C. A. (2008). Multiresolution wavelet
framework models brightness induction effects. Vis. Res., 48(5):733 -- 751.
[Pattadkal et al., 2018] Pattadkal, J. J., Mato, G., van Vreeswijk, C., Priebe, N. J., and Hansel,
D. (2018). Emergent orientation selectivity from random networks in mouse visual cortex. Cell
reports, 24(8):2042 -- 2050.
[Petitot, 2017] Petitot, J. (2017). Elements of Neurogeometry: Functional Architectures of Vision.
Lecture Notes in Morphogenesis. Springer International Publishing.
[Prandi and Gauthier, 2017] Prandi, D. and Gauthier, J.-P. (2017). A semidiscrete version of
the Petitot model as a plausible model for anthropomorphic image reconstruction and pattern
recognition. SpringerBriefs in Mathematics. Springer International Publishing, Cham.
[Purves et al., 2008] Purves, D., Wojtach, W. T., and Howe, C. (2008). Visual illusions: An
Empirical Explanation. Scholarpedia, 3(6):3706. revision #89112.
[Ratliff, 1965] Ratliff, F. (1965). Mach bands: quantitative studies on neural networks. Holden-Day,
San Francisco London Amsterdam.
[Sarti and Citti, 2015] Sarti, A. and Citti, G. (2015). The constitution of visual perceptual units
in the functional architecture of V1. J. comput. neurosc., 38(2):285 -- 300.
[Scholl et al., 2013] Scholl, B., Tan, A. Y., Corey, J., and Priebe, N. J. (2013). Emergence of
orientation selectivity in the mammalian visual pathway. Journal of Neuroscience, 33(26):10616 --
10624.
[Shapiro and Todorovic, 2016] Shapiro, A. G. and Todorovic, D. (2016). The Oxford compendium
of visual illusions. Oxford University Press.
[Smirnakis et al., 1997] Smirnakis, S. M., Berry, M. J., Warland, D. K., Bialek, W., and Meis-
ter, M. (1997). Adaptation of retinal processing to image contrast and spatial scale. Nature,
386(6620):69.
20
[Teich and Qian, 2010] Teich, A. F. and Qian, N. (2010). V1 orientation plasticity is explained
by broadly tuned feedforward inputs and intracortical sharpening. Visual neuroscience, 27(1-
2):57 -- 73.
[Weintraub and Krantz, 1971] Weintraub, D. J. and Krantz, D. H. (1971). The Poggendorff illu-
sion: amputations, rotations, and other perturbations. Attent. Percept. Psycho., 10(4):257 -- 264.
[White, 1979] White, M. (1979). A new effect of pattern on perceived lightness. Perception,
8(4):413 -- 416.
[Wilson, 2003] Wilson, H. R. (2003). Computational evidence for a rivalry hierarchy in vision.
Proceedings of the National Academy of Sciences, 100(24):14499 -- 14503.
[Wilson, 2007] Wilson, H. R. (2007). Minimal physiological conditions for binocular rivalry and
rivalry memory. Vision Research, 47(21):2741 -- 2750.
[Wilson, 2017] Wilson, H. R. (2017). Binocular contrast, stereopsis, and rivalry: Toward a dynam-
ical synthesis. Vision Research, 140:89 -- 95.
[Wilson and Cowan, 1972] Wilson, H. R. and Cowan, J. D. (1972). Excitatory and inhibitory
interactions in localized populations of model neurons. BioPhys. J., 12(1).
[Wilson and Cowan, 1973] Wilson, H. R. and Cowan, J. D. (1973). A mathematical theory of the
functional dynamics of cortical and thalamic nervous tissue. Kybernetik, 13(2):55 -- 80.
21
|
1306.2893 | 1 | 1306 | 2013-06-12T16:56:32 | Resolving structural variability in network models and the brain | [
"q-bio.NC",
"nlin.AO",
"physics.data-an"
] | Large-scale white matter pathways crisscrossing the cortex create a complex pattern of connectivity that underlies human cognitive function. Generative mechanisms for this architecture have been difficult to identify in part because little is known about mechanistic drivers of structured networks. Here we contrast network properties derived from diffusion spectrum imaging data of the human brain with 13 synthetic network models chosen to probe the roles of physical network embedding and temporal network growth. We characterize both the empirical and synthetic networks using familiar diagnostics presented in statistical form, as scatter plots and distributions, to reveal the full range of variability of each measure across scales in the network. We focus on the degree distribution, degree assortativity, hierarchy, topological Rentian scaling, and topological fractal scaling---in addition to several summary statistics, including the mean clustering coefficient, shortest path length, and network diameter. The models are investigated in a progressive, branching sequence, aimed at capturing different elements thought to be important in the brain, and range from simple random and regular networks, to models that incorporate specific growth rules and constraints. We find that synthetic models that constrain the network nodes to be embedded in anatomical brain regions tend to produce distributions that are similar to those extracted from the brain. We also find that network models hardcoded to display one network property do not in general also display a second, suggesting that multiple neurobiological mechanisms might be at play in the development of human brain network architecture. Together, the network models that we develop and employ provide a potentially useful starting point for the statistical inference of brain network structure from neuroimaging data. | q-bio.NC | q-bio | Resolving structural variability in network models and the brain
Florian Klimm1,2,3, Danielle S. Bassett1,4,∗, Jean M. Carlson1, Peter J. Mucha3,5
1Department of Physics, University of California, Santa Barbara, CA 93106 USA
2 Institut fur Physik, Humboldt-Universitat zu Berlin, 12489 Germany
3 Department of Mathematics, University of North Carolina, Chapel Hill, NC 27599 USA
4 Sage Center for the Study of the Mind, University of California, Santa Barbara, CA 93106 USA
5 Department of Applied Physical Sciences, University of North Carolina, Chapel Hill, NC 27599 USA
∗Corresponding author email [email protected]
November 8, 2018
Abstract
Large-scale white matter pathways crisscrossing the cortex create a complex pattern of connec-
tivity that underlies human cognitive function. Generative mechanisms for this architecture have
been difficult to identify in part because little is known in general about mechanistic drivers of struc-
tured networks. Here we contrast network properties derived from diffusion spectrum imaging data
of the human brain with 13 synthetic network models chosen to probe the roles of physical network
embedding and temporal network growth. We characterize both the empirical and synthetic net-
works using familiar graph metrics, but presented here in a more complete statistical form, as scatter
plots and distributions, to reveal the full range of variability of each measure across scales in the
network. We focus specifically on the degree distribution, degree assortativity, hierarchy, topological
Rentian scaling, and topological fractal scaling -- in addition to several summary statistics, including
the mean clustering coefficient, the shortest path-length, and the network diameter. The models
are investigated in a progressive, branching sequence, aimed at capturing different elements thought
to be important in the brain, and range from simple random and regular networks, to models that
incorporate specific growth rules and constraints. We find that synthetic models that constrain the
network nodes to be physically embedded in anatomical brain regions tend to produce distributions
that are most similar to the corresponding measurements for the brain. We also find that network
models hardcoded to display one network property (e.g., assortativity) do not in general simultane-
ously display a second (e.g., hierarchy). This relative independence of network properties suggests
that multiple neurobiological mechanisms might be at play in the development of human brain net-
work architecture. Together, the network models that we develop and employ provide a potentially
useful starting point for the statistical inference of brain network structure from neuroimaging data.
3
1
0
2
n
u
J
2
1
]
.
C
N
o
i
b
-
q
[
1
v
3
9
8
2
.
6
0
3
1
:
v
i
X
r
a
1
1
Introduction
Increasing resolution of noninvasive neuroimaging methods for quantifying structural brain organization
in humans has inspired a great deal of theoretical activity [1, 2, 3, 4], aimed at developing methods
to understand, diagnose, and predict aspects of human development and behavior based on underlying
organizational principles deduced from these measurements [5, 6, 7]. Ultimately, the brain is a network,
composed of neuronal cell bodies residing in cortical grey matter regions, joined by axons, protected by
myelin. Diffusion-weighted magnetic resonance imaging methods trace these white matter connections,
based on the diffusion of water molecules through the axonal fiber bundles. While resolution has not
reached the level of individual neurons and axons, these methods lead to reliable estimates of the density
of connections between regions and fiber path lengths. The result is a weighted adjacency matrix, with
a size and complexity that increases with the resolution of the measurements [8, 9].
The immense complexity of this data makes it difficult to directly deduce the underlying mechanisms
that may lead to fundamental patterns of organization and development in the brain [10]. As a result,
comparison studies with synthetic network models, employing quantitative graph statistics to reduce
the data to a smaller number of diagnostics, have provided valuable insights [11, 12, 13, 14, 15]. These
models and statistics provide a vehicle to compare neuroimaging data with corresponding measurements
for well-characterized network null models. However, the methods are still in development [16, 17, 18],
and vulnerable to the loss of critical information through oversimplification of complex, structured data
sets, by restricting comparisons to coarse measurements that ignore variability [19, 20, 10].
Two critical questions motivate development of network methodologies for the brain. The first
question focuses on predictive statistics: Are there graph metrics that may ultimately be useful in
parsing individual differences and diagnosing diseases? Comparing empirical brain data to benchmark
null models might help establish statistical significance of a topological property [21, 22, 23] or assist in
obtaining a statistical inference about differences in brain network structure between groups [16]. The
second question focuses on network characteristics from a fundamental, development and evolutionary
perspective: What organizational principles underlie growth in the human brain? Here comparing
empirical brain data to simplified model networks that have been created to capture some aspect of,
for example, neurodevelopmental growth rules [24], neuronal functions [11], or physiological constraints
[25] may aid in developing a mechanistic understanding of the brain's network architecture (e.g., [26, 27,
28]). Both efforts require a basic understanding of the topological similarities and differences between
synthetic networks and empirical data.
In this paper, we perform a sequence of detailed, topological comparisons between empirical brain
networks obtained from diffusion imaging data and 13 synthetic network models. The models are
investigated in a tree-like branching order, beginning with the simplest, random or regular graphs, and
progressively adding complexity and constraints (see Figure 1). The objective of this investigation is
to determine, in a controlled, synthetic setting, the impact of additional network properties on the
topological measurements.
At the coarsest level in the model hierarchy, we distinguish between synthetic networks that are
constructed purely based on rules for connectivity between nodes (non-embedded), and those that con-
strain nodes to reside in anatomical brain regions (embedded) (see Figure 1). While non-embedded
models are frequently used for statistical inference, recent evidence has suggested that physical, embed-
ding constraints may have important implications for the topology of the brain's large-scale anatomical
connectivity [26, 27, 2, 28, 8, 22, 29]. By examining both non-embedded and embedded models, we hope
our results will help to guide the use, development, and understanding of more biologically realistic
models for both statistical and mechanistic purposes [30, 23].
A second important classification of the synthetic models in our study separates those obtained from
static ensembles with fixed statistical properties and those generated using mechanistic growth rules
(see Figure 1). While algorithms for generating networks based on static sampling and growth rules
2
Figure 1: Branching Structure of Synthetic Model Examination. We distinguish between
synthetic networks that are constructed based on rules for connectivity between nodes (non-embedded ),
and those that constrain nodes to reside in anatomical brain regions (embedded ). We further distinguish
between synthetic networks that are obtained from static ensembles (static), and those that are obtained
from growth rules (growing).
In the non-embedded case, we explore common benchmark networks
including regular lattice, Erdos-R´enyi, and small-world models as well as a second set of networks that
are based on these benchmarks but that also employ additional constraints. For growing models, we
explore the Bar´abasi-Albert model and introduce an affinity model inspired by preferential attachment-
like properties of neuronal growth. In the embedded case, we distinguish between models that utilize
true or false node locations (i.e., models derived from a spatial embedding independent of the known,
physical node locations) and explore several growing models inspired by hypotheses regarding wiring
minimization in brain development [26, 28, 29].
3
Non-Embedded Model Order (Ring Lattice) Disorder (Erdös-Rényi) Modular (Modular Small-World) With Constraints (Fractal Hierarchical) With Constraints (Gaussian Drop-Off) Preferential Attachment (Barábasi-Albert) With Constraints (Affinity) Embedded Model Growth Models Static Models False Node Locations (Random Geometric) With Constraints (Distance Drop-off) True Node Locations (Minimally Wired) Growing Distance (Distance Drop-Off Growth) Static Models Growth Models Static + Growing (Hybrid Distance Growth) ? With Constraints (Configuration) DDG HDG DD MW RG AF BA CF ER RL GD MS FH ultimately both produce ensembles of fixed graphs for our comparison with data, additional constraints
imposed by underlying growth rules may facilitate understanding of mechanisms for development and
evolution in the brain as well as other biological and technological networks.
To compare the models with brain data, we employ a subset of the many network diagnostics that
have been proposed as measures of network topology [31]. Many network diagnostics can be described as
summary diagnostics, in which a property of the network organization is reduced to a single diagnostic
number. However, the comparison of summary diagnostics between real and model networks can be
difficult to interpret [32] because they often hide the granularity at which biological interpretations can
be made. To maximize the potential for a mechanistic understanding, we therefore study the following
four diagnostic relationships obtained from a distribution of values over network nodes or topological
scales: hierarchy [33], degree assortativity [34], topological Rentian scaling [35, 36], and the topological
fractal dimension [37]. Each of these relational properties have previously been investigated in the
context of anatomical brain networks in humans [38, 39, 28]. In this paper, we use them to examine
the differences between empirically derived anatomical brain networks and synthetic network models.
4
2 Materials and Methods
2.1 Data
We utilize previously published diffusion spectrum imaging data [39] to examine network structure
of anatomical connectivity in the human brain. The direct pathways between N = 998 regions of
interest are estimated using deterministic white matter tractography in 5 healthy human participants
[39]. This procedure results in an N × N weighted undirected adjacency matrix W representing the
network, with elements Wij indicating the (normalized) number of streamlines connecting region i to
region j (see Figure 2). The organization of white matter tracts can be examined at two distinct levels
of detail: topological and weighted. Studies of the topological organization of brain anatomy focus on
understanding the presence or absence of white matter tracts between regions [27, 26, 28], while studies
of the weighted organization focus on understanding the strength of white matter connectivity between
those regions. In this paper, we explore the topological organization of white matter connectivity. In
future work we plan to build additional constraints into our models that will enable a comparison of
model and empirical weighted networks.
To study topological organization, we construct the binary adjacency matrix A in which the element
Aij is equal to 1 if the employed tractography algorithm identifies any tracts (of any strength) linking
region i with region j (i.e., Wij (cid:54)= 0). In this data [39], the adjacency matrix A is relatively sparse,
resulting in a network density of ρ = 2M/[N (N − 1)] ≈ 2.7%, where M = 1
ij Aij is the total number
of connections present. This estimate of brain network sparsity is consistent with estimates extracted
from other similar data sets of comparable network size [8, 47].
(cid:80)
2
Given the potential variability in the topological organization of networks extracted from different
individuals [8, 48, 49, 50, 51], we report results for one individual in the main manuscript and describe
the consistency of these results across subjects in the Supplementary Materials.
2.2 Network Diagnostics
We measure four network properties including degree assortativity, hierarchy, Rentian scaling, and
topological fractal dimension as well as several summary diagnostics, as reported in Table 2.
Assortativity. The number of edges emanating from node i is referred to as its degree, denoted
by ki. The degree assortativity of a network, or more simply 'assortativity' here, is defined as the
correlation between a node's degree and the mean degrees of that node's neighbors which can be
calculated as
r =
,
(1)
M−1(cid:80)
M−1(cid:80)
m
1
2 (j2
m jmkm − [M−1(cid:80)
m) − [M−1(cid:80)
m + k2
m
1
2 (jm + km)]2
m
1
2 (jm + km)]2
where jm, km are the degrees of the nodes at either end of the mth edge, with m = 1 . . . M [52]. The
assortativity measures the likelihood that a node connects to other nodes of similar degree (leading
to an assortative network, r > 0) or to other nodes of significantly different degree (leading to a
disassortative network, r < 0). Social networks are commonly found to be assortative while networks
such as the internet, World-Wide Web, protein interaction networks, food webs, and the neural network
of C. elegans are disassortative [34].
Hierarchy. The hierarchy of a network is defined quantitatively by a relationship between the
node degree and the local clustering coefficient Ci [53]. For each individual node i, Ci is defined as:
Ci =
∆exist
∆possible
(2)
where ∆exist is the ratio of the number of existing triangle subgraphs that include node i, and ∆possible
is the number of node triples containing node i. Using this local definition, the clustering coefficient
5
Figure 2: Adjacency Matrices for Brain and Synthetic Models. Example adjacency matrices
are provided for the brain and for the 13 synthetical network models described in Figure 1. In the
empirical brain data and the non-embedded null models, network nodes are ordered along the x and
y-axes to maximize connectivity along the diagonal, as implemented by the reorderMAT.m function in
the Brain Connectivity Toolbox [43]. In the embedded models, nodes are listed in the same order as
they are in the empirical brain data. Abbreviations are as listed in Table 3.
of the graph C as a whole (a summary diagnostic) is defined as the mean of Ci over all nodes in the
network.
The definition of hierarchy is based on a presumed power law relationship between the local clus-
tering coefficient Ci and the degree ki of all nodes i in the network [33]:
Ci ∼ k
−β
i
.
6
(3)
Non-Embedded Model Embedded Model Growth Models Static Models Static Models Growth Models ? DDG HDG DD MW RG AF BA CF ER RL GD MS FH Brain Embedded Non-Embedded For a given network, the best fit to the scaling exponent β is referred to as the network hierarchy.
Topological Rentian Scaling. In contrast to the physical Rent's rule [35], the topological Rent's
rule is defined as the scaling of the number of nodes n within a topological partition of a network
with the number of connections or edges, e, crossing the boundary of that topological partition. If
the relationship between these two variables is described by a power law (i.e., e ∝ npT ), the network
is said to show topological Rentian scaling, or a fractal topology, and the exponent of this scaling
relationship is known as the topological Rent exponent, pT [54]. Thus higher values of the topological
Rent exponent are indicative of higher dimensional network topology. Pragmatically, to determine pT ,
we follow the procedure outlined in [36] where topological partitions are created by a recursive min-cut
bi-partitioning algorithm that ignores spatial locations of network nodes [28].
Topological Fractal Dimension. The topological Rent's exponent described above is related to
the topological dimension, DT , of the network according to the inequality pT ≥ 1− 1
[54]. To directly
quantify the topological dimension of a network, we evaluate its topological invariance under length-
scale transformations [37]. We employ a box-counting method [55] in which we count the number of
boxes NB of topological size lB that are necessary to cover the network. The fractal dimension of the
network can then be estimated as the exponent dB of the putative power law relationship
DT
NB ≈ l
−dB
N .
(4)
The fractal dimension of a network is a measure of the network's complexity. We note that the process
of tiling the network into boxes of different sizes is non-deterministic. To account for this variability,
we report mean values of dB over 50 different tilings of the a given network.
Additional Quantities of Interest. In Table 2, we list several summary diagnostics of interest
to complement our analysis of relational properties. These include the average path length between
node i and j, defined as the shortest number of edges one would have to traverse to move from node
i to node j [56]. The path length of an entire network, P , is then defined as the average path length
from any node to any other node in the network: P =
ij Pij, while the maximal path length
between any two pairs of nodes is called the diameter D = maxij{Pij}.
(cid:80)
1
N (N−1)
All computational and basic statistical operations (such as t-tests and correlations) were implemented
2.3 Statistics, Software, and Visualization
using MATLAB R(cid:13) (2009b, The MathWorks Inc., Natick, MA) software. Graph diagnostics were esti-
mated using a combination of in-house software, the Brain Connectivity Toolbox [43], and the MATLAB
Boost Graph Library (\protect\vrulewidth0pthttp://www.stanford.edu/∼dgleich/programs/ ). To
perform the recursive topological partitioning employed in the examination of topological Rentian scal-
ing, we used the software tool hMETIS [57].
Several of the network models that we investigate include one or more tunable parameters affecting
the details of the generated graphs. These include the Barab´asi-Albert, affinity, and hybrid distance
growth models. To compare these network models to the data, we optimized parameter values to
minimize the difference between the model network and the empirical brain network. Specifically, we
used the Nelder-Mead simplex method, which is a derivative-free optimization method, that minimizes
the value of a difference metric δm between the two networks. We chose to let δm be the sum of
the absolute relative difference of seven of the network characteristics reported in Table 2 (clustering
coefficient C, path length P , diameter D, degree assortativity r, hierarchical parameter β, topological
Rentian exponent pT , and topological fractal dimension dB), although we note that other choices are
of course possible.
7
Abbreviation Description
Citation
Uniform connection probability
Random rewiring preserving degree dis-
tribution
Fixed degree to k nearest neighbors
Gaussian drop-off in edge density with in-
creasing distance from the diagonal
Fully connected modules linked together
by evenly distributed random connections
Modular structure across n hierarchi-
cal levels; connection density decays as
1/(En)
Network growth by preferential attach-
ment rule
Two-step preferential attachment growth
with hardcoded assortativity and hierar-
chy
[40]
[41]
[40]
[42, 43]
[43]
[43]
[44]
Wire together random node locations
with shortest possible connections
Wire together true node locations with
shortest possible connections
Wire together true node locations with a
probability that drops off with distance
between nodes
[45]
[28, 26, 27]
[46]
Network growth by distance drop-off rule
Minimally wired network that grows with
distance drop-off rule
Model Name
Non-embedded
Static
Erdos-R´enyi
Configuration
Ring Lattice
Gaussian Drop-Off
Modular Small-World
Fractal Hierarchical
Growth
Barab´asi-Albert
Affinity
Embedded
Static
Random Geometric
Minimally Wired
Distance Drop-Off
ER
CF
RL
GD
MS
FH
BA
AF
RG
MW
DD
Growth
Distance Drop-Off Growth DDG
Hybrid Distance Growth
HDG
Table 1: Network Models Names, Abbreviations, Intuitive Descriptions, and Associated
References.
8
3 Results
In this section we individually compare topological network diagnostics calculated for the empirical
brain data to each of the 13 network models that appear in Figures 1 and 2. We proceed through the
catalog of synthetic models along the branches illustrated in Figure 1. We begin with the simplest
models (i.e. non-embedded, static, random and regular), and incrementally add structure, constraints,
growth mechanisms, and embedding in order to isolate how these additional features impact the mea-
sured diagnostics.
For each network we present statistical results for three diagnostics: (i) the degree distribution
P (ki) vs. ki, (ii) the mean node degree of the neighboring nodes vs. node degree ki for each node i
(used to calculate assortativity), and (iii) the local clustering coefficient Ci vs. node degree ki for each
i (used to calculate hierarchy). In Figures 3 -- 6 and 8 -- 9, the results for the empirical brain network are
shown in gray and the corresponding results for each of the synthetic models are shown in a contrasting
color on the same graph to facilitate comparisons. In addition, we illustrate the scaling relationships
used to evaluate Rentian scaling and the topological dimension of each network (see Figures 7 and 10).
For our comparisons, we group the models first into the set of non-embedded models, followed by
the embedded models and we further group results according to the branches of inquiry outlined in
Figures 1 and 2. For each model we briefly describe our method for generating the synthetic network,
followed by a description of the diagnostics compared to the empirical results.
3.1 Non-Embedded Network Models
We begin by comparing the network organization of the brain's anatomical connectivity with that of
8 network models whose structure is not a priori constrained to accommodate a physical embedding
of the nodes onto cortical regions in the brain. (In the next subsection, we will examine 5 embedded
network models.) The non-embedded network models include an Erdos-R´enyi graph, a configuration
model with the same degree distribution as the empirical network, a ring lattice graph, a modular
small-world graph, a fractal hierarchical graph, a Gaussian drop-off model, a Barab´asi-Albert graph,
and an affinity model (see Figure 2 for associated example adjacency matrices for these graphs and
Table 3 for abbreviations of model names). These models range from disordered to ordered (e.g., the
Erdos-R´enyi and regular lattice models) with a range of mesoscale organization for intermediate cases
(e.g., modular small-world and fractal hierarchical models) which influence the network diagnostics,
and (dis)similarities to corresponding measurements for the brain.
3.1.1 Static Non-Embedded Models
Erdos-R´enyi (ER) Model: The Erdos-R´enyi (ER) model is an important benchmark network that
is often used as a comparison null mode for statistical inference. Specifically, we consider the 'G(N, M )
model' where the ER graph is constructed by connecting pairs chosen uniformly at random from N
total nodes until M edges exist in the graph [40]. The degree distribution generated by this procedure
is, as expected, relatively symmetric about the mean degree ρ(N − 1) ≈ 27 (see Figure 3A(i)).
The ER model is a poor fit to brain anatomical connectivity (see Figure 3A). The degree distribution
is much more sharply peaked than the corresponding distribution for the brain. For the ER graph the
variance is approximately equal to the mean degree, while the corresponding data for the brain is
more broadly distributed. As a result, the ER network misses structure associated with both high
degree hubs and low degree nodes. Because edges are placed at random, organizational properties like
assortativity and hierarchy are not observed and -- as expected theoretically -- the clustering coefficient
is smaller and the path length shorter than that of anatomical brain networks (see Table 2).
Configuration (CF) Model: We next consider a modification of the ER graph that is constrained
to have the same degree distribution as the empirical data. We refer to this as the configuration model
9
Figure 3: Comparison between the (i) degree distribution (number f of nodes with a given degree ki),
(ii) assortativity (correlation between a node's degree ki and the mean degree of that node's neighbors
k(cid:48)
i, summarized by parameter r), and (iii) hierarchy (the relationship between the clustering coefficient
Ci and the degree ki over all nodes in the network, summarized by parameter β) of the (A) Erdos-R´enyi
and (B) configuration model with conserved degree distribution models and the same diagnostics of the
brain anatomical data (grey). Black lines indicate best linear fit to the data (dashed) and model (solid)
networks. In panel (B) the lower (nonzero) bound on the clustering coefficient -- which corresponds to
the presence of only one triangle -- as a function of degree is indicated by the red line.
(CF). We generate randomized graphs by an algorithm that chooses two existing connections uniformly
at random (a ←→ b and c ←→ d) and switches their associations (a ←→ d and c ←→ b) [58].
This model agrees with the empirical degree distribution by construction (see Figure 3B(i)). How-
ever, it does not fit the higher order association of a node's degree with that node's mean neighbor
degree (assortativity) (see Figure 3B(ii)). The average clustering coefficient remains small, although
it is larger than that observed in the ER network. In Figure 3B(iii), we observe a small association
between the clustering coefficient and degree (hierarchy) which appears to be driven by nodes of small
degree. To interpret this finding, we note that the nonzero minimum of the clustering coefficient of a
node of degree k is given by
cmin(cid:54)=0(k) =
.
(5)
2
k(k − 1)
Thus, nodes of small degree tend to have a higher minimum non-zero clustering than nodes of high
degree.
In comparison to the ER model, the existence of small degree nodes leads to an increased
diameter of the graph whereas the existence of high degree nodes leads to the maintenance of a short
average path length.
Ring Lattice (RL) Model: In contrast to the two previous models, the ring lattice (RL) model
has a highly ordered topology where each node is connected to its 2M
N ≈ 27 nearest neighbors.
By construction, the degree distribution for the ring lattice is extremely sharply peaked.
If the
number of edges M is divisible by the number of nodes N , then all nodes have equal degree, otherwise
the remainder is distributed uniformly at random throughout the network, resulting is a very narrow
spread in the distribution. The clustering coefficient is close to unity, indicating that most neighbors of
a node are also connected to each other. The restriction to local connectivity results in a large diameter
and long average path length. The small variation in degree induced by the random distribution of
the remaining edges is insufficient to induce assortativity (see Figure 4A). Interestingly, however, this
model displays topological network hierarchy because nodes that have been assigned those remaining
10
1030507090020406080kk'rdata=0.15r=0.02050100f10110210−2100βdata=0.25β=−0.03C1030507090020406080kk'r=−0.01050100fDistribution equal by construction10110210−2100β=0.08cmin(k)CAB(i)(ii)(iii)(i)(ii)(iii)Erdos-Renyi (ER) ModelCon(cid:31)guration (CF) Modeliiiiii Figure 4: Comparison between the (i) degree distribution (number f of nodes with a given degree ki),
(ii) assortativity (correlation between a node's degree ki and the mean degree of that node's neighbors
k(cid:48)
i, summarized by parameter r), and (iii) hierarchy (the relationship between the clustering coefficient
Ci and the degree ki over all nodes in the network, summarized by parameter β) of the (A) ring lattice
and (B) Gaussian drop-off models and the same diagnostics in the brain anatomical data (grey). Black
lines indicate best linear fit to the data (dashed) and model (solid) networks.
edges have a higher than average degree which directly decreases the clustering coefficient of those
nodes. These results underscore the fact that very small amounts of noise in a data set can create the
illusion of the presence of a network property where it does not exist.
Gaussian Drop-Off (GD) Model: Compared to the brain, the random and randomized models
exhibit lower clustering, and the regular ring lattice exhibits higher clustering. An intermediate topology
between these two extremes is obtained by generalizing the concept of local connections from the ring
lattice to a stochastically generated network where the density of connections drops off at rate κ with
increasing distance from the main diagonal of the adjacency matrix.
We chose a value for κ by examining the empirical brain data as follows. First, we reordered the
adjacency matrix such that the connections (represented by nonzero matrix elements) are predominantly
located near the matrix diagonal, using the code reorderMAT.m in the Brain Connectivity Toolbox
[43]. We then fit a Gaussian function to the empirical drop-off of the first 400 off-diagonal rows of the
reordered brain adjacency matrix [43]. We note that the fit provided an R2 value of approximately
0.75.
The very localized structure in this model, similar to that observed in an RL model, is softened by
the presence of a few long-range connections which decreases the path length and brings the average
clustering coefficient closer to that of the data (see Figure 4B). The non-periodic boundary conditions
lead to a small subpopulation of nodes with low degree. Because these nodes are neighbors in the
adjacency matrix, they tend to be connected to one another, leading to an assortative topology. The
same explanation underlies the existence of a hierarchical topology, because these low degree boundary
nodes predominantly connect with one another.
Modular Small-World (MS) Model: Small world networks have received a great deal of at-
tention [53] as a conceptual characterization of structure that combines local order with long range
connections. While the small world concept is sufficiently general that most networks that are not
strictly regular or random fall into this category, small world organization represents more biologically
relevant organization than the previous four cases [59, 60, 28, 8]. In addition to the small-world fea-
ture, biological networks including those extracted from human brain connectome data [61, 8, 62, 63]
11
AB1030507090020406080kk'r=−0.01050100f10110210−2100β=0.55C4871030507090020406080kk'r=0.2050100f10110210−2100β=0.39C 120(i)(ii)(iii)(i)(ii)(iii)Regular Lattice (RL) ModelGaussian Drop-O(cid:31) (GD) Modeliiiiii Figure 5: Comparison between the (i) degree distribution (number f of nodes with a given degree ki),
(ii) assortativity (correlation between a node's degree ki and the mean degree of that node's neighbors
k(cid:48)
i, summarized by parameter r), and (iii) hierarchy (the relationship between the clustering coefficient
Ci and the degree ki over all nodes in the network, summarized by parameter β) of the (A) modular
small-world and the (B) fractal hierarchical models and the same diagnostics in the brain anatomical
data (grey). Black lines indicate best linear fit to the data (dashed) and model (solid) networks.
also often display community structure where set of nodes (modules) tend to be highly and mutually
interconnected with one another combined with some long-distance connections.
For this study, we construct a synthetic small world network that consists of small, fully-connected
modules composed of 4 nodes, randomly linked with one another with enough edges to match the
density of the empirical network. This topology leads to high clustering, short path length, and small
diameter [43]. The randomly distributed inter-module links emanating from relatively high degree
nodes decrease the clustering coefficient of these nodes because nodes in two different modules are
unlikely to be otherwise linked. This structure therefore leads to a hierarchical topology (see Figure
5A(iii)). However, because the inter-module links are randomly distributed, nodes that contain such
links are no more likely to share an edge with another such node than they are to share a link with
any other node in the network. The model therefore does not display any observable assortativity (see
Figure 5A(ii)).
Fractal Hierarchical (FH) Model: Like small world networks, fractal hierarchical topology has
become a popular classification of networks and applies broadly, at least to some extent, to topologies
that are neither regular nor random. Fractal hierarchical structure has been linked to some observed
network structure in the brain [8, 64, 61, 65, 66] and its use in neural network models produces several
behaviors reminiscent of empirical neurobiological phenomena [67, 68, 11].
To construct a fractal hierarchical model [33], we follow the approach outlined in [69]. We begin
with a set of 4-node modules. We connect pairs of these 4-node modules with a probability p1 to form
8-node modules. We connect pairs of 8-node modules with a probability p2 to form 16-node module.
Importantly, the probability p of inter-module connections decreases at each level at a prescribed drop-
off rate; that is, p1 is larger than p2, p2 is larger than p3, etc. The probabilities at each level are
related to one another by a probability drop-off rate. This module-pairing process is repeated until we
have formed a 1024-node fractal hierarchical network. To obtain a N = 998 network comparable to
the empirical brain data, we chose 26 nodes uniformly at random to delete from the network. If the
network contained more (fewer) edges than the empirical network, we repeated the process with an
increased (decreased) probability drop-off rate. The algorithm terminates when we obtain a network
12
AB1030507090020406080kk'r=−0.01050100f10110210−2100β=0.44C1030507090020406080kk'r=0050100f10110210−2100β=0.25C(i)(ii)(iii)(i)(ii)(iii)Modular Small-World (MS) ModelFractal Hierarchical (FH) Modeliiiiii with the correct number of edges.
The fractal hierarchal network yields extremely similar results to the small world network in terms
of the degree distribution, assortativity, and hierarchy (compare Figure 5A with Figure 5B). The
striking similarities are surprising given the differences in how the two networks are constructed. While
the networks share strong 4-node module building blocks, they differ in their coarser structure. The
similarity in the results depicted in Figure 5 suggest that the level-dependent structure in the fractal
hierarchical model is not well-captured by these graph properties. Other types of network properties
that specifically test for multiresolution phenomenon in brain structure might more readily distinguish
between these two synthetic models [70, 71].
Figure 6: Comparison between the (i) degree distribution (number f of nodes with a given degree ki),
(ii) assortativity (correlation between a node's degree ki and the mean degree of that node's neighbors
k(cid:48)
i, summarized by parameter r), and (iii) hierarchy (the relationship between the clustering coefficient
Ci and the degree ki over all nodes in the network, summarized by parameter β) of the (A) Barab´asi-
Albert and (B) affinity models and the same diagnostics in the brain anatomical data (grey). Black lines
indicate best linear fit to the data (dashed) and model (solid) networks. In panel (B), the parameter
values used for the affinity model are the following: γ = 1.94, δ = 3.48, and = 3.36.
3.1.2 Growing Non-Embedded Models
In this section we explore two non-embedded growth models (see Figure 1). The first is the Barab´asi-
Albert preferential attachment model and the second is an affinity model which we design to capture
assortative and hierarchical structure.
Barab´asi-Albert (BA) Model: All models described thus far, with the exception of the config-
uration model, share a common and critical short-coming: the degree distribution is much narrower
than that of the empirical networks. A model that produces a broader distribution of node degrees is
the Barab´asi-Albert model of preferential attachment [44].
To construct a BA network, we begin with a single edge connecting two nodes. Then we iteratively
add a single node to the network by linking the new node to m existing nodes. The probability of
linking the new node to an existing node is given by a preferential attachment function Π(k) = k + k0
with dimensionless parameter k0 tuning the rate of decrease in the degree distribution. Note that as
k0 → ∞, the resultant graph becomes increasingly similar to an ER graph.
To identify a model network in this family that best fits the empirical data, we tune k0 to minimize
the difference between the model topology and the empirical topology as described in Section 2. We
13
AB50100050100kk'r=−0.0205010010110210−2100β=0Cf1030507090020406080kk'r=0.05050100f10110210−2100β=0.18C(i)(ii)(iii)(i)(ii)(iii)Barabasi-Albert (BA) ModelA(cid:31)nity (AF) Modeliiiiii find that networks constructed using k0 = 4 provide the best available fit to the empirical data. The
number of edges m added with each new node is determined by the total number of edges M . This
procedure produces networks with low clustering and broad degree distributions, although the number
of low-degree nodes is underestimated in comparison to the empirical data (see Figure 6A(i)). Despite
the broad degree distribution, the network does not display an assortative or hierarchical topology (see
Figure 6A(ii) -- (iii)).
Affinity (AF) Model: We introduce an extension of the BA model that includes constraints
specifically designed to capture assortative and hierarchical structure. We define the affinity model
by a two step preferential attachment function that does not depend on a node's current degree but
instead depends on a dimensionless affinity parameter α. We begin with N nodes, and to each node we
assign a unique affinity αi distributed uniformly at random in the interval [0,1]. The value of αi remains
unchanged throughout the growth process (see Algorithm 1). We choose a node with probability ∝ αγ
i
and link that node preferentially to another node j with a similar affinity αj. This assortative mixing
for affinity ensures degree assortativity. In addition, we choose a preferential attachment function (see
Algorithm 1, line 6) such that nodes with small values of affinity (e.g. small degree) are relatively
more likely to gain edges with neighbors of similar affinity (and therefore degree) than nodes with large
values of affinity. Small degree nodes therefore are more clustered than their high degree counterparts,
leading to a hierarchical network structure.
Algorithm 1: Growth algorithm for the affinity model.
input : number of nodes N
number of edges M
number of seed edges M0
attachment regulators γ, δ and
output: Adjacency matrix A
1 initialize graph with N nodes;
2 connect M0 pairs of nodes chosen uniformly at random;
3 assign each node an affinity given by αi = i−1
N−1 ;
4 while M(cid:48)= current # of edges < M do
5
6
out of the set of nodes with k > 0 , choose a node i with probability ∝ αγ
connect node i to node j (chosen at uniformly at random) with probability
∝ αi − αjmin{0,−δ+· αi}
i
7 end
To compare this model to the empirical data, we use a derivative-free optimization method to iden-
tify the parameter values for γ, δ, and that minimize the difference between the empirical and model
networks; see Section 2. The affinity model has a very broad degree distribution with a concentration
of low degree nodes and an extremely heavy tail of high degree nodes (see Figure 6B(i)). The network
is both assortative and hierarchical although the average clustering is lower than that found in the
empirical data (see Figure 6B(ii) -- (iii)). The randomly chosen edges connecting nodes of high degree
induce a small diameter and short path length.
It is not surprising that the affinity model provides a better fit for the empirical data for these
specific diagnostics than other synthetic networks we have considered so far, since it was specifically
constructed to do so. This is, however, no guarantee that this algorithm will capture other network
properties of the empirical data. Indeed, the fact that the affinity model also shows a similar topological
dimension to the empirical brain network is surprising and interesting (see next section).
14
3.1.3 Diagnostics Estimating the Topological Dimension:
In this section, we compare topological measures of the empirical data with the set of 8 non-embedded
synthetic networks: 6 static models and 2 growth models.
Using a box-counting method, we estimate the fractal dimension of the empirical and synthetic
model networks (see Section 2) and observe three distinct classes of graphs (see Figure 7, main panel).
The first group, which includes the Erdos-R´enyi and modular small-world models, has a diameter that
is too small to allow an adequate estimation of the fractal dimension of the network using the box-
counting method. The second group, which includes the Gaussian drop-off and ring lattice models, has
a large diameter leading to a small fractal dimension. The third group, which includes the remainder
of the models, has a similar diameter to the empirical network and therefore similar fractal dimension.
By these comparisons, the affinity model is the best fit to the data and the configuration model is the
second best fit.
The Gaussian drop-off and ring lattice models also show distinct topological Rentian scaling in
comparison to the other models (see Figure 7, inset). Above a topological box size of 16 nodes, the
number of inter-box connections does not increase because the edges are highly localized topologically.
All other models display a swifter scaling of the number of edges with the number of nodes in a
topological box in comparison to the empirical data. The affinity model displays the most similar
scaling to that observed in the empirical data.
Figure 7: Diagnostics Estimating the Topological Dimension. (Main Panel) The number of
boxes as a function of the topological size of the box, as estimated using the box-counting method [55]
(see Section 2) for the real and synthetic networks. (Inset) The topological Rentian scaling relationship
between the number of edges crossing the boundary of a topological box and the number of nodes inside
of the box (see Section 2) for the real and synthetic networks. Lines indicate data points included in
fits reported in Table 2.
15
2345102037101001000# of boxes NBbox size lB248163264128256100100010000# of inter−box connections e# of nodes n per boxErdös-Rényi 2.30.96 Con(cid:30)guration 5.20.96 Ring Lattice 2.10.67Barabási-Albert 4.50.96Modular Small World 2.60.95Fractal Hierarchical 6.70.88A(cid:29)nity 5.10.8Gaussian Drop-O(cid:28) 1.90.59Data 4.10.75dBpT 3.2 Embedded Network Models
The non-embedded models described in the previous section necessarily ignore a fundamental property
of the brain:
its embedding in physical space. Spatial constraints likely play an important role in
determining the topological properties of brain graphs [22, 29, 28, 26, 27]. In this section, we explore
the topological properties of spatially embedded graphs in which the probability of connecting any two
nodes in the network depends on the Euclidean distance between them [45]. We explore the same
topological diagnostics as we did in the previous section: degree distribution, assortativity, hierarchy,
and diagnostics estimating the topological dimension of the network. As a whole, we find that spatially
embedded models capture more topological features of the empirical networks than models that lack
the physical embedding constraint.
3.2.1 Static Embedded Models
Random Geometric (RG) Model: A random geometric model can be constructed by distributing
nodes uniformly at random in a volume [45, 72, 73]. We employ a classical neurophysiological embedding
in which the x-axis represents the right-left dimension, the y-axis represents the anterior-posterior
dimension, and the z-axis represents the superior-inferior dimension. We use a rectangular volume
where the length of each side is equal to the maximal Euclidean distance between nodes as measured
along that axis and we distribute N nodes uniformly at random within this volume. The M pairs of
nodes with the shortest between-node distance are each connected by an edge.
The heterogeneity of node placement in the volume leads to a broad degree distribution and high
clustering between spatially neighboring nodes, leading to a large network diameter and long path
length (see Figure 8A(i) and Table 2). Because of the homogeneity of the connection rule, which is
identical across all nodes, nodes with high degree (those in close proximity to other nodes) tend to
connect to other nodes of high degree and nodes of low degree (those far from other nodes) tend to
connect to other low degree nodes, leading to degree assortativity (see Figure 8A(ii)). Nodes at the
edges of spatial clusters will tend to have high degree but low clustering, leading to a hierarchical
topology (see Figure 8A(iii)).
Minimally Wired (MW) Model: As noted above, nodes in the RG model are placed uniformly
at random in a volume. To add additional anatomical constraints to the model, we can construct a
minimally wired model (MW) in which nodes are placed at the center of mass of anatomical brain
regions. The M pairs of nodes with the shortest between-node distance are then each connected by an
edge.
The MW provides an interesting point of comparison to the RG because it allows us to assess what
topological properties are driven by the precise spatial locations of brain regions alone. The degree
distribution in the MW is narrower than it is in either the RG or the empirical brain network, likely
because the brain parcellation used in this study is largely grid-like over the cortex (see Figure 8B(i)).
Like the RG, the MW displays degree assortativity and a hierarchical topology (see Figure 8B(ii) -- (iii)),
and has high clustering and long path length. However, in general the diagnostic relationships extracted
from the MW model do not match those of the empirical brain network as well as those extracted from
the RG model.
Distance Drop-Off (DD) Model: Both the minimally wired and the random geometric models
connect only the M pairs of nodes with the shortest inter-node distance. These models therefore
lack long distance connections which are known to be present in the brain, and have been argued
to enable swift communication between distant brain areas [59]. To include this additional biological
characteristic, we next study the distance drop-off model (DD) [46], in which we place nodes at empirical
brain region locations and then connect pairs of nodes with a probability that depends on the distance
r between nodes: P ∝ g(r). Note that the minimally wired model is a special case of the DD model if
we choose P ∝ g(r) to be a step function with threshold r0. Here, however, we fit a function g(r) to the
16
Figure 8: Comparison between the (i) degree distribution (number f of nodes with a given degree ki),
(ii) assortativity (correlation between a node's degree ki and the mean degree of that node's neighbors
k(cid:48)
i, summarized by parameter r), and (iii) hierarchy (the relationship between the clustering coefficient
Ci and the degree ki over all nodes in the network, summarized by parameter β) of the (A) random
geometric (RG), (B) minimally wired (MW), and (C) distance drop-off (DD) models and the same
diagnostics in the brain anatomical data (grey). Black lines indicate best linear fit to the data (dashed)
and model (solid) networks.
connection probability of the empirical data as a function of distance (see Supplementary Material).
The results of the DD model are similar to those that we observed in the case of the minimally
wired and random geometric models (see Figure 8C). However, longer distance connections are present
in this model which decrease the clustering, path length, diameter, and strength of the assortativity
and hierarchy. In general, the diagnostic relationships extracted from the DD model match those of
the empirical brain network significantly better than the same diagnostics extracted from the RG and
MW models.
3.2.2 Embedded Growth Models
Distance Drop-Off Growth (DDG) Model: The random geometric, minimally wired, and dis-
tance drop-off models all have narrower degree distributions than the empirical data. To expand the
degree distribution while still utilizing the empirical node placement and empirically derived probabil-
ity function P ∝ g(r), we construct a distance drop-off growth model (DDG). We begin with M0 seed
edges which we distribute uniformly at random throughout the network. We then choose a node i with
ki > 0 uniformly at random and create an edge between node i and node j (with no constraint on kj)
according to the probability P ∝ g(r). We continue adding edges in this manner until the number of
edges in the network is equal to M .
The degree distribution and assortativity of the DDG are surprisingly similar to that observed
in the empirical data (see Figure 9A(i) -- (ii)). However, the stochasticity of the growth rule induces
a decrease in clustering and we do not observe a hierarchical topology (see Figure 9A(iii)). Neither
the network diameter nor the path length are significantly altered in comparison to the non-growing
distance drop-off model.
Hybrid Distance Growth (HDG) Model: The values for all summary diagnostics reported in
Table 2 for the above models that are most similar to the data are those calculated for the minimally
In a final model, we combine facets of both models in
wired and distance drop-off growth models.
17
AB1030507090020406080kk'r=0.5050100f10110210−2100β=0.3C1030507090020406080kk'r=0.4050100f10110210−2100β=0.37CC1030507090020406080kk'r=0.25050100f10110210−2100β=0.11C(i)(ii)(iii)(i)(ii)(iii)(i)(ii)(iii)Distance Drop-o(cid:31) (DD) ModelMinimally Wired (MW) ModelRandom Geometric (RG) Modeliiiiiiiii Figure 9: Comparison between the (i) degree distribution (number f of nodes with a given degree ki),
(ii) assortativity (correlation between a node's degree ki and the mean degree of that node's neighbors
k(cid:48)
i, summarized by parameter r), and (iii) hierarchy (the relationship between the clustering coefficient
Ci and the degree ki over all nodes in the network, summarized by parameter β) of the (A) distance
drop-off growth (DDG) and the (B) hybrid distance growth (HDG) models and the same diagnostics
in the brain anatomical data (grey). Black lines indicate best linear fit to the data (dashed) and model
(solid) networks. In panel (B), we use 4000 minimized wired seed edges.
a hybrid distance growth model (HDG). We begin by creating a minimally wired model for the M0
shortest connections. We then use the growing rule of the distance drop-off growth model to add the
remaining M − M0 edges to the network. This process can be interpreted as the creation of strongly
connected functional modules that afterwards are cross-connected and embedded in the full network.
Using a derivative-free optimization method, we estimate that the value of M0 that produces a network
most similar to the empirical network is M0 = 4000; see Section 2.
As expected, this hybrid model produces a degree distribution, assortativity, and hierarchy in
between those produced by the minimally wired and distance drop-off growth models and therefore
similar to those observed in the data (see Figure 9B(i) -- (iii)). However, the clustering, diameter, and
path length remain low in comparison to the empirical data (see Table 2), suggesting that this model
does not contain as much local order as the brain.
3.2.3 Diagnostics Estimating the Topological Dimension
In this section, we compare topological measures of the empirical data with the set of 5 embedded
synthetic networks: 3 static models and 2 growth models.
We observe that the estimates of the topological dimension, using both box-counting and Rentian
scaling methods, derived from the physical network models are more similar to the empirical data
than those derived from the topological network models (see Figures 7 and 10). The two highly
locally clustered networks (the minimally wired and random geometric models) have larger diameters
than the brain, decreasing their estimated fractal dimension in comparison. The distance drop-off
and distance drop-off growth models are higher dimensional than the empirical data while the hybrid
distance growth model displays the same dimension as the empirical data. The hybrid model also
produces Rentian scaling with the most similar exponent to that obtained from the empirical data.
The identified similarities between models and empirical data are somewhat surprising given that none
of these models were explicitly constructed to attain a given topological dimension.
18
AB1030507090020406080kk'r=0.09050100f10110210−2100β=−0.14C1030507090020406080kk'r=0.07050100f10110210−2100β=0.4C(i)(ii)(iii)(i)(ii)(iii)Distance Drop-O(cid:31) Growth (DDG) ModelHybrid Distance Growth (HDG) Modeliiiiii 4 Discussion
We examined graph diagnostics of 13 synthetic network models and compared them to those extracted
from empirically derived brain networks estimated from diffusion imaging data [39]. We found that in
general if a model was hard-coded to display one topological property of the brain (e.g., the degree
distribution or the assortativity), it was unlikely to also display a second topological property, suggesting
that a single mechanism is unlikely to account for the complexity of real brain network topology. We also
observed that those models that employed information about node location and inter-node distances
(e.g., embedded network models) were more likely to display similar topological properties to the
empirical data than those that were constructed based on topological rules alone (e.g., non-embedded
network models). In our examination, three models performed noticeably better than all others: the
hybrid distance growing model, the affinity model, and the distance drop-off model. Together, these
results provide us with important insights into the relationships between multiple topological network
properties. Moreover, these model networks form a catalogue of null tests with a range of biological
realism that can be used for statistical inference in static as opposed to dynamic network investigations
[62, 23].
Figure 11 provides a summary of graph diagnostics extracted from real and synthetic model data.
We measure the relative difference between model and data, normalized by the value obtained from
the model that fits the data the least for each diagnostic: (rmodel − rdata)/ max{rall models}. Models are
placed in descending order, from those with the largest relative difference to the data (left-most side of
the graph) to those with the smallest relative difference to the data (right-most side of the graph). We
observe that embedded models generally have a smaller relative distance to the empirical data than
non-embedded models. This result supports the view that physical constraints likely play an important
role in large-scale properties of neurodevelopment.
Figure 10: Diagnostics Estimating the Topological Dimension. (Main Panel) The number of
boxes as a function of the topological size of the box, estimated using the box-counting method [55]
(see Section 2) for the real and embedded model networks. (Inset) The topological Rentian scaling
relationship between the number of edges crossing the boundary of a topological box and the number
of nodes inside of the box (see Section 2) for the real and embedded model networks. Lines indicate
data points included in fits reported in Table 2.
19
234510101001000Random Geometric 3.30.64Minimally Wired 3.10.64Distance Drop-O(cid:31) 5.70.82Distance Drop-O(cid:31) Growth 5.60.86 Hybrid Distance Growth 4.10.81Data 4.10.75# of boxes NBbox size lBdBp248163264128256100100010000# of inter−box connections e# of nodes n per boxT 4.1 Non-embedded Models
We probe non-embedded models with differing amounts and types of structure. While the Erdos-R´enyi
model provides an important benchmark with a random topology, it bears little resemblance to the brain
network. Although a homogeneous random distribution of links has been suggested to characterize the
small-scale structure of neuron-to-neuron connections [74, 75], the large-scale structure of human and
animal brains instead displays heterogeneous connectivity [59]. Perhaps one of the simplest measures
of this heterogeneity is found in the degree distribution, which displays a predominance of low degree
nodes and a long tail of high degree nodes. In comparing the degree distribution of the brain to that
obtained from a BA model, it is clear that this tail, however, is not well-fit by a power-law, a finding
consistent with previous reports in brain anatomy [21, 38] and function [76, 15]. However, by matching
the empirical data, for example using a configuration model with the same degree distribution, we note
that we do not automatically uncover higher order structures like assortativity, suggesting that the
degree distribution provides only limited insight into the forces constraining brain network development.
Several decades ago, neuroanatomists observed that the pattern of connections in several animal
brains displayed a combination of both densely clustered areas and long range projects between distant
areas [77, 78, 79, 80]. The regular lattice and Gaussian drop-off models are able to capture these densely
connected structures but fail to capture the extent of long-range connectivity observed in the brain. The
small-world modular and fractal hierarchical models contain both properties: dense local connectivity
and long-range interactions. The fractal hierarchical model has the added benefit of containing nested
structures, which have been implicated in the heterogeneity of neuronal ensemble activity [11] and in the
separation and integration of information processing across multiple frequency bands [81]. Moreover,
hierarchical modular structure has been identified in organization of white matter streamlines in human
Figure 11: Comparison of the Network Models and Brain Data. (Top Panel) For each model, we
illustrate how summary network statistics (Assortativity r, hierarchy β, clustering C, Rentian scaling
pT , fractal dimension dB, diameter D, mean path length P ) differ from the same statistics extracted
from empirical data. (Main Panel) The black line indicates the sum of the absolute values of the
relative difference between each model and the data. The color image in the background indicates the
difference between the degree distribution of the model and that of the data: red colors indicate that
the model has too many nodes of a given degree, while blue colors indicate that the model has too few
nodes of a given degree. Less saturated colors indicate more similarity between the degree distributions
of the model and the data.
20
DDG diffusion weighted imaging data [8, 64, 66] and implicated in neurobiological phenomena [67, 68, 11].
None of the non-embedded models discussed earlier in this section simultaneously provide a het-
erogeneous degree distribution, degree assortativity, hierarchical topology, and realistic topological
dimensions. Such a "No Free Lunch" rule is perhaps unsurprising, in that a network that is developed
to directly obtain one property typically fails to also display a second property. However, this result
suggests that the topological properties that we explore here are in some sense independent from one
another. In light of the previously reported correlations between network diagnostics in human brain
networks [82] suggesting the need for methods to identify distinguishing properties [83, 70], our find-
ing interestingly suggests that such correlations are potentially specific to the brain system and not
expected theoretically.
Finally, in our affinity model, we hard-code both degree assortativity and a continuous hierarchical
topology, rather than the discrete hierarchy employed in nested models like the fractal hierarchical
model examined here. Interestingly, however, and in contrast to the other non-embedded models, we
simultaneously obtain a heterogeneous degree distribution, and similar estimates of the topological
dimension. This model fits multiple properties of brain networks that were not explicitly included in
the construction of the network model, but are nevertheless a consequence of a three-parameter fit in
the specific affinity model selected. The affinity model therefore serves as a promising candidate as
both a generative model and statistical null model of brain organization.
4.2 Embedded Models
In an effort to include additional biological constraints, we also explore several models that employ
information regarding either the physical placement of network nodes or that place constraints on the
Euclidean lengths of network edges.
In general, this set of networks outperforms most of the non-
embedded network models that we studied, supporting the notion that physical constraints might play
important roles in brain network development and structure [26, 27, 29, 25, 28, 8, 84, 75].
It is important to preface the discussion of our results by mentioning the fact that the properties
of empirically derived brain networks display a heterogeneity that could at least in part stem from
the peculiar physical properties of the organ. Brains are symmetric objects, with the two hemispheres
being connected with one another via tracts in the corpus callosum and via subcortical structures. This
separation allows for a very different topology within a hemisphere than between hemispheres. Moreover,
cortical areas (gray matter) form a shell around the outer edges of the brain while their connections
(white matter) compose the inner volume. Finally, brain areas are inherently heterogeneous in physical
volume, making their distances from one another far from homogeneous. While the morphology of the
brain constrains its potential topological properties, evidence also suggests that the lengths of tracts
connecting brain areas follow a heavy tailed distribution, with short tracts being relatively common
and long tracts being relatively rare [26, 27]. These findings are in concert with the idea that energy
efficiency -- to develop, maintain, and use neuronal wiring -- remains a critical factor in brain evolution
and development [85, 29].
In this study, we begin with a random geometric model, whose nodes are placed uniformly at
random in a volume but whose edges selectively link nodes that are nearby in physical space. In light
of the simplicity of this model, it is somewhat surprising that we obtain such good agreement with the
empirical degree distribution, the presence of assortativity, and the presence of a hierarchical topology.
In the minimally wired graph we employ a similar connection rule but also fix node placement to be
identical to that in the empirical brain network, following previous studies [28]. However, neither of
these two models are able to capture the extent of long-distance connections in the empirical data. By
employing the distance drop-off model, we can fix a connection probability that varies with distance,
rather than simply a connection threshold. This connection probability, however, is not enough to
provide a realistically broad degree distribution. Our distance drop-off growth model combines the
21
strengths of each of these models by laying down a set of seed edges uniformly at random in a volume
and then iteratively adding edges between pairs of nodes according to a probability that falls off with
inter-node distance. The resulting degree distribution and assortativity properties are the best match
to the empirical data of the models that we studied. A hybrid between the minimally wired model and
the distance drop-off growth model does not perform significantly better in matching these properties
and shows a hierarchical structure that is more pronounced than the data.
Importantly, the embedded network models examined here are purposely simplistic. While arbitrar-
ily more complex models could be constructed, our goal was to isolate individual drivers of topology
and probe their relationship to observed network diagnostics. Other studies of interest in relation to
these findings include those that explore the effects of geometric folding [75], radial surface architectures
[84], and the effects of wiring minimization on functional networks [25].
4.3 Biological Interpretations
While the construction of network models is genuinely critical in providing null tests for statistical
inference of brain structure from data, this avenue of research also has the potential to provide key
insights into the neurobiological mechanisms of brain development and function if performed with
appropriate caution. In light of this second use, we note that several of the network models discussed
in this paper employ rules that are reminiscent of -- or even directly inspired by -- known biological
phenomena. For example, physical models that place constraints on the length of connections in
Euclidean space are consistent with the known distribution of connection lengths in the brain and the
modern understanding of metabolic constraints on the development, maintenance, and use of long wires
[85, 29, 26, 27, 28].
However, even topological constraints that link nodes that have similar sets of neighbors can be
interpreted as favoring links between neurons or regions that share similar excitatory input [25]. As an
example, our affinity model hard-codes two inter-node relationships. First, nodes with a similar degree
are more likely to be connected to one another by an edge, leading to degree assortativity throughout
the network. This behavior can be thought of as a mathematical representation of the intuitive principle
of spatial homophily: large neurons with expansive projections (e.g., pyramidal or basket cells) are more
likely to connect to one another because they densely innervate tissue over large distances. Network
assortativity can also stem from the temporal homophily that occurs during development: neurons
that migrate over longer distances during development are more likely to come into contact with -- and
therefore generate a synapse with -- one another than neurons that migrate over shorter distances. The
second topological relationship hard-coded into the affinity model is the prevalence of clustering in local
neighborhoods, a property consistent with physical constraints on network development. As neurons
develop, it is intuitively more likely for them to create synapses with neighboring neurons than non-
neighboring neurons, thereby closing topological loops in close geographic proximity. While we have
only provided a few examples here, links between topological rules and biological phenomena provide
potentially critical neurophysiological context for the development and assessment of synthetic network
models.
4.4 Future Directions
The perspective that we have taken in choosing synthetic network models is one of parsimonious
pragmatism. We seek to identify models with simplistic construction rules or growth mechanisms
to isolate topological (non-embedded) and physical (embedded) drivers of network topology. One
alternative perspective would be to begin with a certain graph topology (for example, an Erdos-R´enyi
graph), and iteratively rewire edges to maximize or minimize a network diagnostic or set of network
diagnostics [25]. However, this approach requires prior hypotheses about which network diagnostics are
most relevant for brain network development, a choice that is complicated by the observed correlations
22
between such diagnostics [82]. Another approach is to employ exponential random graph models [16, 86,
19], which provide a means to generate ensembles of networks with a given set of network properties but
do not provide a means to isolate mechanistic drivers of those network properties. A third approach is
to construct a mechanistic model based on particle-particle collisions, which might serve as a physical
analogy to the biological phenomena of neuronal migration through chemical gradients [87, 88].
In
each of these cases, a perennial question remains: at what spatial scale should we construct these
models to gain the most insight into the relevant biology? Important future directions could include the
development of multiscale growth models, enabling us to bridge the scales between neuronal mechanisms
and large-scale structure.
4.5 Methodological Limitations
There remain important limitations to our work. First, the development of high resolution imaging
methods and robust tractography algorithms to resolve crossing fibers are fast-evolving areas of research.
Novel imaging techniques have for example recently identified the existence of 90 degree turns in white
matter tracts [89], a biological marker that we are not sensitive to in our data. Secondly, we have
focused on understanding the (binary) topology of brain network architecture rather than its weighted
connection strengths in this first study. Our choice was informed by three factors: 1) An understanding
of the relationship between synthetic network models and brain network topology could be useful for
informing a similar investigation into network geometry, 2) In these particular networks, node degree
(binary) and node strength (weighted by the number of streamlines) are strongly correlated (Pearson's
correlation coefficient r = 0.41, p = 1 × 10−41) and therefore topology serves as a proxy for weighted
connectivity, and 3) The choice of how to weight the edges in an anatomical network derived from
diffusion imaging is an open one [90], and therefore investigations independent of these choices are
particularly useful. Finally, our models could be extended to include additional physical features of the
human brain such as bilateral symmetry, the topological ramifications of which are not well understood.
4.6 Conclusion
In this paper, we have examined the mechanistic drivers of network topologies by employing and devel-
oping a range of synthetic network models governed by both topological (non-embedded) and physical
(embedded) rules and comparing them to empirically derived brain networks. We hope that these tools
will be useful in the statistical inference of anatomical brain network structure from neuroimaging data
and that future work can build on these findings to identify neurobiologically relevant mechanisms for
healthy brain architecture and its alteration in disease states.
Acknowledgments
This work was supported by the Errett Fisher Foundation, Sage Center for the Study of the Mind, Ful-
bright Foundation, Templeton Foundation, David and Lucile Packard Foundation, PHS Grant NS44393,
German Academic Exchange Service and the Institute for Collaborative Biotechnologies through con-
tract no. W911NF-09-D-0001 from the U.S. Army Research Office, and the National Science Foundation
(DMS-0645369). The funders had no role in study design, data collection and analysis, decision to pub-
lish, or preparation of the manuscript. We thank Olaf Sporns for giving us access to the human diffusion
spectrum imaging data used in this work.
23
0
0
1
r
o
f
d
e
t
r
o
p
e
r
s
i
n
o
i
t
a
i
v
e
d
d
r
a
d
n
a
t
s
d
n
a
n
a
e
m
e
h
T
.
s
l
e
d
o
m
k
r
o
w
t
e
n
c
i
t
e
h
t
n
y
s
d
n
a
l
a
e
r
n
i
y
g
o
l
o
p
o
t
k
r
o
w
t
e
n
f
o
s
c
i
t
s
o
n
g
a
i
D
:
2
e
l
b
a
T
l
a
n
o
i
t
i
d
d
a
r
o
f
l
a
i
r
e
t
a
M
y
r
a
t
n
e
m
e
l
p
p
u
S
e
e
s
(
a
t
a
d
n
i
a
r
b
e
h
t
r
o
f
d
e
t
r
o
p
e
r
s
i
e
u
l
a
v
l
a
c
i
r
i
p
m
e
e
h
t
d
n
a
l
e
d
o
m
k
r
o
w
t
e
n
c
i
t
e
h
t
n
y
s
h
c
a
e
f
o
s
n
o
i
t
a
z
i
l
a
e
r
d
n
a
c
i
t
s
i
n
i
m
r
e
t
e
d
t
o
n
e
r
a
g
n
i
l
a
c
s
n
a
i
t
n
e
R
d
n
a
n
o
i
s
n
e
m
d
i
l
a
t
c
a
r
f
g
n
i
i
n
m
r
e
t
e
d
r
o
f
s
m
h
t
i
r
o
g
l
a
e
h
t
t
a
h
t
e
t
o
N
.
)
s
t
n
a
p
i
c
i
t
r
a
p
r
e
h
t
o
r
o
f
s
t
l
u
s
e
r
.
s
e
m
i
t
e
l
p
i
t
l
u
m
k
r
o
w
t
e
n
e
m
a
s
e
h
t
o
t
s
m
h
t
i
r
o
g
l
a
e
h
t
g
n
i
y
l
p
p
a
y
b
d
e
n
i
a
t
b
o
e
b
n
a
c
s
e
t
a
m
i
t
s
e
e
l
p
i
t
l
u
m
e
r
o
f
e
r
e
h
t
averageshortestpathlengthP
1
.
3
9
7
7
7
0
0
0
0
2
.
.
.
.
.
0
3
0
0
0
0
±
±
±
±
±
±
3
2
4
4
4
7
8
1
.
4
4
4
.
2
.
.
.
8
2
2
2
1
1
3
.
.
0
0
±
±
4
5
.
.
2
2
6
5
1
.
.
.
0
0
0
±
±
±
2
5
7
.
.
.
4
5
2
2
1
.
.
0
0
±
±
6
6
.
.
2
2
diameterD
6
4
6
4
1
5
.
.
.
.
.
0
0
0
0
0
±
±
±
±
±
7
2
8
0
5
.
.
.
.
.
7
8
3
5
3
3
3
4
1
4
.
.
0
0
±
±
0
8
.
.
4
4
5
.
0
±
4
.
0
1
3
.
0
1
±
1
9
.
4
5
5
.
.
0
0
±
±
0
6
.
.
5
4
topologicalfractaldimensiondB
RentianscalingpT
globalclusteringC[%]
hierarchyβ
assortativityr
widthofdegreedistributionminki -- maxki
modedegree(#)
1
.
0
±
7
.
3
3
0
0
.
0
±
5
4
7
.
0
5
.
1
4
7
4
2
.
0
9
4
1
.
0
7
8
--
0
)
0
4
(
5
2
&
3
2
3
0
2
1
1
.
.
.
.
0
2
0
0
0
1
±
±
±
±
±
±
2
3
0
3
1
9
.
.
.
0
4
0
4
.
1
7
3
.
.
0
0
±
±
4
6
.
.
3
3
3
5
1
3
3
2
0
0
0
0
0
0
0
0
0
0
0
0
.
.
.
.
.
.
0
0
0
0
0
0
±
±
±
±
±
±
2
3
6
2
3
8
5
5
8
6
5
9
9
9
3
4
9
8
.
.
.
.
.
.
0
0
0
0
0
0
4
1
0
0
0
.
.
0
0
±
±
6
4
8
5
.
9
0
.
0
4
5
8
5
0
2
0
0
0
1
.
.
.
.
.
.
0
0
0
0
0
0
±
±
±
±
±
±
0
8
1
0
2
2
.
9
.
7
3
6
5
.
5
.
.
.
5
1
2
4
3
7
9
0
.
1
0
±
±
3
6
1
4
.
3
2
1
3
.
.
.
0
0
0
±
±
±
4
7
4
.
.
.
2
2
4
5
6
.
.
0
0
±
±
3
5
.
.
4
4
6
8
4
2
0
0
0
0
0
0
.
.
.
0
0
0
±
±
±
4
9
5
4
1
3
6
8
6
.
.
.
0
0
0
6
4
0
0
0
0
.
.
0
0
±
±
9
1
4
2
8
8
.
.
0
0
4
.
0
±
9
.
3
5
2
.
0
6
±
6
.
4
4
5
.
6
1
6
3
.
.
0
0
±
±
4
7
.
.
6
7
1
1
7
3
2
2
5
7
0
0
0
0
0
0
.
0
.
.
.
.
.
0
0
0
0
0
±
±
±
±
±
±
1
5
5
4
1
6
1
1
5
1
4
1
.
0
.
.
.
.
.
0
0
0
0
0
−
4
8
0
0
.
.
0
0
±
±
3
5
0
1
.
.
0
0
3
0
.
0
±
0
3
.
0
4
0
.
7
0
5
±
6
3
.
2
0
1
.
0
9
4
0
0
.
0
.
0
±
±
0
1
1
3
.
0
.
0
−
9
9
8
0
1
0
0
1
1
0
0
0
0
0
0
.
.
0
0
.
.
.
.
0
0
0
0
±
±
±
±
±
±
2
1
2
1
2
2
0
0
0
1
0
0
0
.
0
0
0
.
.
.
0
0
0
−
.
.
0
0
−
0
1
3
0
0
.
0
.
0
±
±
5
7
3
0
0
.
.
0
0
−
)
2
±
5
4
(
--
)
2
±
2
1
(
)
2
)
)
)
±
2
1
8
2
±
±
4
(
4
)
--
7
4
7
(
)
4
3
--
6
8
(
(
)
.
(
--
--
0
5
--
)
)
.
±
)
0
1
1
0
±
±
±
5
(
.
4
4
2
8
2
1
1
.
7
(
(
(
1
(
)
5
)
±
9
±
3
7
0
(
8
--
(
)
--
1
)
.
0
2
±
±
4
0
(
.
1
(
)
6
±
6
8
(
1
±
7
2
)
8
)
)
±
)
)
5
5
0
0
6
±
±
4
2
3
(
±
1
1
3
5
(
9
9
0
2
(
(
8
7
.
1
1
4
&
0
(
±
±
±
7
3
6
6
2
3
2
2
2
.
7
2
)
)
4
5
±
±
5
6
5
4
(
(
2
6
±
±
3
7
2
3
0
.
0
±
3
5
.
0
1
0
.
1
0
7
±
9
3
.
3
0
2
.
0
2
0
1
.
.
0
0
±
±
1
7
0
.
0
.
0
)
3
±
0
5
(
--
)
1
±
4
(
9
4
--
7
)
3
±
3
5
(
--
)
6
.
0
±
3
.
1
(
)
5
)
6
6
±
±
6
1
6
3
(
2
--
(
)
--
1
)
1
±
±
7
(
2
(
)
4
±
1
5
(
)
4
±
)
8
5
6
6
(
(
3
±
9
2
2
8
2
±
6
2
)
)
5
4
±
±
0
5
4
5
(
(
2
6
±
±
4
6
2
1
k
r
o
w
t
e
N
l
a
c
i
r
i
p
m
E
n
i
a
r
B
d
e
d
d
e
b
m
e
-
n
o
N
c
i
t
a
t
S
R
E
F
C
L
R
D
G
S
M
H
F
h
t
w
o
r
G
A
B
F
A
d
e
d
d
e
b
m
E
c
i
t
a
t
S
G
R
W
M
D
D
h
t
w
o
r
G
G
D
D
G
D
H
24
Supplementary Materials
Extracting Empirical Connection Probability Drop-off
In the distance drop-off (DD), distance drop-off growth (DDG), and hybrid distance growth (HDG)
models we used a connection density g that depends on the Euclidean distance r between a pair of
nodes. To tune our models, we estimated the empirical relationship between connection density and
distance, hereafter referred to as connection probability drop-off, from the brain data. We observed
that the probability drop-off of connections that lay within a single hemisphere displayed significantly
different behavior than the probability drop-off of connections that lay between hemispheres. We did
not observe systematic differences between the connection probability drop-off in the two hemispheres,
and therefore treat all within-hemispheric connections identically without distinguishing connections
that lay in the left hemisphere from connections that lay in the right hemisphere.
To estimate the connection probability drop-off function g(r) from the empirical data, we used
an adaptive binning algorithm that determines
# possible connections for each bin width ∆r, where ∆r
is chosen to ensure that each bin contains on average 50 connections. To fit the data with sufficient
precision, we first define a truncated power-law function using the form
# connections
f (x) = cxα exp (−λx)
(6)
where c is a constant, x is the minimum physical distance of connections in the bin, α is the power-law
exponent, and λ is the natural exponent. However, we observed that this single truncated power-law
fit was unable to adequately estimate the preponderance of long distance connections in the data. We
therefore used a piecewise function with two truncated power-law functions of the form
(cid:40)
g(x) =
f (x)
1
2 f (x) + 1
2 f (x0) x
x0
−γ
if x > x0
if x < x0
(7)
where f (x) is as defined in Equation 6, λ is the power-law exponent for all bins in which x < x0, x0
is the minimum physical distance at which the truncated power-law function begins to fit the data,
and γ is the power-law exponent for all bins in which x > x0. To minimize boundary and resolution
effects, we excluded the bin with the smallest minimum physical distance and the bin with the largest
minimum physical distance.
Algorithmically, we first fit a single truncated power-law function to the data to obtain estimates
for the parameters c, α, and λ (see dashed line in Figure 12). We then used these estimates as initial
parameters in the fit of Equation 7 to the data (see solid line in Figure 12). To obtain boundary values
for the function, we performed a linear interpolation from the bin with the smallest minimum physical
distance to the boundary value g(0) = 1 (see green line in Figure 12).
The results of this model, which contains 5 tunable parameters (c,α,λ,x0,γ), are shown in Figure
12, where we observe a good agreement between the fit and the data. The estimated parameter values
are provided in Table 3.
Type
Intra-Hemisphere
Inter-Hemisphere
c
0.1490
0.0033
α
1.6608
3.4492
λ
0.2188
0.2695
x0
30.8774
36.1939
γ
0.8890
4.0198
Table 3: Parameter Estimates for Empirical Connection Density Drop-Off for the fits of
Equation 7 to intra- and inter-hemispheric data.
25
Figure 12: Empirical Connection Probability Drop-Off with Physical Distance The connection
probability drop-off g(r) for (A) intra- and (B) inter-hemispheric connections. Empirical brain data
is given by the data points: red indicates bins that were not utilized in the fits, blue indicates bins
in which x < x0, cyan indicates bins in which x > x0, green indicates outlier bins excluded from fit.
Fits are given by the lines: dotted line indicates the initial single truncated power-law fit, solid black
line indicates the piecewise truncated power-law fit, and solid green indicates the piecewise truncated
power-law fit with the interpolation to g(0) = 1.
26
Physical DistanceConnection Probability Drop-O(cid:31)10−210010210−210−1100Intra-Hemispheric Connections10−210010210−210−1100Inter-Hemispheric ConnectionsPhysical DistanceABConnection Probability Drop-O(cid:31) Figure 13: Reliability of Relational Properties Across Data Sets. The (i) degree distribution
(number f of nodes with a given degree ki), (ii) assortativity (correlation between a node's degree
ki and the mean degree of that node's neighbors k(cid:48)
i, summarized by parameter r), and (iii) hierarchy
(the relationship between the clustering coefficient Ci and the degree ki over all nodes in the network,
summarized by parameter β) for each of the six data sets separately shown in panels (A)-(F). In panel
(A), data set 1 shown in grey was used in the visualizations provided in the main manuscript.
27
AB1030507090020406080Data set 1kk'r=0.15050100f10110210−2100β=0.25CData Set 11030507090020406080Data set 1kk'r=0.15050100f10110210−2100β=0.36CData Set 2C1030507090020406080Data set 1kk'r=0.17050100f10110210−2100β=0.24CData Set 3DEF1030507090020406080Data set 1kk'r=0.17050100f10110210−2100β=0.31C1030507090020406080Data set 1kk'r=0.2050100f10110210−2100β=0.32C1030507090020406080Data set 1kk'r=0.26050100f10110210−2100β=0.27CData Set 4Data Set 5Data Set 6(i)(ii)(iii)(i)(ii)(iii)(i)(ii)(iii)(i)(ii)(iii)(i)(ii)(iii)(i)(ii)(iii)iiiiiiiiiiiiiiiiii Figure 14: Reliability of the Topological Dimension Estimates Across Data Sets. (Main
Panel) The number of boxes as a function of the topological size of the box, estimated using the
box-counting method [55] (see Methods) for the six empirical brain data sets. (Inset) The topological
Rentian scaling relationship between the number of edges crossing the boundary of a topological box
and the number of nodes inside of the box (see Methods) for the six empirical brain data sets.
28
234510101001000 Data Set 1 4.20.76 Data Set 2 4.00.72 Data Set 3 4.20.75 Data Set 4 3.80.75 Data Set 5 4.40.74 Data Set 6 4.10.75# of boxes NBbox size lBdBp248163264128256100100010000# of interbox connections e# of nodes n per box References
[1] Sporns O (2013) Making sense of brain network data. Nat Methods 10: 491 -- 493.
[2] Kaiser M (2011) A tutorial in connectome analysis: topological and spatial features of brain
networks. Neuroimage 57: 892 -- 907.
[3] Alstott J, Breakspear M, Hagmann P, Cammoun L, Sporns O (2009) Modeling the impact of
lesions in the human brain. PLoS Comput Biol 5: e1000408.
[4] Honey CJ, Sporns O, Cammoun L, Gigandet X, Thiran JP, et al. (2009) Predicting human resting-
state functional connectivity from structural connectivity. Proc Natl Acad Sci U S A 106: 2035 -- 40.
[5] Griffa A, Baumann PS, Thiran JP, Hagmann P (2013) Structural connectomics in brain diseases.
Neuroimage Epub Ahead of Print.
[6] Fornito A, Zalesky A, Pantelis C, Bullmore ET (2012) Schizophrenia, neuroimaging and connec-
tomics. Neuroimage Epub ahead of print.
[7] Behrens TE, Sporns O (2012) Human connectomics. Curr Opin Neurobiol 22: 144 -- 153.
[8] Bassett DS, Brown JA, Deshpande V, Carlson JM, Grafton ST (2011) Conserved and variable
architecture of human white matter connectivity. Neuroimage 54: 1262 -- 1279.
[9] Cammoun L, Gigandet X, Meskaldji D, Thiran JP, Sporns O, et al. (2012) Mapping the human
connectome at multiple scales with diffusion spectrum MRI. J Neurosci Methods 203: 386 -- 397.
[10] Hermundstad AM, Bassett DS, Brown KS, Aminoff EM, Clewett D, et al. (2013) Structural foun-
dations of resting-state and task-based functional connectivity in the human brain. Proc Natl
Acad Sci U S A 110: 6169 -- 6174.
[11] Rubinov M, Sporns O, Thivierge JP, Breakspear M (2011) Neurobiologically realistic determinants
of self-organized criticality in networks of spiking neurons. PLoS Comput Biol 7: e1002038.
[12] Bassett DS, Meyer-Lindenberg A, Weinberger DR, Coppola R, Bullmore E (2009) Cognitive fitness
of cost-efficient brain functional networks. Proc Natl Acad Sci USA 106: 11747-52.
[13] Supekar K, Menon V, Rubin D, Musen M, Greicius MD (2008) Network analysis of intrinsic
functional brain connectivity in Alzheimer's disease. PLoS Comput Biol 4: e1000100.
[14] He Y, Chen ZJ, Evans AC (2007) Small-world anatomical networks in the human brain revealed
by cortical thickness from mri. Cereb Cortex 17: 2407-2419.
[15] Achard S, Salvador R, Whitcher B, Suckling J, Bullmore E (2006) A resilient, low-frequency, small-
world human brain functional network with highly connected association cortical hubs. J Neurosci
26: 63-72.
[16] van Wijk BC, Stam CJ, Daffertshofer A (2010) Comparing brain networks of different size and
connectivity density using graph theory. PLoS One 5: e13701.
[17] Fornito A, Zalesky A, Breakspear M (2013) Graph analysis of the human connectome: Promise,
progress, and pitfalls. Neuroimage Epub ahead of print.
[18] Johansen-Berg H (2013) Human connectomics - what will the future demand? Neuroimage Epub
Ahead of Print.
29
[19] Simpson SL, Moussa MN, Laurienti PJ (2012) An exponential random graph modeling approach
to creating group-based representative whole-brain connectivity networks. Neuroimage 60: 1117 --
1126.
[20] Rubinov M, Sporns O (2011) Weight-conserving characterization of complex functional brain net-
works. Neuroimage 56: 2068 -- 2079.
[21] Humphries MD, Gurney K, Prescott TJ (2006) The brainstem reticular formation is a small-world,
not scale-free, network. Proc Biol Sci B 273: 503-511.
[22] Bullmore ET, Bassett DS (2011) Brain graphs: graphical models of the human brain connectome.
Annu Rev Clin Psychol 7: 113 -- 140.
[23] Bassett DS, Porter MA, Wymbs NF, Grafton SF, Carlson JM, et al. (2013) Robust detection of
dynamic community structure in networks. Chaos 23.
[24] Cahalane DJ, Clancy B, Kingsbury MA, Graf E, Sporns O, et al. (2011) Network structure implied
by initial axon outgrowth in rodent cortex: empirical measurement and models. PLoS One 6:
e16113.
[25] V´ertes PE, Alexander-Bloch AF, Gogtay N, Giedd JN, Rapoport JL, et al. (2012) Simple models
of human brain functional networks. Proc Natl Acad Sci U S A 109: 5868 -- 5873.
[26] Kaiser M, Hilgetag CC (2006) Non-optimal component placement, but short processing paths, due
to long-distance projections in neural systems. PLoS Comput Biol 2: e95.
[27] Chen BL, Hall DH, Chklovskii DB (2006) Wiring optimization can relate neuronal structure and
function. Proc Natl Acad Sci U S A 103: 4723 -- 4728.
[28] Bassett DS, Greenfield DL, Meyer-Lindenberg A, Weinberger DR, Moore S, et al. (2010) Effi-
cient physical embedding of topologically complex information processing networks in brains and
computer circuits. PLoS Comput Biol 6: e1000748.
[29] Bullmore E, Sporns O (2012) The economy of brain network organization. Nat Rev Neurosci 13:
336 -- 349.
[30] Bassett DS, Gazzaniga MS (2011) Understanding complexity in the human brain. Trends Cogn
Sci 15: 200 -- 209.
[31] Newman MEJ (2010) Networks: An Introduction. Oxford University Press.
[32] Alexander-Bloch A, Lambiotte R, Roberts B, Giedd J, Gogtay N, et al. (2012) The discovery of
population differences in network community structure: new methods and applications to brain
functional networks in schizophrenia. Neuroimage 59: 3889 -- 3900.
[33] Ravasz E, Barab´asi AL (2003) Hierarchical organization in complex networks. Phys Rev E 67:
026112.
[34] Newman MEJ (2006) Modularity and community structure in networks. Proc Natl Acad Sci USA
103: 8577 -- 8582.
[35] Christie P, Stroobandt D (2000) The interpretation and application of Rent's Rule. IEEE Trans
VLSI Syst 8: 639-648.
[36] Greenfield DL (2010) Communication Locality in Computation: Software, Chip Multiprocessors
and Brains. Ph.D. thesis, University of Cambridge.
30
[37] Song C, Havlin S, Makse HA (2005) Self-similarity of complex networks. Nature 433: 392-395.
[38] Bassett DS, Bullmore ET, Verchinski BA, Mattay VS, Weinberger DR, et al. (2008) Hierarchical
organization of human cortical networks in health and schizophrenia. J Neurosci 28: 9239-9248.
[39] Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ, et al. (2008) Mapping the structural
core of human cerebral cortex. PLoS Biol 6: e159.
[40] Bollob´as B (2001) Random Graphs. Cambridge University Press.
[41] Newman MEJ (2003) The sturcture and function of complex networks. SIAM Review 45: 167 -- 256.
[42] May PJC, Tiitinen H (2011) Speech recognition based on the processing solutions of auditory
cortex. In: Honkela T, Duch W, Girolami M, Kaski S, editors, Artificial Neural Networks and
Machine Learning -- ICANN 2011, Springer Berlin Heidelberg, volume 6792 of Lecture Notes in
Computer Science. pp. 421 -- 428.
[43] Rubinov M, Sporns O (2009) Complex network measures of brain connectivity: uses and interpre-
tations. Neuroimage Epub Ahead of Print.
[44] Albert R, Barab´asi AL (2002) Statistical mechanics of complex networks. Reviews of Modern
Physics 74.
[45] Barthelemy M (2010) Spatial networks. arXiv .
[46] Avin C (2008) Distance graphs: From random geometric graphs to bernoulli graphs and between.
In: Proceedings of the fifth international workshop on Foundations of mobile computing. pp. 71-78.
[47] Zalesky A, Fornito A, Harding IH, Cocchi L, Yucel M, et al. (2010) Whole-brain anatomical
networks: does the choice of nodes matter? Neuroimage 50: 970 -- 983.
[48] Dennis EL, Jahanshad N, Toga AW, McMahon KL, de Zubicaray GI, et al. (2012) Test-retest
reliability of graph theory measures of structural brain connectivity. Med Image Comput Comput
Assist Interv 15: 305 -- 312.
[49] Owen JP, Ziv E, Bukshpun P, Pojman N, Wakahiro M, et al. (2013) Test-retest reliability of
computational network measurements derived from the structural connectome of the human brain.
Brain Connect 3: 160 -- 176.
[50] Cheng H, Wang Y, Sheng J, Kronenberger WG, Mathews VP, et al. (2012) Characteristics and
variability of structural networks derived from diffusion tensor imaging. Neuroimage 61: 1153 --
1164.
[51] Vaessen MJ, Hofman PA, Tijssen HN, Aldenkamp AP, Jansen JF, et al. (2010) The effect and
reproducibility of different clinical DTI gradient sets on small world brain connectivity measures.
Neuroimage 51: 1106 -- 1116.
[52] Newman ME (2002) Assortative mixing in networks. Phys Rev Lett 89: 208701.
[53] Watts DJ, Strogatz SH (1998) Collective dynamics of 'small-world' networks. Nature 393: 440-442.
[54] Ozaktas HM (1992) Paradigms of connectivity for computer circuits and networks. Opt Eng 31:
1563-1567.
[55] Concas G, Locci MF, Marchesi M, Pinna S, Turnu I (2006) Fractal dimension in software networks.
Europhys Lett 76: 1221-1227.
31
[56] Dijkstra EW (1959) A note on two problems in connexion with graphs. Numerische Mathematik
1: 269 -- 271.
[57] Karypis G, Aggarwal R, Kumar V, Shekhar S (1997) Multilevel hypergraph partitioning: Appli-
cations in vlsi domain. In: 34th Design and Automation Conference. pp. 526 - 529.
[58] Maslov S, Sneppen K (2002) Specificity and stability in topology of protein networks. Science 296:
910-913.
[59] Sporns O (2010) Networks of the Brain. MIT Press.
[60] Bullmore E, Sporns O (2009) Complex brain networks: graph theoretical analysis of structural
and functional systems. Nat Rev Neurosci 10: 186 -- 198.
[61] Meunier D, Lambiotte R, Fornito A, Ersche KD, Bullmore ET (2009) Hierarchical modularity in
human brain functional networks. Front Neuroinformatics 3: 37.
[62] Bassett DS, Wymbs NF, Porter MA, Mucha PJ, Carlson JM, et al. (2011) Dynamic reconfiguration
of human brain networks during learning. Proc Natl Acad Sci U S A 10: 7641 -- 7646.
[63] Bassett DS, Wymbs NF, Rombach MP, Porter MA, Mucha PJ, et al. (2013) Core-periphery or-
ganisation of human brain dynamics In revision.
[64] Gallos LK, Sigman M, Makse HA (2012) The conundrum of functional brain networks: small-world
efficiency or fractal modularity. Front Physiol 3.
[65] Meunier D, Lambiotte R, Bullmore ET (2010) Modular and hierarchically modular organization
of brain networks. Front Neurosci 4.
[66] Zhou C, Zemanov´a L, Zamora G, Hilgetag CC, Kurths J (2006) Hierarchical organization unveiled
by functional connectivity in complex brain networks. Phys Rev Lett 97: 238103.
[67] Wang SJ, Hilgetag CC, Zhou C (2011) Sustained activity in hierarchical modular neural networks:
self-organized criticality and oscillations. Front Comput Neurosci 5.
[68] Kaiser M, Hilgetag CC (2010) Optimal hierarchical modular topologies for producing limited
sustained activation of neural networks. Front Neuroinform 4.
[69] Sporns O (2006) Small-world connectivity, motif composition, and complexity of fractal neuronal
connections. Biosystems 85: 55 -- 64.
[70] Onnela JP, Fenn DJ, Reid S, Porter MA, Mucha PJ, et al. (2012) Taxonomies of networks from
community structure. Phys Rev E 86: 036104.
[71] Lohse C, Bassett DS, Carlson JM (2013) Multiresolution analysis of cohesive mesoscale structures
in human brain networks. PLoS Comp Biol In Submission.
[72] Penrose M (2003) Random Geometric Graphs. Oxford University Press.
[73] Dall J, Christensen M (2002) Random geometric graphs. Phys Rev E 66: 016121.
[74] Braitenberg V, Schuz A (1998) Cortex: Statistics and Geometry of Neural Connectivity. Springer,
Berlin.
[75] Henderson JA, Robinson PA (2011) Geometric effects on complex network structure in the cortex.
Phys Rev Lett 107: 018102.
32
[76] van den Heuvel MP, Stam CJ, Boersma M, Hulshoff Pol HE (2008) Small-world and scale-free
organization of voxel-based resting-state functional connectivity in the human brain. Neuroimage
43: 528 -- 539.
[77] Scannell JW, Blakemore C, Young MP (1995) Analysis of connectivity in the cat cerebral cortex.
J Neurosci 15: 1463 -- 1483.
[78] Scannell JW, Burns GA, Hilgetag CC, O'Neil MA, Young MP (1999) The connectional organization
of the cortico-thalamic system of the cat. Cerebral Cortex 9: 277 -- 99.
[79] Young MP, Scannell JW, O'Neill MA, Hilgetag CC, Burns G, et al. (1995) Non-metric multidi-
mensional scaling in the analysis of neuroanatomical connection data and the organization of the
primate cortical visual system. Philos Trans R Soc Lond B Biol Sci 348: 281 -- 308.
[80] Felleman DJ, Van Essen DC (1991) Distributed hierarchical processing in primate visual cortex.
Cerebral Cortex 1: 1 -- 47.
[81] He BJ, Zempel JM, Snyder AZ, Raichle ME (2010) The temporal structures and functional sig-
nificance of scale-free brain activity. Neuron 66: 353 -- 369.
[82] Lynall ME, Bassett DS, Kerwin R, McKenna P, Muller U, et al. (2010) Functional connectivity
and brain networks in schizophrenia. J Neurosci : Epub ahead of print.
[83] Bounova G, de Weck O (2012) Overview of metrics and their correlation patterns for multiple-
metric topology analysis on heterogeneous graph ensembles. Phys Rev E 85: 016117.
[84] Raj A, Chen YH (2011) The wiring economy principle: connectivity determines anatomy in the
human brain. PLoS One 6: e14832.
[85] Attwell D, Laughlin SB (2001) An energy budget for signaling in the grey matter of the brain. J
Cereb Blood Flow Metab 21: 1133 -- 1145.
[86] Simpson SL, Hayasaka S, Laurienti PJ (2011) Exponential random graph modeling for complex
brain networks. PLoS One 6: e20039.
[87] Gonz´alez MC, Lind PG, Herrmann HJ (2006) System of mobile agents to model social networks.
Phys Rev Lett 96: 088702.
[88] Lind PG, Herrmann HJ (2007) New approaches to model and study social networks. New J Phys
9.
[89] Wedeen VJ, Rosene DL, Wang R, Dai G, Mortazavi F, et al. (2012) The geometric structure of
the brain fiber pathways. Science 335: 1628 -- 1634.
[90] Rubinov M, Bassett DS (2011) Emerging evidence of connectomic abnormalities in schizophrenia.
J Neurosci 31: 6263-5.
33
|
1509.03162 | 4 | 1509 | 2017-03-02T08:03:15 | Fundamental activity constraints lead to specific interpretations of the connectome | [
"q-bio.NC"
] | The continuous integration of experimental data into coherent models of the brain is an increasing challenge of modern neuroscience. Such models provide a bridge between structure and activity, and identify the mechanisms giving rise to experimental observations. Nevertheless, structurally realistic network models of spiking neurons are necessarily underconstrained even if experimental data on brain connectivity are incorporated to the best of our knowledge. Guided by physiological observations, any model must therefore explore the parameter ranges within the uncertainty of the data. Based on simulation results alone, however, the mechanisms underlying stable and physiologically realistic activity often remain obscure. We here employ a mean-field reduction of the dynamics, which allows us to include activity constraints into the process of model construction. We shape the phase space of a multi-scale network model of the vision-related areas of macaque cortex by systematically refining its connectivity. Fundamental constraints on the activity, i.e., prohibiting quiescence and requiring global stability, prove sufficient to obtain realistic layer- and area-specific activity. Only small adaptations of the structure are required, showing that the network operates close to an instability. The procedure identifies components of the network critical to its collective dynamics and creates hypotheses for structural data and future experiments. The method can be applied to networks involving any neuron model with a known gain function. | q-bio.NC | q-bio | Fundamental activity constraints lead to
specific interpretations of the connectome
Jannis Schuecker1,Y
, Maximilian Schmidt 1,Y
, Sacha J. van Albada1
, Markus
Diesmann1,2,3
, and Moritz Helias1,3
1Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced Simulation (IAS-6)
and JARA BRAIN Institute I, Julich Research Centre, 52428 Julich, Germany
2Department of Psychiatry, Psychotherapy and Psychosomatics, Medical Faculty, RWTH Aachen
University, 52062 Aachen, Germany
3Department of Physics, Faculty 1, RWTH Aachen University, 52062 Aachen, Germany
YBoth authors contributed equally to this work.
7
1
0
2
r
a
M
2
]
.
C
N
o
i
b
-
q
[
4
v
2
6
1
3
0
.
9
0
5
1
:
v
i
X
r
a
Correspondence to: Jannis Schuecker, Wilhelm-Johnen-Strasse, 52425 Julich
[email protected]
Activity constraints on the connectome
Schuecker, Schmidt et al.
Abstract
The continuous integration of experimental data into coherent models of the brain is an increasing chal-
lenge of modern neuroscience. Such models provide a bridge between structure and activity, and identify
the mechanisms giving rise to experimental observations. Nevertheless, structurally realistic network
models of spiking neurons are necessarily underconstrained even if experimental data on brain connec-
tivity are incorporated to the best of our knowledge. Guided by physiological observations, any model
must therefore explore the parameter ranges within the uncertainty of the data. Based on simulation
results alone, however, the mechanisms underlying stable and physiologically realistic activity often re-
main obscure. We here employ a mean-field reduction of the dynamics, which allows us to include
activity constraints into the process of model construction. We shape the phase space of a multi-scale
network model of the vision-related areas of macaque cortex by systematically refining its connectivity.
Fundamental constraints on the activity, i.e., prohibiting quiescence and requiring global stability, prove
sufficient to obtain realistic layer- and area-specific activity. Only small adaptations of the structure are
required, showing that the network operates close to an instability. The procedure identifies components
of the network critical to its collective dynamics and creates hypotheses for structural data and future
experiments. The method can be applied to networks involving any neuron model with a known gain
function.
Author summary
The connectome describes the wiring patterns of the neurons in the brain, which form the substrate
guiding activity through the network. The influence of its constituents on the dynamics is a central topic
in systems neuroscience. We here investigate the critical role of specific structural links between neuronal
populations for the global stability of cortex and elucidate the relation between anatomical structure and
experimentally observed activity. Our novel framework enables the evaluation of the rapidly growing
body of connectivity data on the basis of fundamental constraints on brain activity and the combination
of anatomical and physiological data to form a consistent picture of cortical networks.
Introduction
The neural wiring diagram, the connectome, is gradually being filled through classical techniques com-
bined with innovative quantitative analyses (Markov et al., 2014a,b) and new technologies (Wedeen et al.,
2008; Axer et al., 2011). The connectivity between neurons is considered to shape resting-state and task-
related collective activity (Cole et al., 2014; Shen et al., 2015). For simple networks, clear relationships
with activity are known analytically, e.g., a dynamic balance between excitatory and inhibitory inputs
in inhibition-dominated random networks leads to an asynchronous and irregular state (van Vreeswijk &
Sompolinsky, 1996; Amit & Brunel, 1997a; Tetzlaff et al., 2012). Structures and mechanisms underlying
1
Activity constraints on the connectome
Schuecker, Schmidt et al.
Figure 1 The integrative loop between experiment, model, and theory. Black arrows represent the
classical forward modeling approach: Experimental anatomical data are integrated into a model structure,
which gives rise to the activity via simulation. The model activity is compared with experimental functional
data. The usual case of disagreement leads to the need to change the model definition. By experience and
intuition the researcher identifies the parameters to be changed, proceeding in an iterative fashion. Once the
model agrees well with a number of experimental findings, it may also generate predictions on the outcome of
new experiments. Red arrows symbolize our new approach: informed by functional data, an analytical theory
of the model identifies critical connections to be modified, refining the integration of data into a consistent
model structure and generating predictions for anatomical studies.
large-scale interactions have been identified by means of dynamical models (Deco & Jirsa, 2012; Cabral
et al., 2014) and graph theory (Markov et al., 2013; Goulas et al., 2014). Furthermore, the impact of
local network structure on spike-time correlations is known in some detail (Ostojic et al., 2009; Pernice
et al., 2011; Trousdale et al., 2012). Under certain conditions, there is a one-to-one mapping between cor-
relations in neuronal network activity and effective connectivity, a measure that depends on the network
structure and on its activity (van Albada et al., 2015). In conclusion, anatomy does not uniquely deter-
mine dynamics, but dynamical observations help constrain the underlying anatomy. Despite advances in
understanding simple networks, a complete picture of the relationship between structure and dynamics
remains elusive.
Fig. 1 visualizes the integrative loop between experiment, model, and theory that guides the inves-
tigation of structure and dynamics. In the traditional forward modeling approach, structural data from
experimental studies determine the connectivity between model neurons. Combined with the specification
of the single-neuron dynamics and synaptic interactions, simulations yield a certain network dynamics.
There is a fundamental problem with this approach.
Despite the impressive experimental progress in determining network parameters, any neural network
model is underdetermined, because of the sheer complexity of brain tissue and inevitable uncertainties in
the data. For instance, counting synapses on a large scale presently takes a prohibitive amount of time, and
no available technique allows determining precise synaptic weights for entire neural populations. Although
it may be possible to quantify the full connectome of an individual in future, inter-individual variability
will require modelers to express connectivity as connection probabilities or rules of self-organization if
they want to learn about general principles of the brain. In modeling studies, parameters are usually
tuned manually to achieve a satisfactory state of activity, which becomes unfeasible for high-dimensional
2
Activity constraints on the connectome
Schuecker, Schmidt et al.
models due to the size of the parameter space. In particular, it is a priori unclear how parameters of the
model influence its activity. In consequence, modifications cannot be performed in a targeted fashion,
and it is difficult to find a minimal set of modifications necessary for aligning a model with experimental
activity data.
Overcoming this problem requires a shift of perspective. Instead of regarding the model exclusively
in a forward manner, generating predicted activity from structure, we in addition consider the system
in a reverse manner, predicting the structure necessary to explain the observed activity. Our theory,
starting from a mean-field description, provides a direct link between structure and activity. In contrast
to simulations, the theory is invertible, which we exploit to identify connections critical for the dynamics
and to find a minimal set of modifications to the structure yielding a realistic set point of activity. The
predicted alterations generate hypotheses on brain structure, thus feeding back to anatomical experiments.
Applying the method to a multi-scale network model of the vision-related areas of macaque cortex, we
derive targeted modifications of a set of critical connections, bringing the model closer to experimental
observations of cortical activity. Based on the global properties of the bistable phase space of the model,
the method refines the model's construction principles within experimental uncertainties and identifies
the connections that decisively shape the dynamics. Preliminary results have been presented in abstract
form (Schmidt et al., 2015).
Results
Global stability in a simple network
We consider networks with neurons structured into N interconnected populations. A neuron in popula-
tion i receives Kij incoming connections from neurons in population j, each with synaptic efficacy Jij.
Additionally, each neuron is driven by Kext Poisson sources with rate νext and synaptic efficacy Jext.
All neurons in one population have identical parameters, so that we can describe the network activity in
terms of population-averaged firing rates νi.
Our method is applicable if the employed neuron model has a known gain function, either analytically
or as a function obtained by interpolating numerical results from simulations. In this study, we model
single cells as leaky integrate-and-fire model (LIF) neurons (see Methods). The possible stationary states
of these networks are characterized by the firing rates that are equilibria of
ν :=
dν
ds
= Φ(ν, A) − ν,
(1)
where s is a pseudo-time. The gain function Φ is known analytically and A indicates its dependence on
the model parameters {K, J , νext, . . .} (see Methods).
The input-output relationship Φ typically features a non-linearity which, in combination with feedback
connections, can cause multi-stability in the network. In particular excitatory-excitatory loops cause the
3
Activity constraints on the connectome
Schuecker, Schmidt et al.
system defined by (1) to exhibit multi-stable behavior in the stationary firing rates. A necessary condition
for the bistability is that the transfer function has an inflection point. The LIF neuron model can have such
an inflection point, originating from the interplay of its leak term and the threshold. Less realistic neuron
models, such as the perfect integrate-and-fire model, do not have such an inflection point. To illustrate
its origin, we first consider the noiseless case (Dayan & Abbott, 2001) without absolute refractoriness
(τr = 0). The transfer function initially grows from zero with infinite slope due to the threshold and
crosses the identity line (Fig. 2C). For larger input the leak term can be neglected and the transfer
function approaches a linear function with finite slope 1
τm
R
Vθ −Vr
(see, e.g., (Kriener et al., 2014), eq.
11), equivalent to a perfect integrator. This is only possible with a negative curvature at intermediate
rates, i.e., a reduction in the slope, which makes the transfer function cross the identity line another
time, causing the bistability. Network-generated noise only affects the low-rate regime where it smears
out the kink causing the transfer function to grow from zero with positive curvature (see, e.g., (Brunel,
2000), eq. 22). Importantly, the qualitative picture, i.e., the bistable behavior, is not altered. A finite
refractory period only has an effect for very high input values where the transfer function saturates at
1/τr = 500 spikes/s for the given parameters.
The basic problem is that there is a trade-off between excitation at the fixed points and their stability.
In particular, exciting the model to bring a fixed point closer to experimental observations requires a
method to preserve its stability. We achieve this by controlling the influence of the excitatory-excitatory
loops on the phase space of the network.
As an illustration we first study the mechanism using the simple network architecture depicted in
Fig. 2A: a population of excitatory neurons is coupled to itself with indegree K and is driven by external
Poisson sources with the same indegree Kext = K and rate νext. All connections have identical synaptic
weights J. An increase in the external drive shifts the input-output relationship Φ(ν, νext) of this one-
dimensional system (Fig. 2C) to the left. The bifurcation diagram is shown in Fig. 2E: for low νext there
is only one fixed point with low activity (LA). When increasing νext, an additional pair of fixed points of
which one is stable and the other is unstable emerges via a saddle-node bifurcation, leading to a bistable
system. The second stable fixed point exhibits high firing rates, denoted as the high-activity (HA) state.
For even higher values of νext, the LA state loses stability.
The equilibria, given by the zeros of the velocity ν in the bistable case, are shown in Fig. 2F. An
increase of the drive on the one hand increases the firing rate of the LA fixed point (inset) but on the
other hand reduces its basin of attraction, indicated by the colored bars in the top left corner. For
illustrative purposes, we extend the problem to two dimensions by splitting the excitatory population
into two subpopulations of equal size (Fig. 2B), mimicking a loop between excitatory populations in the
model of the vision-related areas of macaque cortex. The corresponding (symmetric) two-dimensional
phase space is shown in Fig. 2D. The basin of attraction for the LA fixed point, limited by the separatrix
(Strogatz, 1994), is reduced with increasing external drive.
Since we have a bistable system, there must be at least one unstable fixed point on the separatrix at
4
Activity constraints on the connectome
Schuecker, Schmidt et al.
A
B
E
E
E
s shown in gray.
Figure 2 Activity flow in an illustrative network example. Left column: Global stability analysis in the
single-population network. A Illustration of network architecture. C Upper panel: Input-output relationship
Φ(ν, νext) for external Poisson drive νext = 160 1
In addition Φ(ν, νext) for τr = 0 for the
noiseless case (blue) and the noisy case (red). The inset shows the gray curve over a larger input range. Lower
panel: Φ(ν, νext) for different rates of the external Poisson drive νext = [150, 160, 170] 1
s from black to light
gray. Intersections with the identity line (dashed) mark fixed points of the system, which are shown in E as a
function of νext. F Flux ν in the bistable case for Φ(νext = 160 1
s , K) in blue,
s , K ′) in red. Intersections with zero (dashed line) mark fixed points. The
and modified system Φ(ν ′
inset shows an enlargement close to the LA fixed point. Horizontal bars at top of figure denote the size of
the basin of attraction for each of the three settings. Right column: Global stability analysis in the network
of two mutually coupled excitatory populations. B Illustration of network architecture. D Flow field and
nullclines (dashed curves) for Φ(νext = 160 1
s , K) and separatrices (solid lines), LA fixed point (rectangle),
HA fixed point (cross) and unstable fixed points (circles) for Φ(νext = 160 1
s , K)
in blue, and Φ(νext = 161 1
s , K ′) in red. The red separatrix and the red unstable fixed point coincide with the
black ones. G Enlargement of D close to the LA fixed points. Flow field of original system shown in black, of
modified system in red.
s , K) in black, Φ(νext = 161 1
ext = 161 1
s , K) in black, Φ(ν ′
ext = 161 1
5
Activity constraints on the connectome
Schuecker, Schmidt et al.
the intersection of the nullclines, i.e., the subspace for which the velocity νi in direction i vanishes. We
use the unstable fixed point to preserve the basin of attraction when the external drive νext is increased.
For this purpose, we modify the connectivity K → K ′ to reverse the shift of the unstable fixed point due
to the parameter change νext → ν ′
ext (see Methods for a detailed derivation). Since the separatrix follows
the unstable fixed point, this approximately restores the original basin of attraction.
The resulting velocity of the system Φ(ν ′
ext, K ′) (Φ defines the system (1)) is shown in Fig. 2F. The
firing rate in the LA fixed point is increased as desired (inset), and the unstable fixed point coincides with
that obtained in the original system Φ(νext, K). This pattern of fixed points is also indicated by the zero
vectors of the velocity field ν (Fig. 2D). The separatrix follows the unstable fixed point, and the basin of
attraction in the system Φ(ν ′
ext, K ′) is restored to that in the original system Φ(νext, K). Fig. 2G shows
the behavior of the LA fixed point in more detail. The modification of K does not noticeably change the
location of the LA fixed point. In conclusion, the method allows us to increase the firing rates in the LA
fixed point without modifying its basin of attraction.
The purely excitatory network is the simplest model to explain a phase space configuration with a LA
and a HA fixed point. Inhibitory feedback is not necessary for this bistability, but it would certainly alter
the input-output relationship. For example the classical excitatory-inhibitory network (Brunel, 2000)
in the balanced regime has an input-output relationship with a negative slope and thus only one fixed
point exists. However, if a pair of such balanced E-I networks is coupled by sufficiently strong mutually
excitatory connections, these connections cause a bistability in a similar manner. Thus the mechanism
shown in the purely excitatory network can also lead to the emergence of a HA attractor in more complex
networks with inhibition.
Bistability in the multi-scale network model
We investigate a multi-scale network model of cortical areas to understand the structural features essential
for a realistic state of baseline activity. The model extends and adapts the microcircuit model presented
in (Potjans & Diesmann, 2014), which covers 1 mm2 surface area of early sensory cortex (Fig. 3A), to all
vision-related areas of macaque cortex (Fig. 3B). Based on the microcircuit model, an area is composed
of 4 layers (2/3, 4, 5, and 6) each having an excitatory (E) and an inhibitory (I) neural population, except
parahippocampal area TH, which consists of only 3 layers (2/3, 5, and 6). A detailed description of the
data integration is given in (Schmidt et al., 2016).
Simulations of the model (Fig. 3C) reveal that, though realistic levels of activity can be achieved for
populations in layers 2/3 and 4, populations 5E and 6E of the majority of areas show vanishingly low
or zero activity in contrast to empirical data (Swadlow, 1988; de Kock & Sakmann, 2009). Inputs from
subcortical and non-visual cortical areas are modeled as Poissonian spike trains, whose rate νext is a free,
global parameter. To elevate the firing rates in the excitatory populations in layers 5 and 6, we enhance
the external Poisson drive onto these populations parametrized by κ (see Methods). However, already
6
Activity constraints on the connectome
Schuecker, Schmidt et al.
Figure 3 Bistability of the model. A Sketch of the microcircuit model serving as a prototype for the areas of
the multi-area model (figure and legend adapted from figure 1 of Potjans and Diesmann (Potjans & Diesmann,
2014), with permission). B Sketch of the most common laminar patterns of cortico-cortical connectivity of
the multi-area model. C Population-averaged firing rates encoded in color for a spiking network simulation of
the multi-area model with low external drive (κ = 1.0). D As C but for increased external drive (κ = 1.125).
The color bar refers to both panels. Areas are ordered according to their architectural type along the horizontal
axis from V1 (type 8) to TH (type 2) and populations are stacked along the vertical axis. The two missing
populations 4E and 4I of area TH are marked in black and firing rates < 10−2 Hz in gray. E Histogram
of population-averaged firing rates shown in C for excitatory (blue) and inhibitory (red) populations. The
horizontal axis is split into linear- (left) and log-scaled (right) ranges. F as E corresponding to state shown in
D.
7
Activity constraints on the connectome
Schuecker, Schmidt et al.
a perturbation of a few percent leads to a state with unrealistically high rates (Fig. 3D), caused by the
reduced basin of attraction of the low-activity state similar to Fig. 2D. Our aim is to improve the working
point of the model such that all populations exhibit spiking activity ? 0.05 spikes/s while preventing
the model from entering a state with unrealistically high rates of ? 30 spikes/s (figure 13 of (Swadlow,
1988), (de Kock & Sakmann, 2009)). The employed technique exposes the mechanism giving rise to the
observed instability and identifies the circuitry responsible for this dynamical feature.
Targeted modifications preserve global stability
We apply the procedure derived in Methods and find targeted modifications to the connectivity K that
preserve the global stability of the low-activity fixed point for increased values of the external drive κ.
In the following we choose the inactive state ν(0) = (0, . . . , 0)T as the initial condition. The exact
choice is not essential since we are only interested in the fixed points of the system. Fig. 4B shows the
integration of (1) over pseudo-time s for different levels of the external drive to populations 5E and 6E
parametrized by κ. For low values of κ, the integration converges to the LA fixed point shown in Fig. 4D,
and is in agreement with the activity emerging in the simulation (Fig. 3C). For increased values of κ, the
system settles in the HA fixed point (Fig. 4E), again in agreement with the simulation. The population-
specific firing rates in the HA state found in the mean-field predictions (Fig. 4E) are also close to those in
the simulation (Fig. 3D), but minor deviations occur due to the violation of the assumptions made in the
diffusion approximation. In particular, at these pathologically high rates, the neurons fire regularly and
close to the reciprocal of their refractory period, while in the mean-field theory we assume Poisson spike
statistics. Still, the mean-field theory predicts the bistability found in the simulation. Since the theory
yields reliable predictions in both stable fixed points, we assume that also the location of the unstable
fixed point in between these two extremes is accurately predicted by the theory.
To control the separatrix we need to find the unstable fixed point of the system. This is nontrivial
since the numerical integration of (1) for finding equilibria by construction only converges to stable fixed
points.
If the unstable fixed point has only one repelling direction (Fig. 5A), it constitutes a stable
attractor on the N − 1 dimensional separatrix. The separatrix is a stable manifold (Strogatz, 1994), and
therefore a trajectory originating in its vicinity but not near an unstable fixed point initially stays in the
neighborhood. If an initial condition just outside the separatrix is close to the basin of attraction of a
particular unstable fixed point, the trajectory initially approaches the latter. Close to the fixed point
the velocity is small. Ultimately trajectories diverge from the separatrix in the fixed point's unstable
direction, as illustrated in Fig. 4A. In conclusion, we expect a local minimum in the velocity along the
trajectories close to the unstable fixed point. To estimate the location of the unstable fixed point in this
manner, we need to find initial conditions close to the separatrix. Naively, we would just fix the value of κ
and vary the initial condition. However, due to the high dimensionality of our system this is not feasible
in practice. Instead, we vary κ for a fixed initial condition. Fig. 4B shows the firing rate averaged across
8
Activity constraints on the connectome
Schuecker, Schmidt et al.
Stable FP
Unstable FP
Separatrix
Figure 4 Application of the mean-field theory to the multi-area model. A Trajectories of (1) starting
inside the separatrix converge to an unstable fixed point. Trajectories starting close to the separatrix are
initially attracted by the unstable fixed point but then repelled in the fixed point's unstable direction and
finally converge to a stable fixed point. B Firing rate averaged across populations over time. Integration of
(1) leads to convergence to either the low-activity (LA) or the high-activity (HA) attractor for different choices
of the external input factor κ, with κ = 1 the original level of external drive. We show eight curves with κ
varying from 1.0 to 1.014 in steps of 0.002 and two additional curves for κ = 1.007662217, 1.007662218. The
curves for the largest factor (κ = 1.007662217) that still leads to the LA state and for the smallest factor
(κ = 1.007662218) that leads to the HA state are marked in blue and red, respectively. The four curves with
κ ≤ 1.006 coincide with the blue curve. C Euclidean norm of the velocity vector in the integration of (1) for
the different choices of κ. The vertical dashed line indicates the time sc of the last local minimum in the blue
curve. D Stationary firing rate in the different areas and layers of the model in a low-activity state for κ = 1.0
as predicted by the mean-field theory (same display as in Fig. 3). E As D, but showing the high-activity state
for κ = 1.125.
9
Activity constraints on the connectome
Schuecker, Schmidt et al.
Figure 5 Eigenspectrum analysis of network stability. A Eigenvalue spectrum of the effective connectivity
matrix M for the first (blue squares) and second (red dots) iteration. The dashed vertical line marks the
edge of stability at a real part of 1. B Contribution ηl ((13)) of an individual eigenprojection l to the shift of
the unstable fixed point versus the relative change in indegrees associated with l for the first iteration. The
data point corresponding to λ(1)
is marked in red. The inset shows the relative angles between δν ∗ and the
eigenvectors ul. The red line corresponds to the critical eigendirection. C Entries of the eigenvector u(1)
associated with λ(1)
in the populations of the model. The affected areas are 46 and FEF. D Same as C for
the second iteration.
c
c
c
Figure 6 Unstable fixed points in subsequent iterations. Population firing rates at the unstable fixed
point as predicted by the mean-field theory encoded in color for iterations 1 (A) and 2 (B). Same display as
in Fig. 3.
populations for two trajectories starting close to the separatrix, where the first one converges to the LA
fixed point and the second one to the HA state. The trajectories diverge near the unstable fixed point
and thus we define the last local minimum of the Euclidean norm of the velocity vector as the critical
time sc at which we assume the system to be close to the unstable fixed point (Fig. 4C). We find four
relevant and distinct unstable fixed points, of which two are shown in Fig. 6.
To counteract the shift of the separatrix caused by the increase in κ, we follow the procedure described
in Methods. We subject the modifications of connectivity to the additional following constraints.
In
line with the anatomical literature, we do not allow for changes of the connectivity that would lead to
cortico-cortical connections originating in the granular layer 4 (Felleman & Van Essen, 1991), and we
also disallow inhibitory cortico-cortical connections, as the vast majority of long-range connections are
known to be excitatory (Salin & Bullier, 1995; Tomioka & Rockland, 2007). In addition, we naturally
10
Activity constraints on the connectome
Schuecker, Schmidt et al.
restrict indegrees to positive values. We find that four iterations (numbered by index j) corresponding
to the four distinct unstable fixed points suffice to preserve the basin of attraction of the LA state with
respect to an increase of the external drive up to κ = 1.15. In the following we concentrate on iterations
1 and 2, where the second one is also representative for iterations 3 and 4, which are qualitatively alike.
To derive the required modifications of the indegree matrix, we decompose K into its N eigenmodes
and quantify the contribution of each eigenmode to the shift of the unstable fixed point (see Methods).
This allows us to identify the most effective eigendirection:
in each iteration j there is exactly one
unstable eigendirection with an eigenvalue Re(cid:16)λ(j)
c (cid:17) > 1 (Fig. 5A). The associated critical eigenvector is
approximately anti-parallel to the shift of the fixed point, δν ∗ (inset of Fig. 5B), and of similar length.
The critical eigendirection (red dot in Fig. 5B) constitutes the most effective modification, giving the
largest contribution to the desired shift while requiring only a small change of 2.3 % in average total
indegrees. In the chosen space of eigenmodes, the modifications are minimal in the sense that only this
most effective eigenmode is changed.
The associated eigenvector u(1)
c predominately points into the direction of populations 4E and 5E of
areas FEF and 46 (Fig. 5C), while u(2)
c has large entries in the 5E populations of two areas (Fig. 5D). The
high rates of these populations at the unstable fixed points (cf. Fig. 6A,B with Fig. 5C,D) reflect that
the instability is caused by increased rates in excitatory populations, particularly in population 5E. Each
iteration shifts the transition to the HA state (the value of κ for which the separatrix crosses the initial
condition) to higher values of κ and increases the attainable rates of populations 5E and 6E in the LA
state (Fig. 7A). After all four iterations, the average total indegrees (summed over source populations)
of the system are changed by 11.3%. The first iteration mainly affects connections within and between
areas 46 and FEF (Fig. 7B). In particular, the excitatory loops between the two areas are reduced in
strength, especially those involving layer 5 (Fig. 7C). We thus identify two areas forming a critical loop.
In the remaining iterations, the changes are spread across areas and especially connections originating
in layer 5 are weakened (Fig. 7D). In conclusion, the method identifies critical structures in the model
both on the level of areas and on the level of layers and populations, and leads to a small but specific
structural change of the model.
Analysis of the modifications
In the following we analyze the modifications of the connectivity with respect to the internal and inter-
area connections in detail. The intrinsic circuits of the areas are modified in different directions, as
shown for two exemplary areas V4 and CITv in Fig. 8A. Despite this heterogeneity, significant changes
affect mostly excitatory-excitatory connections (Fig. 8A, bottom panel) with connections from popula-
tion 5E experiencing the most significant changes (top panel of Fig. 8A). In fact, the anatomical data
(Binzegger et al., 2004) underlying the microcircuit model (Potjans & Diesmann, 2014) contain only
two reconstructed excitatory cells from layer 5, but considerably more for other cell types, indicating a
11
Activity constraints on the connectome
Schuecker, Schmidt et al.
Figure 7 Altered phase space and modified connections. A Firing rates averaged across populations 5E
and 6E and across areas for different stages from the original model (black) to iteration 4 (light gray) as a
function of κ, predicted by the mean-field theory. B Relative changes in the indegree δKAB/PB KAB between
areas A, B in the first iteration. C Layer-specific relative changes δKln/Pn Kln in the connections within
in population-specific indegrees summed over target populations, Pl δKln/Pl Kln, combined for iterations
and between areas FEF and 46, for the first iteration. Populations are ordered from 2/3E (left) to 6I (right) on
the horizontal axis and from 6I (bottom) to 2/3E (top) on the vertical axis as in panel D. D Relative changes
two, three and four.
higher uncertainty for layer 5 connections. Fig. 8B shows the correlation between intrinsic connectivity
changes for all pairs of areas, with areas ordered according to a hierarchical clustering using a farthest
point algorithm (Voorhees, 1986) on the correlation matrix. We find four clusters each indicating a group
of areas which undergo changes with similar patterns. The groups are displayed in different colors in
the histogram in Fig. 8B. The areas of the model are categorized into architectural types based on cell
densities and laminar thicknesses (see Methods). Areas with architectural type 4, 5 and 6 are distributed
over several clusters. We can interpret this as a differentiation of these types into further subtypes.
The resulting changes of the intra-areal connectivity are small (Fig. 8A), but still significant for network
stability.
Connections between areas can be characterized by their F LN and SLN (see Methods). The F LN
reflects the overall strength of an inter-areal connection and is only weakly affected across connections
(Fig. 8C), with a correlation between original and modified logarithms of F LN of rPearson = 0.79. Sig-
nificant variations in the F LN occur mostly for very weak connections that are likely to have substantial
relative uncertainties in the experimental data. The two overlapping red dots in Fig. 8C represent the
connections between areas 46 and FEF, which are modified in the first iteration (Fig. 7C). The SLN
determines the laminar pattern of the location of source neurons for cortico-cortical connections. Over-
all, data are available for 24% of the inter-areal connections in the parcellation of Felleman & van Essen
(Felleman & Van Essen, 1991), while the SLN for the rest are derived from the sigmoidal law. The
majority of connections undergo small changes in their laminar source pattern (Fig. 8D) and connections
with large modifications (δSLN > 0.5) are weak (average F LN = 6· 10−4 compared to F LN = 10−2 in
12
Activity constraints on the connectome
Schuecker, Schmidt et al.
AA/Pj K ij
AA/Pi K ij
indegrees summed over target populations and averaged across areas, hPi δK ij
Figure 8 Analysis of changes in connectivity. A Top panel: relative changes in population-specific intrinsic
AAiA. Bottom panel:
changes in the indegrees within and between exemplary areas V4 and CITv relative to the total indegrees of the
target populations, i.e., δK ij
AA. Populations ordered as in Fig. 7C. B Pearson correlation coefficient
of the changes of the internal indegrees δK ij
AA between all pairs of the 32 areas. Areas ordered according to
hierarchical clustering using a farthest point algorithm (Voorhees, 1986). The heights of the bars on top of
the matrix indicate the architectural types of the areas (types 1 and 3 do not appear in the model) with color
representing the respective clusters. C F LN of the modified connectivity after 4 iterations versus the original
F LN of the model. Only F LN > 10−3 are shown for a better overview. The overlapping red dots represent
the connections between areas 46 and FEF. Unity line shown in gray. D Histogram of the cumulative changes
in SLN over all four iterations (δSLN = SLN (4) − SLN (0)).
13
Activity constraints on the connectome
Schuecker, Schmidt et al.
the model as a whole). Because weak connections are represented by low counts of labeled neurons, they
have a relatively large uncertainty in their laminar patterns, justifying larger adjustments. Spearman's
rank correlation between the SLN of the original model that were directly taken from experiments and
the logarithmic ratios of cell densities is ρ = −0.63 (p = 3 · 10−11, p-value of a two-sided test for un-
correlated data). For the modified model, we take the SLN of all connections into account and obtain
ρ = −0.40 (p = 6· 10−20), indicating a reduced, but still significant, monotonic dependence between SLN
and the logarithmic ratios of cell densities.
To judge the size of the modification to the connectivity, we compare it to the variability of measured
cortico-cortical connection densities (Markov et al., 2014a). We quantify the latter as the average inter-
individual standard deviation of the logarithmic F LN , i.e., σ = (cid:28)q(cid:0)log F LN − log F LN(cid:1)2(cid:29), where the
overbar · denotes the average over injections and h·i the average over connections. This variability equals
2.17 while the average modification of the logarithmic F LN is 1.34. The main experimental connection
probabilities used to construct the intra-areal connectivity of the model have an average relative standard
deviation of 30% across electrophysiological experiments (cf. Table 1 of (Potjans & Diesmann, 2014))
while the intra-areal connection probabilities of the model are modified by 9% on average. The authors of
(Scannell et al., 2000) report even greater variability in their review on cortico-cortical and thalamocortical
connectivity. These considerations show that on average, the changes applied to the connectivity are well
within the uncertainties of the data. Overall, 35 out of 603 connections were removed from the network.
In the CoCoMac database, 83 % of these are indicated by only a single tracer injection, while the overall
proportion of connections measured by a single injection is 59 %.
For the modified connectivity and κ = 1.125, which we choose to avoid being too close to the transition
(Fig. 7A), the theory predicts average rates in populations 5E and 6E of 1.3 and 0.18 spikes/s, which is
closer to experimentally observed rates compared to the original model. Furthermore we find that the
modified connectivity allows us to decrease the inhibition in the network to g = −11. Simulating the full
spiking network model then results in reasonable rates across populations and areas (Fig. 9B, D). The
average rates in populations 5E and 6E are increased compared to a simulation of the original model
from 0.09 and 2 · 10−5 spikes/s to 3.0 and 0.4 spikes/s, respectively. All populations exhibit firing rates
within a reasonable range of 0.05 to 30 spikes/s (Fig. 9D), as opposed to the original state in which a
considerable fraction of excitatory neurons is silent (Fig. 3E). The theoretical prediction is in excellent
agreement with the rates obtained in the simulation (Fig. 9A, C). Small discrepancies are caused by
violations of the employed assumptions, i.e., Poissonian spiking statistics (Fourcaud & Brunel, 2002).
Differences between theory and simulation are small, and negligible for the central aim of the study: the
integration of activity constraints into the data-driven construction of multi-scale neuronal networks.
14
Activity constraints on the connectome
Schuecker, Schmidt et al.
Figure 9 Improved low-activity fixed point of the model. Population-averaged firing rates for κ = 1.125
encoded in color (A) predicted by the analytical theory and (B) obtained from the full simulation of the spiking
network. Same display as in Fig. 3. C Analytical versus simulated firing rates (black dots) and identity line
(gray). D Histogram of population-averaged simulated firing rates. Same display as in Fig. 3.
Discussion
This study investigates the link between experimentally measured structural connectivity and neuronal
activity in a multi-scale spiking network model of the vision-related areas of macaque cortex (Schmidt
et al., 2016). We devise a theoretical method that systematically combines anatomical connectivity with
physiological activity constraints. Already weak constraints, demanding the activity to neither vanish nor
be pathologically high, yield a set of specific but small structural modifications necessary to increase the
model's excitation to a realistic level. We do not fit the model parameters to experimental activity data
in the sense of minimizing an error function, since the sparse relevant experimental data do not allow for
defining such a function in a meaningful way. Nevertheless, we considerably improve the network activity
with a change in only a small set of parameters. The procedure constrains the experimentally obtained
connectivity maps to a realization that is compatible with physiological experiments. This establishes a
path from experimentally observed activity to specific hypotheses about the anatomy and demands for
further experiments.
Connections are modified both inside and between cortical areas, on average well within the known
uncertainties of the underlying data. The model areas are based on the early sensory cortex model pre-
sented in (Potjans & Diesmann, 2014). This circuit is adapted to individual areas by taking into account
neuronal densities and laminar thicknesses. The model definition renders areas with equal architectural
type similar in their internal connectivity, a drastic but inevitable simplification due to the lack of more
detailed experimental data. The proposed method softens this assumption: small adaptations of internal
connectivity distinguish the architectural types into further subtypes. These modifications are signif-
icant for the global stability of the network. Thus, our approach enables purely anatomy-based area
categorizations to be refined with dynamical information.
15
Activity constraints on the connectome
Schuecker, Schmidt et al.
Connections between areas are changed in terms of total strength and laminar patterns. Overall, the
changes are small, but significant for specific connections. The loop formed by areas 46 and FEF is critical
to the global stability of the network. Both areas have been investigated in (Markov et al., 2014a), albeit
in a different parcellation scheme than the scheme used here (Felleman & Van Essen, 1991). Our method
suggests a weaker coupling of these two areas than found in the anatomical data set. Uncertainties,
partly due to the mapping between parcellations, leave room for this interpretation. Areas 46 and FEF
belong to prefrontal cortex and are multimodal, indicating that the influence of other parts of cortex
could stabilize them, a mechanism outside the scope of the present model of vision-related areas. Both
explanations can be tested either in an experimental study or with an extended model.
A few weak connections undergo large changes in their laminar patterns. With the present activity
constraints, the method hereby weakens hierarchical relations in the model structure, indicating that pre-
serving these relations requires additional dynamical constraints such as layer-specific coherence between
areas (?Bastos et al., 2015). Conversely, it is possible to achieve satisfactory population rates with a less
pronounced hierarchical structure.
Our analysis reveals that layer 5 excitatory cells play a critical role in the model's dynamics, in
line with the observed ongoing activity in mammalian neocortex (Sanchez-Vives & McCormick, 2000;
Beltramo et al., 2013). This critical role is often attributed to single-neuron properties, with a subset of
layer 5 neurons displaying pacemaker activity (Le Bon-Jego & Yuste, 2007; Lorincz et al., 2015; Neske
et al., 2015). In addition, we here find that the network architecture itself already explains the strong
impact of layer 5 on the phase space of the network, suggesting that single-neuron properties and network
structure jointly enable layer 5 to exert its dominant influence.
The cortico-cortical connectivity of the model is compiled from the extensive dataset of (Markov
et al., 2014a) combined with the CoCoMac database (Stephan et al., 2001; Bakker et al., 2012), which
collects data from hundreds of tracing studies. One could consider alternative methods for combining
this information into one connectivity graph, for instance taking into account how consistently a given
connection is reported across studies (Schmitt et al., 2014), and compare different methods by analyzing
the resulting network dynamics. The presented mean-field theory could then be used to estimate the
firing rates of each network instance without performing time-consuming simulations. However, here we
first choose the integrative approach that accumulates all available experimental evidence into a single
model and afterwards modify the resulting connectivity with our analytical procedure, thereby effectively
discarding uncertain connections.
We restrict this study to networks of leaky integrate-and-fire model neurons, consistent with the key
concept of the models we consider: individual cells are modeled in a simple manner to expose the impact
of structural connectivity on the network dynamics. Moreover, the current-based leaky integrate-and-fire
neuron can reproduce in-vivo like activity (Rauch et al., 2003; Jolivet et al., 2006) and is analytically
tractable, which enables the identification of mechanisms underlying specific network effects. More com-
plex neuron models can be incorporated into the method by replacing the gain function of each neuronal
16
Activity constraints on the connectome
Schuecker, Schmidt et al.
population with an analytical expression or an interpolated function obtained from spiking single-neuron
simulations. For example, one could use a conductance-based point-neuron model for which the network
dynamics can be described by population rate models (Shriki et al., 2003), featuring a non-monotonic
gain function: the gain is reduced if excitatory and inhibitory inputs are increased in a balanced manner
(Kuhn et al., 2004). Generally, this renders a system more stable. However, the bistability considered
in our work is caused by excitatory inputs. Since conductance-based models also have a monotonically
increasing gain function in dependence of the excitatory conductance alone, we expect the bistability to
occur for such models as well.
Importantly, our method only requires a description of the system's fixed points and their dependence
on model parameters. The employed theory uses the diffusion approximation to derive a self-consistency
equation for the stationary population rates valid for high indegrees and small synaptic efficacies. These
requirements are fulfilled in the multi-area model and therefore the theoretical prediction agrees with the
stationary activity of the simulation. Moreover, the theory predicts the bistability of the model, which is
non-trivial, as the mean-field assumption of Poisson statistics of the activity is generally violated in the
high-activity state. Nevertheless, since the activity in this state is mostly driven by strong mean inputs
and the theory converges to the noiseless solution in the mean-driven limit, its predictions still provide
viable approximations. The firing rates in the unstable fixed points are predominantly low, while the
exceptions with very high rates are again mean-dominated. Presumably this explains the accuracy by
which the theory predicts the locations of these fixed points and the resulting global stability properties.
Since the high-activity attractor of the model under consideration is unrealistic, we aim to prevent a
transition to this state. However, in the high-dimensional system it is not a priori clear in which direction
the separatrix has to be shifted to ensure stable dynamics. We therefore choose a pragmatic approach and
shift the separatrix back to its initial location, inverting the shift which reduced the global stability of the
low-activity fixed point. We achieve this to a good approximation by preserving the location of unstable
fixed points on the separatrix. To this end, we use a linearization around these locations and an eigenmode
decomposition to identify the set of connections to adjust. In the multi-area model, this linearization is
justified because the system operates close to an instability so that only minor modifications are required.
The method can be generalized to larger modifications by changing parameters iteratively in small steps.
Biological networks have various stabilization mechanisms not considered here, which render them
less critical. For instance, during growth, homeostatic mechanisms guide the system toward the right
structure. Furthermore, short-term synaptic plasticity (reviewed in (Morrison et al., 2008)), homeostatic
synaptic scaling (Turrigiano et al., 1998) and spike-frequency adaptation (e.g. summarized in (Benda &
Herz, 2003)) may prevent the system from entering the high-activity state. However, introducing these
self-organizing mechanisms increases model complexity, causing a more intricate relation between struc-
ture and activity. Therefore, we start from a mean-field description on the level of neuronal populations,
ignoring details of synaptic dynamics. Mild constraints on the activity lead to a network structure within
the anatomical range of parameters. This network yields globally stable activity, suggesting that addi-
17
Activity constraints on the connectome
Schuecker, Schmidt et al.
tional stabilization mechanisms are not required to achieve this. Nonetheless, they can potentially render
the network more robust against external stimulation.
The observed inter-individual variability may reflect that mechanisms of self-organization and home-
ostasis find structurally different implementations of the same function. Thus, in studies across individuals
we cannot expect that progress in experimental technology narrows down the variability of parameters in-
definitely. The combined ranges of parameters rather specify the solution space, and our method provides
a way to find a particular solution.
In principle, the method applies to any model parameter. It would be possible to modify, for example,
synaptic weights. Since experimental data on synaptic weights are sparse, this is another natural choice.
Moreover, such an analysis may provide hints about suitable synaptic plasticity rules that dynamically
stabilize the model. The method can be applied to networks with more complex sets of attractors
compared to the bistable case considered here. Though in high-dimensional systems such as our multi-
area model, a larger number of attractors would make it more challenging to find all relevant unstable
fixed points, the underlying idea of preserving the location of a separatrix is general. In contrast to the
model considered here, transitions between fixed points can have a functional meaning in certain neuronal
networks with multiple attractors. The specific location of the separatrix is then functionally relevant.
Our method exposes the sensitivity of the location of the separatrix to certain model parameters and
allows controlling its location in a specific manner.
In this work, we analyzed the global stability properties of the neuronal network on the population
level. In contrast, Ostojic (Ostojic, 2014) performs a local stability analysis on the level of single neu-
rons of an initially stable fixed point in a system with only one attractor. The author investigates the
point at which the real parts of the eigenvalues of the Jacobian matrix evaluated at this fixed point
become positive, i.e., the fixed point turns spectrally unstable and the system undergoes a transition
to a heterogeneous asynchronous state. Analyzing the spectral stablilty on the single-neuron level does
not reveal the global stability properties required in the current work: While a local stability analysis
only considers infinitesimal perturbations, studying the basin of attraction gives information about the
size of fluctuations against which the fixed point is stable. We expect both attractors to be spectrally
stable because they do not show strong rate fluctuations and the mean-field theory predicts the activity
accurately in both cases. Nevertheless, a heterogeneous state could occur if the synaptic weights were
increased or the external drive was stronger. However, (Goedeke et al., 2016) show that the transition
in stochastic systems that quantitatively resemble the spiking network (Grytskyy et al., 2013), does not
coincide with the loss of spectral stability.
One striking feature of the heterogeneous state is bursty spiking behavior of individual cells. Bursty
spiking is also observed in the multi-area model for increased synaptic weights of inter-area connections
(Schmidt et al., 2016). The fixed points cannot be accurately described in this case because the fluctua-
tions need to be taken into account (Mastroguiseppe & Ostojic, 2016). Simulations show (figures 4 and
5 of (Schmidt et al., 2016)) that the modifications obtained in this study are still able to prevent the
18
Activity constraints on the connectome
Schuecker, Schmidt et al.
system from a transition to a HA attractor also in the presence of bursting neurons. This indicates that
the phase space does not change qualitatively and our results are robust against such bursting behavior.
Experimental data on stationary activity in cortex are sparse. We therefore restrict ourselves to
fundamental constraints and increase the drive to the model in an area-unspecific way to fulfill them.
The resulting heterogeneity of the firing rates across areas is thus not imposed by the method, but rather
arises from the connectivity that remains strongly informed by anatomical data. Alternatively, one could
predefine a desired state and investigate the parameter changes necessary to achieve it.
The presented analytical method that combines anatomy and activity data into a consistent model
is restricted to stationary firing rates. In future studies, also higher-order statistical measures of activity
can be used as constraints. Resting-state fMRI, for example, provides information on the functional
connectivity between areas as a second-order measure. When combined with analytical predictions of
functional connectivity, the method may shed light on the anatomical connection patterns underlying
inter-area communication.
Methods
In this study, we model single cells as leaky integrate-and-fire model neurons with exponentially decaying
postsynaptic currents. Table 1 specifies the model parameters.
In the diffusion approximation, the
Table 1 Specification of the neuron and synapse parameters.
Synapse parameters
Name
J ± δJ
g
de ± δde
di ± δdi
d ± δd
vt
Name
τm
τr
τs
Cm
Vr
θ
EL
Value
87.8 ± 8.8 pA
−16 (Fig. 3)
−11 (Fig. 9)
1.5 ± 0.75 ms
0.75 ± 0.375 ms
d = s/vt ± 1
2 s/vt
3.5 m/s
Value
10 ms
2 ms
0.5 ms
250 pF
−65 mV
−50 mV
−65 mV
Description
excitatory synaptic strength
relative inhibitory synaptic strength
local excitatory transmission delay
local inhibitory transmission delay
inter-areal transmission delay, with s the
distance between areas
transmission speed
Neuron model
Description
membrane time constant
absolute refractory period
postsynaptic current time constant
membrane capacity
reset potential
fixed firing threshold
leak potential
dynamics of the membrane potential V and synaptic current Is is (Fourcaud & Brunel, 2002)
τm
τs
dV
dt
dIs
dt
= −V + Is(t)
= −Is + µ + σ√τmξ(t),
19
(2)
Activity constraints on the connectome
Schuecker, Schmidt et al.
where τm is the membrane time constant and τs the synaptic time constant, respectively. The membrane
resistance τm/Cm has been absorbed into the definition of the current. The input spike trains are
approximated by a white noise current with fluctuations ∝ σ2 and mean value µ. Here ξ is a centered
Gaussian white process satisfying hξ(t)i = 0 and hξ(t)ξ(t′)i = δ(t−t′). Whenever the membrane potential
V crosses the threshold θ, the neuron emits a spike and V is reset to the potential Vr, where it is clamped
during τr. All neurons in one population have identical parameters, so that we can describe the network
activity in terms of population-averaged firing rates νi that depend on population-dependent input µi, σi
determined by the connectivity. Using the Fokker-Planck formalism, the stationary firing rates for each
population i are given by (Fourcaud & Brunel, 2002)
1
νi
= τr + τm√π
=: 1/Φi(ν, A)
θ−µi (A)
σi(A) +γ√ τs
σi(A) +γ√ τs
Vr −µi (A)
τm
τm
ex2
(1 + erf(x)) dx
µi(A) = τmXj
i (A) = τmXj
σ2
KijJ ijνj + τmKextJextνext
KijJ 2
ijνj + τmKextJ 2
extνext,
(3)
(4)
(5)
which is correct up to linear order in pτs/τm and where γ = ζ(1/2)/√2, with ζ denoting the Rie-
mann zeta function (Abramowitz & Stegun, 1974). Here, A is chosen from the set of model param-
eters {K, J , νext, . . .}.
by lower case, i.e., a = (a00, a01, . . . , a0N , a10, . . . , a1N , aN 0 . . . , aN N ) = vec(AT)T following (Magnus &
If A is a matrix, we vectorize it by concatenating its rows and indicate this
Neudecker, 1995). If the chosen parameter is a scalar we denote it with a.
We find the fixed points of (3) by solving the first-order differential equation (1) (Wong & Wang,
2006)
ν :=
dν
ds
= Φ(ν, A) − ν,
using the classical fourth-order Runge-Kutta method (RK4) with step size h = 0.01, where s denotes a
dimensionless pseudo-time. The same approach can be used to solve the activity on a single neuron level
(Sadeh & Rotter, 2015). Note that (1) does not reflect the real time evolution of the population rates,
but rather is a mathematical method to obtain the system's fixed points. In contrast to (Wong & Wang,
2006) we do not only search for stable fixed points, but also use (1) to obtain unstable attractors (cf.
Results), an idea originating from the study of simple attractor networks ((Amit & Brunel, 1997b) esp.
their figure 2 and eq. 12).
In a bistable situation, the initial condition of (1) determines which fixed point the system settles
in. However, studying the behavior for a particular initial condition is of minor interest, since the actual
spiking network is a stochastic system which fluctuates around the fixed points of the deterministic
system defined by (1). Even if we knew that (1) would relax to the LA fixed point for one particular
20
Activity constraints on the connectome
Schuecker, Schmidt et al.
initial condition, this would not necessarily imply that this state is indefinitely stable. Global stability is
determined by the size of the basin of attraction of the LA fixed point.
In the following, we derive the equations leading us to targeted modifications of a parameter b neces-
sary to compensate for the changes in the global stability induced by the change of another parameter a.
To this end, we study the behavior of the fixed points with respect to an infinitesimal change δa = a′ − a
in the chosen model parameter. Let ν ∗(a) and ν ∗(a′) be the corresponding locations of the fixed points
and δν ∗ = ν ∗(a′) − ν ∗(a) their separation. We can then expand ν ∗(a′) into a Taylor series up to first
order in δa and obtain
ν ∗(a′) = ν ∗(a) + ∆aδa
⇔ δν ∗ = ∆aδa ,
with
∆a,ij =
=
=
dνi (µi, σi)
daj
dΦi (µi, σi)
daj
dµi
daj
∂Φi
∂µi
{z}Si
+
dσ2
i
daj
,
∂Φi
∂σi
1
2σi
Ti
{z }
(6)
(7)
where we notice that Si and Ti only depend on the target population i. We accordingly define two diagonal
matrices S and T with Sii = Si and Tii = Ti. We further define the connectivity matrix W = K ⊛ J,
where ⊛ denotes element-wise multiplication, also called the Hadamard product (see (Cichocki et al.,
2009), for a consistent set of symbols for operations on matrices). The derivatives with respect to aj have
the compact expressions
dµi
daj
=
∂µi
∂aj
+Xn
∂µi
∂νn
dνn
daj
= (Daµ)ij + τmXn
(KinJin
)∆a,nj ,
=Win
{z }
with the Jacobian (Daf )ij := ∂fi
∂aj
of some vector-valued function f and
dσ2
i
daj
= (cid:0)Daσ2(cid:1)ij + τmXn
)∆a,nj ,
(KinJ 2
in
=W2,in
{z }
where we use the Hadamard product again to define the matrix W2 := K ⊛ J ⊛ J. Inserting the total
21
Activity constraints on the connectome
Schuecker, Schmidt et al.
derivatives into (7), we derive the final expression for ∆a, reading
∆a = S [Daµ + τmW ∆a] + T (cid:2)Daσ2 + τmW2∆a(cid:3)
⇔ ∆a =
(cid:2)SDaµ + T Daσ2(cid:3)
}
1 − τm(SW + T W2)
}
{z
=: ¯∆a
−1
=:M
{z
,
(8)
where we use 1 for the identity matrix and define the effective connectivity matrix M and the matrix
¯∆a. The latter has dimensionality N × P , where P is the dimension of a (for example, P = N 2 for
a = k the vector of indegrees). With the aid of (8), evaluating (6) at the unstable fixed point predicts
the shift of the separatrix (Fig. 2D) to linear order. We now consider an additional parameter b which is
modified to counteract the shift of the unstable fixed point caused by the change in parameter a, i.e.,
∆aδa
!= −∆bδb
⇔ ¯∆aδa = − ¯∆bδb,
(9)
where we used that the inverse of 1 − M appears on both sides of the equation and hence drops out.
Note that the tuple (a, b) may represent any combination of model parameters, for example external
input, indegrees, synaptic weights, etc. For a particular choice of (a, b) we solve (9) for δb. For the
illustrative example shown in Fig. 2, where a = νext and b = K, (9) simplifies to
δK =
¯∆νext δνext
¯∆K
=
Kextδνext
ν
,
since S and T appearing in the respective ¯∆′s cancel each other.
To determine critical connections in a more complex model, we choose b = k, i.e.
the vector of
indegrees, and solve (9) with a decomposition into eigenmodes. We can write the right-hand side as
−(cid:0) ¯∆kδk(cid:1)i = −Xj
τm(cid:0)SiJij + TiJ 2
ij(cid:1) νjδKij .
(10)
The equation holds because µi, σi are only affected by connections to population i, and therefore their
derivatives ∂µi/∂kl, ∂σi/∂kl and hence ¯∆k,il, vanish for l /∈ [(i − 1)N + 1, iN ]. We now make the ansatz
δKij = Xl
ǫl
τm
(cid:0)ulvlT(cid:1)ij
(SJ + T (J ⊛ J )) ij
,
(11)
which decomposes the changes δK into eigenmodes of the effective connectivity. The ul and vl are the
l-th right and left eigenvectors of M as defined in (8), fulfilling the bi-orthogonality vlTun = δln and the
completeness relation Pl ulvlT = 1. Inserting (10) with (11) into (9) yields
¯∆aδa = −Xl
ǫlulvlTν.
22
(12)
Activity constraints on the connectome
Schuecker, Schmidt et al.
Thus we can solve for ǫl by multiplying from the left with vnT
vnT ¯∆aδa
=:an
{z
= −Xl
an
νn .
}
ǫn = −
ǫl vnTul
δnl
{z }
vlTν
{z}=:νl
Our goal is to find a set of connections which dominate the size of the basin of attraction of the LA fixed
point. Inserting (12) into (6) leads to
δν ∗ = Xl
= Xl
= Xl
(1 − M )−1ǫlulvlTν
ǫl
1 − λl
−al
1 − λl
ul νl
ul,
where λl are the eigenvalues of M , which are either real or complex conjugate pairs since M ∈ RN ×N .
To determine the influence of each eigenmode on the shift of the fixed point, we project the eigenvectors
ul onto the fixed-point shift δν ∗ by multiplying each side with δν ∗δν ∗T and solve again for δν ∗ to
δν ∗,
(13)
obtain
δν ∗ = Xl
−al
1 − λl
δν ∗Tul
δν ∗Tδν ∗
=: ηl
{z
}
where we define the (possibly complex-valued) coefficients ηl. We aim at a decomposition of δν ∗ into
real components. If Im(λl) = 0, ηl is real, so we can work directly with ηl := ηl. Complex eigenvalues
Im(λl) 6= 0 and corresponding eigenvectors come in conjugate pairs, so in this case we combine the
l , to have all contributions ηl ∈ R and Pl ηl = 1 by construction
corresponding coefficients ηl := ηl + η∗
(13). Each ηl quantifies how much of the total fixed-point shift can be attributed to the l-th eigenmode,
which allows identification of the most effective eigendirection (see Results), where we apply the ansatz
(11) to a multi-area, multi-layer model of the vision-related areas of macaque cortex.
The spiking simulations of the network model were carried out on the JUQUEEN supercomputer
(Julich Supercomputing Centre, 2015). All simulations were performed with NEST version 2.8.0 (Eppler
et al., 2015) with optimizations for the use on the supercomputer which will be included in a future
release. The simulations use a time step of 0.1 ms and exact integration for the subthreshold dynamics
of the leaky integrate-and-fire neuron model (reviewed in (Plesser & Diesmann, 2009)).
23
Activity constraints on the connectome
Schuecker, Schmidt et al.
Multi-area model
The multi-area model of the vision-related areas of macaque cortex uses the microcircuit model of (Potjans
& Diesmann, 2014) as a prototype for all 32 areas in the FV91 parcellation (Felleman & Van Essen, 1991)
and customizes it based on experimental findings on cortical structure. From anatomical studies, it is
known that cortical areas in the macaque monkey are heterogeneous in their laminar structure and can
be roughly categorized into 8 different architectural types based on cell densities and laminar thicknesses.
This distinction was originally developed for prefrontal areas (Barbas & Rempel-Clower, 1997), and then
extended to the entire cortex (Hilgetag et al., 2016). The visual cortex, and thus the model, comprises
areas of categories 2, 4, 5, 6, 7 and 8. Precise layer-specific neuron densities are available for a number
of areas, while for other areas, the neuron density is estimated based on their architectural type (see
(Schmidt et al., 2016) for details).
The inter-areal connectivity is based on binary data from the CoCoMac database (Stephan et al.,
2001; Bakker et al., 2012; Felleman & Van Essen, 1991; Rockland & Pandya, 1979; Barnes & Pandya,
1992) indicating the existence of connections, and quantitative data from (Markov et al., 2014a). The
latter are retrograde tracing data where connection strengths are quantified by the fraction of labeled
neurons (F LN ) in each source area. The original analysis of the experimental data was performed in the
M132 atlas (Markov et al., 2014a). Both the FV91 and the M132 parcellations have been registered to
F99 space (Van Essen, 2002), a standard macaque cortical surface included with the software tool Caret
(Van Essen et al., 2001). This enables mapping between the two parcellations.
On the target side, we use the exact coordinates of the injections to identify the equivalent area
in the FV91 parcellation. To map the data on the source side from the M132 atlas to the FV91 par-
cellation, we count the number of overlapping triangles on the F99 surface between any given pair of
regions and distribute data proportionally to the amount of overlap using the F99 region overlap tool
on http://cocomac.g-node.org. In the model, this F LN is mapped to the indegree KAB the target area
A receives from source area B divided by its total indegree, i.e., F LNAB = KAB/PB ′ KAB ′. Here,
KAB is defined as the total number of synapses between A and B divided by the total number of neu-
rons in A. On the source side, laminar connection patterns are based on CoCoMac (Felleman & Van
Essen, 1991; Barnes & Pandya, 1992; Suzuki & Amaral, 1994; Morel & Bullier, 1990; Perkel et al., 1986;
Seltzer & Pandya, 1994) and on fractions of labeled neurons in the supragranular layers (SLN ) (Markov
et al., 2014b). Gaps in these data are bridged exploiting a sigmoidal relation between SLN and the
logarithmized ratio of overall cell densities of the two areas, similar to (Beul et al., 2015). We map
the SLN to the ratio between the number of synapses originating in layer 2/3 and the total number of
synapses between the two areas, assuming that only excitatory populations send inter-area connections,
i.e., SLNAB = Pi K i,2/3E
AB N i
A/Pi,j K ij
ABN i
A, where the indices i and j go over the different populations
within area A and B, respectively. In the context of the model, we use the terms F LN and SLN to refer
to the relevant relative indegrees given here. On the target side, the CoCoMac database provides data
24
Activity constraints on the connectome
Schuecker, Schmidt et al.
from anterograde tracing studies (Felleman & Van Essen, 1991; Rockland & Pandya, 1979; Jones et al.,
1978; Seltzer & Pandya, 1991).
Missing inputs in the model, i.e., from subcortical and non-visual cortical areas, are replaced by
Poissonian spike trains, whose rate νext is a free, global parameter. In the original model all populations
of a particular area receive the same indegree of external inputs Kext. The only exception to this rule
is area TH where the absence of granular layer 4 is compensated by an increase of the external input
to populations 2/3E and 5E by 20 %. To elevate the firing rates in the excitatory populations in layers
5 and 6, we increase the external drive onto these populations. The possibility of a higher drive onto
these populations is left open by the sparseness of the corresponding experimental data. We enhance the
external Poisson drive of the 5E population, parametrized by the K5E,ext incoming connections per target
neuron (indegree), in all areas by increasing κ = K5E,ext/Kext. The simultaneous increase in the drive of
6E needs to be stronger, since the firing rates in population 6E of the original model (Fig. 3C) are even
lower than the rates of 5E (averaged across areas: 0.09 spikes/s for 5E compared to 2 · 10−5 spikes/s for
6E). We thus scale up K6E,ext linearly with κ such that κ = 1.15 results in K6E,ext/Kext = 1.5.
Acknowledgement
The authors gratefully acknowledge the computing time granted (jinb33) by the JARA-HPC Vergabegremium
and provided on the JARA-HPC Partition part of the supercomputer JUQUEEN at Forschungszentrum
Julich. Partly supported by Helmholtz Portfolio Supercomputing and Modeling for the Human Brain
(SMHB), the Helmholtz young investigator group VH-NG-1028, EU Grant 269921 (BrainScaleS). This
project received funding from the European Union's Horizon 2020 research and innovation programme
under grant agreement No. 720270. All network simulations carried out with NEST (http://www.nest-
simulator.org).
25
Activity constraints on the connectome
Schuecker, Schmidt et al.
References
Abramowitz, M & Stegun, IA (1974). Handbook of Mathematical Functions: with Formulas, Graphs,
and Mathematical Tables. (New York: Dover Publications).
Amit, DJ & Brunel, N (1997a). Dynamics of a recurrent network of spiking neurons before and following
learning. Network: Comput Neural Systems 8:373 -- 404.
Amit, DJ & Brunel, N (1997b). Model of global spontaneous activity and local structured activity during
delay periods in the cerebral cortex. Cereb Cortex 7:237 -- 252.
Axer, M, Grassel, D, Kleiner, M, Dammers, J, Dickscheid, T, Reckfort, J, Hutz, T, Eiben, B, Pietrzyk,
U, Zilles, K, et al. (2011). High-resolution fiber tract reconstruction in the human brain by means of
three-dimensional polarized light imaging. Front Neuroinformatics 5.
Bakker, R, Thomas, W, & Diesmann, M (2012). CoCoMac 2.0 and the future of tract-tracing databases.
Front Neuroinformatics 6.
Barbas, H & Rempel-Clower, N (1997). Cortical structure predicts the pattern of corticocortical connec-
tions. Cereb Cortex 7:635 -- 646.
Barnes, CL & Pandya, DN (1992). Efferent cortical connections of multimodal cortex of the superior
temporal sulcus in the rhesus monkey. J Compar Neurol 318:222 -- 244.
Bastos, AM, Vezoli, J, Bosman, CA, Schoffelen, JM, Oostenveld, R, Dowdall, JR, De Weerd, P, Kennedy,
H, & Fries, P (2015). Visual areas exert feedforward and feedback influences through distinct frequency
channels. Neuron 85:390 -- 401.
Beltramo, R, D'Urso, G, Maschio, MD, Farisello, P, Bovetti, S, Clovis, Y, Lassi, G, Tucci, V, Tonelli,
DDP, & Fellin, T (2013). Layer-specific excitatory circuits differentially control recurrent network
dynamics in the neocortex. Nat Neurosci 16:227 -- 234.
Benda, J & Herz, V (2003). A universal model for spike-frequency adaptation. Neural Comput 15:2523 --
2546.
Beul, SF, Barbas, H, & Hilgetag, CC (2015). A predictive structural model of the primate connectome.
arXiv preprint arXiv:1511.07222.
Binzegger, T, Douglas, RJ, & Martin, KAC (2004). A quantitative map of the circuit of cat primary
visual cortex. J Neurosci 39:8441 -- 8453.
Brunel, N (2000). Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons.
J Comput Neurosci 8:183 -- 208.
Cabral, J, Kringelbach, ML, & Deco, G (2014). Exploring the network dynamics underlying brain activity
during rest. Prog Neurobiol 114:102 -- 131.
26
Activity constraints on the connectome
Schuecker, Schmidt et al.
Cichocki, A, Zdunek, R, Phan, AH, & Amari, Si (2009). Nonnegative matrix and tensor factorizations:
applications to exploratory multi-way data analysis and blind source separation. (John Wiley & Sons).
Cole, MW, Bassett, DS, Power, JD, Braver, TS, & Petersen, SE (2014). Intrinsic and task-evoked network
architectures of the human brain. Neuron 83:238 -- 251.
Dayan, P & Abbott, LF (2001). Theoretical Neuroscience. (Cambridge: MIT Press).
de Kock, CPJ & Sakmann, B (2009). Spiking in primary somatosensory cortex during natural whisking
in awake head-restrained rats is cell-type specific. Proc Natl Acad Sci USA 106:16446 -- 16450.
Deco, G & Jirsa, VK (2012). Ongoing cortical activity at rest: Criticality, multistability, and ghost
attractors. J Neurosci 32:3366 -- 3375.
Eppler, JM, Pauli, R, Peyser, A, Ippen, T, Morrison, A, Senk, J, Schenck, W, Bos, H, Helias, M, Schmidt,
M, et al. (2015). NEST 2.8.0.
Felleman, DJ & Van Essen, DC (1991). Distributed hierarchical processing in the primate cerebral cortex.
Cereb Cortex 1:1 -- 47.
Fourcaud, N & Brunel, N (2002). Dynamics of the firing probability of noisy integrate-and-fire neurons.
Neural Comput 14:2057 -- 2110.
Goedeke, S, Schuecker, J, & Helias, M (2016). Noise dynamically suppresses chaos in neural networks.
arXiv. 1603.01880v1 [q-bio.NC].
Goulas, A, Bastiani, M, Bezgin, G, Uylings, HB, Roebroeck, A, & Stiers, P (2014). Comparative anal-
ysis of the macroscale structural connectivity in the macaque and human brain. PLoS Comput Biol
10:e1003529.
Grytskyy, D, Tetzlaff, T, Diesmann, M, & Helias, M (2013). A unified view on weakly correlated recurrent
networks. Front Comput Neurosci 7:131.
Hilgetag, CC, Medalla, M, Beul, S, & Barbas, H (2016). The primate connectome in context: principles
of connections of the cortical visual system. NeuroImage p. (in press).
Jolivet, R, Rauch, A, Luscher, HR, & Gerstner, W (2006). Predicting spike timing of neocortical pyra-
midal neurons by simple threshold models. J Comput Neurosci 21:35 -- 49.
Jones, E, Coulter, J, & Hendry, S (1978). Intracortical connectivity of architectonic fields in the somatic
sensory, motor and parietal cortex of monkeys. J Compar Neurol 181:291 -- 347.
Julich Supercomputing Centre (2015). JUQUEEN: IBM Blue Gene/Q supercomputer system at the
Julich Supercomputing Centre. Journal of large-scale research facilities 1.
27
Activity constraints on the connectome
Schuecker, Schmidt et al.
Kriener, B, Helias, M, Rotter, S, Diesmann, M, & Einevoll, G (2014). How pattern formation in ring
networks of excitatory and inhibitory spiking neurons depends on the input current regime. Front
Comput Neurosci 7:1 -- 21.
Kuhn, A, Aertsen, A, & Rotter, S (2004). Neuronal integration of synaptic input in the fluctuation-driven
regime. J Neurosci 24:2345 -- 2356.
Le Bon-Jego, M & Yuste, R (2007). Persistently active, pacemaker-like neurons in neocortex. Front
Neurosci 1:123 -- 129.
Lorincz, ML, Gunner, D, Bao, Y, Connelly, WM, Isaac, JT, Hughes, SW, & Crunelli, V (2015). A distinct
class of slow ( 0.2 -- 2 Hz) intrinsically bursting layer 5 pyramidal neurons determines UP/DOWN state
dynamics in the neocortex. J Neurosci 35:5442 -- 5458.
Magnus, JR & Neudecker, H (1995). Matrix differential calculus with applications in statistics and
econometrics. (John Wiley & Sons).
Markov, NT, Ercsey-Ravasz, M, Van Essen, DC, Knoblauch, K, Toroczkai, Z, & Kennedy, H (2013).
Cortical high-density counterstream architectures. Science 342.
Markov, NT, Ercsey-Ravasz, MM, Ribeiro Gomes, AR, Lamy, C, Magrou, L, Vezoli, J, Misery, P, Falchier,
A, Quilodran, R, Gariel, MA, et al. (2014a). A weighted and directed interareal connectivity matrix
for macaque cerebral cortex. Cereb Cortex 24:17 -- 36.
Markov, NT, Vezoli, J, Chameau, P, Falchier, A, Quilodran, R, Huissoud, C, Lamy, C, Misery, P, Giroud,
P, Ullman, S, et al. (2014b). Anatomy of hierarchy: Feedforward and feedback pathways in macaque
visual cortex. J Compar Neurol 522:225 -- 259.
Mastroguiseppe, F & Ostojic, S (2016).
Intrinsically-generated fluctuating activity in excitatory-
inhibitory networks. arXiv p. 1605.04221.
Morel, A & Bullier, J (1990). Anatomical segregation of two cortical visual pathways in the macaque
monkey. Visual neuroscience 4:555 -- 578.
Morrison, A, Diesmann, M, & Gerstner, W (2008). Phenomenological models of synaptic plasticity based
on spike-timing. Biol Cybern 98:459 -- 478.
Neske, GT, Patrick, SL, & Connors, BW (2015). Contributions of diverse excitatory and inhibitory
neurons to recurrent network activity in cerebral cortex. J Neurosci 35:1089 -- 1105.
Ostojic, S (2014). Two types of asynchronous activity in networks of excitatory and inhibitory spiking
neurons. Nat Neurosci 17:594 -- 600.
Ostojic, S, Brunel, N, & Hakim, V (2009). How connectivity, background activity, and synaptic properties
shape the cross-correlation between spike trains. J Neurosci 29:10234 -- 10253.
28
Activity constraints on the connectome
Schuecker, Schmidt et al.
Perkel, DJ, Bullier, J, & Kennedy, H (1986). Topography of the afferent connectivity of area 17 in the
macaque monkey: A double-labelling study. J Compar Neurol 253:374 -- 402.
Pernice, V, Staude, B, Cardanobile, S, & Rotter, S (2011). How structure determines correlations in
neuronal networks. PLoS Comput Biol 7:e1002059.
Plesser, HE & Diesmann, M (2009). Simplicity and efficiency of integrate-and-fire neuron models. Neural
Comput 21:353 -- 359.
Potjans, TC & Diesmann, M (2014). The cell-type specific cortical microcircuit: Relating structure and
activity in a full-scale spiking network model. Cereb Cortex 24:785 -- 806.
Rauch, A, La Camera, G, Luscher, H, Senn, W, & Fusi, S (2003). Neocortical pyramidal cells respond
as integrate-and-fire neurons to in vivo like input currents. J Neurophysiol 90:1598 -- 1612.
Rockland, KS & Pandya, DN (1979). Laminar origins and terminations of cortical connections of the
occipital lobe in the rhesus monkey. Brain Res 179:3 -- 20.
Sadeh, S & Rotter, S (2015). Orientation selectivity in inhibition-dominated networks of spiking neurons:
effect of single neuron properties and network dynamics. PLoS Comput Biol 11:e1004045.
Salin, PA & Bullier, J (1995). Corticocortical connections in the visual system: structure and function.
Physiol Rev 75:107 -- 154.
Sanchez-Vives, MV & McCormick, D (2000). Cellular and network mechanisms of rhythmic recurrent
activity in neocortex. Nat Neurosci 3:1027 -- 1034.
Scannell, JW, Grant, S, Payne, BR, & Baddeley, R (2000). On variability in the density of corticocortical
and thalamocortical connections. Phil Trans R Soc B 355:21 -- 35.
Schmidt, M, Bakker, R, Diesmann, M, & van Albada, S (2016). Full-density multi-scale account of
structure and dynamics of macaque visual cortex. arXiv preprint arXiv:1511.09364v4.
Schmidt, M, Schuecker, J, Diesmann, M, & Helias, M (2015). Shaping phase space of neural networks
via connectivity. In Proceedings of the 11th Gottingen Meeting of the German Neuroscience Society
p. T26 7C.
Schmitt, O, Eipert, P, Kettlitz, R, Lessmann, F, & Wree, A (2014). The connectome of the basal ganglia.
Brain Structure and Function pp. 1 -- 62.
Seltzer, B & Pandya, DN (1991). Post-rolandic cortical projections of the superior temporal sulcus in
the rhesus monkey. J Compar Neurol 312:625 -- 640.
Seltzer, B & Pandya, DN (1994). Parietal, temporal, and occipita projections to cortex of the superior
temporal sulcus in the rhesus monkey: A retrograde tracer study. J Compar Neurol 343:445 -- 463.
29
Activity constraints on the connectome
Schuecker, Schmidt et al.
Shen, K, Hutchison, RM, Bezgin, G, Everling, S, & McIntosh, AR (2015). Network structure shapes
spontaneous functional connectivity dynamics. J Neurosci 35:5579 -- 5588.
Shriki, O, Hansel, D, & Sompolinsky, H (2003). Rate models for conductance-based cortical neuronal
networks. Neural Comput 15:1809 -- 1841.
Stephan, K, Kamper, L, Bozkurt, A, Burns, G, Young, M, & Kotter, R (2001). Advanced database
methodology for the collation of connectivity data on the macaque brain (CoCoMac). Phil Trans R
Soc B 356:1159 -- 1186.
Strogatz, SH (1994). Nonlinear Dynamics and Chaos: with Applications to Physics, Biology, Chemistry,
and Engineering. (Reading, Massachusetts: Perseus Books).
Suzuki, WL & Amaral, DG (1994). Perirhinal and parahippocampal cortices of the macaque monkey:
cortical afferents. J Compar Neurol 350:497 -- 533.
Swadlow, HA (1988). Efferent neurons and suspected interneurons in binocular visual cortex of the awake
rabbit: Receptive fields and binocular properties. J Neurophysiol 59:1162 -- 1187.
Tetzlaff, T, Helias, M, Einevoll, G, & Diesmann, M (2012). Decorrelation of neural-network activity by
inhibitory feedback. PLoS Comput Biol 8:e1002596.
Tomioka, R & Rockland, KS (2007). Long-distance corticocortical GABAergic neurons in the adult
monkey white and gray matter. J Compar Neurol 505:526 -- 538.
Trousdale, J, Hu, Y, Shea-Brown, E, & Josic, K (2012). Impact of network structure and cellular response
on spike time correlations. PLoS Comput Biol 8:e1002408.
Turrigiano, GG, Leslie, KR, Desai, NS, Rutherford, LC, & Nelson, SB (1998). Activity-dependent scaling
of quantal amplitude in neocortical neurons. Nature 391:892 -- 896.
van Albada, SJ, Helias, M, & Diesmann, M (2015). Scalability of asynchronous networks is limited by
one-to-one mapping between effective connectivity and correlations. PLoS Comput Biol 11:e1004490.
Van Essen, DC (2002). Windows on the brain: the emerging role of atlases and databases in neuroscience.
Curr Opin Neurobiol 12:574 -- 579.
Van Essen, DC, Drury, HA, Dickson, J, Harwell, J, Hanlon, D, & Anderson, CH (2001). An integrated
software suite for surface-based analyses of cerebral cortex. Journal of the American Medical Informatics
Association 8:443 -- 459.
van Vreeswijk, C & Sompolinsky, H (1996). Chaos in neuronal networks with balanced excitatory and
inhibitory activity. Science 274:1724 -- 1726.
Voorhees, EM (1986). Implementing agglomerative hierarchic clustering algorithms for use in document
retrieval. Information Processing & Management 22:465 -- 476.
30
Activity constraints on the connectome
Schuecker, Schmidt et al.
Wedeen, VJ, Wang, R, Schmahmann, JD, Benner, T, Tseng, W, Dai, G, Pandya, D, Hagmann, P,
D'Arceuil, H, & de Crespigny, AJ (2008). Diffusion spectrum magnetic resonance imaging (dsi) trac-
tography of crossing fibers. NeuroImage 41:1267 -- 1277.
Wong, KF & Wang, XJ (2006). A recurrent network mechanism of time integration in perceptual
decisions. J Neurosci 26:1314 -- 1328.
31
|
1410.5610 | 3 | 1410 | 2017-12-13T15:00:09 | Logarithmic distributions prove that intrinsic learning is Hebbian | [
"q-bio.NC",
"cs.NE"
] | In this paper, we present data for the lognormal distributions of spike rates, synaptic weights and intrinsic excitability (gain) for neurons in various brain areas, such as auditory or visual cortex, hippocampus, cerebellum, striatum, midbrain nuclei. We find a remarkable consistency of heavy-tailed, specifically lognormal, distributions for rates, weights and gains in all brain areas examined. The difference between strongly recurrent and feed-forward connectivity (cortex vs. striatum and cerebellum), neurotransmitter (GABA (striatum) or glutamate (cortex)) or the level of activation (low in cortex, high in Purkinje cells and midbrain nuclei) turns out to be irrelevant for this feature. Logarithmic scale distribution of weights and gains appears to be a general, functional property in all cases analyzed. We then created a generic neural model to investigate adaptive learning rules that create and maintain lognormal distributions. We conclusively demonstrate that not only weights, but also intrinsic gains, need to have strong Hebbian learning in order to produce and maintain the experimentally attested distributions. This provides a solution to the long-standing question about the type of plasticity exhibited by intrinsic excitability. | q-bio.NC | q-bio |
Logarithmic distributions prove that intrinsic learning is
Hebbian1
Gabriele Scheler
Carl Correns Foundation for Mathematical Biology
1030 Judson Dr., Mountain View, Ca 94040
[email protected]
Abstract
In this paper, we present data for the lognormal distributions of spike rates, synaptic weights
and intrinsic excitability (gain) for neurons in various brain areas, such as auditory or visual
cortex, hippocampus, cerebellum, striatum, midbrain nuclei. We find a remarkable consistency
of heavy-tailed, specifically lognormal, distributions for rates, weights and gains in all brain ar-
eas examined. The difference between strongly recurrent and feed-forward connectivity (cortex
vs. striatum and cerebellum), neurotransmitter (GABA (striatum) or glutamate (cortex)) or
the level of activation (low in cortex, high in Purkinje cells and midbrain nuclei) turns out to
be irrelevant for this feature. Logarithmic scale distribution of weights and gains appears to be
a general, functional property in all cases analyzed. We then created a generic neural model
to investigate adaptive learning rules that create and maintain lognormal distributions. We
conclusively demonstrate that not only weights, but also intrinsic gains, need to have strong
Hebbian learning in order to produce and maintain the experimentally attested distributions.
This provides a solution to the long-standing question about the type of plasticity exhibited
by intrinsic excitability.
1 Introduction
Individual neurons have very different, but mostly stable, mean spike rates under a variety of
conditions [1, 2]. To report on behavioral results, spike counts are often normalized with respect
to the mean for each neuron. But this obscures an important question: Why do neurons within
a tissue operate at radically different levels of output frequency? In order to answer this question
our approach is twofold: (a) we try to document this phenomenon for different neural tissues and
behavioral conditions. We also examine neural properties for their distribution, namely intrinsic
gains and synaptic weights; (b) we build a very generic neural model to explore the conditions for
generating and maintaining these distributions. First, we give examples for the distribution of mean
spike rates for principal neurons under spontaneous conditions, as well as in response to stimuli.
We furthermore document distributions for intrinsic excitability [3, 4, 5] for cortical and striatal
neurons, as well as synaptic weight distributions [6, 7, 8, 9, 10, 11].
1This paper appeared in F1000Research 2017, 6:1222. The GNN simulation software is available at https:
//github.com/gscheler/GNN.
1
With the current data, we show that the distribution of spike rates within any neural tissue
follows a power-law distribution, i.e. a distribution with a 'heavy tail'. There is also a small
number of very low-frequency neurons, so that we have a lognormal distribution [2]. This lognormal
distribution is present in spontaneous spike rates as well as under behavioral stimulation. For
each neuron, the deviation from the mean rate attributable to a stimulus is small (CV = 0.3–1,
standard deviation = 1–4 spikes/s), when compared to the variability in mean spike rate over the
whole population (5-7-fold), cf. Tables 1 and3.
This work refers back to data initially reported in [1]. At the time, we only had data on spike
rates of cortical neurons available, plus independent evidence on intrinsic properties of striatal
neurons. The observation on cortical data was taken up by [12, 2], and led to a number of papers
[13, 14] focusing on the power-law distribution of spike rates as a cortical phenomenon, seeking
explanations in the recurrent excitatory connectivity of cortical tissue [2, 15, 13, 16]. However,we
find the same spike rate distributions for midbrain nuclei, medium striatal neurons and cerebellar
Purkinje cells, which do not have this kind of connectivity. It has even been found in the spinal motor
networks of turtles [17]. We then extended the data search for intrinsic excitability and found that
lognormal distributions are ubiquitous there as well, at least in cortical as well as in striatal tissues.
Finally, lognormal distributions have also been found for synaptic weights [6, 7, 8, 9, 10, 11, 18].
The explanation for this universal phenomenon must lie elsewhere.
For this purpose we constructed a generic model for neuronal populations with adaptable weights
and gains. We initialized both weights and gains with uniform, Gaussian or lognormal distribu-
tions.We then employed either Hebbian or homeostatic adaptation rules on both. Under a variety
of conditions we could show that lognormal distributions develop from any initial distribution
only with Hebbian (positive) adaptivity. Additional homeostatic adaptation stabilized learning but
erased the lognormal distributions if it was stronger than Hebbian adaptation. We could even show
that the widths of the distributions from the model match with the experimental data for rates,
weights, and gains (Tables 1 and 3) that we have available. Lognormal distributions can only be
maintained by positive, Hebbian-type learning rules [15], while homeostatic plasticity alone destroys
lognormal distributions [19].
There are a number of different learning rules and variants which all follow the 'Hebbian' princi-
ple: strong activation leads to strengthening, weak activation leads to weakening [20]. STDP rules
are a variant of Hebbian learning for spiking neurons, which emphasize temporal sequence, but have
the same positive learning effect [21, 22, 23]. It has been noticed that positive learning rules lead
to run-away activation and unstable network behavior, and that they need to be counteracted by
homeostatic processes [24, 25] . We present a generalized model of synaptic learning which consists
of both Hebbian (positive) and homeostatic (negative) adaptation rules [24], and show that positive
(Hebbian) learning is necessary to establish a lognormal synaptic weight distribution.
For intrinsic learning it has often been assumed that it may implement purely homeostatic
adaptation [26, 27, 28, 29, 30], but see also [31]. Experimental results are often inconsistent [32,
33, 34, 4, 35, 36, 37, 38]. We will present results for a lognormal gain distribution in a number
of tissues. It will be shown by simulation that the same principle holds: only a Hebbian, positive
learning rule is capable of maintaining lognormal distributions, while homeostatic adaptation serves
to establish stability.
This finally answers the question that experimental researchers have investigated for some time:
Is intrinsic plasticity mostly homeostatic, i.e. adjusts values inversely to use, or is there Hebbian,
positive learning involved: when a neuron fires, does its gain increase? The answer is that the
attested distribution of intrinsic gain can only derive from a Hebbian style adjustment rule, even
2
though additional homeostatic adaptation is possible. Intrinsic plasticity is Hebbian.
2 Methods
In this section we report on data collection for spike rates, intrinsic properties and synaptic weights.
Secondly, we explain the simulation model we constructed to explore the generation and persistence
of the attested distributions.
2.1 Experimental data
We analyze five data sets for spike rates from principal neurons under behavioral activation:
1. inferior temporal (IT) cortex from monkeys [39]
2. primary auditory cortex (A1) from monkeys [40]
3. primary auditory cortex (A1) of rats [2]
4. Purkinje cells in cerebellum [41, 42, 43, 44]
5. midbrain principal cells from inferior colliculus (IC) from the guinea pig [45]
In monkey IT, single unit activity was recorded over 200ms for passive viewing of 77 different
natural stimuli for 100 neurons, each stimulus shown 10 times [39]. This yielded 770 spike rate
response data points per neuron. For these data, we show mean spike rate, standard deviation,
max-min values, coefficient of variation (CV) and Fano factor (FF) (Figure 1), (cf.
[1]). What is
remarkable is that the dispersion for each neuron (variance related to mean, FF) is fairly constant,
and not related to the rank of a neuron as high or low-frequency. In other words, neurons have
roughly similar behavioral responses relative to their average spike rate. For this reason,many
behavioral experiments have reported percentage of increase/decrease of spiking as the relevant
parameter.
Figure 1: Spike rate data for neurons from inferotemporal cortex (IT) in monkeys [39].
100 neurons, passive viewing of 77 stimuli, 10 trials (770 data points per neuron), data collected
over 200ms. The data are shown for each neuron, where neurons are sorted by mean spike count.
A: Mean spike rates (blue), standard deviation (red), and minimum/maximum absolute values
(green). B: Mean spike rates histogram shows a lognormal distribution (σ*=2.32). C: Distributions
of standard deviation (green) and CV (blue) have linear slopes, with small variation. The Fano
factor (red), measuring the dispersion for each neuron, is nearly constant at about 2.
Additionally, we show data from primary auditory cortex (A1) from awake monkeys, which were
recorded for spike responses to a 50ms, 100ms, or 200ms pure tone ([40], Figure 2A and B) and
3
010203040506070809010002468101214161820neuron numberspikes/second010203040506070809010001234CV [σ/µ] F [σ2/µ]R=0.9R=0.79R=0.110102030405060708090100246 Std Devneuron number0246810121402468101214spikes/secondnumber of neuronsR=0.894ABC Figure 2: Responses to pure tones in primary auditory cortex from awake monkeys
[40] and firing rates from primary auditory cortex in rats [2]. A: Distribution of spike
rates to a 50ms tone (n = 119, red) 100ms tone (n = 115, blue), or 200ms tone (n = 23, green) in
primary auditory cortex in monkeys [40]. B: Histogram for spike rate distribution for the 100ms
tone response (n = 115) fitted by an exponential (red) or lognormal (blue) distribution [40]. C:
Spontaneous spike rate distribution from primary auditory cortex in unanesthetized rats [2] fitted
by an exponential (red) or lognormal (blue) distribution. Note that the spontaneous firing rates
are much lower and narrower distributed than evoked spikes in response to stimuli at short time
scales (B), but that they still follow a lognormal distribution.
data for spike rates from the primary auditory cortex of rats for four different conditions, which
were recorded as cell-attached in vivo recordings ([2], Figure 2C).
For midbrain nuclei neurons (IC), we re-analyzed spike rates in response to tones (for 200ms
after stimulus onset) under variations of binaural correlation [45]. The frequency ranking of neurons
by mean spike rate, standard deviation, min-max values, CV and FF are shown in (Figure 3A-C).
CV and FF are similar to the cortical data.
Figure 3: Neuronal response to binaural stimulation for inferior colliculus of the guinea
pig [45] (n = 30), data collected over 100 ms, 200-500 trials. The data are shown for each
neuron, with neurons sorted by mean spike count. A: Mean spike rates (blue), standard deviation
(red), and minimum/maximum absolute values (green). B: Mean spike rates histogram shows a
lognormal distribution. C: Distributions of standard deviation (green), CV (blue) and FF (red).
Again, the dispersion is fairly constant.
Data from GABAergic cerebellar Purkinje cells offer some difficulty for this analysis since they
have regular single spikes at high frequencies, and in addition, calcium-based complex spikes [43].
4
020406080100120020406080100120140160neuron numberspikes/secondspikes/second020406080100120number of neurons0246810121416R=0.925R=0.62spikes/second0102030405060number of neurons0102030405060708090R=0.971R=0.82ABC051015202530050100150200250neuron numberspikes/second020406080100120140012345678910spikes/secondnumber of neuronsR=0.9250510152025300121020304050CV [σ/µ] F [σ2/µ] R=0.56R=0.57R=0.2051015202530020406080 Std Devneuron numberABC Complex spike rates however are low (< 1Hz). This can therefore be regarded as a form of
multiplexing, with two separate codes, where single spike rates can be separately assessed in their
distribution. Here we report data for single spikes from in vivo recordings in anesthetized rats [43]
(Figure 4A)and data from spontaneous spiking (in the absence of synaptic stimulation) under in
vitro conditions ([41, 42], Figure 4B and C).
In order to show values for standard deviation and variance, data for two Purkinje cells from a
behavioral experiment, [44], i.e. single spike rates during arm movements from monkeys, have been
added to the ranking of spontaneously firing neurons by mean spike rate from [41] (Figure 4D).
Figure 4: Spike rates for cerebellar Purkinje cells from rats [41, 42, 43, 44]. A: Data for
single spikes for Purkinje cells recorded from anesthetized rats (spontaneous in vivo) [43](n = 346).
B: Data for spike frequencies of isolated cell bodies of mouse Purkinje cells in vitro [41] (n = 34).
C: Spontaneous spike rates for Purkinje cells in slices (n = 106) [42]. D: Spike counts per neuron
from [41] (n = 34), together with variability data from [44] (n = 2).
The logarithmic (heavy-tailed) distribution of spike rates is evident under all conditions.
The distribution of spike rates for neurons spiking in the absence of synaptic input shows that
there are differences in the intrinsic excitability of neurons. To further explore this we looked at
three additional datasets, which report the action potential firing of a cell due to injected current,
such as by a constant pulse. This can be defined as the neuronal gain parameter (spike rate divided
by current, [Hz/nA]).
1. medium striatal neurons in slices from rat dorsal striatum and nucleus accumbens shell (NAcb
shell) [3], cf. [1], Figure 5.
2. cortical neurons in cat area 17 in vivo [46], Figure 6A and B
5
D12366084108132051015202530spikes/secondnumber of neuronsR=0.980020406080100120140160024681012spikes/secondnumber of neuronsR=0.87205101520253035020406080100120140160180neuron numberspikes/secondABC020406080100120140160180010203040506070spikes/secondnumber of neuronsR=0.862 3. striatal neurons from globus pallidus (GP) from awake rats [5], Figure 6C
In [1], we already presented the data from striatum, which show that the spike response to a
constant current follows a heavy-tailed distribution [3]. Figure 5A shows the spike rate in response
to current pulses of different magnitude in two different areas, nucleus accumbens (NAcb) shell
and dorsal striatum. Figure 5B and C show the distribution of rheobase (current-to-threshold) for
dorsal striatal and NAcb shell neurons. Distributions appear mostly lognormal, with the exception
of the 200pA current pulse response and the data in Figure 5C, which appear normally distributed.
Figure 5: Spike rate and gain distributions in basal ganglia [3]. A: Spike rate in response
to a 300ms constant current pulse at 180pA (blue), 200pA (green), and 220pA (red) for neurons
in dorsal striatum (n = 28, solid lines) and nucleus accumbens shell (n = 24, dashed lines). B:
Gain (Hz/nA) for neurons in dorsal striatum (n = 28). C: Gain (Hz/nA) for neurons in NAcb shell
(n = 24).
We extend this dataset by recordings from different types of cortical neurons in cat area 17 in
vivo ([46], Figure 6A and B) and from GP in awake rats ([5], Figure 6C).
A lognormal distribution of intrinsic gain is clearly apparent, except for fast-spiking interneurons,
which, however, may be a sampling error (n=33).
Figure 6: Gain [Hz/nA] for cortical and striatal neurons. A: Gain for all types of cortical
neurons in vivo (n = 220) [46]. B: Gain for fast spiking cortical neurons only (n = 33) [46]. C:
Gain for neurons in globus pallidus (GP) in response to a +100pA current pulse (n = 145) [5].
Synaptic weight distributions have been investigated starting with [10] in hippocampus by mea-
suring EPSC magnitude [47, 6, 48, 49, 50] (Figure 7). There is also a review paper available [51]
to summarize the findings. Recently, the expression of AMPA receptor subunit GluA1, which is
6
0510152025300102030405060spikes/secondneuron numbergain [Hz/nA]050100150200250number of neurons01234567R=0.8175010015020025001234567gain [Hz/nA]number of neuronsR=0.795ABC01002003004005006007008000510152025303540gain [Hz/nA]number of neuronsR=0.972010020030040050060070080090000.511.522.533.544.55gain [Hz/nA]number of neuronsR=0.597050100150200250300350024681012141618gain [Hz/nA]number of neuronsR=0.921ABC correlated with spine size, has also been measured [52], Figure 8. We used five datasets from cortex,
hippocampus and cerebellum:
1. EPSPs for deep-layer pyramidal-pyramidal cell connections in rat visual cortex [47, 6]
2. EPSP amplitudes for deep-layer excitatory neuron connections in somatosensory cortical slices
of juvenile rats [49]
3. EPSP amplitudes for CA1 to CA3 connections in guinea pig hippocampal slices [10]
4. EPSCs for granule cells to Purkinje cells in adult rat cerebellar slices [50]
5. labeled GluA1 AMPA receptor subunit in mouse somatosensory barrel cortex [52]
Figure 7: Strengths of EPSPs in cortex, hippocampus and cerebellum [47, 6, 49, 10, 50].
A: Cortex: Deep-layer (L5) pyramidal-pyramidal cell connections [47, 6]. B: Cortex: Deep-layer
(L5) pyramidal-pyramidal cell connections [49]. C: Hippocampus: CA1 to CA3 connections [10].
D: Cerebellum: Granule cells to Purkinje cells [50].
In [6], EPSP magnitude was measured for L5 pyramidal neurons in slices from rat visual cortex,
averaged over 45-60 responses, and peak amplitude recorded, (Figure 7A). Similar data were used in
[49] for slices from a single barrel column in rats (Figure 7B). A 30-fold variation of coupling strength
was noted. In [10], EPSPs between CA3 and CA1 in hippocampal slices were recorded, by detecting
7
D0123456700.10.20.30.40.50.60.70.80.9weight [mV]probability densityR=0.92800.10.20.30.40.50.60.7024681012weight [mV]probability densityR=0.95100.050.10.150.20.250.30.3502468101214weight [mV]probability densityR=0.889ABC00.511.522.533.5400.20.40.60.811.2weight [mV]probability densityR=0.964 somatic membrane potential changes in response to presynaptic neuron stimulation (Figure 7C).
There are also synaptic weight data on granule cell to Purkinje cell connections [50, 11, 51], which
show a similar distribution, but have an order of magnitude weaker connections than cortical
connections (Figure 7D). Finally, a different type of evidence was obtained in [52], namely labeling
for a subunit of AMPA receptors in layer 2/3 mouse barrel cortex in vivo both before and after
whisker stimulation. The AMPA intensity is distributed lognormally over the spines, corresponding
to the observations on the strengths of EPSPs. It is noticeable that stimulation leads to increase
of on average 200% (two-fold) in about 30% of spines [52]. Yet as we know, over time the overall
distribution of synaptic strengths remains stable. For synaptic weights, just as for intrinsic gains
Figure 8: AMPA subunit distribution as a marker of synaptic weight. A: Expression of
labeled GluA1 AMPA receptor subunit in layer 2/3 mouse barrel cortex in vivo follows a lognormal
distribution (σ∗ = 2.59, µ∗ = 0.32, n = 560). B: GluA1 density for control (black) and after 1 hour
whisker stimulation (red). Stimulation leads to an increase of GluA1 in ≈ 30% of neurons [52].
and spike rates, lognormal distributions have been found for both EPSPs and AMPA receptor
distribution in a highly consistent manner.
In many cases, the data were only available in the form of histograms. The parameters of the
lognormal distribution were then obtained by fitting the data using a Nelder-Mead optimization
method. A number of parameters were derived from these fits and reproduced in Tables 1–3, cf. 3.2.
2.2 Simulation model
Given are two neuron populations I, J each with n = 1000 neurons and variable random connectivity
C between I and J. C determines the density between I and J. The input population I always
has excitatory output onto J.
Inhibitory input to J is modeled by a population H with n =
200. The output neuron population J may also have recurrent excitatory connectivity. Figure 9
shows the architecture for the generic neural network used. The model (GNN) was programmed in
Matlab, and is available in the public repository github (https://github.com/gscheler/GNN, DOI:
https://doi.org/10.5281/zenodo.829949).
The input population I is modeled according to [2] for pyramidal cortical neurons with a spike
rate distribution of µ∗ = 4.95 and σ∗ = 1.98 (Table 1). The goal is to generate a spike rate
distribution RJ for J, given a gain distribution G for the target neurons, and the weight distribution
WIJ such that RJ is similar to RI .
8
Spine sGluA1 intensity hour 0 (a.u.)0246810Number of spines0102030SameUpDownPercentage of spines01020304050607080AB Figure 9: Generic neural network model with neuron populations I, J (excitatory, blue
arrows) and H (inhibitory, red arrows). Lognormal distributions occur for gain G, rates RI ,
RJ , RH , and weight distributions WIJ . J may have recurrent connectivity.
For each neuron j in J the spike rate rj is calculated by applying its gain gj to the weighted
sum of its connected excitatory inputs and its inhibition. Cj is the set of neurons from I that have
excitatory connections to neuron j.
(cid:88)
i∈Cj
rj = gj(
wijri − rH
j )
(1)
where the rate ri is taken from the distribution RI , rH
from RH , wij from WIJ , and gj from G.
j
gj is modeled as a factor for a linear gain function. It is possible to use a sigmoidal gain function
instead, but this makes no difference for the conclusions from the model (Section 3.3). The output
RJ may be used as input to I with a matrix WI,J for tests of the adaptation rules.
For the adaptation of weights WIJ and gains G, we use Hebbian or homeostatic rules, as
described in Section 3.3.
The system described in this way is sufficient for all the calculations on the shape of distributions
used in this paper.
3 Results
3.1 Universality of lognormal distributions
We have documented the distribution of spike rates, gains, and weights for different types of neurons
(Figures 1- 8).
The distribution in all cases follows a lognormal shape.
In some cases, we had data on the
variability of spike rates and analyzed them for dispersion (CV, FF) under behavioral stimulation.
While the fold-change from low spiking neurons to high spiking neurons is high, 5- to 7-fold, the
9
0510152025300102030405060gain [Hz/pA]number of neurons0510152025300102030405060gain [Hz/pA]number of neurons0510152025300102030405060gain [Hz/pA]number of neurons0510152025303540020406080100120140spikes/secondnumber of neuronsRRIJ020406080100120140160180200020406080100120140spikes/secondnumber of neuronsHWHGIJGRIJGGGGGGG variability for each neuron is comparatively low. It also seems to be adequately described by a
percentage change over the whole population. This means that a low spiking neuron never reaches
the same rate as a high spiking neuron, even when fully activated.
The similarities across neural systems are striking. For instance, in a midbrain nucleus (inferior
colliculus) which is essentially an 'output' site for auditory and somatosensory cortex, spike rates are
high overall [45], nonetheless the distribution of mean spike rate, and the variability are comparable
to cortical data (Figure 3). Hippocampus, cerebellum and cortex vary in degree of bursting and
spike irregularity [53], but the rate distribution is constant. The distribution of mean spike rate is
also essentially the same under spontaneous and under behavioral conditions.
Lognormal distributions were obtained by fit to the histograms obtained from data (goodness
of linear fit, mean ≈ 0.92, s. figures 1-8). The lognormal distribution is a very simple statistical
distribution [54], almost as simple and as universal in the description of natural processes as a
Gaussian distribution (with which it is identical for small σ∗). Even though the datasets were
occasionally fairly small, and more data could be added to obtain greater precision, the conclusion
seems warranted that the underlying natural process is as simple and general as the multiplication
of independent variables [55], rather than assuming more complex processes which may lead to
other exponential-family distributions.
Lognormal rate distributions appear to be an essential property of neural tissues that occur in
areas with very different neuron types and connectivity, and different absolute spike frequencies.
They are present during spontaneous activity, and under activation of a network, in vivo as well
as in vitro. They have a counterpart in a lognormal distribution of intrinsic excitability, and
lognormal synaptic connectivity. This type of distribution seems to be an essential component
of the functional structure of a mature network, which is not altered by learning, plasticity, or
processing of information.
3.2 Data analysis for distributions
A lognormal distribution is characterized by parameters µ∗and σ∗. µ∗ = eµ is the median, a scale
parameter, which determines the height of the distribution. σ∗ = eσ is the multiplicative standard
deviation, a shape parameter which determines the width of the distribution. For distributions with
small σ∗(approximately σ∗ < 1.2 or σ < 0.182) a lognormal distribution is essentially identical to a
normal distribution. (The coefficient of variation CV ∼ σ∗ − 1, so that for CV < 0.18, a lognormal
equals normal distribution.) We collected data on spike rate, gain and synaptic weight distribution
for a number of tissues in different experimental conditions (Tables 1 -3). For the height of the
spike rate distribution, there are known differences, e.g. with lower values for cortex (µ∗ ≈ 4.5) and
higher values for Purkinje cell (µ∗ ≈ 30) and midbrain nuclei (cf. Table 1). In other words, spike
rates differ between brain areas such as cortex and cerebellum by a factor of 10.
In contrast to that, the width of spike rate distributions is more similar, with an average at
σ∗ ≈ 2.2, with one outlier. The gain has a smaller σ∗, i.e. a more normal, less heavy-tailed
distribution than the spike rate. Minus the outlier (3.46), the mean for σ∗ is only 1.86, considerably
lower than the width of the spike rate distribution (Table 2). For weight distributions (Table 3),
the width σ∗ is consistently larger, with an average of almost 3 (2.91). The synaptic strength (µ∗)
varies over at least one order of magnitude between cortex and cerebellum.
It turns out that σ∗ values are significantly different for rates, gains and weights, lowest for gains
(σ∗ ≈ 1.8), higher for rates (σ∗ ≈ 2.2) and highest,(σ∗ ≈ 3), for weights. The data that we have are
not precise enough to draw quantitative conclusions, but no large distinctions are apparent between
10
the tissues (Tables 1 - 3). We use a generic neural network to recreate lognormal distributions by
adaptation rules and we will also show that distribution widths are structural properties which
follow from general network properties.
Tissue
IT cortex [39]
A1 cortex [2]
A1 cortex [40]
Purkinje in vitro [42]
Purkinje in vitro [41]
Purkinje in vivo [43]
Inferior colliculus [45]
Mean Median Variance Width Mode
eµ−σ2
2.2
3.1
13.6
27.6
30.0
20.7
11.13
µ∗
4.50
4.95
27.11
31.82
31.19
33.12
27.39
σ2
0.71
0.47
0.69
0.14
0.198
0.47
0.90
σ∗
2.32
1.98
2.29
1.46
1.56
1.98
2.58
µ
1.5
1.6
3.3
3.46
3.44
3.5
3.31
n
100
145
263
106
34
319
30
Table 1: Statistics of spike rate distributions in different tissues.
Tissue
Peak
Dorsal striatum [3]
NAcb shell [3]
GP in vivo [5]
GP model [5]
Cortical [46]
[Hz/nA] µ[Hz/nA]
4.24
4.67
3.4
4.0
4.96
48
65
6.6
37
105
Mean Median Variance Width Mode
eµ−σ2
48.3
65.4
6.4
36.6
104.6
µ∗
69.41
106.70
29.96
54.60
142.59
σ∗
1.82
2.01
3.46
1.88
1.75
σ2
0.36
0.49
1.54
0.40
0.31
n
28
24
146
10000
220
Table 2: Statistics of intrinsic excitability (gain) in different tissues.
Tissue
Cortex L23 [7]
Cortex L23 [8]
Cortex L23 [9]
Cortex L5 [47]
Cortex L5 [49]
Hippocampus [10]
Cerebellum GC-PC [11]
Cortex in vivo [52]
Mean Median Variance Width Mode
eµ−σ2
0.17
0.31
0.08
0.13
0.41
0.05
0.01
0.13
µ
-0.99
0.25
-0.94
-0.56
-0.31
-2.61
-2.70
-1.14
σ∗
2.39
3.28
3.46
3.36
2.14
1.93
3.85
2.59
µ∗
0.37
1.28
0.39
0.57
0.73
0.07
0.07
0.32
σ2
0.76
1.41
1.54
1.47
0.58
0.43
1.82
0.90
n
48
35
61
1004
26
71
104
560
Table 3: Statistics of synaptic weight distributions in different tissues.
3.3 Generating lognormal distributions with generic neural networks
Since not only mean spike rates but also both components, intrinsic excitability and synaptic
weights, have lognormal distribution, this raises the question of how the functional system that we
11
observe is generated. It is obvious, if the data are accurate, that these are basic parameters of any
simulation and need to be reproduced in any model to make it biologically realistic.
We set up a generic neural network model (cf. Section 2.2) to explore the mechanisms of
generating and maintaining rate, weight and gain distributions. The model consists of a source
neuron group I, a target group J, a population of inhibitory neurons H, which are connected with
J, and potentially recurrent excitation in the target group J. The spike rate distribution RI acts
through a weight distribution W onto a gain distribution G, where inhibition H is subtracted, and
a spike rate output distribution RJ is produced (Figure 9).
In the simplest case, we look at two sets of neurons, the source and the target. The source sends
excitatory connections to the target, and exhibits variable weights at outgoing synapses. The input
that a target neuron receives is fed through a linear filter G to produce an output rate RJ according
to Eq (1). The distribution for RJ depends on G and W as well as on RI . The system is sufficient
for calculations on the shape of distributions, as well as the effects of Hebbian and homeostatic
plasticity.
We have explored the dependencies between gain, weight and rate distributions in simulations.
First, we found that the width of the output spike rate distribution RJ depends heavily on the gain
distribution, but only slightly on the input weight distribution (Figure 10). It does depend on the
overall connectivity C, where σ∗
RJ is wider for lower connectivity, but not very much (Figure 10).
Secondly, the width of the output distribution RJ does not depend on RI or RH either (Figure 11).
Figure 10: The width of the rate distribution for J, σ∗
G, but not on the weight distribution σ∗
σ∗
sheet C=5%,lower sheet C=10%). (µ∗
W = 0.7, µ∗
RJ , depends heavily on the gain
W . There is a slight effect of connectivity (upper
G = 30, N = 1000, σ∗
RI = 2.74, µ∗
RI = 4.5.)
The most important factor for a spike rate distribution remains the gain σ∗
G.
12
Figure 11: The width of the rate distribution for J, σ∗
(µ∗
G = 30, C = 10%, N = 1000, µ∗
W = 0.7, µ∗
RJ , does not depend on RI or RH .
RI = 4.5.)
3.4 Adaptation
We may now ask, where do lognormal spike rate distributions come from? How is the system set
up, i.e. what rules of adaptation generate lognormal distributions in weights and gains?
In the case of cortical networks, there are excitatory recurrent interactions that constitute a
significant part of total input. In the case of cerebellar or striatal neurons, there are no recurrent
excitatory interactions, only inhibitory interneurons and excitatory input. The generation of log-
normal distributions must therefore be independent of recurrent excitation. It requires a system
where continuous input shapes the weights and gains of a target network J.
We start with the system that we described before, with random assignment of weights and gains.
We employ adaptivity for weights, and also for gains, by positive Hebbian learning, or by negative
homeostatic learning. The output of I is fed into J, and W and G are adaptive. Additionally, J
may have excitatory recurrent connectivity, and learning takes place within the network J.
From any given initial spike rate distribution (Gaussian, uniform, lognormal) for I, we calculate
W assuming a positive learning (Hebbian) adjustment rule, which is dependent on input and output
frequencies. Each individual weight wij is updated by
w(cid:48)
ij = wij + λwij(rI
j − µ)
i rJ
We use parameters λ and µ, such that the generated spike rate output RJ is compatible in strength
with the input rate RI .
Using Hebbian learning, we generate a weight distribution W that is lognormally distributed,
independent of the initial configuration or the distribution of the gains in the system (Figure 12).
13
4.23.73.22.7σ*R(H)2.21.71.21.21.72.2σ*R(I)2.73.23.72.72.63.22.82.933.14.2σ*R(J) The lognormal distribution also develops independently of the rate distribution of the inputs, it
only develops faster with lognormal rather than normally distributed spike rate input (not shown).
It makes no difference whether we use a recurrent system J, or a non-recurrent population J with
input from a population I with a spike rate distribution, as long as we use a Hebbian weight
adaptation rule. For the shape of the distribution, it also does not matter whether we route the
output of J back to I, or whether we use local or no recurrence.
Figure 12: Hebbian learning results in lognormal weight distribution independent of
gain distribution. Given is a lognormal input rate σ∗
RI = 4.95. A: Initial weight
configurations: Gaussian (grey) or uniform (blue). B: After Hebbian learning using Gaussian gain
distribution (grey, blue as before). C: After Hebbian learning using uniform gain distribution (grey,
blue as before). D: After Hebbian learning using lognormal gain distribution. (σ∗
G =
32.7) (grey, blue as before).
G = 1.37, µ∗
RI = 2, µ∗
To show the effect of the adaptation rule, we also used homeostatic synaptic plasticity to adjust
the weights. This means that the weight is adjusted inversely to the spike rate of input and output
neurons.
j − µ)
i rJ
ij = wij − λwij(rI
w(cid:48)
14
DWij00.511.522.533.54number×10500.511.522.53Wij00.511.522.533.54number×10500.511.522.53Wij00.511.522.533.54number×10500.511.522.53ABCWij00.511.522.533.54number×10500.511.522.53 In this case, it is very clear that with any input or initial configuration and any gain distribution,
only a normal distribution of weights results (Figure 13). Again, a lognormal input spike rate slows
the process of adaptation, but the end result is the same, a normal distribution.
Figure 13: Homeostatic learning results in Gaussian weight distributions, indepen-
dent of gain distribution. Given is a lognormal input rate σ∗
RI = 4.95. A: Initial
Configuration: Gaussian (grey) or uniform (blue). B: Homeostatic weight learning using Gaussian
gain distribution (grey, blue as before). C: Homeostatic weight learning using uniform gain distri-
bution (grey, blue as before). D: Homeostatic weight learning using lognormal gain distribution
(σ∗
RI = 2, µ∗
G = 1.37, µ∗
G = 32.7) (grey, blue as before).
Since gain distributions are also lognormal, we may ask in the same way how they develop and
are maintained by plasticity rules. We adapt the linear gain G by either Hebbian or homeostatic
learning. Each gain can be adjusted by a Hebbian rule
or a homeostatic rule
g(cid:48)
j = gj + λgj(rJ
j − µ)
j = gj − λgj(rJ
g(cid:48)
j − µ)
15
DWij00.511.522.533.54number×104012345678910Wij00.511.522.533.54number×104012345678910Wij00.511.522.533.54number×104012345678910ABCWij00.511.522.533.54number×104012345678910 with parameters λ and µ.
We start with uniform or normally distributed G in an environment where W is lognormal,
normal or uniform, and RI is normal or lognormal. If we adapt only G for any initial configuration,
using any distribution for RI , including the lognormal distribution, and a lognormal or normal
weight distribution, we arrive at a normal distribution for G with homeostatic learning and a
lognormal distribution with Hebbian learning (Figure 14).
Lognormal distributions develop from Hebbian plasticity, and homeostatic plasticity generates
only normal distributions. The explanation lies in the nature of random statistical events, which
generate normal distributions when the underlying mechanisms are sums of many small events, but
lognormal distributions when the underlying mechanisms are multiplicative [54]. We also wanted
RI = 2, µ∗
Figure 14: Hebbian or homeostatic gain learning determine lognormal or Gaussian
outcome. Given is a lognormal input rate σ∗
RI = 4.95. A: Initial Configuration: Gaussian
(grey) or uniform (blue). B and C: Hebbian learning using lognormal or Gaussian weights, resulting
gain distribution is lognormal. D and E: Homeostatic learning using lognormal or Gaussian
weights, resulting gain distribution is Gaussian.
to understand the observed widths of the distributions. We hypothesized that the differences for σ∗
between W , R and G result from the network structure. Accordingly we started a simulation with
initial uniform values for G and W and Hebbian update rules using the same learning rate λ for
both (Figure 15). We find that gain, rate and weight distributions match the experimental values,
and that this is true for any tested constellation. We also found that Hebbian learning alone quickly
escalates values, which develop exponentially, and that additional rounds of homeostatic adaptation
are required to stabilize the system. Homeostatic learning pushes the system back towards a normal
distribution.
Our data, in the most general way, allowing for various conditions and architectures, show that
16
gain20406080100120number0100200gain20406080100120number0100200gain305070number100200300400500gain305070number100200300400500gain305070number1002003001000DEACB Figure 15: Experimental and generated distribution widths for spike rates, gains and
weights. Grey, experimental measurements (s. Tables); red, generated with 100% Hebbian learn-
ing; or blue, 80% Hebbian and 20% homeostatic learning combined. The basic distinction in distri-
bution width for gains, rates, weights is reproduced with Hebbian learning, additional homeostatic
learning matches experimental values best.
Hebbian learning is required both for intrinsic gain and for weights in order to generate the attested
lognormal distributions. This is an interesting result, because it shows that we need prominent
Hebbian intrinsic learning to explain the gain distributions that we find experimentally. Intrinsic
learning is not just homeostatic adaptation, it follows the same rules as synaptic weight learning.
4 Discussion
4.1 Universality of lognormal distributions
Spike rates of neurons seem to be universally distributed according to a lognormal distribution, with
many neurons at low spike rates, and a small number at successively higher spike rates (heavy-tail)
[1]. The same distributions are found for synaptic weights [6], and intrinsic properties associated
with excitability (gain) [1]. The neurons that we reported on are of very different types, and they
are embedded in different kinds of connectivity. Medium spiny neurons and Purkinje cells are
GABAergic (inhibitory), while cortical and IC neurons are glutamatergic (excitatory), but this is
not reflected in a distinct spike rate distribution. They also fire with very different average spike
rates. IC neurons operate at very high frequencies, and Purkinje neurons at much higher frequencies
than cortical or striatal projection neurons. But they all have the same spike rate distribution. It
has been suggested [2] that lognormal spike distributions are a feature of cortical tissue and arise
from strong excitatory recurrent connectivity, but this is experimentally not substantiated nor is
it theoretically necessary. While cortical pyramidal neurons exist in a heavily excitatory recurrent
17
gainspike rateweightdistribution σ*11.522.533.54 environment, medium spiny neurons, cerebellar Purkinje cells and IC neurons act mostly in a
feed-forward way, i.e. they don't have significant recurrent excitatory (glutamatergic) connectivity.
Beyond spike rate distribution, we also gathered data on weight and gain distributions. Again
the observation of lognormal distributions is ubiquitous. We find synaptic weight distributions
for cortex [6] and cerebellum that are lognormal, with characteristic width of distributions. For
intrinsic properties, striatal projection neurons and cortical neurons [46] show responses to constant
current and current-to-threshold (gain) distributions, which again appear lognormally distributed,
with smaller widths than spike rate distributions.
Our models show that lognormal distributions arise even in a purely input-output environment,
and that they are a result of Hebbian learning of weights and gains, quite independent of the overall
magnitude of the spike rates.
4.2 Generating lognormal distributions
Mean spike rates, as well as intrinsic excitability and synaptic weights, have lognormal distributions.
It has often been assumed that variability in intrinsic excitability is a source of noise in neural
computation [56], even though others have argued that intrinsic variability contributes to neural
coding [57, 58] and that intrinsic plasticity follows certain rules [19]. An excellent overview of the
experimentally attested forms of intrinsic plasticity is contained in [35], cf. [59, 60, 61]. Many other
detailed observations are contained in [36, 37, 38, 62].
Recently, Mahon and Charpier [4] have shown that intrinsic excitability is stable in individual
neurons under control conditions, while stimulation protocols (e.g. in barrel cortex of anesthetized
rats) change intrinsic excitability by at least 50-100%. However, the conclusions drawn from the
experimental research are often contradictory. Intrinsic plasticity is sometimes assumed to act in
a negative, homeostatic way, i.e. opposite to synaptic plasticity [4], but sometimes in a 'Hebbian',
positive way, i.e. cooperative with synaptic plasticity [38, 63]. There is evidence for (short-term)
negative or homeostatic plasticity, which has been previously investigated [4].
Our work has now shown that any kind of neural system with linear gains requires positive,
Hebbian intrinsic plasticity to produce and maintain a lognormal distribution of gains. We also
could show that the observed widths of the distributions, i.e. the differences for σ∗ between W , R
and G, naturally result from the network structure and are built into the system simply by Hebbian
adaptation.
Lognormal distributions may arise as stable properties of the system during early development
(the set-up of the system), i.e. before actual pattern storage or event memory develops, and they are
maintained during processing by a Hebbian type of positive adaptation events. Homeostatic plas-
ticity consists in downregulating gains or weights with increases in firing rates. Purely homeostatic
learning results in normal distributions, and erases existing lognormal distributions. By combining
homeostatic and Hebbian adaptation we can achieve and maintain stable lognormal distributions.
4.3 Why logarithmic coding schemes
A lognormal distribution means that values are normally distributed on a logarithmic scale. From
an engineering perspective, basic Hebbian plasticity for synapses and intrinsic properties is sufficient
to generate stable logarithmic distributions. If there is random variation of multiplicative events,
as in Hebbian plasticity, a lognormal distribution will be the result [54].
18
This is related to principles of sensory coding, where logarithmic scale signal processing enhances
perception of weak signals, while also being able to respond to large signals - effectively increasing
the perceptual range compared to linear coding [14].
In an interconnected network logarithmic
coding may turn into a property for the access of representations. Feature clusters, or event traces
could be accessed by targeted connections to the top-level neurons, which then activate lower level
neurons in their immediate vicinity. By accessing high frequency neurons preferentially, a whole
feature area can be reached, and local diffusion will provide any additional computation. Similarly,
the results of a local computation can be efficiently distributed by high frequency neurons to other
areas. Fast point-to-point communication using only high frequency neurons may be sufficient for
fast responses in many cases. Scale-free networks in general support synchronization, which is also
a useful feature for rapid information transfer and access [64].
Recently, publications [65, 66] have shown that there is indeed a difference between high-
frequency and low-frequency neurons in their connectivity: high-frequency neurons have short
delays, strong connections, and directed targets, while low-frequency neurons have long delays,
weak connections and diffuse targets.
The lognormal distribution of spike rates has significant implications for neural coding. Log-
arithmic spike rates are coupled with linear variance for responses to behavioral stimulation. In
other words, the greatest part of the coding results already from the frequency rank of the neuron
itself, such that high frequency neurons have the largest impact. A fixed mean rate for each neuron
allows stable expectation values for network computations.
Logarithmic, hierarchical coding does not need to be sparse. The low frequency neurons may
matter the most in terms of input response. With lognormal synaptic weight distributions, if
strong synapses are kept stable, they may transmit an input neuron's mean firing rate to targets
and in this way provide stability to the system. All other synapses could be arbitrary. This would
allow for continued pattern learning to be implemented by the bulk of low weight synapses, while
the framework of neuronal interactions, e.g., the ensemble structure, could be unchanged. Such
a division of labor between strong synapses and weaker ones could have many advantages in a
complex, modular network.
Experimental data have often shown that sampling of neuronal responses from a large population
(105 or more neurons), which become activated at 30% or more, yields accuracy for a stimulus
already for small samples (100-200 neurons or 1-2%) (e.g., [67]). We may suggest that this happens
when we sample from a highly modular structure, and we have been able to replicate the effect
with lognormal networks [68].
5 Conclusions
In our earlier work [1], we found that intrinsic excitability manifested by spike response to current
injection and rheobase in vitro for dorsal striatal and nucleus accumbens neurons seems to have the
same distribution as the firing rate in cortex under in vivo conditions. Approximately at the same
time, [6] had observed a heavy-tailed distribution of synaptic weights in cortical tissue.
In this article, we have done three things: (a) collected data to show that rate, weight and gain
distributions in different brain areas all follow a heavy-tailed, specifically a lognormal distribution;
(b) created a generic neural network model to show that these distributions arise from Hebbian
learning, and specifically that intrinsic plasticity must be Hebbian as well; and (c) shown that the
width of the distributions, as experimentally attested, arise naturally from the network structure
and the role of its components, in a very robust way. We have also discussed what the lognormal
19
distribution means for neural coding: a division of labor between fast transmission by high-frequency
neurons and low-level computation by low-frequency neurons in a modular structure, and possibly
a division of labor between stable components (strong synapses, high-frequency neurons) and more
variable components (weak synapses, low-frequency neurons).
References
[1] G Scheler and J Schumann. Diversity and stability in neuronal output rates. In Soc Neurosci
Meeting, 2006.
[2] T Hrom´adka, MR DeWeese, and AM Zador. Sparse representation of sounds in the unanes-
thetized auditory cortex. PLoS Biol, 6(1):e16, January 2008.
[3] FW Hopf, MM Moran, ML Mohamedi, GF Gonzales, W Mailliard, JK Seamans, PW Kalivas,
and A Bonci. Inhibition of the slow calcium-dependent potassium channel in the lateral dorsal
striatum enhances action potential firing in slice and enhances performance in a habit memory
task. In Soc Neurosci Meeting, 2005.
[4] S Mahon and S Charpier. Bidirectional plasticity of intrinsic excitability controls sensory inputs
efficiency in layer 5 barrel cortex neurons in vivo. J Neurosci, 32(33):11377–89, August 2012.
[5] C Gunay, JR Edgerton, and D Jaeger. Channel density distributions explain spiking variability
in the globus pallidus: a combined physiology and computer simulation database approach. J
Neurosci, 28(30):7476–91, July 2008.
[6] S Song, PJ Sjostrom, M Reigl, S Nelson, and DB Chklovskii. Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS biology, 3(3):e68, March 2005.
[7] A Mason, A Nicoll, and K Stratford. Synaptic transmission between individual pyramidal
neurons of the rat visual cortex in vitro. J Neurosci, 11:72–84, 1991.
[8] D Feldmeyer, J Lubke, and B Sakmann. Efficacy and connectivity of intracolumnar pairs of
layer 2/3 pyramidal cells in the barrel cortex of juvenile rats. J Physiol, 575:583–602, 2006.
[9] C Holmgren, T Harkany, B Svennenfors, and Y Zilberter. Pyramidal cell communication within
local networks in layer 2/3 of rat neocortex. J Physiol, 551(Pt 1):139–53, 2003.
[10] RJ Sayer, MJ Friedlander, and SJ Redman. The time course and amplitude of EPSPs evoked at
synapses between pairs of hippocampal Ca3/Ca1 neurons in the hippocampal slice. J Neurosci,
10:826-836, 1990.
[11] N Brunel, V Hakim, P Isope, JP Nadal, and B Barbour. Optimal information storage and the
distribution of synaptic weights: perceptron versus Purkinje cell. Neuron, 43(5):745–57, 2004.
[12] M Shafi, Y Zhou, J Quintana, C Chow, J Fuster, and M Bodner. Variability in neuronal
activity in primate cortex during working memory tasks. J Neurosci, 146(3):1082–108, 2007.
[13] A Roxin, N Brunel, D Hansel, G Mongillo, and C van Vreeswijk. On the distribution of firing
rates in networks of cortical neurons. J Neurosci, 31(45):16217–26, 2011.
Cerebral Cortex, 23:293–304, 2013.
20
[14] G Buzsaki and K Mizuseki. The log-dynamic brain: how skewed distributions affect network
operations. Nat Rev Neurosci, 15(4):264–278, 2014.
[15] AA Koulakov, T Hrom´adka, and AM Zador. Correlated connectivity and the distribution of
firing rates in the neocortex. J Neurosci, 29(12):3685–94, 2009.
[16] A Wohrer, MD Humphries, and CK Machens. Population-wide distributions of neural activity
during perceptual decision-making. Progr Neurobiol, 103:156–93, April 2013.
[17] PC Petersen and RW Berg, Lognormal firing rate distribution reveals prominent fluctuation-
driven regime in spinal motor networks. eLife, 5, e18805, 2016.
[18] Y Ikegaya, T Sasaki, D Ishikawa, N Honma, K Tao, N Takahashi, G Minamisawa, S Ujita,
and N Matsuki, Interpyramid spike transmission stabilizes the sparseness of recurrent network
activity.
[19] G Scheler. Learning intrinsic excitability in medium spiny neurons. F1000Res, 2:88, 2014.
[20] M Gilson, C Savin, and F Zenke. Editorial: Emergent Neural Computation from the Interaction
of Different Forms of Plasticity. Front Comput Neurosci, 9:145, November 2015.
[21] S Song, KD Miller, and LF Abbott. Competitive Hebbian learning through spike-timing-
dependent synaptic plasticity. Nat Neurosci, 3(9):919–26, September 2000.
[22] W Gerstner and WM Kistler. Mathematical formulations of Hebbian learning. Biol Cybern,
87(5-6):404–415, December 2002.
[23] N Caporale and Y Dan. Spike Timing–Dependent Plasticity: A Hebbian Learning Rule. Annual
Review of Neuroscience, 31(1):25–46, July 2008.
[24] F Zenke and W Gerstner. Hebbian plasticity requires compensatory processes on multiple
timescales. Philos Trans R Soc Lond B Biol Sci, 372(1715):20160259, March 2017.
[25] C Tetzlaff, C Kolodziejski, I Markelic, and F Worgotter. Time scales of memory, learning, and
plasticity. Biol Cybern, 106(11-12):715–726, December 2012.
[26] NS Desai. Homeostatic plasticity in the CNS: synaptic and intrinsic forms. J Physiol Paris,
97(4-6):391–402, July 2003.
[27] J Naude, B Cessac, H Berry, and B Delord. Effects of Cellular Homeostatic Intrinsic Plasticity
on Dynamical and Computational Properties of Biological Recurrent Neural Networks. J
Neurosci, 33(38):15032–15043, September 2013.
[28] J Cannon and P Miller. Stable Control of Firing Rate Mean and Variance by Dual Homeostatic
Mechanisms. J Math Neurosci, 7(1):1, December 2017.
[29] V Kazantsev, S Gordleeva, S Stasenko, and A Dityatev. A Homeostatic Model of Neuronal
Firing Governed by Feedback Signals from the Extracellular Matrix. PLoS ONE, 7(7):e41646,
July 2012.
[30] RH Cudmore and NS Desai. Intrinsic plasticity. Scholarpedia, 3(2):1363, 2008.
21
[31] PJ Tully, MH Hennig, and A Lansner. Synaptic and nonsynaptic plasticity approximating
probabilistic inference. Front Synaptic Neurosci, 6:8, April 2014.
[32] M Sehgal, C Song, VL Ehlers, and JR Moyer. Learning to learn – Intrinsic plasticity as a
metaplasticity mechanism for memory formation. Neurobiol Learn Mem, 105:186–199, October
2013.
[33] LR Whitaker, BL Warren, M Venniro, TC Harte, KB McPherson, J Beidel, JM Bossert,
Y Shaham, A Bonci, and BT Hope. Bidirectional Modulation of Intrinsic Excitability in Rat
Prelimbic Cortex Neuronal Ensembles and Non-Ensembles after Operant Learning. Journal
Neurosci, 37(36):8845–8856, September 2017.
[34] SD Greenhill, A Ranson, and K Fox. Hebbian and Homeostatic Plasticity Mechanisms in
Regular Spiking and Intrinsic Bursting Cells of Cortical Layer 5. Neuron, 88(3):539–552,
November 2015.
[35] JT Paz, S Mahon, P Tiret, S Genet, B Delord, and S Charpier. Multiple forms of activity-
dependent intrinsic plasticity in layer V cortical neurons in vivo. J Physiol, 587(13):3189–3205,
2009.
[36] E Campanac, C Gasselin, A Baude, S Rama, N Ankri, and D Debanne. Enhanced intrinsic
excitability in basket cells maintains excitatory-inhibitory balance in hippocampal circuits.
Neuron, 77(4):712–722, 2013.
[37] E Campanac, G Daoudal, N Ankri, and D Debanne. Downregulation of dendritic IH in Ca1
pyramidal neurons after LTP. J Neurosci, 28(34):8635–8643, 2008.
[38] A Frick, J Magee, and D Johnston. LTP is accompanied by an enhanced local excitability of
pyramidal neuron dendrites. Nat Neurosci, 7(2):126–135, 02 2004.
[39] CP Hung, G Kreiman, T Poggio, and JJ DiCarlo. Fast readout of object identity from macaque
inferior temporal cortex. Science, 310(5749):863–6, 2005.
[40] X Wang, T Lu, RK Snider, and L Liang. Sustained firing in auditory cortex evoked by preferred
stimuli. Nat Letters, 435:341–345, 2005.
[41] IM Raman and BP Bean. Ionic currents underlying spontaneous action potentials in isolated
cerebellar Purkinje neurons. J Neurosci, 19(5):1663–74, 1999.
[42] M Hausser and BA Clark. Tonic synaptic inhibition modulates neuronal output pattern and
spatiotemporal synaptic integration. Neuron, 19(3):665–678, 1997.
[43] C de Solages, G Szapiro, N Brunel, V Hakim, P Isope, P Buisseret, C Rousseau, B Barbour,
and C L´ena. High-frequency organization and synchrony of activity in the Purkinje cell layer
of the cerebellum. Neuron, 58(5):775–88, 2008.
[44] AV Roitman, S Pasalar, MTV Johnson, and TJ Ebner. Position, direction of movement, and
speed tuning of cerebellar Purkinje cells during circular manual tracking in monkey. J Neurosci,
25(40):9244–9257, 2005.
22
[45] O Zohar, TM Shackleton, AR Palmer, and M Shamir. The effect of correlated neuronal firing
and neuronal heterogeneity on population coding accuracy in guinea pig inferior colliculus.
PloS one, 8(12):e81660, 2013.
[46] LG Nowak, R Azouz, MV Sanchez-Vives, CM Gray, and DA McCormick. Electrophysiological
classes of cat primary visual cortical neurons in vivo as revealed by quantitative analyses. J
Neurophysiol, 89(3):1541–66, 2003.
[47] PJ Sjostrom, GG Turrigiano, and SB Nelson. Rate, timing, and cooperativity jointly determine
cortical synaptic plasticity. Neuron, 32(6):1149–1164, 2014/08/03 2001.
[48] MCW van Rossum, GQ Bi, and GG Turrigiano. Stable Hebbian Learning from Spike Timing-
Dependent Plasticity. J Neurosci, 20(23):8812–8821, 2000.
[49] A Frick, D Feldmeyer, M Helmstaedter, and B Sakmann. Monosynaptic connections between
pairs of L5a pyramidal neurons in columns of juvenile rat somatosensory cortex. Cerebral
Cortex, 18(2):397–406, 2008.
[50] P Isope and B Barbour. Properties of unitary granule cell → Purkinje cell synapses in adult
rat cerebellar slices. J Neurosci, 22(22):9668–9678, 2002.
[51] B Barbour, N Brunel, V Hakim, and JP Nadal. What can we learn from synaptic weight
distributions? Trends Neurosci, 30(12):622–9, 2007.
[52] Y Zhang, RH Cudmore, DT Lin, DJ Linden, and RL Huganir. Visualization of NMDA receptor-
dependent AMPA receptor synaptic plasticity in vivo. Nat Neurosci, 18(3):402–7, 2015.
[53] Y Mochizuki, T Onaga, H Shimazaki, T Shimokawa, Y Tsubo, R Kimura, A Saiki, Y Sakai,
Y Isomura, S Fujisawa, K Shibata, D Hirai, T Furuta, T Kaneko, S Takahashi, T Nakazono,
S Ishino, Y Sakurai, T Kitsukawa, JW Lee, H Lee, MW Jung, C Babul, PE Maldonado,
K Takahashi, FI Arce-McShane, CF Ross, BJ Sessle, NG Hatsopoulos, T Brochier, A Riehle,
P Chorley, S Grun, H Nishijo, S Ichihara-Takeda, S Funahashi, K Shima, H Mushiake, Y Ya-
mane, H Tamura, I Fujita, N Inaba, K Kawano, S Kurkin, K Fukushima, K Kurata, M Taira,
KI Tsutsui, T Ogawa, H Komatsu, K Koida, K Toyama, BJ Richmond, and S Shinomoto.
Similarity in neuronal firing regimes across mammalian species. J Neurosci, 36(21):5736–5747,
2016.
[54] E Limpert, W Stahel, and M Abbt. Log-normal distributions across the sciences: Keys and
clues. BioScience, 51(5):341–352, 2001.
[55] EL Crow and K Shimizu, editors. Lognormal Distributions:Theory and Applications. Dekker,
1988.
[56] M Rudolph and A Destexhe. The discharge variability of neocortical neurons during high-
conductance states. Neuroscience, 119(3):855–873, 2003.
[57] G Scheler. Regulation of neuromodulator receptor efficacy–implications for whole-neuron and
synaptic plasticity. Prog Neurobiol, 72(6):399–415, 2004.
[58] M Stemmler and C Koch. How voltage-dependent conductances can adapt to maximize the
information encoded by neuronal firing rate. Nat Neurosci, 2(6):521–527, 06 1999.
23
[59] D Debanne. Plasticity of neuronal excitability in vivo. J Physiol, 587(13):3057–3058, 2009.
[60] E Campanac and D Debanne. Plasticity of neuronal excitability: Hebbian rules beyond the
synapse. Arch Ital Biol, 145(3–4):277–287, 2007.
[61] G Doudal and D Debanne. Long-term plasticity of intrinsic excitability:
learning rules and
mechanisms. Learn Mem, 10(6):456–465, 2003.
[62] TP Carvalho and DV Buonomano. Differential effects of excitatory and inhibitory plasticity
on synaptically-driven neuronal input-output functions. Neuron, 61(5):774–785, 03 2009.
[63] S Mahon, JM Deniau, and S Charpier. Various synaptic activities and firing patterns in
cortico-striatal and striatal neurons in vivo. J Physiol-Paris, 97(4–6):557 – 566, 2003.
[64] G Scheler. Network topology influences synchronization and intrinsic read-out. arXiv:q-
bio/0507037, 2005.
[65] Y Omura, MM Carvalho, K Inokuchi, and T Fukai. A lognormal recurrent network model for
burst generation during hippocampal sharp waves. J Neurosci, 35(43):14585–14601, 2015.
[66] S Nigam, M Shimono, S Ito, FC Yeh, N Timme, M Myroshnychenko, CC Lapish, Z Tosi,
P Hottowy, WC Smith, SC Masmanidis, AM Litke, O Sporns, and JM Beggs. Rich-club
organization in effective connectivity among cortical neurons. J Neurosci, 36(3):670–684, 2016.
[67] DH O'Connor, SP Peron, D Huber, and K Svoboda. Neural activity in barrel cortex underlying
Vibrissa-based object localization in mice. Neuron, 67(6):1048–61, September 2010.
[68] G. Scheler. Extreme pattern compression in lognormal networks [version 1; not peer reviewed].
F1000Research 2016, 5:2177 (poster) (doi: 10.7490/f1000research.1113011.1)
24
|
1807.11935 | 1 | 1807 | 2018-07-31T17:51:16 | Network models in neuroscience | [
"q-bio.NC"
] | From interacting cellular components to networks of neurons and neural systems, interconnected units comprise a fundamental organizing principle of the nervous system. Understanding how their patterns of connections and interactions give rise to the many functions of the nervous system is a primary goal of neuroscience. Recently, this pursuit has begun to benefit from the development of new mathematical tools that can relate a system's architecture to its dynamics and function. These tools, which are known collectively as network science, have been used with increasing success to build models of neural systems across spatial scales and species. Here we discuss the nature of network models in neuroscience. We begin with a review of model theory from a philosophical perspective to inform our view of networks as models of complex systems in general, and of the brain in particular. We then summarize the types of models that are frequently studied in network neuroscience along three primary dimensions: from data representations to first-principles theory, from biophysical realism to functional phenomenology, and from elementary descriptions to coarse-grained approximations. We then consider ways to validate these models, focusing on approaches that perturb a system to probe its function. We close with a description of important frontiers in the construction of network models and their relevance for understanding increasingly complex functions of neural systems. | q-bio.NC | q-bio |
Network models in neuroscience
Danielle S. Bassett1,2,3,4,7, Perry Zurn5, and Joshua I. Gold6
1Department of Bioengineering, School of Engineering and Applied Sciences,
University of Pennsylvania, Philadelphia, PA, 19104
2Department of Physics & Astronomy, College of Arts and Sciences,
University of Pennsylvania, Philadelphia, PA, 19104
3Department of Electrical & Systems Engineering, School of Engineering and Applied Sciences,
University of Pennsylvania, Philadelphia, PA, 19104
4Department of Neurology, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA, 19104
5Department of Philosophy, American University, Washington, DC, 20016
6Department of Neuroscience, Perelman School of Medicine,
University of Pennsylvania, Philadelphia, PA, 19104 and
7To whom correspondence should be addressed: [email protected]
From interacting cellular components to networks of neurons and neural systems, interconnected
units comprise a fundamental organizing principle of the nervous system. Understanding how their
patterns of connections and interactions give rise to the many functions of the nervous system is a
primary goal of neuroscience. Recently, this pursuit has begun to benefit from the development of
new mathematical tools that can relate a system's architecture to its dynamics and function. These
tools, which are known collectively as network science, have been used with increasing success to
build models of neural systems across spatial scales and species. Here we discuss the nature of
network models in neuroscience. We begin with a review of model theory from a philosophical
perspective to inform our view of networks as models of complex systems in general, and of the
brain in particular. We then summarize the types of models that are frequently studied in network
neuroscience along three primary dimensions: from data representations to first-principles theory,
from biophysical realism to functional phenomenology, and from elementary descriptions to coarse-
grained approximations. We then consider ways to validate these models, focusing on approaches
that perturb a system to probe its function. We close with a description of important frontiers
in the construction of network models and their relevance for understanding increasingly complex
functions of neural systems.
INTRODUCTION
The brain is composed of intricate networks that operate
at many different levels of organization. For example, at
small spatial scales gene regulatory networks direct neu-
ronal cell fate, and both chemical and electrical synapses
define the accessible routes of information transmission
between neurons [1]. At intermediate spatial scales, lam-
inar architecture in cortex is accompanied by stereotyped
inter-laminar connectivity thought to support ensemble
dynamics and resultant computations [2]. At even larger
spatial scales, the anatomical locations of inter-areal pro-
jections display a precise spatial arrangement associated
with a diverse repertoire of functional processes [3, 4].
Although networks are fundamental to brain structure,
the complexity of these networks poses challenges to un-
derstanding their function. Unlike a sphere, which we can
quickly guess to have the capacity to be rolled or thrown,
a network -- with its tangle of wires -- defies any simple
conspectus. Thus, even though more than a century has
passed since Camillo Golgi, Santiago Ramon y Cajal, and
other neuroanatomists introduced to the world the intri-
cate beauty of the networks of neurons that comprise
our nervous system, our understanding of how those net-
works give rise to perception, learning, memory, cogni-
tion, action, and other aspects of brain function remains
incomplete.
One fruitful set of approaches for understanding re-
lationships between the brain's networked architecture
and its many functions that has emerged in recent years
is network science. This discipline deals with the study
of systems whose structure, function, or dynamics de-
pend upon the pattern of interconnections between units
[5]. Network science is inherently interdisciplinary, draw-
ing on and integrating among recent advances in mathe-
matics, physics, computer science, and engineering [6, 7].
Although early work in the field was largely devoted to
the study of social systems, efforts over the last decade
have focused increasingly on the study of neural systems
across spatial scales, temporal scales, and species [8 -- 10].
These newer efforts, collectively referred to as network
neuroscience, model neural systems as networks to dis-
till the dependence of brain function and dysfunction on
interconnection architecture [11].
In this chapter, we review recent work in network
neuroscience that has been applied to understanding the
brain, emphasizing the diversity of approaches that now
fall under this general framework [12]. Because network
neuroscience is fundamentally a modelling endeavor,
we begin with a broad perspective on model theory
from philosophy. We then consider how networks are
models, before turning to a discussion of the types of
network models that are commonly used in neuroscience.
Motivated by approaches to validate network models via
prediction, we discuss the importance of perturbation-
based techniques for understanding network function.
We close by outlining important directions for future
work to build, use, and validate network models in
neuroscience.
MODEL THEORY - A PHILOSOPHICAL
PERSPECTIVE
The term "model" can bring quite different pictures to
the mind's eye. A small, inexact replica of a 1910 Schacht
Roadster. A miniature cityscape coarsely true to the
form of Cambridge, England. Rodin's "The Thinker". A
contemporary reworking of a piece from Greek mythology
[13]. A person one wishes to emulate, or an ideal one
hopes to become. IBM's Watson. A cerebral organoid
(miniature organ in vitro [14]) or a human blinking eye-
on-a-chip [15]. A set of interdependent partial differential
equations producing dynamics reminiscent of a real-world
system. In mentally traversing these diverse examples,
one immediately realizes that the space of model types
is exceedingly large, and one wonders whether it is even
possible to define what a model actually is, or what it is
not.
The question of how to define the term "model" is the
focus of a branch of philosophy known as model theory,
which aims to identify the essential elements that make
models what they are and to disambiguate the character-
istics that distinguish different types of models from one
another [16]. At its most basic level, a model is a repre-
sentation of one or more aspects of the world. It aims to
better understand what something is by measuring and
imaging what something does. As such, models inherit
a basic philosophical conundrum [17]: what precisely is
their relationship to the target systems they model? And
from whence do they derive their truth value? Must they
simply evidence functional coherence and have pragmatic
purchase, or must they also meaningfully correspond to
what they represent, and, if so, how is that meaningful-
ness determined?
In science, at least four types of models have been rec-
ognized that each provide different answers to these ques-
tions. These types are: scalar, idealized, analogical, and
phenomenological [18, 19]. Scalar models, much like the
Roadster replica, either magnify or reduce their target
systems.
Idealized models abstract and isolate a lim-
ited set of features from their target systems. Analogical
models highlight relevant similarities between two tar-
get systems, whether those similarities are shared proper-
ties or comparable structures. Finally, phenomenological
models represent only the observable elements of their
target systems, without postulating any theoretical ex-
planation as to why those elements are what they are.
Recent work in model theory explores the intriguing
possibility that all of these forms of scientific models are
heuristic devices not unlike literary fictions [20 -- 22]. The
2
models-as-fictions theory reconceptualizes models as fic-
tional entities that aim to narrativize certain features of
a target system [23 -- 31]. According to this framework,
models -- much like fictions -- may imaginatively isolate
and abstract or distort and exaggerate certain features of
the world in such a way as to facilitate epistemic access
[32]. They may creatively instantiate either analogical or
phenomenological substructures of their target systems
in order to crystallize insight. Scientific models are there-
fore subject to evaluation at the level of both artistry
(clarity, elegance, originality) and function. Moreover,
just as the literary tradition provides new fiction with
meaningful constraints in advance, so the scientific com-
munity provides the parameters within which new mod-
els are developed and applied. Scientific models are, in
this sense, accountable to the scientific communities that
use them in the exploration of target systems that are
already of particular value and interest [33]. Finally,
different sorts of models can be used together to build
a multi-scalar narrative architecture, modeling comple-
mentary features of a target system beside one another.
Whether used in isolation or in conjunction, scientific
models illuminate the overarching structure of a target
system precisely through the practice and provocation of
creative imagination.
NETWORKS AS MODELS
The incipient challenge in modeling biological systems
is to identify the most meaningful characteristics of the
system that are distilled into a sensible representation
[34]. That is, biological models are inherently idealized
models of complex systems, and their construction re-
quires identifying first the form and degree of abstraction
to use. This process requires a set of value judgments
("What are the most meaningful characteristics?") and
a commitment to epistemic cleanliness ("What details
of biology can we defensibly ignore?"). These principles
of valuation and purposeful ignorance are manifest even
when exercising simple visual depictions, which arguably
comprise the most impoverished of modeling approaches
[35]. One similarly faces choices of what to depict and
what not to depict when building any simple mathemat-
ical representation of the system. For example, when
building a differential equation to represent a system,
one must choose which processes to encapsulate and not
encapsulate in a variable.
The fundamental assumption of network neuroscience
is that idealized models of the brain should be con-
structed using analogical principles that focus on the net-
worked architecture of the nervous system. As Cajal saw
under his microscope, the nervous system is composed
of individual neurons that are interconnected in complex
ways. Accordingly, the earliest network models were ide-
alized versions of this network structure, with nodes rep-
resenting neurons and edges representing the connections
between them. More recently, network models have been
developed within and across multiple spatial and tem-
poral scales, at not just the level of interconnected neu-
rons but also networks of subcellular components, multi-
cellular systems, or both. As detailed below, these mod-
els also can be phenomenological, based on measured el-
ements of the nervous system, or more theory driven.
Despite this diversity, these models retain key features
that can be understood in terms of their basic idealized
and analogical structure: an architecture built using in-
terconnected units.
Such an architecture is typically encoded in a graph:
an object composed of nodes, which represent units of the
system, and edges, which represent interactions or links
between those units [36, 37]. Studies of graphs can be
neatly separated into two categories: those that consider
artificial graphs with arbitrary wiring principles [38], and
those that consider graphs that reflect the architecture of
a real system [39]. In both cases, one seeks to describe
the mathematical properties of the graph with the goal of
understanding the function of the system. The patterns
of which units can and cannot (or do and do not) interact
with one another can allow one to deduce where informa-
tion might be relatively more densely or relatively more
sparsely located, where vulnerability might exist to in-
jury or perturbation, and where circumscribed instances
of collective dynamics might emerge [40 -- 43].
In simple graphs, all units are represented by identi-
cal nodes, and all edges are represented as either existing
or not existing (Fig. 1). These representations can be
encoded using a binary weighting scheme. Furthermore,
interactions are assumed to be bidirectional: if an edge
exists between node i and node j, then an edge also ex-
ists between node j and node i. The very first formal
network models of neural systems employed such binary,
undirected graphs [44 -- 49]. Nonetheless, it is relatively
straightforward to adapt this encoding to a continuous
weighting scheme, as well as to specify distinct weights
for the edge from node i to node j, and for the edge
from node j to node i. With the continued refinement
of empirical measurement techniques, the inclusion of
edge weights in network models has become increasingly
prevalent, providing richer insights into system function
and dynamics [3, 50 -- 53].
Network models are simple constructs. They can be
used effectively to study social, biological, technologi-
cal, and physical systems [6]. Yet this flexibility is a
marked reminder that the intuitions that one gains from
a network model depend strongly on what the nodes and
edges are chosen to represent. A structural motif in a
network of humans interlinked by friendships can mean
something quite different than the same structural motif
in a network of neurons interlinked by synapses. Thus,
in any endeavor that translates a complex system into
a network model, it is critical to specify exactly what
the nodes and edges (or more complicated model compo-
nents) represent, and to ensure that interpretations are
drawn in accordance with those choices [54].
3
TYPES OF NETWORK MODELS IN
NEUROSCIENCE
Network neuroscience as a field is devoted to build-
ing, exercising, and validating network models of neural
systems with the explicit goal of better understanding
brain structure and function, as well as cognition, be-
havior, and disease [55 -- 59]. The types of network models
that are built share a similar analogical basis that em-
phasizes the importance of network-based architectures
across spatial and temporal scales. However, these mod-
els differ from one another in many important ways that
directly impact the sorts of inferences that can be justi-
fiably drawn from them. Here we briefly describe recent
efforts to systematize the study of network models in neu-
roscience by organizing these similarities and differences
according to three dimensions that reflect the model cat-
egories described above [12]: first, their phenomenologi-
cal basis, ranging from representations of measured phe-
nomena to first-principles theory; second, their target
of idealization, from biophysical to functional features;
and third, their scalar focus, ranging from elementary
descriptions to coarse-grained approximations (Fig. 2).
The dimension from data representation to first-
principles theory is arguably the most fundamental to
network modeling efforts in neuroscience [60]. Modeling
efforts of the former type begin with empirically acquired
data. They then seek to build a representation of those
data by stipulating which part of the data to represent as
a network node, and which part of the data to represent
as a network edge. Intuitively, the data representation
provides an abstract, non-visual depiction or description
of the system (for a few early examples, see [45 -- 49, 61 --
66]).
In contrast, to make a prediction about system
behavior either now or in the future, one must turn to
models that instantiate first-principles theories. These
models combine a network with a mathematical expres-
sion specifying the dynamics of network nodes, network
edges, or collections of nodes and/or edges (see, for ex-
ample, [67 -- 75]). Data-driven network models enjoy the
benefits of biological realism, whereas theory-based mod-
els have the capacity to make predictions and unearth
function.
The dimension from biophysically to functionally de-
fined features differentiates models that are physical in
nature from those that are statistical in nature. Network
models with biophysical realism are composed of nodes
that represent physical units, including neurons, cortical
columns, or Brodmann areas; and of edges that represent
physical links, including synapses, projections, or white-
matter tracts (see, for example [48, 49, 53, 76 -- 78]). Net-
work models addressing functional phenomenology are
comprised of nodes and edges that are not necessarily
physically instantiated but may instead be defined as sta-
tistical abstractions [47, 65, 79 -- 82]. Common examples
of such abstract edges are those that offer estimates of ef-
fective connectivity or functional connectivity, the latter
of which are also referred to as noise correlations [83 -- 85].
4
FIG. 1. Schematic of network models. (Top Left Corner) The simplest network model for neural systems is one that
represents the pattern of connections (edges) between neural units (nodes). More sophisticated network models can be con-
structed by adding edge weights and node values, or explicit functional forms for their dynamics. Multilayer networks can
be used to represent a set of interconnected networks, and dynamic networks can be used to understand the reconfiguration
of network systems over time. (Bottom Right Corner) Common measures of interest include: degree, which is the number of
edges emanating from a node; clustering, which is related to the prevalence of triangles; cavities, which describe the absence
of edges; hubness, which is related to a node's influence; paths, which determine the potential for information transmission;
communities, or local groups of densely interconnected nodes; shortcuts, which are one possible marker of global efficiency of
information transmission; and core-periphery structure, which facilitates local integration of information gathered from or sent
to more sparsely connected areas. Figure adapted with permission from [12].
It is often important to distinguish between these two
types of models because they have distinct utility in as-
sessing a network's physical constitution versus inferring
its functional capacities.
The dimension from elementary descriptions to coarse-
grained approximations is critical to support a multiscale
understanding of brain structure, function, and dynam-
ics. In general, network models can encode the organiza-
tion of interconnections among cells, ensembles, cortical
columns or subcortical nuclei, and large-scale brain ar-
eas. As evidenced by the diversity of scales represented in
current empirical and theoretical investigations, no single
level of description can provide a complete explanation
for cognitive function and behavior. Yet in many cases,
it is worthwhile or at least practical to consider a single
scale for a given study and to use insights gained at that
scale to inform larger theories of multiscale function. The
challenge in developing an appropriate network model at
a particular scale is to ensure that network nodes rep-
resent well-defined, discrete, non-overlapping units, and
that network edges represent organic, irreducible rela-
tions [54]. Whereas models at the final spatial scale con-
sider elementary building blocks [4, 86 -- 92], models at the
coarse spatial scale consider emergent functions [67].
Together, these complementary dimensions define a
three-dimensional space of network models that can be
used to enhance our understanding of brain structure
and function (Fig. 3). Notably, prior modeling efforts
have not been pursued with equal vigor in all volumes
of this space, partly reflecting historical factors and the
changing state of neurotechnologies. Early work focused
on large-scale network models of anatomy, which repre-
sent coarse-grained data representations with biophys-
ical realism [50, 93, 94]. Less well-studied are first-
principles theories of functional phenomenology, partic-
ularly among elementary units. With a marked increase
DynamictimeDegreeClusteringHubnessCavitiesCommunitiesPathsShortcutsCore-peripheryEdge WeightsNode Valuesl1l2l3MultilayerNodeEdge 5
FIG. 2. Three dimensions of network model types. (a) We posit that efforts to understand mechanisms of brain structure,
function, development, and evolution in network neuroscience can be organized along three key dimensions of model types.
(b) The first dimension extends from elementary descriptions to coarse-grained approximations. (c) The second dimension
extends from biophysical realism to functional phenomenology. (d) The third dimension extends from data representation to
first-principles theory. Figure adapted with permission from [12].
in the pace of data acquisition and the capacities for data
analysis, we anticipate increasing success developing net-
work models at the center of the volume. In other words,
we anticipate a wider array of studies that inform first-
principles theories with data, complement physical mod-
els with statistical inferences on informational capacity,
and build explicit multiscale accounts of network func-
tion. These multi-faceted network models will enhance
our ability to explain different parts, processes, or prin-
ciples of the nervous system.
MODELING PERTURBATIONS TO NETWORKS
When building network models, we are often concerned
with demonstrating their validity [12]. In prior work, we
followed canonical principles for validating animal mod-
els of disease to suggest that network models can display
three distinct types of validity [95 -- 98]: descriptive, ex-
planatory, and predictive. Intuitively, descriptive validity
requires that the model resemble the system under study,
a concept akin to face validity in animal models [99]. For
example, a model with descriptive validity might accu-
SimilarityConnectomesDynamical modelsTime seriesabelementarycoarse structuralfunctionaltheorydataRegionsPopulationsNeuronsChromatin001001101001100ObservationsRepresentationsPredictionsBiological modelings1s2cdf(n,e) 6
or edge removal. At times referred to as virtual lesioning,
this approach was developed to quantify the robustness
of a network by estimating the difference between the
value of a graph statistic estimated before node or edge
removal, and the value of that same graph statistic esti-
mated after node or edge removal (Fig. 1b) [102]. When
nodes are removed at random, the approach is referred
to as a random attack. When nodes are removed based
on their topological role in the network, estimated us-
ing the values of various graph statistics such as degree
and betweenness centrality, the approach is referred to
as a targeted attack [47]. These approaches have recently
been used to better understand the impact of regional
dysfunction in schizophrenia and stroke [103 -- 105].
Other approaches address how a perturbation of the
activity of a node or edge can change the activity of
other parts of the network. This approach is central to
network control theory and built on the foundations of
linear systems theory [106, 107]. Here one considers the
pattern of interconnections between units as well as a
model of dynamics that specifies how the activity at one
node can travel along edges to other nodes in the graph
[101]. By modeling both the connectivity and the dynam-
ics, one can identify "driver" nodes with time-dependent
control that can guide the system's activity [108]. A re-
cent application of these techniques to the connectome of
C. elegans demonstrated that a particular network model
had striking validity in predicting the effects of single cell
ablasions on the organism [75]. The approach can also be
extended beyond the identification of drivers controlling
all dynamics to the identification of drivers controlling
specific dynamics [109]. The specificity of this extension
allows for the study of unique control strategies within
neural systems. A recent application of this technique
provided an explanation for the anatomical location of
areas of the brain involved in executive function, as those
most capable of enacting modal controllability [73].
Despite their analytical tractability, point perturba-
tions can be the most difficult to enact and interpret in
the context of real neural systems. On a conceptual level,
a single, functional node or edge used in a model might
not have an obvious, well-defined anatomical substrate
in the brain to target. On a practical level, even given
a well-defined target, it may not be possible to cleanly
perturb just that target given the lack of complete speci-
ficity associated with current microstimulation, optoge-
netic, and pharmacological methods. Nonetheless, point
perturbations represent a useful starting point in consid-
ering the validation of network models.
Moving beyond point perturbations, it is also of inter-
est to consider perturbation to multiple points in the net-
work, or to entire areas or volumes of neural systems. In-
tuitively, multi-point control is a natural reflection of cir-
cuit activity, where several areas may be activated simul-
taneously to orchestrate a change in communication or
dynamics [110]. Multi-point control could also be fruit-
fully applied to the development of stimulation therapies
to quiet seizure dynamics using implantable devices [111 --
FIG. 3. Density of study in the three dimensional
space of network models. Density plot showing varying
levels of study across the 3-dimensional space of model types
in network neuroscience. The center of the 3-dimensional
space (aubergine) is becoming increasingly accessible due
to empirical and computational advances over the past few
decades. The least well-studied space (periwinkle) represents
first-principles theories of functional phenomenology at the el-
ementary level of description. Clear volumes indicate spaces
that are most commonly studied.
rately reflect the specific pattern of nodes and edges ob-
served in anatomical or functional data. By contrast,
explanatory validity requires that the model can be used
to define statistical tests, for example by assessing causal
relations based on the network's architecture. Finally,
predictive validity is attained when there is a correlation
between a network model's response to perturbation and
an organism's response to that same perturbation [97].
Such perturbative studies can be operationalized using
stimulation, lesion, ablation, or drugs.
Predictive validity is often the final goal of any scien-
tific domain of inquiry [100]. Because predictive validity
depends on understanding its response to a perturbation,
it is of interest to consider the different ways in which a
network model can be perturbed. Recent work in the
physics and engineering communities has begun to focus
on how the architecture of a network determines how per-
turbations affect its function. A simple way in which to
parse these studies is to consider separately perturbations
applied: (i) to a single node or to a single edge ("point
perturbations"), (ii) to a set of nodes or to a set of edges,
and (iii) across a fixed area or volume of the network's
topology. In the context of network neuroscience, these
different types of perturbations may be accessible to dis-
tinct empirical techniques and collectively could be used
to better understand both endogenous and exogenous
control, thereby informing clinical intervention [101].
From a modeling perspective, point perturbations are
perhaps the simplest type to study. Initial modeling ap-
proaches focused point perturbations in the form of node
Newly explorableFrontiersNetwork Model Spacecoarseelementaryfunctionaldatastructuraltheory 114]. Fortunately, the general network control framework
is readily extended to account for the activity of multiple
control points simultaneously, and can be used to directly
model the propagation of stimulation along white mat-
ter tracts to predict distant effects on regional activity
[115, 116]. Extending these tools to affect control over
continuous areas or volumes of a network is more difficult,
and remains an important area for future work. Progress
in this area is critical for extending network models to
account for other chemical mechanisms of transcellular
communication and the effects of glia, neuromodulatory
systems, and other mechanisms on brain function and
behavior [117 -- 120].
MODELING NETWORK GROWTH AND
EVOLUTION
The study of network perturbations, while useful for
understanding endogenous and exogenous mechanisms
of control, is also pertinent to an understanding of how
neural systems came to be, how they develop, and how
they age. A change in gene expression can alter the nat-
ural progression of cell fate from pluripotent stem cell
through neuroprogenitor cell and eventually neuron [92].
A fluctuation in chemical gradients can comprise a per-
turbation that alters the course of neuronal migration,
and by extension the location and density of synapses
[122]. In fully developed adult neurons, Hebb's rule es-
sentially postulates that perturbations to neuronal firing
can alter cellular-level network architecture [123]. Even
at the large scale in humans, long-term training can in-
duce changes in white matter architecture evident in non-
invasive neuroimaging [124]. Understanding how pertur-
bations of the organism or part of the organism over both
short and long time scales affect network growth and evo-
lution is an important open area of research.
Some progress has been made over the last few years in
constructing so-called generative network models. Such
models stipulate a wiring rule in the hope of producing
a network architecture that displays topological or func-
tional properties that are similar to those observed in the
networks representing real systems [121]. The basic idea
is that a wiring rule that produces a network topology
similar to that observed in the real system is a candi-
date mechanism for network generation. The inference
is made stronger if the wiring rule also displays charac-
teristics thought to be consistent with biology, such as
parsimony and efficiency. A common way of testing the
pragmatic utility of the generative network model is to
determine if it can be used to make out-of-sample pre-
dictions about held-out network data.
Generative network models tend to be built in one of
three types: single-shot models, growth models, and de-
velopmental models (Fig. 4). A single-shot model spec-
ifies a form for the connection probabilities, from which
all edges and their weights are then drawn [125 -- 130]. A
growth model specifies a time-dependent wiring rule that
7
indicates how nodes and possibly even edges are added
over time [131]. Developmental models extend the biolog-
ical realism of the effort even farther by specifying wiring
rules in which the time scales of the model match the time
scales of development in the organism under study [77].
Together, these three types of generative network mod-
els vary in the time scale over which they operate and in
their neurobiological plausibility.
Recently, generative network models have been devel-
oped and applied to explain neurophysiological and neu-
roanatomical data across both elementary descriptions
and coarse-grained approximations [127, 128]. For ex-
ample, a particularly striking single-shot model of the
neural connectome of the nematode C. elegans demon-
strated that a wiring rule based on the random outgrowth
of axons, in combination with a competition for available
space at the target neuron, was able to recapitulate the
empirical network's edge length distribution [132]. A de-
velopmental model of the same organism combined in-
formation regarding the pattern of interconnectivity be-
tween neurons with information regarding the birth times
of neurons and their spatial locations [77]. This study
provided compelling evidence for a trade-off between the
network's topology and cost that appears to be differen-
tially negotiated over different developmental time peri-
ods. With a few exceptions [125], most generative net-
work models have focused on data representations more
so than first-principles theories, and biophysical realism
more so than functional phenomenology. Expanding ef-
forts to fill the full space of model types will be an im-
portant area for future work in generative modeling.
FUTURE DIRECTIONS
In considering the future utility of network models
for advancing our understanding of neural systems, it
is worth pointing out that the models used to date are
relatively simple from a mathematical perspective.
It
remains an open question whether more complex net-
work models might prove useful or merely obfuscate in-
ference. To address this question, one could rationally
assess whether some aspect of a known neurophysiolog-
ical process remains unaccounted for by existing mod-
els. For example, to increase descriptive validity, one
might wish to build an annotated network, where nodes
can be assigned values or properties, reflecting for ex-
ample cerebral glucose metabolism estimates from flu-
orodeoxyglucose (FDG)-positron emission tomography
(PET), blood-oxygen-level dependent (BOLD) contrast
imaging, magnetoencephalographic (MEG) or electroen-
cephalographic (EEG) power, gray matter volume or cor-
tical thickness, or cytoarchitectonic properties [133, 134].
Furthermore, to increase explanatory validity, one might
wish to build multilayer networks where the nodes and
edges in each layer are obtained from different types of
measurements [135 -- 137], and where the architecture of
the network is allowed to vary over time in concert with
8
FIG. 4. Distinct classes of generative network models. Generative network models exist in three main classes that differ
in the timescales over which they operate. (Left) Single-shot models specify a functional form for the probability with which
any two nodes are linked with one another. (Middle) Growth models specify rules by which nodes and/or edges are added to
the network over time, which is commonly discretized in arbitrary units. (Right) Developmental models specify wiring rules
with fixed timescales in actual units of seconds, minutes, days, months, years, etc. in an effort to better match the true growth
mechanisms of an organism [77]. Figure reproduced with permission from [121].
system function [67, 138 -- 141]. Such richer models could
allow one to test how network dynamics at one time point
or in one modality might cause a change in network dy-
namics at another time point or in another modality.
A second way in which to address the question of
whether more complex network models might prove use-
ful
is to consider whether the testing of a particu-
lar hypothesis requires a novel network model. For
example, recent efforts have provided initial evidence
that some higher-order, non-pairwise interactions oc-
cur between neurons and between large-scale brain ar-
eas [142, 143]. Critically, all of the network models we
have discussed here are based on pairwise interactions
and cannot directly account for non-pairwise interactions
[144, 145]. Tools that have been developed in the ap-
plied mathematics community that can encode and char-
acterize higher-order relations include hypergraphs (an
edge can link any number of vertices) and simplicial com-
plexes (higher-order interaction terms become fundamen-
tal units) [146, 147]. These generalizations of graphs
may be critical for an accurate understanding of neu-
ronal codes and associated computations both at micro-
and macro-scales [148 -- 150]. As the field moves beyond
univariate accounts to postulate more network-based hy-
potheses, richer network models may be required.
CONCLUSION
From cellular to regional scales, neural circuitry is an
interconnected system. In such a system, network mod-
eling is a particularly useful approach for distilling inter-
connection patterns into tractable mathematical objects
that are amenable to theory. Here we discussed the na-
ture of network models, which share a similar networked
architecture that is justified in terms of its analogies to
brain structure but then differ along several dimensions:
from data representations to first-principles theory, ab-
stractions that emphasize biophysical or functional fea-
tures, and different scales from elementary descriptions
to coarse-grained approximations. We paid particular
attention to models that have been developed to better
understand the response of networked systems to per-
turbation enacted at a single-point, at multiple points,
or across extended areas or volumes of the organism.
We also offered an extended discussion of generative net-
work models that seek to identify candidate wiring mech-
anisms for circuit evolution or development. We suggest
that network models are particularly appropriate for neu-
ral systems. Accordingly, future advances in our under-
standing of computation and cognition will depend upon
the expansion of these models in mathematical sophis-
tication, and the development of richer, network-based
hypothesis of brain structure, function, and dynamics.
Model timescalet = 1t = 2t = 3t = 4Growth modelsijijincreasing probabilityof connection forming"Single shot"Developmental modelsNeurobiologically plausible ACKNOWLEDGEMENTS
We thank Blevmore Labs and Ann E. Sizemore for ef-
forts in graphic design. We also thank D. Lydon-Staley,
A. E. Sizemore, E. Cornblath and D. Zhou for help-
ful comments on an earlier version of this manuscript.
D.S.B. acknowledges support from the John D. and
Catherine T. MacArthur Foundation, the Alfred P. Sloan
Foundation, the ISI Foundation, the Paul Allen Foun-
dation, the Army Research Laboratory (W911NF-10-2-
0022), the Army Research Office (Bassett-W911NF-14-
1-0679, Grafton-W911NF-16-1-0474, DCIST-W911NF-
17-2-0181), the Office of Naval Research, the National
Institute of Mental Health (2-R01-DC-009209-11, R01-
MH112847, R01-MH107235, R21-M MH-106799), the
National Institute of Child Health and Human Devel-
[1] Francis, M. M., Mellem, J. E. & Maricq, A. V. Bridging
the gap between genes and behavior: recent advances
in the electrophysiological analysis of neural function
in Caenorhabditis elegans. Trends Neurosci 26, 90 -- 99
(2003).
[2] Sherman, M. A. et al. Neural mechanisms of transient
neocortical beta rhythms: Converging evidence from
humans, computational modeling, monkeys, and mice.
Proc Natl Acad Sci U S A 113, E4885 -- E4894 (2016).
[3] Betzel, R. F. & Bassett, D. S. The specificity and ro-
bustness of long-distance connections in weighted, in-
terareal connectomes. Proc Natl Acad Sci U S A Epub
Ahead of print (2018).
[4] Betzel, R. F., Medaglia, J. D. & Bassett, D. S. Di-
versity of meso-scale architecture in human and non-
human connectomes. Nature Communications In Press
(2018).
[5] Albert, E. & Barabasi, A.-L.
Statistical mechanics
of complex networks. Reviews of Modern Physics 74
(2002).
9
opment (1R01HD086888-01), National Institute of Neu-
rological Disorders and Stroke (R01 NS099348), and
the National Science Foundation (BCS-1441502, BCS-
1430087, NSF PHY-1554488 and BCS-1631550). J.I.G.
acknowledges support from the National Science Founda-
tion (nsf-ncs 1533623), the National Eye Institute (R01-
EY015260), and the National Institute of Mental Health
(R01-MH115557). The content is solely the responsibil-
ity of the authors and does not necessarily represent the
truth.
COMPETING INTERESTS
The authors declare no competing interests.
Front Neurol 8, 626 (2017).
[15] Chan, Y. K. et al. In vitro modeling of emulsification of
silicone oil as intraocular tamponade using microengi-
neered eye-on-a-chip.
Invest Ophthalmol Vis Sci 56,
3314 -- 3319 (2015).
[16] Gelfert, A. How to Do Science with Models: A Philo-
sophical Primer (2016).
[17] Papineau, D. Reality and Representation (Blackwell,
1987).
[18] Hartmann, S. & Frigg, R. Scientific models. In et al.,
S. S. (ed.) The Philosophy of Science: An Encyclopedia,
vol. 2, 740 -- 749 (Routledge, 2012).
[19] Frigg, R. & Hartmann, S.
in science
(2012). URL https://plato.stanford.edu/entries/
models-science.
Models
[20] Toon, A. Models as Make-Believe: Imagination, Fic-
tion, and Scientific Representation (Palgrave, 2012).
[21] Frigg, R. & Hunter, M. C. Beyond Mimesis and Con-
vention: Representation in Art and Science (Springer,
2010).
[6] Newman, M. E. J. Networks: An Introduction (MIT
[22] Suarez, M. Fictions in Science: Philosophical Essays
Press, 2010).
[7] Newman, M. E. J. Complex systems: A survey. Am. J.
Phys. 79, 800 -- 810 (2011).
[8] Bullmore, E. & Sporns, O. Complex brain networks:
graph theoretical analysis of structural and functional
systems. Nat Rev Neurosci 10, 186 -- 198 (2009).
[9] Fornito, A., Zalesky, A. & Bullmore, E. T. Brain Net-
on Modelling and Idealization (Routledge, 2009).
[23] Barbrousse, A. & Ludwig, P. Fictions and models. In
Suarez, M. (ed.) Fictions in Science: Philosophical Es-
says on Modelling and Idealization, 56 -- 75 (Routledge,
2009).
[24] Frigg, R. Models and fiction. Synthese 172, 251 -- 268
(2010).
work Analysis (Oxford University Press, 2016).
[25] Frigg, R. & Nguyen, J. The fiction view of models
[10] van den Heuvel, M. P., Bullmore, E. T. & Sporns, O.
Comparative connectomics. Trends Cogn Sci 20, 345 --
361 (2016).
reloaded. The Monist 99, 225 -- 242 (2016).
[26] Godfrey-Smith, P. Models and fictions in science. Philo-
sophical Studies 143, 101 -- 116 (2009).
[11] Bassett, D. S. & Sporns, O. Network neuroscience. Nat
[27] Toon, A. Models as Make-Believe: Imagination, Fic-
Neurosci 20, 353 -- 364 (2017).
[12] Bassett, D. S., Zurn, P. & Gold, J. I. On the nature and
use of models in network neuroscience. Nature Reviews
Neuroscience In Press (2018).
[13] Eugenides, J. Middlesex (Farrar, Straus and Giroux,
2002).
[14] Huang, J. et al. Tranylcypromine causes neurotox-
icity and represses BHC110/LSD1 in human-induced
pluripotent stem cell-derived cerebral organoids model.
tion, and Scientific Representation (Palgrave, 2012).
[28] Frigg, R. Fiction and scientific representation. In Frigg,
R. & Hunter, M. C. (eds.) Beyond Mimesis and Conven-
tion: Representation in Art and Science, 97 -- 138 (2010).
[29] Frigg, R. Fiction in science. In Woods, J. (ed.) Fictions
and Models: New Essays, 247 -- 287 (Philosophia Verlag,
2010).
[30] Garcia-Carpintero, M. Fictional entities, theoretical
In Frigg, R. & Hunter,
models, and figurative truth.
[31] Frigg, R. & Nguyen, J.
M. C. (eds.) Beyond Mimesis and Convention: Repre-
sentation in Art and Science, 139 -- 168 (Springer, 2010).
Scientific representation is
representation-as. In Chao, H. K. & Reiss, J. (eds.) Phi-
losophy of Science in Practice, 149 -- 179 (Spring, 2017).
[32] Elgin, C. Z. Telling instances. In Frigg, R. & Hunter,
M. C. (eds.) Beyond Mimesis and Convention: Repre-
sentation in Art and Science, 1 -- 17 (Springer, 2010).
[33] Almeder, R. Pragmatism and philosophy of science: A
critical survey. International Studies in the Philosophy
of Science 12, 171 -- 195 (2007).
[34] Bellomo, N., Elaiw, A., Althiabi, A. M. & Alghamdi,
M. A. On the interplay between mathematics and bi-
ology: hallmarks toward a new systems biology. Phys
Life Rev 12, 44 -- 64 (2015).
[35] Tufte, E. R. The visual display of quantitative informa-
tion (2001).
[36] Bollobas, B. Graph Theory: An Introductory Course
(Springer-Verlag, 1979).
[37] Bollobas, B. Random Graphs (Academic Press, 1985).
[38] Harary, F. Graph Theory (Addison-Wesley, 1969).
[39] Cohen, R. & Havlin, S. Complex Networks: Structure,
Robustness and Function (Cambridge University Press,
2010).
[40] Gomez-Gardenes, J., Moreno, Y. & Arenas, A. Paths
to synchronization on complex networks. Phys Rev Lett
98, 034101 (2007).
[41] Cisneros, L., Jimenez, J., Cosenza, M. G. & Parravano,
A. Information transfer and nontrivial collective behav-
ior in chaotic coupled map networks. Phys Rev E Stat
Nonlin Soft Matter Phys 65, 045204 (2002).
[42] Simonsen, I., Buzna, L., Peters, K., Bornholdt, S. &
Helbing, D. Transient dynamics increasing network vul-
nerability to cascading failures. Phys Rev Lett 100,
218701 (2008).
[43] Albert, R., Jeong, H. & Barabasi, A. L. Error and attack
tolerance of complex networks. Nature 406, 378 -- 382
(2000).
[44] Felleman, D. J. & Van Essen, D. C. Distributed hierar-
chical processing in the primate cerebral cortex. Cereb
Cortex 1, 1 -- 47 (1991).
[45] Young, M. P., Scannell, J. W., Burns, G. A. & Blake-
more, C. Analysis of connectivity: neural systems in
the cerebral cortex. Rev Neurosci 5, 227 -- 250 (1994).
[46] Scannell, J. W., Blakemore, C. & Young, M. P. Analysis
of connectivity in the cat cerebral cortex. J Neurosci 15,
1463 -- 1483 (1995).
[47] Achard, S., Salvador, R., Whitcher, B., Suckling, J. &
Bullmore, E. A resilient, low-frequency, small-world hu-
man brain functional network with highly connected as-
sociation cortical hubs. J Neurosci 26, 63 -- 72 (2006).
[48] Sporns, O., Tononi, G. & Kotter, R. The human con-
nectome: a structural description of the human brain.
PLoS Comput Biol 1, e42 (2005).
[49] Kaiser, M. & Hilgetag, C. C. Nonoptimal component
placement, but short processing paths, due to long-
distance projections in neural systems. PloS Compu-
tational Biology 2, e95 (2006).
[50] Bassett, D. S. & Bullmore, E. T.
brain networks revisited.
1073858416667720 (2016).
Small-world
Neuroscientist Sep 21,
[51] Rubinov, M. & Sporns, O. Weight-conserving charac-
terization of complex functional brain networks. Neu-
roimage 56, 2068 -- 2079 (2011).
10
[52] Markov, N. T. et al. The role of long-range connections
on the specificity of the macaque interareal cortical net-
work. Proceedings of the National Academy of Sciences
110, 5187 -- 5192 (2013).
[53] Oh, S. W. et al. A mesoscale connectome of the mouse
brain. Nature 508, 207 -- 214 (2014).
[54] Butts, C. T. Revisiting the foundations of network anal-
ysis. Science 325, 414 -- 416 (2009).
[55] Medaglia, J. D., Lynall, M. E. & Bassett, D. S. Cogni-
tive network neuroscience. J Cogn Neurosci 27, 1471 --
1491 (2015).
[56] Sporns, O. Contributions and challenges for network
models in cognitive neuroscience. Nat Neurosci 17, 652 --
660 (2014).
[57] Stam, C. J. Modern network science of neurological
disorders. Nat Rev Neurosci 15, 683 -- 695 (2014).
[58] Braun, U. et al. From maps to multi-dimensional net-
work mechanisms of mental disorders. Neuron 97, 14 -- 31
(2018).
[59] Fornito, A., Bullmore, E. T. & Zalesky, A. Opportu-
nities and challenges for psychiatry in the connectomic
era. Biol Psychiatry Cogn Neurosci Neuroimaging 2,
9 -- 19 (2017).
[60] Abbott, L. F. Theoretical neuroscience rising. Neuron
60, 489 -- 495 (2008).
[61] Hilgetag, C. C., O'Neill, M. A. & Young, M. P. Hierar-
chical organization of macaque and cat cortical sensory
systems explored with a novel network processor. Philos
Trans R Soc Lond B Biol Sci 355, 71 -- 89 (2000).
[62] Stam, C. J. Functional connectivity patterns of human
magnetoencephalographic recordings: a 'small-world'
network? Neurosci Lett 355, 25 -- 28 (2004).
[63] De Vico Fallani, F. et al. Brain connectivity structure in
spinal cord injured: evaluation by graph analysis. Conf
Proc IEEE Eng Med Biol Soc 1, 988 -- 991 (2006).
[64] Micheloyannis, S. et al.
Small-world networks and
disturbed functional connectivity in schizophrenia.
Schizophr Res 87, 60 -- 66 (2006).
[65] Bettencourt, L. M., Stephens, G. J., Ham, M. I. &
Gross, G. W. Functional structure of cortical neuronal
networks grown in vitro. Phys Rev E 75, 021915 (2007).
[66] Watts, D. J. & Strogatz, S. H. Collective dynamics of
'small-world' networks. Nature 393, 440 -- 442 (1998).
[67] Breakspear, M. Dynamic models of large-scale brain
activity. Nat Neurosci 20, 340 -- 352 (2017).
[68] Melozzi, F., Woodman, M. M., Jirsa, V. K. & Bernard,
C. The virtual mouse brain: A computational neuroin-
formatics platform to study whole mouse brain dynam-
ics. eNeuro 4, ENEURO.0111 -- 17.2017 (2017).
[69] Bezgin, G., Solodkin, A., Bakker, R., Ritter, P. &
McIntosh, A. R. Mapping complementary features of
cross-species structural connectivity to construct realis-
tic "Virtual Brains". Hum Brain Mapp 38, 2080 -- 2093
(2017).
[70] Ritter, P., Schirner, M., McIntosh, A. R. & Jirsa, V. K.
The virtual brain integrates computational modeling
and multimodal neuroimaging. Brain Connect 3, 121 --
145 (2013).
[71] Roy, D. et al. Using the virtual brain to reveal the role of
oscillations and plasticity in shaping brain's dynamical
landscape. Brain Connect 4, 791 -- 811 (2014).
[72] Falcon, M. I., Jirsa, V. & Solodkin, A. A new neuroin-
formatics approach to personalized medicine in neurol-
ogy: The Virtual Brain. Curr Opin Neurol 29, 429 -- 436
11
(2016).
055533 (2017).
[73] Gu, S. et al. Controllability of structural brain networks.
[93] Bassett, D. S. & Bullmore, E. Small-world brain net-
Nat Commun 6, 8414 (2015).
[74] Kim, J. Z. et al. Role of graph architecture in con-
trolling dynamical networks with applications to neu-
ral systems. Nature Physics Epub Ahead of Print
(2018).
[75] Yan, G. et al. Network control principles predict neu-
ron function in the Caenorhabditis elegans connectome.
Nature 550, 519 -- 523 (2017).
[76] Varshney, L. R., Chen, B. L., Paniagua, E., Hall,
D. H. & Chklovskii, D. B. Structural properties of the
Caenorhabditis elegans neuronal network. PLoS Com-
put Biol 7, e1001066 (2011).
[77] Nicosia, V., V´ertes, P. E., Schafer, W. R., Latora, V.
& Bullmore, E. T. Phase transition in the economi-
cally modeled growth of a cellular nervous system. Pro-
ceedings of the National Academy of Sciences USA 110,
7880 -- 7885 (2013).
[78] Bassett, D. S. et al. Efficient physical embedding of
topologically complex information processing networks
in brains and computer circuits. PLoS Comput Biol 6,
e1000748 (2010).
[79] van Diessen, E. et al. Opportunities and methodological
challenges in EEG and MEG resting state functional
brain network research. Clin Neurophysiol 126, 1468 --
1481 (2015).
[80] Chu, C. J. et al. Emergence of stable functional net-
works in long-term human electroencephalography. J
Neurosci 32, 2703 -- 2713 (2012).
[81] Burns, S. P. et al. Network dynamics of the brain and
influence of the epileptic seizure onset zone. Proc Natl
Acad Sci U S A 111, E5321 -- E5330 (2014).
[82] Khambhati, A. N., Davis, K. A., Lucas, T. H., Litt,
B. & Bassett, D. S. Virtual cortical resection reveals
push-pull network control preceding seizure evolution.
Neuron 91, 1170 -- 1182 (2016).
[83] Friston, K. J. Functional and effective connectivity: a
review. Brain Connect 1, 13 -- 36 (2011).
works. Neuroscientist 12, 512 -- 523 (2006).
[94] Liao, X., Vasilakos, A. V. & He, Y. Small-world human
brain networks: Perspectives and challenges. Neurosci
Biobehav Rev 77, 286 -- 300 (2017).
[95] Shmueli, G. To explain or to predict? Statistical Science
25, 289 -- 310 (2010).
[96] Willner, P. The validity of animal models of depression.
Psychopharmacology 83, 1 -- 16 (1984).
[97] Belzung, C. & Lemoine, M. Criteria of validity for an-
imal models of psychiatric disorders: Focus on anxiety
disorders and depression. Biology of Mood and Anxiety
Disorders 1, 9 (2011).
[98] McKinney, W. T. J. & Bunney, W. E. J. Animal model
implications for
of depression. I. Review of evidence:
research. Arch Gen Psychiatry 21, 240 -- 248 (1969).
[99] Willner, P. Reliability of the chronic mild stress model
of depression: A user survey. Neurobiol Stress 6, 68 -- 77
(2017).
[100] Shneider, A. M. Four stages of a scientific discipline;
four types of scientist. Trends Biochem Sci 34, 217 -- 223
(2009).
[101] Tang, E. & Bassett, D. S. Control of dynamics in brain
networks. Reviews of Modern Physics In Press (2018).
[102] Dong, G. et al. Robustness of network of networks under
targeted attack. Phys Rev E Stat Nonlin Soft Matter
Phys 87, 052804 (2013).
[103] Lynall, M. E. et al. Functional connectivity and brain
networks in schizophrenia. J Neurosci 30, 9477 -- 9487
(2010).
[104] Lo, C. Y. et al. Randomization and resilience of brain
functional networks as systems-level endophenotypes of
schizophrenia. Proc Natl Acad Sci U S A 112, 9123 --
9128 (2015).
[105] Alstott, J., Breakspear, M., Hagmann, P., Cammoun,
L. & Sporns, O. Modeling the impact of lesions in the
human brain. PLoS Comput Biol 5, e1000408 (2009).
[106] Motter, A. E. Networkcontrology. Chaos 25, 097621
[84] Brody, C. D. Correlations without synchrony. Neural
(2015).
Comput 11, 1537 -- 1551 (1999).
[85] Brody, C. D. Disambiguating different covariation
types. Neural Comput 11, 1527 -- 1535 (1999).
[86] Feldt, S., Bonifazi, P. & Cossart, R. Dissecting func-
tional connectivity of neuronal microcircuits: experi-
mental and theoretical insights. Trends Neurosci 34,
225 -- 236 (2011).
[87] Kim, S. Y. & Lim, W. Fast sparsely synchronized brain
rhythms in a scale-free neural network. Phys Rev E 92,
022717 (2015).
[88] Sautois, B., Soffe, S. R., Li, W. C. & Roberts, A. Role
of type-specific neuron properties in a spinal cord motor
network. J Comput Neurosci 23, 59 -- 77 (2007).
[89] Teller, S. et al. Emergence of assortative mixing between
clusters of cultured neurons. PLoS Comput Biol 10,
e1003796 (2014).
[90] Tang, A. et al. A maximum entropy model applied to
spatial and temporal correlations from cortical networks
in vitro. J Neurosci 28, 505 -- 518 (2008).
[91] Kaiser, M. Mechanisms of connectome development.
Trends Cogn Sci 21, 703 -- 717 (2017).
[92] Mahadevan, A. R., Grandel, N. E., Robinson, J. T. &
Qutub, A. A. Living neural networks: Dynamic network
analysis of developing neural progenitor cells. bioRxiv
[107] Kailath, T. Linear Systems (Prentice-Hall, Inc., 1980).
[108] Liu, Y.-Y., Slotine, J.-J. & Barab´asi, A.-L. Controllabil-
ity of complex networks. Nature 473, 167 -- 173 (2011).
[109] Pasqualetti, F., Zampieri, S. & Bullo, F. Controlla-
bility metrics, limitations and algorithms for complex
networks.
IEEE Transactions on Control of Network
Systems 1, 40 -- 52 (2014).
[110] Palmigiano, A., Geisel, T., Wolf, F. & Battaglia, D.
Flexible information routing by transient synchrony.
Nat Neurosci 20, 1014 -- 1022 (2017).
[111] De Ridder, D., Perera, S. & Vanneste, S. State of the
art: Novel applications for cortical stimulation. Neuro-
modulation 20, 206 -- 214 (2017).
[112] Ehrens, D., Sritharan, D. & Sarma, S. V. Closed-loop
control of a fragile network: application to seizure-like
dynamics of an epilepsy model. Front Neurosci 9, 58
(2015).
[113] Taylor, P. N. et al. Optimal control based seizure abate-
ment using patient derived connectivity. Front Neurosci
9, 202 (2015).
[114] Jobst, B. C. et al. Brain-responsive neurostimulation
in patients with medically intractable seizures arising
from eloquent and other neocortical areas. Epilepsia
58, 1005 -- 1014 (2017).
[115] Muldoon, S. F. et al. Stimulation-based control of dy-
namic brain networks. PLoS Comput Biol 12, e1005076
(2016).
[116] Stiso, J. et al. White matter network architecture guides
direct electrical stimulation through optimal state tran-
sitions. arXiv 1805, 01260 (2018).
[117] Borroto-Escuela, D. O. et al. The role of transmitter
diffusion and flow versus extracellular vesicles in volume
transmission in the brain neural-glial networks. Philos
Trans R Soc Lond B Biol Sci 370 (2015).
[118] Savtchouk, I. & Volterra, A. Gliotransmission: Beyond
black-and-white. J Neurosci 38, 14 -- 25 (2018).
[119] Safaai, H., Neves, R., Eschenko, O., Logothetis, N. K.
& Panzeri, S. Modeling the effect of locus coeruleus
firing on cortical state dynamics and single-trial sensory
processing. Proc Natl Acad Sci U S A 112, 12834 -- 12839
(2015).
[120] Bruinsma, T. J. et al.
The relationship between
dopamine neurotransmitter dynamics and the blood-
oxygen-level-dependent (BOLD) signal: A review of
pharmacological functional magnetic resonance imag-
ing. Front Neurosci 12, 238 (2018).
[121] Betzel, R. F. & Bassett, D. S. Generative models for
network neuroscience: prospects and promise. J R Soc
Interface 14, 20170623 (2017).
[122] Wrobel, M. R. & Sundararaghavan, H. G. Directed mi-
gration in neural tissue engineering. Tissue Eng Part B
Rev 20, 93 -- 105 (2014).
[123] Bi, G. & Poo, M. Synaptic modification by correlated
activity: Hebb's postulate revisited 24, 139 -- 166 (2001).
[124] Scholz, J., Klein, M. C., Behrens, T. E. & Johansen-
Berg, H. Training induces changes in white-matter ar-
chitecture. Nat Neurosci 12, 1370 -- 1371 (2009).
[125] Vertes, P. E. et al. Simple models of human brain func-
tional networks. Proc Natl Acad Sci U S A 109, 5868 --
5873 (2012).
[126] Vertes, P. E., Alexander-Bloch, A. & Bullmore, E. T.
Generative models of rich clubs in Hebbian neuronal
networks and large-scale human brain networks. Philos
Trans R Soc Lond B Biol Sci 369, 1653 (2014).
[127] Betzel, R. F. et al. Generative models of the human
connectome. Neuroimage 124, 1054 -- 1064 (2016).
[128] Beul, S. F., Grant, S. & Hilgetag, C. C. A predictive
model of the cat cortical connectome based on cytoar-
chitecture and distance. Brain Struct Funct 220, 3167 --
3184 (2015).
[129] Beul, S. F., Barbas, H. & Hilgetag, C. C. A predictive
structural model of the primate connectome. Sci Rep 7,
43176 (2017).
[130] Hilgetag, C. C., Medalla, M., Beul, S. F. & Barbas,
H. The primate connectome in context: Principles of
connections of the cortical visual system. Neuroimage
134, 685 -- 702 (2016).
[131] Klimm, F., Bassett, D. S., Carlson, J. M. & Mucha, P. J.
Resolving structural variability in network models and
the brain. PLoS Computational Biology 10, e1003491
(2014).
[132] Kaiser, M., Hilgetag, C. C. & van Ooyen, A. A simple
rule for axon outgrowth and synaptic competition gen-
12
erates realistic connection lengths and filling fractions.
Cereb Cortex 19, 3001 -- 3010 (2009).
[133] Newman, M. E. J. & Clauset, A. Structure and infer-
ence in annotated networks. Nature Communications 7,
11863 (2016).
[134] Murphy, A. C. et al. Explicitly linking regional activa-
tion and function connectivity: Community structure of
weighted networks with continuous annotation. arXiv
1611, 07962 (2016).
[135] Kivel, M. et al. Multilayer networks. J. Complex Netw.
2, 203 -- 271 (2014).
[136] Tewarie, P. et al. Structural degree predicts functional
network connectivity: a multimodal resting-state fMRI
and MEG study. Neuroimage 97, 296 -- 307 (2014).
[137] Yu, Q. et al. Building an EEG-fMRI multi-modal brain
graph: A concurrent EEG-fMRI study. Front Hum Neu-
rosci 10, 476 (2016).
[138] Holme, P. & Saramaki, J. Temporal networks. Phys.
Rep. 519, 97 -- 125 (2012).
[139] Khambhati, A. N., Sizemore, A. E., Betzel, R. F. & Bas-
sett, D. S. Modeling and interpreting mesoscale network
dynamics. Neuroimage S1053-8119, 30500 -- 1 (2017).
[140] Sizemore, A. E. & Bassett, D. S. Dynamic graph met-
rics: Tutorial, toolbox, and tale. Neuroimage S1053-
8119, 30564 -- 5 (2017).
[141] Kopell, N. J., Gritton, H. J., Whittington, M. A. &
Kramer, M. A. Beyond the connectome: the dynome.
Neuron 83, 1319 -- 1328 (2014).
[142] Ganmor, E., Segev, R. & Schneidman, E. Sparse low-
order interaction network underlies a highly correlated
and learnable neural population code. Proc Natl Acad
Sci U S A 108, 9679 -- 9684 (2011).
[143] Lord, L. D. et al. Insights into brain architectures from
the homological scaffolds of functional connectivity net-
works. Front Syst Neurosci 10, 85 (2016).
[144] Sizemore, A. E., Phillips-Cremins, J., Ghrist, R. & Bas-
sett, D. S. The importance of the whole:
topologi-
cal data analysis for the network neuroscientist. arXiv
1806 (2018).
[145] Petri, G. et al. Homological scaffolds of brain functional
networks. J R Soc Interface 11, 20140873 (2014).
[146] Bassett, D. S., Wymbs, N. F., Porter, M. A., Mucha,
P. J. & Grafton, S. T. Cross-linked structure of network
evolution. Chaos 24, 013112 (2014).
[147] Giusti, C., Ghrist, R. & Bassett, D. S. Two's company,
three (or more) is a simplex: Algebraic-topological tools
for understanding higher-order structure in neural data.
J Comput Neurosci 41, 1 -- 14 (2016).
[148] Sizemore, A. E. et al. Cliques and cavities in the hu-
man connectome. J Comput Neurosci Epub Ahead of
Print (2017).
[149] Reimann, M. W. et al. Cliques of neurons bound into
cavities provide a missing link between structure and
function. Front Comput Neurosci 11, 48 (2017).
[150] Curto, C., Itskov, V., Veliz-Cuba, A. & Youngs, N. The
neural ring: an algebraic tool for analyzing the intrinsic
structure of neural codes. Bull Math Biol 75, 1571 -- 1611
(2013).
|
1906.06189 | 1 | 1906 | 2019-06-14T13:06:31 | The scientific case for brain simulations | [
"q-bio.NC"
] | A key element of the European Union's Human Brain Project (HBP) and other large-scale brain research projects is simulation of large-scale model networks of neurons. Here we argue why such simulations will likely be indispensable for bridging the scales between the neuron and system levels in the brain, and a set of brain simulators based on neuron models at different levels of biological detail should thus be developed. To allow for systematic refinement of candidate network models by comparison with experiments, the simulations should be multimodal in the sense that they should not only predict action potentials, but also electric, magnetic, and optical signals measured at the population and system levels. | q-bio.NC | q-bio | The scientific case for brain
simulations
Gaute T. Einevoll1,2,*, Alain Destexhe3,4, Markus Diesmann5,6,7, Sonja
Grün5,8, Viktor Jirsa9, Marc de Kamps10, Michele Migliore11, Torbjørn V.
Ness1, Hans E. Plesser1,5, Felix Schürmann12
1Faculty of Science and Technology, Norwegian University of Life Sciences, 1432 Ås,
Norway
2Department of Physics, University of Oslo, 0316 Oslo, Norway
3Paris-Saclay Institute of Neuroscience (NeuroPSI), Centre National de
la Recherche Scientifique, 91198 Gif-sur-Yvette, France
4European Institute for Theoretical Neuroscience, 75012 Paris, France
5Institute of Neuroscience and Medicine (INM-6) and Institute for Advanced
Simulation (IAS-6) and JARA-Institut Brain Structure-Function Relationships (INM-10),
Jülich Research Centre, 52425 Jülich, Germany
6Department of Psychiatry, Psychotherapy and Psychosomatics, RWTH Aachen
University, 52074 Aachen, Germany
7Department of Physics, RWTH Aachen University, 52074 Aachen, Germany
8Theoretical Systems Neurobiology, RWTH Aachen University, 52074 Aachen,
Germany
9Institut de Neurosciences des Systèmes (INS), Inserm, Aix Marseille Univ, 13005
Marseille, France
10Institute for Artificial and Biological Intelligence, School of Computing, Leeds LS2 9JT,
United Kingdom
11Institute of Biophysics, National Research Council, 90146 Palermo, Italy
12 Blue Brain Project, Ecole polytechnique federale de Lausanne (EPFL), Campus
Biotech, 1202 Geneva, Switzerland
*Corresponding Author: Gaute T. Einevoll, Faculty of Science and Technology, Norwegian
University of Life Sciences, PO Box 5003, 1432 Ås, Norway; [email protected]
In Brief
A key element of several large-scale brain research projects such as
the EU Human Brain Project is simulation of large networks of
neurons. Here it is argued why such simulations are indispensable for
bridging the neuron and system levels in the brain.
Abstract
A key element of the European Union's Human Brain Project (HBP)
and other large-scale brain research projects is simulation of large-
scale model networks of neurons. Here we argue why such
simulations will likely be indispensable for bridging the scales between
the neuron and system levels in the brain, and a set of brain
simulators based on neuron models at different levels of biological
detail should thus be developed. To allow for systematic refinement of
candidate network models by comparison with experiments, the
simulations should be multimodal in the sense that they should not
only predict action potentials, but also electric, magnetic, and optical
signals measured at the population and system levels.
Keywords:
brain simulation, neuron, network, model, simulation
1 Introduction
Despite decades of intense research efforts investigating the brain at the molecular, cell,
circuit and system levels, the operating principles of the human brain, or any brain, remain largely
unknown. Likewise, effective treatments for prevalent serious psychiatric disorders and dementia
are still lacking (Hyman, 2012; Masters et al., 2015). In broad terms one could argue that we now
have a fairly good understanding of how individual neurons operate and process information, but
that the behavior of networks of such neurons is poorly understood. Following the pioneering work
of Hubel and Wiesel mapping out receptive fields in the early visual system (Hubel and Wiesel,
1959), similar approaches have been used to explore how different types of sensory input and
behaviour are represented in the brain. In these projects the statistical correlation between
recorded neural activity, typically action potentials from single neurons, and sensory stimulation or
behaviour of the animal is computed. From this, so-called descriptive mathematical models have
been derived, accounting for, say, how the firing rate of a neuron in the visual system depends on
the visual stimulus, see e.g., Dayan and Abbott (2001, Ch. 2).
The qualitative insights gained by obtaining these descriptive receptive-field models should
not be underestimated, but these models offer little insight into how networks of neurons give rise
to the observed neural representations. Such insight will require mechanistic modeling where
neurons are explicitly modeled and connected in networks. Starting with the seminal work of
Hodgkin and Huxley who developed a mechanistic model for action-potential generation and
propagation in squid giant axons (Hodgkin and Huxley, 1952), biophysics-based modeling of neurons
is now well established (Koch, 1999; Dayan and Abbott, 2001; Sterratt et al., 2011). Numerous
mechanistic neuron models tailored to model specific neuron types have been constructed, for
example, for cells in mammalian sensory cortex (Hay et al., 2011; Markram et al., 2015; Pozzorini et
al., 2015), hippocampus (Migliore et al., 1995) and thalamus (McCormick and Huguenard, 1992;
Halnes et al., 2011).
At the level of networks, most mechanistic studies have focused on generic properties and
have considered stylized models with a single or a handful of neuronal populations consisting of
identical neurons with statistically identical connection properties. Such studies have given
invaluable qualitative insights into the wide range of possible network dynamics (see Brunel (2000)
for an excellent example), but real brain networks have heterogeneous neural populations and
more structured synaptic connections. For small networks, excellent models aiming to mimic real
neural networks have been developed, a prominent example being the circuit in the crustacean
stomatogastric nervous system comprising a few tens of neurons (Marder and Goaillard, 2006).
However, even though pioneering efforts to construct comprehensive networks with tens of
thousands of neurons mimicking cortical columns in mammalian sensory cortices, have been
pursued, e.g., Traub et al. (2005); Potjans and Diesmann (2014); Markram et al. (2015); Schmidt et
al. (2018a); Arkhipov et al. (2018), mechanistic modelling of biological neural networks mimicking
specific brains, or brain areas, is still in its infancy.
A cubic millimetre of cortex contains several tens of thousands of neurons, and until
recently, limitations in computer technology have prohibited the mathematical exploration of
neural networks mimicking cortical areas even in the smallest mammals. With the advent of modern
supercomputers, simulations of networks comprising hundreds of thousands or millions of neurons
are becoming feasible. Thus several large-scale brain projects, including the EU Human Brain Project
(HBP) and MindScope at the Allen Brain Institute, have endeavoured to create large-scale network
models for mathematical exploration of network dynamics (Kandel et al., 2013). In the HBP, where
all authors of this paper participate, the goal is not so much to create models for specific brain
areas, but rather to create general-purpose brain simulators. These brain simulators, which also
aptly are called brain-simulation engines, will not be tied to specific candidate models but rather be
applicable for execution of many candidate models, both current and future candidates. As such
their use by the scientific community for mathematical exploration of brain function is expected to
go well beyond the planned end of the HBP project in 2023.
In this article, we present the scientific case for brain simulations, in particular the
development and use of multi-purpose brain simulators, and argue why such simulators will be
indispensable in future neuroscience. Furthermore, the long-term maintenance and continued
development of such simulators are not feasible for individual researchers, nor individual research
groups. Rather, community efforts as exemplified by the brain-simulator developments in HBP are
required.
2
Brain simulations
Brain function relies on activity on many spatial scales, from the nanometer scale of atoms
and molecules to the meter scale of whole organisms, see, e.g., Devor et al. (2013). And unlike, for
example, in a canister of gas, these scales are intimately connected. While the replacement of a
single gas molecule with another has no effect on the overall behaviour of the gas, a change in a
DNA molecule can change the brain dramatically, like in Huntington's disease (Gusella et al., 1983).
Mechanistic models can act as `bridges between different levels of understanding' (Dayan and
Abbott, 2001, Preface) as for example in the Hodgkin-Huxley model where axonal action-potential
propagation is explained in terms of the properties of ion channels, that is, molecules (proteins)
embedded in the cell membrane. Today's most impressive multiscale simulations are arguably the
weather simulations that provide, with increased accuracy year by year, our weather forecasts
(Bauer et al., 2015). These physics- and chemistry-based simulations bridge scales from tens of
meters to tens of thousands of kilometers, the size of our planet, and are in computational
complexity comparable to whole-brain simulations (Koch and Buice, 2015).
2.1
Brain network simulations
Large-scale brain-simulation projects have until now predominantly focused on linking the
neuron level to the network level, that is, simulating synaptically connected networks of hundreds,
thousands or more neurons. One obvious reason is that at present such networks, whose properties
presumably lie at the heart of our cognitive abilities, are particularly difficult to understand with
qualitative reasoning alone, that is, without the aid of mathematics. Another reason is that starting
with the seminal works of Hodgkin and Huxley (1952) and Rall (Segev et al., 1994), we now have a
biophysically well-founded scheme for simulating how individual neurons process information, that
is, how they integrate synaptic inputs from other neurons and generate action potentials. This
scheme is covered in all textbooks in computational neuroscience (see, e.g., Koch (1999); Dayan and
Abbott (2001); Sterratt et al. (2011)) and typically also in computational neuroscience courses given
at universities. Numerous neuron models are now available for reuse and further development and
can be downloaded from databases such as ModelDB (senselab.med.yale.edu/modeldb/), the
Neocortical Microcircuit Collaboration (NMC) Portal (bbp.epfl.ch/nmc-portal), the Brain Observatory
at
Source Brain
(opensourcebrain.org). Mathematical models for synaptic function, including synaptic plasticity,
have also been developed, and all the necessary building blocks for creating models for networks of
neurons are thus available.
(observatory.brain-map.org), and Open
the Allen Brain
Institute
Some large-scale network models have been based on morphologically detailed neuron
models (Reimann et al., 2013; Markram et al., 2015; Arkhipov et al., 2018), some have used stylized
spatially-extended neuron models (Traub et al., 2005; Tomsett et al., 2015; Migliore et al., 2015),
some have used point neurons of the integrate-and-fire type (Lumer et al., 1997; Izhikevich and
Edelman, 2008; Potjans and Diesmann, 2014; Hagen et al., 2016; van Albada et al., 2018; Schmidt et
al., 2018a,b), and some have used firing-rate units representing population activity (Schirner et al.,
2018). More biological detail does not by itself mean that the model is more realistic. In fact, point
neurons, that is, neuron models where the membrane potential is assumed to be the same across
dendrites and soma, have been found to be excellently suited to reproduce experimentally recorded
action potentials following current stimulation (Jolivet et al., 2008; Pozzorini et al., 2015). The
various neuron models have different pros and cons, and the choice of which to use depends on the
question asked (Herz et al., 2006). We thus argue that a set of brain simulators for simulation of
models at different levels of biological detail should be developed.
For weather simulations the goal is clear, that is, to accurately predict temperature,
precipitation and wind at different geographical locations. Likewise, brain simulations should predict
what can be experimentally measured, not only action potentials, but also population-level
measures such as local field potentials (LFP), electrocorticographic signals (ECoG) and voltage-
sensitive dye imaging (VSDI) signals, as well as systems-level measurements such as signals recorded
by electroencephalography (EEG) or magnetoencephalography (MEG) (Brette and Destexhe, 2012),
cf. Figure 1. For these electrical, magnetic and optical measures the `measurement physics' seems
well established, that is, mathematical models for the biophysical link between electrical activity in
neurons and what is measured by such recordings have been developed, see references in caption
of
BIONET
(alleninstitute.github.io/bmtk/bionet.html) for prediction of such electrical and magnetic signals
from simulated network activity, both using biophysically-detailed multicompartment models
(Lindén et al., 2014; Gratiy et al., 2018; Hagen et al., 2018) and point-neuron models of the
integrate-and-fire type (Hagen et al., 2016), are now publically available. For functional magnetic
resonance imaging (fMRI) the biophysical link between activity in individual neurons and the
recorded BOLD signal is not yet established (but see Uhlirova et al. (2016b,a)), and a mechanistic
forward-modeling procedure linking microscopic brain activity to the measurements is not yet
available.
(lfpy.github.io)
Simulation
tools
Figure
1.
and
such
as
LFPy
Figure 1: Electric and magnetic signals to be computed in brain network simulations.
Measures of neural activity in cortical populations: (i) spikes (action potentials) and
LFP from a linear microelectrode inserted into cortical grey matter, (ii) ECoG from
electrodes positioned on the cortical surface, (iii) EEG from electrodes positioned on
the scalp, and (iv) MEG measuring magnetic fields stemming from brain activity by
means of SQUIDs placed outside the head. For reviews on the biophysical origin and
link between neural activity and the signals recorded in the various measurements, see
Hämäläinen et al. (1993); Nunez and Srinivasan (2006); Brette and Destexhe (2012);
Buzsaki et al. (2012); Einevoll et al. (2013); Pesaran et al. (2018); Hagen et al. (2018).
Figure 2 illustrates the use of brain network simulators for a so-called barrel column in
somatosensory cortex. Each such column primarily processes sensory information from a single
whisker on the snouts of rodents, and in rats a barrel column contains some tens of thousands of
neurons. A column can be modeled as a network of interconnected neurons based on biophysically-
detailed multicompartment models (here referred to as level I), point-neuron models of the
integrate-and-fire type (level II), or firing-rate units where each unit represents activity in a neuronal
population (level III). Regardless of the underlying neuron type, the simulator should preferably be
multimodal, that is, simultaneously predict many types of experimental signals stemming from the
same underlying network activity. Ideally the neuron models at the different levels of detail should
be interconnected in the sense that the simpler neuron models should be possible to reduce from
(or at least be compatible with) the more detailed neuron models. The field of statistical physics
addresses such scale bridging. A prime example of its application is the development of the
thermodynamic ideal-gas law describing the macroscopic properties of gases in terms of variables
like temperature or pressure from the microscopic Newtonian dynamics of the individual gas
molecules. As a neuroscience example, several projects have aimed to derive firing-rate models
(level III) from spiking neuron models (level II), see, e.g., de Kamps et al. (2008); Deco et al. (2008);
Ostojic and Brunel (2011); Bos et al. (2016); Schwalger et al. (2017); Heiberg et al. (2018).
Figure 2: Illustration of multimodal modeling with brain simulators. Network
dynamics in a cortical column (barrel) processing whisker stimulation in rat
somatosensory cortex (left) can be modelled with units at different levels of detail. In
the present example we have a level organization with biophysically detailed neuron
models (level I), simplified point-neuron models (level II), and firing-rate models with
neuron populations as fundamental units (level III). Regardless of level, the network
simulators should aim to predict the contribution of the network activity to all
available measurement modalities. In addition to the electric and magnetic
measurement modalities illustrated in Figure 1, the models may also predict optical
signals, for example, signals from voltage-sensitive dye imaging (VSDI) signals and
two-photon calcium imaging (Ca im.).
One simulator - many models
2.2
When discussing simulations, it is important to distinguish between the model and the simulator.
Here
• model refers to the equations with all parameters specified,
• simulator refers to the software tool that can execute the model (like NEURON, NEST, and
The Virtual Brain used in HBP), and
• simulation refers to the execution of a model in a simulator.
In some fields of science, simulators are intimately tied to solving a particular model. One
example is atomic physics where there is consensus both about what equation to solve (the
Schrödinger equation) and the numerical values of the few parameters involved (electron mass,
Planck's constant, ...). In contrast, for simulations of brain networks we can and should have a clear
separation. The simulators used in HBP are accordingly designed to execute many different models,
just like calculus can be used in many different physics calculations. Also, if possible, one should as a
control execute the same model on different simulators to check for consistency of the results, see,
e.g., van Albada et al. (2018); Shimoura et al. (2018). To facilitate this, software packages for
simulator-independent specification of neuronal network models are developed (Davison et al.,
2008).
Network simulators not tailored to specific brain-function hypotheses
3
In experimental neuroscience the method is often intimately tied to the hypothesis being tested. In
electrophysiological experiments, for example, the experimental set-up and execution protocols are
tailored to most efficiently answer the biological question asked. In contrast, brain network
simulators should not be designed to test a particular hypothesis about brain function, rather they
should be designed so that they can test many existing and future hypotheses.
Discovery of Newton's law of gravitation - an analogy
3.1
As the tools for testing a hypothesis can easily be confused with the hypothesis itself, we here
present an analogy from physics, Isaac Newton's discovery of the law of gravitation. While one
could argue that the establishment of a mechanistic understanding of brain function, that is,
understanding our cognitive abilities on the basis of neuronal action, would be a breakthrough of
similar magnitude as discovering this law, this is not the point here. Here this example is only used
to illustrate the role of network simulators in brain research.
Prior to Newton's theory, the planetary orbits had since ancient times been described by
the Ptolemaic model. This model assumed the Earth to be at center of the universe and the planets
to move in trajectories described by a complicated arrangement of circles within circles (so-called
epicycles). The model predicted the planetary orbits accurately and was used for more than 1500
years to make astronomical charts for navigation. It was thus a successful descriptive model, but it
shed little light on the underlying physical mechanisms governing the planetary movement. As such
it had a similarly useful role as the present descriptive, receptive-field-like models accounting for
neural representations in the brain.
Newton's theory of gravitation provided a mechanistic understanding of planetary
movement based on his two hypotheses on (i) how masses attract each other and (ii) how the
movement of masses is changed when forces are acting on them. But the theory went beyond
planetary orbits in that it, for example, also successfully predicted known kinematic laws of falling
apples, trajectories of cannon balls, and high and low tides due to gravitational attraction between
the Moon and the water in our oceans.
The first hypothesis of Newton was that two masses 𝑚 and 𝑀 with a separation distance 𝑟
attract each other with a force 𝐹𝑔 given by
𝐹𝑔 = 𝐺
𝑚𝑀
𝑟2 , (1)
where 𝐺 is the gravitational constant. The second hypothesis was that when a force 𝐹, in this case
𝐹𝑔 in Eq. (1), is acting on a mass 𝑚, the mass will be accelerated with an acceleration 𝑎 according to
𝑎 =
𝐹𝑔
𝑚
. (2)
To test the validity of these hypotheses, Newton had to compare with experiments, that is, available
measurements of planetary orbits. However, the connection between the mathematically
formulated hypotheses in Eqs. (1-2) and shapes of predicted planetary orbits is not obvious. In fact,
Newton developed a new type of mathematics, calculus, to make testable predictions from his
theory to allow for its validation (Leibniz independently developed calculus around the same time).
Without the appropriate mathematics it would have been impossible for Newton to test whether
nature behave according to his hypotheses.
Comparison with experiments demonstrated that Newtons's hypotheses were correct, and
Newton's theory of gravitation is now one of the pillars of physics. However, if it had turned out
that the predictions of planetary orbits were not in accordance with the observational data, he
could have tried other hypotheses, that is, other mathematically-formulated hypotheses than those
in Eqs. (1-2). Then, with the aid of his newly invented calculus, he could have made new planetary
orbit predictions and check whether these were in better agreement with measurements.
The point is that calculus was a tool to test Newton's hypotheses about the movement of
masses, it was not a part of the hypotheses themselves. Likewise, we argue that brain network
simulators should, in analogy to calculus, be designed to be tools for making precise predictions for
brain measurements for any candidate hypothesis for how brain networks are designed and
operate.
3.2
Hypothesis underlying brain network simulators
At present we do not have any well-grounded, and certainly not generally accepted, theory
about how networks of millions or billions of neurons work together to provide the salient brain
functions in animals or humans. We do not even have a well-established model for how neurons in
primary visual cortex of mammals work together to form the intriguing neuronal representations
with, for example, orientation selectivity and direction selectivity that were discovered by Hubel
and Wiesel sixty years ago (Hubel and Wiesel, 1959). Moreover, we do not have an overview over all
neuron types in the brain. However, we do know the biophysical principles for how to model
electrical activity in neurons and how neurons integrate synaptic inputs from other neurons and
generate action potentials. These principles, which go back to the work of Hodgkin and Huxley
(1952) and Rall (Segev et al., 1994) and are described in numerous textbooks (see, e.g., Koch (1999);
Dayan and Abbott (2001); Sterratt et al. (2011)), are the only hypotheses underlying the
construction of brain network simulators. This is the reason why many models can be represented
in the same simulator and why it is possible to develop generally applicable simulators for network
neuroscience.
However, while we know the principles for how to model neuronal activity, we do not a
priori know all the ingredients needed to fully specify network models. In order to construct
candidate network models,
information on the anatomical structure, electrophysiological
properties, and spatial positions of neurons, as well as information on how these neurons are
connected, are needed. The MindScope project at the Allen Brain Institute as well as the HBP are
gathering such data, and the first large-scale models are constructed on the basis of these and other
sources (Arkhipov et al., 2018). Although the primary goal in the HBP is to create general-purpose
brain simulators, such initial models are needed to guide the construction of these simulators and
to demonstrate their performance and potential usefulness. However, given the present lack of
data on, for example, the strength and plasticity of synaptic connections between the neurons, it is
clear that these initial models can be nothing more than plausible skeleton models to be used as
starting points for further explorations. Experimental data is thus collected to have a starting point
for mathematical exploration, not in the belief that brain function will be understood just by
collecting these data and 'putting them into a large simulator'.
Each candidate network model with specified neuron models, network structure and
synaptic connections precisely defined by a set of model parameters, can be thought of as a
candidate hypothesis. Brain network simulators should be designed to allow for the computation of
predictions of relevant experimental measures from any such candidate model (see Figure 2) so that
the merit of each model can be assessed by comparison with experiments.
In passing, we note that the use of mathematical simulators has a proud history in
neuroscience: The accurate model prediction of the speed and shape of propagating action
potentials in the squid giant axon by Hodgkin and Huxley in the early 1950s required the numerical
solution of the equations on a hand-operated calculation machine, since the newly installed
Cambridge computer was down for six months in 1951 (Hodgkin, 1976).
4
4.1
Use of brain network simulators
Biological imitation game
When the physicist Richard Feynman died in 1988, a statement on his blackboard read:
'What I cannot create, I do not understand.' In the present context an interpretation of this is that
unless we can create mechanistic mathematical models mimicking the behaviour in real brains, our
understanding will have to remain limited. An obvious use of brain network simulators is to
contribute towards building such models. In particular, the simulators should test candidate
network models against experiments so that over time network models improve and get closer to
the networks that are realized in real biological systems. This would amount to identifying the
models that perform best in the 'biological imitation game' (Koch and Buice, 2015), that is, the
models whose predictions best mimic experimental recordings of the same system.
In general a unique winner of such an imitation game will not be found, that is, a specific
network model with a specific set of model parameters. Rather, classes of candidate models with
similar structures and model parameters will likely do equally well, but as more experiments
become available the class of models jointly leading this game will expectedly be reduced in size. At
all times the leading models can be considered as the currently most promising hypotheses for how
the specific biological network is designed and operates, to be challenged by new experiments and
new candidate models.
For Newton it was clear what should be compared: the observed planetary orbits and the
corresponding orbits predicted by his theory. In brain science this is less clear. Action potentials are
clearly the key carrier of information, but what aspects of the trains of action potentials should be
mimicked by brain simulations? Detailed temporal sequence of actions potentials from individual
neurons, coefficients of variation, or firing rates of individual neurons (Jolivet et al., 2008; Gutzen et
al., 2018)? Or maybe the target only should be the average firing rates of populations of neurons?
Likewise, it is unclear what aspects of the LFP or VSDI signals should be compared, the full temporal
signals or maybe the power spectral densities? The question of what criteria should be used to
select the `best' model cannot be fully settled at present. The answer will also depend, for example,
on whether one believes information is coded in firing rates or in the detailed temporal structure of
action-potential trains. And maybe realistic behaviour of a robot following motor commands
produced by a model network could be one success criterium when such models become available?
However, such uncertainties regarding modeling targets should not preclude the initiation of a
'biological imitation game', they only mean that different rules of the game may be considered or
that the rules might change over time.
Overarching ideas on how computations are performed by the brain can also inspire
candidate models. For example, predictive coding (Rao and Ballard, 1999) has emerged as a
contender to the more traditional idea that the brain integrates information from the outside world
from feature detectors through predominantly feedforward processes. Instead, predictive coding
suggests that the brain is constantly updating hypotheses about the world and predicting sensory
information by feedback mechanisms. These two competing ideas could, when instantiated as
specific network models, make different predictions about neurophysiological experiments.
4.2
Validation of data-analysis methods
Another important application of brain simulations is to create benchmarking data for
validation of methods used to analyse experimental data. This approach has already been used to
generate benchmarking data for testing of automatic spike-sorting algorithms (Hagen et al., 2015),
methods for detecting putative synfire chains (Schrader et al., 2008), as well as for testing of
methods used to estimate current-source densities (CSDs) from recorded LFPs (Pettersen et al.,
2008). Several statistical methods have been developed for estimating, for example, functional
connectivities between neural populations and cortical areas based on population-level and
systems-level measures such as LFP, EEG, MEG and fMRI signals (Einevoll et al., 2013; Pesaran et al.,
2018). These statistical analysis methods should be validated on `virtual' benchmarking data
computed by brain network simulators where the ground-truth neuron and network activity are
known (Denker et al., 2012). Even if these model-based benchmarking data do not correspond in
detail to any specific biological system, data-analysis methods claimed to be generally applicable
should also perform well on these simulated data.
4.3
Use by wider research community
Anyone who has tried, knows that learning the calculus needed to derive planetary orbits
from Newton's hypotheses is demanding, and for most a formal training in mathematics is required.
Likewise, the development of brain simulators requires extensive training in mathematics, computer
science and physics, as well as a significant coordinated work effort involving many developers.
Fortunately, just like a practicing neuroscientist does not need to construct, say, a confocal
microscope in order to use it in research, simulators can be used without knowing all the technical
inner workings. Some simulators like NEURON (Carnevale and Hines, 2006) even come with a
graphical user interface, and plug-and-play programs where neural networks can be created and
simulated by pulling elements with the finger onto a canvas, have also been made (Dragly et al.,
2017).
While the complexity of many neural network models will make the use of solely graphical
user interfaces difficult, it should nevertheless be a goal to design brain simulators so that they also
can be used by the general neuroscience community. One aspect of this is that the developers of
widely used simulators should regularly offer tutorial and training courses, for example, in
connection with major neuroscience conferences. Further, high-quality user-level documentation
and support systems for personal inquiries by users must be set up. However, even the best user-
level documentation will not enable the general neuroscience community to easily set up large-
scale network models. Extracting all necessary neurobiological data from experiments, literature
and databases and specifying reliable executable model descriptions, generally require many years
of effort as exemplified by recent studies, e.g., Potjans and Diesmann (2014); Markram et al. (2015);
Schmidt et al. (2018a); Arkhipov et al. (2018). It is thus essential that executable descriptions of such
models are made publicly available as examples and starting points for the community.
Note that the publishing of executable model descriptions may require procedures and
tools that go beyond standard scientific publishing practices. For example, the recent publication of
a comprehensive multi-area model of macaque visual cortex (Schmidt et al., 2018a,b) was
accompanied by detailed model descriptions expressed by technologies like GitHub (https://inm-
6.github.io/multi-area-model/) and Snakemake (Köster and Rahmann, 2012) and accompanied by
an introductory video (https://youtu.be/YsH3BcyZBcU). Further, the authors also provided the
digitized workflow leading from the underlying experimental data to the model specification.
With a plausible candidate network model for, say, a part of V1 in a mouse as a starting
point, scientists with modest training in mathematics, physics, and computer science should be able
to use brain simulations to ask questions like:
• What is the predicted spiking response of various neuron types to different types of visual
stimuli?
• What is the predicted effect on network activity with pharmacological blocking of a
particular ion channel in a particular neuron type?
• What does the visually-evoked LFP signal recorded inside the cortex look like, and what
neuron populations are predicted to contribute most to this signal?
• Can the EEG signal recorded by electrodes on the scalp positioned outside the visual cortex
distinguish between two candidate network models for V1?
The application of brain simulators can be computationally demanding and require use of
supercomputers, especially if many model parameter combinations are to be investigated. Not
everyone has access to such supercomputers, nor the experience to install or maintain large
computer programs. One option would then be to make brain network simulators available through
web-based services so that all computations are done remotely on centralized supercomputer
centers, as is the plan for HBP.
In the long run, network neuroscience can only approach a mechanistic systems-level
understanding if we overcome the complexity barrier by learning how to build on the work of
others, that is, by eventually combining models of smaller brain networks components to larger
structures of more relevance for cognition. Newton said that he had seen further than others
because he was "standing on the shoulders of giants". Likewise, we here argue that we need to find
a way to "standing on the shoulders of each other's mathematical models" to have a hope for a
detailed understanding of the functioning of brain networks.
Discussion and outlook
5
We have here presented arguments for why brain network simulators are not only useful, but likely
also critical for advancing systems neuroscience. By drawing the analogy to Newton's discovery of
the law of gravity, we have argued that brain simulators should not be made to test specific
hypotheses about brain function. Rather, like Newton's development of calculus to allow for testing
of the validity of his physical hypotheses regarding planetary movement (Eqs. 1-2), brain simulators
should be viewed as `mathematical observatories' to test various candidate hypotheses. A brain
simulator is thus a tool, not a hypothesis, and can as such be likened to tools used to image brain
structure or brain activity.
In computational neuroscience one has to 'learn to compute without knowing all the
numbers' (as quoted from talk by John Hopfield at conference in Sigtuna, Sweden some twenty
years ago). What is meant by this is that unlike in, say, quantum-mechanical computations of atomic
properties where the handful of model parameters are known to many digits, the model
parameters specifying brain networks are numerous, uncertain and may also change over time. The
effects of uncertain model parameters and the uncertainty of model predictions can be
systematically studied, though such uncertainty quantification requires repeated evaluations of the
model of interest and is typically computationally demanding (Tennøe et al., 2018).
We need a set of different brain network simulators describing the neurons at different
levels of resolution, that is, different levels of biological detail as exemplified by the three levels
depicted in Figure 2. Simulators based on biophysically detailed, multicompartmental neuron
models (level I) can explore in detail how the dendritic structures affect integration of synaptic
inputs and consequently the network dynamics (Reimann et al., 2013; Markram et al., 2015;
Migliore et al., 2015; Arkhipov et al., 2018). Simulators based on point neurons of the integrate-and-
fire type (level II) are much less computationally demanding so that larger networks can be studied
(Potjans and Diesmann, 2014). Further, the number of model parameters is much smaller and the
fitting of single-neuron models to experimental recordings easier (Pozzorini et al., 2015). Population
firing-rate models (level III) model the dynamics of entire populations which makes the models
computationally, and often also conceptually, much easier (de Kamps et al., 2008; Cain et al., 2016;
Schwalger et al., 2017). With population firing-rate models describing a small patch of cortex, so-
called neural-mass models, one can derive spatially extended models, so-called neural-field models,
covering cortical areas and even complete human cortices (Deco et al., 2008; Ritter et al., 2013;
Sanz Leon et al., 2013; Breakspear, 2017; Schirner et al., 2018). While at present most neural-field
models are based on largely phenomenological neural-mass models (Jansen and Rit, 1995; Deco et
al., 2008), the future goal should be to derive such neural-mass building blocks from population
network models based on individual neurons (Zerlaut et al., 2018), or from fitting to experiments
(Blomquist et al., 2009).
Brain network simulation is still in its infancy, and the simulators and the associated
infrastructure should be developed to allow for the study of larger networks and fully exploit the
capabilities of modern computer hardware (see, e.g., Akar et al. (2019) and Kumbhar et al (2019)).
They should also allow for the study of longer time-scale processes such as homeostatic and
synaptic plasticity (Turrigiano and Nelson, 2004; Keck et al., 2017). With plausible biophysics-based
rules for, for example, spike-timing-dependent long-term synaptic plasticity included in the models,
studies of learning will also be possible. Brain simulators should eventually also be extended to go
beyond the modeling of networks of neurons alone to also incorporate extracellular space and
interaction with glia cells (Solbrå et al., 2018). Likewise, they should allow for studies of effects of
electrical or magnetic stimulation of the brain, either with intracranial electrodes like in deep-brain
stimulation (Perlmutter and Mink, 2006), with surface electrodes (Bosking et al., 2017), or
transcranially (Wassermann et al., 2008).
The present paper has focused on brain simulators for studying networks of neurons. While
not addressed here, there is clearly also a need for simulators for studying brain activity at the
subcellular scale, both for modelling molecular signaling pathways governed by reaction-diffusion
dynamics (Bhalla and Wils, 2010) and for modeling molecular dynamics by Newtonian mechanics
(Rapaport, 2004). We have also focused on the bottom-up-type network models typically pursued in
the computational neuroscience community where model predictions can be compared directly
with physiological experiments. However, network models can also be very useful for concisely and
precisely representing ideas on how the brain may implement cognitive processes. An early
example of such work is the so-called Hopfield model describing how associative memory can be
achieved in recurrent networks of binary neurons (Hopfield, 1982). Over the last decades such
modeling work has, for example, grown to include visual attention (Reynolds and Desimone, 1999;
Deco and Rolls, 2004), language representation (van der Velde and de Kamps, 2006), decision
making (Gold and Shadlen, 2007), and learning (Brader et al., 2007). While none of these works
immediately predict specific detailed outcomes of neurophysiological experiments, they state ideas
about cognitive phenomenon in a concise manner that allows scrutiny and critique.
The development of high-quality brain simulators requires long-term commitment of
resources. Both NEURON and NEST, two key brain simulators in the Human Brain Project, have been
developed over a time period of more than 25 years. Likewise, the continued development,
maintenance and user support of key brain simulators used by the research community will require
long-term funding. These simulation tools can be likened to other joint research infrastructures such
as astronomical observatories or joint international facilities for studies of subatomic particles.
While the expenses for the operation of brain simulators will be much smaller than these
experimental facilities, they should nevertheless be considered as necessary research infrastructure
and preferably be funded as such.
Until we learn how the wide range of spatial and temporal scales involved in brain function
are connected, our understanding of our brains will be limited. Bridging these scales with
mathematical modeling will be a daunting challenge, but encouragingly there are examples from
other branches of science where many scales have been bridged, the most visible likely being
numerical weather prediction (Bauer et al., 2015). Another impressive example of scale bridging is
the engineering underlying smart phones. Here tailored materials made of selected semiconductor
and metal atoms are assembled into numerous transistors (in some sense analogous to neurons)
connected in networks on a chip (`brain'), which together with other components make up the
smart phone (`organism'). These examples have been totally dependent on mathematics and
simulations to bridge models at different scales. So a natural question is: Do we have a chance of
ever understanding brain function without brain simulations?
Acknowledgments
Funding was received from the European Union Seventh Framework Program (FP7/20072013)
under grant agreement No. 604102 (HBP), the European Union's Horizon 2020 Framework
Programme for Research and Innovation under Grant Agreements No. 720270 (Human Brain Project
SGA1), No. 785907 (Human Brain Project SGA2) and No. 754304 (DEEP-EST), the Research Council of
Norway (DigiBrain 248828, CoBra 250128), the Deutsche Forschungsgemeinschaft Grants GR
1753/4-2 and DE 2175/2-1 of the Priority Program (SPP 1665), the Helmholtz Association Initiative
and Networking Fund under project numbers ZT-I-0003 (HAF) and SO-902 (ACA), RTG 2416 'Multi-
senses Multi-scales', VSR computation time grant Brain-scale simulations JINB33, and the Swiss ETH
Domain for the Blue Brain Project.
References
Akar, A.N., Cumming, B., Karakasis, V., Küsters, A., Klijn, W., Peyser, A., Yates, S. (2019). Arbor -- a
morphologically-detailed neural network simulation library for contemporary high-performance computing
architectures, arXiv e-prints, https://arxiv.org/pdf/1901.07454.pdf
Arkhipov, A., Gouwens, N. W., Billeh, Y. N., Gratiy, S., Iyer, R., Wei, Z., Xu, Z., Abbasi-Asl,
R., Berg, J., Buice, M., et al. (2018). Visual physiology of the layer 4 cortical circuit in
silico. PLoS Comput. Biol. 14, e1006535. doi:10.1371/journal.pcbi.1006535
Bauer, P., Thorpe, A., and Brunet, G. (2015). The quiet revolution of numerical weather
prediction. Nature 525, 47 -- 55. doi:10.1038/nature14956
Bhalla, U. S. and Wils, S. (2010). Computational modeling methods for neuroscientists (MIT
Press), chap. Reaction-Diffusion Modeling. 61 -- 92
Blomquist, P., Devor, A., Indahl, U. G., Ulbert, I., Einevoll, G. T., and Dale, A. M. (2009).
Estimation of thalamocortical and intracortical network models from joint thalamic single
electrode and cortical laminar-electrode recordings in the rat barrel system. PLoS Comput.
Biol. 5, e1000328. doi:10.1371/journal.pcbi.1000328
Bos, H., Diesmann, M., and Helias, M. (2016). Identifying anatomical origins of coexisting
oscillations in the cortical microcircuit. PLoS Comput. Biol. 12, e1005132. doi:10.1371/
journal.pcbi.1005132
Bosking, W. H., Sun, P., Ozker, M., Pei, X., Foster, B. L., Beauchamp, M. S., and Yoshor,
D. (2017). Saturation in phosphene size with increasing current levels delivered to human
visual cortex. J. Neurosci. 37, 7188 -- 7197. doi:10.1523/JNEUROSCI.2896-16.2017
Brader, J. M., Senn, W., and Fusi, S. (2007). Learning real-world stimuli in a neural network
with spike-driven synaptic dynamics. Neural Comput. 19, 2881 -- 2912. doi:10.1162/neco.
2007.19.11.2881
Breakspear, M. (2017). Dynamic models of large-scale brain activity. Nat. Neurosci. 20,
340 -- 352. doi:10.1038/nn.4497
Brette, R. and Destexhe, A. (eds.) (2012). Handbook of Neural Activity Measurement
(Cambridge University Press)
Brunel, N. (2000). Dynamics of sparsely connected networls of excitatory and inhibitory
neurons. Comput. Neurosci. 8, 183 -- 208. doi:10.1016/S0928-4257(00)01084-6
Buzsaki, G., Anastassiou, C., and Koch, C. (2012). The origin of extracellular fields and
currents -- EEG, ECoG, LFP and spikes. Nat. Rev. Neurosci. 13, 407 -- 420
Cain, N., Iyer, R., Koch, C., and Mihalas, S. (2016). The computational properties of a simplified cortical column
model. PLoS Comput. Biol. 12, e1005045. doi:10.1371/journal.pcbi.1005045
Carnevale, N. T. and Hines, M. L. (2006). The NEURON Book (Cambridge University Press)
Davison, A. P., Brderle, D., Eppler, J., Kremkow, J., Muller, E., Pecevski, D., Perrinet, L.,
and Yger, P. (2008). Pynn: A common interface for neuronal network simulators. Front.
Neuroinf. 2, 11. doi:10.3389/neuro.11.011.2008
Dayan, P. and Abbott, L. (2001). Theoretical neuroscience (MIT Press, Cambridge)
de Kamps, M., Baier, V., Drever, J., Dietz, M., Mösenlechner, L., and van der Velde, F. (2008).
The state of miind. Neural networks 21, 1164 -- 1181. doi:10.1016/j.neunet.2008.07.006
Deco, G., Jirsa, V. K., Robinson, P. A., Breakspear, M., and Friston, K. (2008). The dynamic
brain: from spiking neurons to neural masses and cortical fields. PLoS Comput. Biol. 4,
e1000092. doi:10.1371/journal.pcbi.1000092
Deco, G. and Rolls, E. T. (2004). A neurodynamical cortical model of visual attention and
invariant object recognition. Vision Res. 44, 621 -- 642
Denker, M., Einevoll, G., Franke, F., Grün, S., Hagen, E., Kerr, J., Nawrot, M., Ness, T. B.,
Wachtler, T., and Wojcik, D. (2012). Report from 1st INCF Workshop on Validation of
Analysis Methods. Tech. rep., International Neuroinformatics Coordinating Facility (INCF)
Devor, A., Bandettini, P. A., Boas, D. A., Bower, J. M., Buxton, R. B., Cohen, L. B., Dale,
A. M., Einevoll, G. T., Fox, P. T., Franceschini, M. A., et al. (2013). The challenge of
connecting the dots in the B.R.A.I.N. Neuron 80, 270 -- 274. doi:10.1016/j.neuron.2013.09.008
Dragly, S.-A., Hobbi Mobarhan, M., Våvang Solbrå, A., Tennøe, S., Hafreager, A., Malthe-
Sørenssen, A., Fyhn, M., Hafting, T., and Einevoll, G. T. (2017). Neuronify: An educational
simulator for neural circuits. eNeuro 4. doi:10.1523/ENEURO.0022-17.2017
Einevoll, G., Kayser, C., Logothetis, N., and Panzeri, S. (2013). Modelling and analysis of
local field potentials for studying the function of cortical circuits. Nat. Rev. Neurosci. 14,
770 -- 785. doi:10.1038/nrn3599
Gold, J. I. and Shadlen, M. N. (2007). The neural basis of decision making. Annu. Rev.
Neurosci. 30, 535 -- 574. doi:10.1146/annurev.neuro.29.051605.113038
Gratiy, S. L., Billeh, Y. N., Dai, K., Mitelut, C., Feng, D., Gouwens, N. W., Cain, N., Koch,
C., Anastassiou, C. A., and Arkhipov, A. (2018). BioNet: A python interface to neuron for
modeling large-scale networks. PloS ONE 13, e0201630. doi:10.1371/journal.pone.0201630
Gusella, J. F., Wexler, N. S., Conneally, P. M., Naylor, S. L., Anderson, M. A., Tanzi, R. E.,
Watkins, P. C., Ottina, K., Wallace, M. R., and Sakaguchi, A. Y. (1983). A polymorphic
dna marker genetically linked to huntington's disease. Nature 306, 234 -- 238
Gutzen, R., von Papen, M., Trensch, G., Quaglio, P., Grün, S., and Denker, M. (2018).
Reproducible neural network simulations: Statistical methods for model validation on the
level of network activity data. Front. Neuroinf. 12, 90. doi:10.3389/fninf.2018.00090
Hagen, E., Dahmen, D., Stavrinou, M. L., Linden, H., Tetzlaff, T., van Albada, S. J., Grun, S.,
Diesmann, M., and Einevoll, G. T. (2016). Hybrid scheme for modeling local field potentials
from point-neuron networks. Cereb. Cortex 26, 4461 -- 4496. doi:10.1093/cercor/bhw237
Hagen, E., Naess, S., Ness, T. V., and Einevoll, G. T. (2018). Multimodal modeling of neural
network activity: computing LFP, ECoG, EEG and MEG signals with LFPy 2.0.
Front. Neuroinf. 12, 92. doi:10.3389/fninf.2018.00092
Hagen, E., Ness, T. V., Khosrowshahi, A., Sørensen, C., Fyhn, M., Hafting, T., Franke, F.,
and Einevoll, G. T. (2015). ViSAPy: A python tool for biophysics-based generation of virtual
spiking activity for evaluation of spike-sorting algorithms. J. Neurosci. Methods 45, 182 -- 204
Halnes, G., Augustinaite, S., Heggelund, P., Einevoll, G. T., and Migliore, M. (2011). A
Multi-Compartment Model for Interneurons in the Dorsal Lateral Geniculate Nucleus. PLoS
Comput. Biol. 7. doi:10.1371/journal.pcbi.1002160
Hämäläinen, M., Hari, R., Ilmoniemi, R., Knuutila, J., and Lounasmaa, O. V. (1993).
Magnetoencephalography - theory, instrumentation, and applications to noninvasive studies of the
working human brain. Rev. Mod. Phys. 65, 413 -- 497
Hay, E., Hill, S., Schürmann, F., Markram, H., and Segev, I. (2011). Models of Neocortical
Layer 5b Pyramidal Cells Capturing a Wide Range of Dendritic and Perisomatic Active
Properties. PLoS Comput. Biol. 7, e1002107. doi:10.1371/journal.pcbi.1002107
Heiberg, T., Kriener, B., Tetzlaff, T., Einevoll, G. T., and Plesser, H. E. (2018). Firing-rate
models for neurons with a broad repertoire of spiking behaviors. J. Comput. Neurosci.
doi:10.1007/s10827-018-0693-9
Herz, A. V. M., Gollisch, T., Machens, C. K., and Jaeger, D. (2006). Modeling single-neuron
dynamics and computations: a balance of detail and abstraction. Science 314, 80 -- 85.
doi:10.1126/science.1127240
Hodgkin, A. L. (1976). Chance and design in electrophysiology: an informal account of certain
experiments on nerve carried out between 1934 and 1952. J. Physiol. 263, 1 -- 21
Hodgkin, A. L. and Huxley, A. F. (1952). A quantitative description of membrane current and
its application to conduction and excitation in nerve. J. Physiol. 117, 500 -- 544
Hopfield, J. J. (1982). Neural networks and physical systems with emergent collective
computational abilities. PNAS 79, 2554 -- 2558
Hubel, D. H. and Wiesel, T. N. (1959). Receptive fields of single neurones in the cat's striate
cortex. J. Physiol. 148, 574 -- 591
Hyman, S. E. (2012). Revolution stalled. Sci. Transl. Med. 4, 155. doi:10.1126/
scitranslmed.3003142
Izhikevich, E. M. and Edelman, G. M. (2008). Large-scale model of mammalian thalamocortical
systems. PNAS 105, 3593 -- 8. doi:10.1073/pnas.0712231105
Jansen, B. H. and Rit, V. G. (1995). Electroencephalogram and visual evoked potential
generation in a mathematical model of coupled cortical columns. Biol. Cybern. 73, 357 -- 366
Jolivet, R., Schürmann, F., Berger, T. K., Naud, R., Gerstner, W., and Roth, A. (2008).
The quantitative single-neuron modeling competition. Biol. Cybern. 99, 417 -- 426. doi:
10.1007/s00422-008-0261-x
Kandel, E. R., Markram, H., Matthews, P. M., Yuste, R., and Koch, C. (2013). Neuroscience
thinks big (and collaboratively). Nat. Rev. Neurosci. 14, 659 -- 664. doi:10.1038/nrn3578
Keck, T., Toyoizumi, T., Chen, L., Doiron, B., Feldman, D. E., Fox, K., Gerstner, W., Haydon,
P. G., Hübener, M., Lee, H.-K., et al. (2017). Integrating hebbian and homeostatic plasticity:
the current state of the field and future research directions. Phil. Trans. R. Soc. B 372.
doi:10.1098/rstb.2016.0158
Koch, C. (1999). Biophysics of Computation (Oxford Univ Press, Oxford)
Koch, C. and Buice, M. A. (2015). A biological imitation game. Cell 163, 277 -- 280. doi:
10.1016/j.cell.2015.09.045
Köster, J. and Rahmann, S. (2012). Snakemake -- a scalable bioinformatics workflow engine.
Bioinformatics 28, 2520 -- 2522. doi:10.1093/bioinformatics/bts480
Kumbhar, P., Hines M., Fouriaux, J., Ovcharenko, A., King, J., Delalondre, F., Schürmann, F. (2019).
CoreNEURON - An Optimized Compute Engine for the NEURON Simulator, arXiv e-prints,
https://arxiv.org/pdf/1901.10975.pdf
Linden, H., Hagen, E., Ł¸eski, S., Norheim, E. S., Pettersen, K. H., and Einevoll, G. T. (2014).
LFPy: A tool for biophysical simulation of extracellular potentials generated by detailed
model neurons. Front Neuroinf. 7. doi:10.3389/fninf.2013.00041
Lumer, E. D., Edelman, G. M., and Tononi, G. (1997). Neural dynamics in a model of the
thalamocortical system. I. Layers, loops and the emergence of fast synchronous rhythms.
Cereb. Cortex 7, 207 -- 227
Marder, E. and Goaillard, J.-M. (2006). Variability, compensation and homeostasis in neuron
and network function. Nat. Rev. Neurosci. 7, 563 -- 574. doi:10.1038/nrn1949
Markram, H., Muller, E., Ramaswamy, S., Reimann, M. W., Abdellah, M., Sanchez, C. A.,
Ailamaki, A., Alonso-Nanclares, L., Antille, N., Arsever, S., et al. (2015). Reconstruction and
simulation of neocortical microcircuitry. Cell 163, 456 -- 492. doi:10.1016/j.cell.2015.09.029
Masters, C. L., Bateman, R., Blennow, K., Rowe, C. C., Sperling, R. A., and Cummings, J. L.
(2015). Alzheimer's disease. Nat Rev. Dis. Primers 1, 15056. doi:10.1038/nrdp.2015.56
McCormick, D. A. and Huguenard, J. R. (1992). A model of the electrophysiological properties
of thalamocortical relay neurons. J. Neurophysiol. 68, 1384 -- 1400
Migliore, M., Cavarretta, F., Marasco, A., Tulumello, E., Hines, M. L., and Shepherd, G. M.
(2015). Synaptic clusters function as odor operators in the olfactory bulb. PNAS, 201502513
Migliore, M., Cook, E. P., Jaffe, D. B., Turner, D. A., and Johnston, D. (1995). Computer
simulations of morphologically reconstructed ca3 hippocampal neurons. J. Neurophysiol.
73, 1157 -- 1168. doi:10.1152/jn.1995.73.3.1157
Nunez, P. L. and Srinivasan, R. (2006). Electric fields of the brain: The Neurophysics of EEG
(Oxford University Press, Inc.), 2nd ed.
Ostojic, S. and Brunel, N. (2011). From spiking neuron models to linear-nonlinear models.
PLoS Comput. Biol. 7, e1001056. doi:10.1371/journal.pcbi.1001056
Perlmutter, J. S. and Mink, J. W. (2006). Deep brain stimulation. Annu. Rev. Neurosci. 29,
229 -- 257. doi:10.1146/annurev.neuro.29.051605.112824
Pesaran, B., Vinck, M., Einevoll, G. T., Sirota, A., Fries, P., Siegel, M., Truccolo, W.,
Schroeder, C. E., and Srinivasan, R. (2018). Investigating large-scale brain dynamics
using field potential recordings: analysis and interpretation. Nat. Neurosci. 21, 903 -- 919.
doi:10.1038/s41593-018-0171-8
Pettersen, K. H., Hagen, E., and Einevoll, G. T. (2008). Estimation of population firing rates
and current source densities from laminar electrode recordings. J. Comput. Neurosci. 24,
291 -- 313. doi:10.1007/s10827-007-0056-4
Potjans, T. C. and Diesmann, M. (2014). The cell-type specific cortical microcircuit: relating
structure and activity in a full-scale spiking network model. Cereb. Cortex 24, 785 -- 806.
doi:10.1093/cercor/bhs358
Pozzorini, C., Mensi, S., Hagens, O., Naud, R., Koch, C., and Gerstner, W. (2015). Automated
high-throughput characterisation of single neurons by means of simplified spiking models.
PLoS Comput. Biol. 11, e1004275. doi:10.1371/journal.pcbi.1004275
Rao, R. P. and Ballard, D. H. (1999). Predictive coding in the visual cortex: a functional
interpretation of some extra-classical receptive-field effects. Nat. Neurosci. 2, 79 -- 87. doi:
10.1038/4580
Rapaport, D. C. (2004). The art of molecular dynamics simulation (Cambridge University
Press)
Reimann, M. W., Anastassiou, C. A., Perin, R., Hill, S. L., Markram, H., and Koch, C. (2013).
A biophysically detailed model of neocortical local field potentials predicts the critical role
of active membrane currents. Neuron 79, 375 -- 390. doi:10.1016/j.neuron.2013.05.023
Reynolds, J. H. and Desimone, R. (1999). The role of neural mechanisms of attention in solving
the binding problem. Neuron 24, 19 -- 29, 111 -- 25
Ritter, P., Schirner, M., McIntosh, A. R., and Jirsa, V. K. (2013). The virtual brain integrates
computational modeling and multimodal neuroimaging. Brain Connect. 3, 121 -- 145. doi:
10.1089/brain.2012.0120
Sanz Leon, P., Knock, S. A., Woodman, M. M., Domide, L., Mersmann, J., McIntosh, A. R.,
and Jirsa, V. (2013). The virtual brain: a simulator of primate brain network dynamics.
Front. Neuroinf. 7, 10. doi:10.3389/fninf.2013.00010
Schirner, M., McIntosh, A. R., Jirsa, V., Deco, G., and Ritter, P. (2018). Inferring multi-scale
neural mechanisms with brain network modelling. eLife 7. doi:10.7554/eLife.28927
Schmidt, M., Bakker, R., Shen, K., Bezgin, G., Diesmann, M., and van Albada, S. J. (2018a).
Multi-scale account of the network structure of macaque visual cortex. Brain Struct. Funct.
223, 1409 -- 1435. doi:10.1007/s00429-017-1554-4
Schmidt, M., Bakker, R., Shen, K., Bezgin, G., Diesmann, M., and van Albada, S. J. (2018b).
A multi-scale layer-resolved spiking network model of resting-state dynamics in macaque
visual cortical areas. PLoS Comput. Biol. 14, e1006359. doi:10.1371/journal.pcbi.1006359
Schrader, S., Grün, S., Diesmann, M., and Gerstein, G. L. (2008). Detecting synfire chain
activity using massively parallel spike train recording. J. Neurophysiol. 100, 2165 -- 2176.
doi:10.1152/jn.01245.2007
Schwalger, T., Deger, M., and Gerstner, W. (2017). Towards a theory of cortical columns:
From spiking neurons to interacting neural populations of finite size. PLoS Comput. Biol.
13, e1005507. doi:10.1371/journal.pcbi.1005507
Segev, I., Rinzel, J., and Shepherd, G. M. (eds.) (1994). Theoretical Foundations of Dendritic
Function: The Collected Papers of Wilfrid Rall with Commentaries (MIT Press)
Shimoura, R. O., Kamiji, N. L., de Oliveira Pena, R. F., Cordeiro, V. L., Ceballos, C. C.,
Romaro, C., and Roque, A. C. (2018). Reimplementation of the potjans-diesmann cortical
microcircuit model: from nest to brian. ReScience 4, 2
Solbrå, A., Bergersen, A., van den Brink, J., Malthe-Sørenssen, A., Einevoll, G., and Halnes,
G. (2018). A kirchhoff-nernst-planck framework for modeling large scale extracellular
electrodiffusion surrounding morphologically detailed neurons. PLoS Comput. Biol. 14,
e1006510
Sterratt, D., Graham, B., Gillies, A., and Willshaw, D. (2011). Principles of computational
modelling in neuroscience (Cambridge University Press)
Tennøe, S., Halnes, G., and Einevoll, G. T. (2018). UncertainPy: A python toolbox for
uncertainty quantification and sensitivity analysis in computational neuroscience. Front.
Neuroinf. 12, 49. doi:10.3389/fninf.2018.00049
Tomsett, R. J., Ainsworth, M., Thiele, A., Sanayei, M., Chen, X., Gieselmann, M. A.,
Whittington, M. A., Cunningham, M. O., and Kaiser, M. (2015). Virtual electrode recording
tool for extracellular potentials (vertex): comparing multi-electrode recordings from simulated
and biological mammalian cortical tissue. Brain Struct. Funct. 220, 2333 -- 2353.
doi:10.1007/s00429-014-0793-x
Traub, R. D., Contreras, D., Cunningham, M. O., Murray, H., LeBeau, F. E. N., Roopun,
A., Bibbig, A., Wilent, W. B., Higley, M. J., and Whittington, M. a. (2005). Single-column
thalamocortical network model exhibiting gamma oscillations, sleep spindles, and
epileptogenic bursts. J. Neurophysiol. 93, 2194 -- 232. doi:10.1152/jn.00983.2004
Turrigiano, G. G. and Nelson, S. B. (2004). Homeostatic plasticity in the developing nervous
system. Nat. Rev. Neurosci. 5, 97 -- 107. doi:10.1038/nrn1327
Uhlirova, H., Kılıç, K., Tian, P., Sakadžic, S., Gagnon, L., Thunemann, M., Desjardins,
M., Saisan, P. A., Nizar, K., Yaseen, M. A., et al. (2016a). The roadmap for estimation
of cell-type-specific neuronal activity from non-invasive measurements.
Phil. Trans. R. Soc. B 371, 20150356. doi:10.1098/rstb.2015.0356
Uhlirova, H., Kilic, K., Tian, P., Thunemann, M., Desjardins, M., Saisan, P. A., Sakadzic, S.,
Ness, T. V., Mateo, C., Cheng, Q., et al. (2016b). Cell type specificity of neurovascular
coupling in cerebral cortex. eLife 5. doi:10.7554/eLife.14315
van Albada, S. J., Rowley, A. G., Senk, J., Hopkins, M., Schmidt, M., Stokes, A. B., Lester,
D. R., Diesmann, M., and Furber, S. B. (2018). Performance comparison of the digital
neuromorphic hardware spinnaker and the neural network simulation software nest for a
full-scale cortical microcircuit model. Front. Neurosci. 12, 291. doi:10.3389/fnins.2018.00291
van der Velde, F. and de Kamps, M. (2006). Neural blackboard architectures of combinatorial
structures in cognition. Behav. Brain Sci. 29, 37 -- 70; discussion 70 -- 108. doi:10.1017/
S0140525X06009022
Wassermann, E., Epstein, C., Ziemann, U., Walsh, V., Paus, T., and Lisanby, S. (2008). Oxford
handbook of transcranial stimulation (Oxford University Press)
Zerlaut, Y., Chemla, S., Chavane, F., and Destexhe, A. (2018). Modeling mesoscopic cortical
dynamics using a mean-field model of conductance-based networks of adaptive exponential
integrate-and-fire neurons. J. Comput. Neurosci. 44, 45 -- 61. doi:10.1007/s10827-017-0668-2
|
1511.00673 | 3 | 1511 | 2016-08-08T10:07:59 | Flow-based network analysis of the Caenorhabditis elegans connectome | [
"q-bio.NC",
"physics.soc-ph"
] | We exploit flow propagation on the directed neuronal network of the nematode Caenorhabditis elegans to reveal dynamically relevant features of its connectome. We find flow-based groupings of neurons at different levels of granularity, which we relate to functional and anatomical constituents of its nervous system. A systematic in silico evaluation of the full set of single and double neuron ablations is used to identify deletions that induce the most severe disruptions of the multi-resolution flow structure. Such ablations are linked to functionally relevant neurons, and suggest potential candidates for further in vivo investigation. In addition, we use the directional patterns of incoming and outgoing network flows at all scales to identify flow profiles for the neurons in the connectome, without pre-imposing a priori categories. The four flow roles identified are linked to signal propagation motivated by biological input-response scenarios. | q-bio.NC | q-bio |
Flow-based network analysis of the Caenorhabditis elegans connectome
Karol A. Bacik,1, ∗ Michael T. Schaub,1, 2, 3, † Mariano
Beguerisse-D´ıaz,1, ‡ Yazan N. Billeh,4 and Mauricio Barahona1, §
1Department of Mathematics, Imperial College London, London SW7 2AZ, United Kingdom
2naXys & Department of Mathematics, University of Namur, B-5000 Namur, Belgium
3ICTEAM, Universit´e catholique de Louvain, B-1348 Louvain-la-Neuve, Belgium
4Computation and Neural Systems Program, California Institute of Technology, CA 91125 Pasadena, USA
(Dated: August 9, 2016)
We exploit flow propagation on the directed neuronal network of the nematode C. elegans
to
reveal dynamically relevant features of its connectome. We find flow-based groupings of neurons at
different levels of granularity, which we relate to functional and anatomical constituents of its nervous
system. A systematic in silico evaluation of the full set of single and double neuron ablations is used
to identify deletions that induce the most severe disruptions of the multi-resolution flow structure.
Such ablations are linked to functionally relevant neurons, and suggest potential candidates for
further in vivo investigation. In addition, we use the directional patterns of incoming and outgoing
network flows at all scales to identify flow profiles for the neurons in the connectome, without
pre-imposing a priori categories. The four flow roles identified are linked to signal propagation
motivated by biological input-response scenarios.
AUTHOR SUMMARY
One of the goals of systems neuroscience is to eluci-
date the relationship between the structure of neuronal
networks and the functional dynamics that they imple-
ment. An ideal model organism to study such interac-
tions is the roundworm C. elegans, which not only has a
fully mapped connectome, but has also been the object
of extensive behavioural, genetic and neurophysiological
experiments. Here we present an analysis of the neu-
ronal network of C. elegans from a dynamical flow per-
spective. Our analysis reveals a multi-scale organisation
of the signal flow in the network linked to anatomical
and functional features of neurons and identifies differ-
ent neuronal roles in relation to signal propagation. We
use our computational framework to explore biological
input-response scenarios as well as exhaustive in silico
ablations, which we relate to experimental findings re-
ported in the literature.
I.
INTRODUCTION
The nematode Caenorhabditis elegans has been used as
a model organism in the life sciences for half a century [1],
and considerable effort has been devoted to elucidate the
properties of its nervous system in relation to functional
behaviour. The C. elegans connectome was originally
charted in 1986 by White et al [2] and has been further
refined by analysis and experiments [3], most recently
in the work of Varshney et al [4]. Using experimental
∗ [email protected]
† [email protected]
‡ [email protected]
§ [email protected]
techniques such as laser ablations, calcium imaging, op-
togenetics and sonogenetics, researchers have examined
functional properties of individual neurons in connection
with motion, learning, or information processing and in-
tegration [5–9]. Other studies have quantified the char-
acteristics of the motion of C. elegans, and how these
change upon genetic mutations [10–12].
With the increased availability of data from such ex-
periments, there is a need to integrate current knowledge
about individual neurons into a comprehensive picture of
how the network of neurons operates [2, 4, 13]. A num-
ber of studies have reported network characteristics of the
C. elegans connectome: it is a small-world network [14]
satisfying mathematical criteria of efficiency [15], with
a heavy-tailed degree distribution [16] and a core-set of
highly-connected, 'rich club' neurons [17]. Furthermore,
the analysis of modules in the network has shown that
certain strongly coupled clusters of neurons can be linked
to biological functions [18–22]. Such observations suggest
that a system-wide analysis of the connectome can pro-
vide valuable functional information. However, finding
simplified mesoscale descriptions that coherently aggre-
gate how information propagates in the directed connec-
tome across multiple scales remains a challenge [23].
In this work, rather than focusing on structural fea-
tures of the network, we analyse the directed and
weighted C. elegans connectome from a dynamics-based
(more specifically, flow-based) perspective. Using dy-
namics to probe the relationship between the structure
and function of a system has become a valuable tool in
many settings [23–25].
In particular, dynamics-based
approaches have been successfully used to study brain
networks (e.g., fMRI and DSI data [26–28]). For an in
depth discussion of network-theoretic methods in neu-
roscience see the extensive reviews [23, 29, 30]. For an
overview on dynamical methods for network analysis we
refer the reader to Refs. [24, 25, 31] and the literature
cited therein.
Our methods use diffusive processes on graphs as a
simple means to link features of the directed network
and propagation dynamics. While diffusive flow is a sim-
plification of the actual propagation in the nervous sys-
tem of C. elegans, we can still gain insight into network
properties of dynamical interest [4]. We exploit these
ideas in two ways. Firstly, we investigate flow-based
partitions of the connectome across multiple scales us-
ing the Markov Stability (MS) framework for community
detection [24, 31, 32]. Our analysis detects subgroups
of neurons that retain diffusive flows over particular time
scales [33] taking into account edge directionality [24, 34].
We then mimic neuronal ablations computationally, and
check all possible single and double ablations in the
connectome to detect those that are most disruptive of
the flow organisation. Secondly, we extract alternative
information of the directed network flows through the
Role Based Similarity (RBS) framework [35–37]. With-
out pre-imposing categories a priori, RBS classifies neu-
rons into flow roles, i.e., classes of neurons with similar
asymmetric patterns of incoming and outgoing network
flows at all scales, which are directly extracted from the
network. Finally, we mimic 'stimulus-response' experi-
ments [5, 7, 38], in which signals propagate through the
network starting from well-defined sets of input neurons
linked to particular biological stimuli. The ensuing time
courses of neuronal flows reveal features of information
processing in C. elegans, in relation to the obtained flow
roles. Our computational analyses are consistent with ex-
perimental findings, suggesting that our framework can
provide guidance towards the identification of potential
neuronal targets for further in vivo experiments.
II. RESULTS
Our analysis uses the C. elegans data published
in Ref. [4] (see www.wormatlas.org/neuronalwiring.
html). To represent the C. elegans connectome, we use
the two-dimensional network layout given by [4], i.e., neu-
rons are placed on the plane according to their normalised
Laplacian eigenvector (x-axis) and processing depth (y-
axis), as seen in Fig. 1 (top panel). We study the largest
weakly-connected component of this network, which con-
tains 279 neurons with 6394 chemical synapses (directed)
and 887 gap junctions (bidirectional). Reference [4] also
provides the position of the soma of each neuron along
the body of the worm, and classifies each neuron as either
sensory (S), interneuron (I) or motor (M).
A. Flow-based partitioning reveals multi-scale
organisation of the connectome
To reveal the multi-scale flow organisation of the C. el-
egans connectome, we use the Markov Stability (MS)
framework described in Sec. 'IV C'. Conceptually, MS
can be understood as follows. Imagine that a drop of ink
2
(signal) is placed on a node and begins to diffuse along
the edges of the graph. If the graph lacks structural or-
ganisation (e.g., random), the ink diffuses isotropically
and rapidly reaches its stationary distribution. However,
the graph might contain subgraphs in which the flow is
trapped for longer than expected, before diffusing out to-
wards stationarity. These groups of nodes constitute dy-
namical, flow-retaining communities in the graph, usually
signifying a strong dynamic coherence within the group
and a weaker coherence with the rest of the network. If
we allow the ink to diffuse just for a short time, then only
small communities are detected, for the diffusion cannot
explore the whole extent of the network. If we observe
the process for a longer time, the ink reaches larger parts
of the network, and the flow communities thus become
larger. By employing dynamics, and in particular by
scanning across time, MS can thus detect cohesive node
groupings at different levels of granularity [24, 33, 39].
In this sense, the time of the diffusion process, denoted
Markov time in the following, acts as a resolution param-
eter.
The flow-based community structure of the C. ele-
gans connectome at medium to coarse levels of resolu-
tion is shown in Fig. 1. The full scan across all Markov
times is shown in Fig. S1 and section SI1. As described
above, the partitions become coarser as the Markov time
t increases, from the finest possible partition, in which
each node forms its own community, to the dominant
bi-partition at long Markov times. The sequence of par-
titions exhibits an almost hierarchical structure, with a
strong spatial localisation linked to functional and organ-
isational circuits (see Fig. 2 and Fig. S2). These findings
are in agreement with the spatial localisation of func-
tional communities often found in brain networks [23], as
well as the hierarchical modularity exhibited by the C. el-
egans connectome as reported in Ref. [40]. We remark
that our community detection method does not enforce a
hierarchical agglomeration of communities: the observed
quasi-hierarchy and spatial localisation is an intrinsic fea-
ture of the C. elegans connectome. In Fig. S2 we quantify
the deviation of the community structure from a strict
hierarchy.
At long Markov times, we find robust partitions con-
taining 6 to 2 communities, denoted A to E in Fig. 1.
Partition A comprises six communities of varying sizes
(from 9 to 104 neurons), well localised along the body
of the worm, as seen in Fig. 2 (c.f.
Section 2.2 in
www.wormatlas.org/neuronalwiring.html). The two
large communities (A1 and A2) have head ganglia neu-
rons of all three functional types (S, I, M). In partic-
ular, A1 contains ring motor neurons and interneurons
as well as the posterior neurons ALN and PLN, whereas
A2 specifically gathers amphid neurons (e.g., AWAL/R,
ASKL/R, ASIL/R, AIYL/R) which feature prominently
in the navigation circuit responsible for exploratory be-
haviour [41]. Communities A3, A4 and A5 in Parti-
tion A consist predominantly of ventral cord motor neu-
rons, differentiated by their soma position along the body
3
Figure 1. Flow-based multiscale partitioning of the connectome of C. elegans. Using Markov Stability, we detect
flow-based partitions in this directed network at all scales. Here we show the medium to coarse Partitions A to E (top panel),
found as optimal at the indicated Markov time intervals (see Fig. S1 for the full sweep of Markov times). The Markov Time
intervals corresponding to different robust partitions are indicated by different colour boxes. Partitions A-E are persistent, as
signalled by their robustness over extended time plateaux in VI(t, t(cid:48)) (heatmap in bottom panel), and robust with respect to
the optimisation, as signalled by dips in the variation of information (cid:104)VI(t)(cid:105) (grey line in bottom panel).
(Fig. 2): A3 contains frontal motor neurons (e.g. VD1
to VD3); A4 consists of mid-body motor neurons (e.g.
VD4 to VD8); A5 comprises posterior motor neurons
(e.g. VD9 and VD10). Such partitioning is consistent
with the motor neuron segmentation model proposed for
C. elegans in Ref. [42]. Finally, A6 contains highly cen-
tral neurons such as AVAL/R or PVCL/R, which have
been found to belong to a rich-club [17], as well as in-
terneurons linked to mechanosensation and tap with-
drawal functional circuits [20].
The coarser Partitions B and C are quasi-hierarchical
merges of A (Figs. 1 and 2). For instance, Partition
C has three groupings: head ganglia (merged A1 and
A2), frontal motor neurons (merged A3 and A4), and
a tail subgroup (merged A5 and A6).
Interestingly,
at later Markov times, we obtain the distinct, coarser
3-community Partition D, which exemplifies how our
method does not enforce a strict hierarchy in the multi-
scale structure. The three groups in Partition D include
a notable community of only three nodes (interneurons
AVFL/R and AVHR), which appear as a cohesive group
only at this particular timescale. Prominent functional
roles of AVF and AVH neurons have been noted pre-
viously [4, 43]: both AVF neurons are responsible for
coordination of egg-laying and locomotion [44]. In addi-
tion, spectral analyses of the gap-junction Laplacian have
shown that AVF, AVH, PHB and C-type motor neurons
are strongly coupled [4]. Finally, the two communities in
the coarsest Partition E split the connectome anatomi-
cally into a group with head and tail ganglia (red), and
another group predominantly with motor neurons (cyan).
B. The effect of single and double neuron ablations
on flow-based communities
Laser ablation experiments are invaluable to probe the
functional role of neurons [5–7], but are time consuming
and technically challenging. We have used our compu-
tational framework to assess the effect that an ablation
of a single neuron, or of a pair of neurons, has on the
signal flow in the connectome. To this end, we compare
the flow-based partitions obtained for the ablated con-
nectome against the original network. If an ablation cre-
ates large distortions in the flow structure, the partitions
of the ablated network will change drastically or become
less robust compared to those found in the unablated net-
work. We have carried out a systematic computational
Number of Communities0.020.080.1426102.884.5722.91Markov Time,Markov Time,64332Partitions found at different resolutionsEigenvector 2Processing deptht=2.88t=4.57t=5.75t=16.60t=22.910.30.250.20.150.10.0510104536215.7516.60 4
Figure 2. Community structure and biological features. Left: As indicated by the dendrogram, the partitions obtained
have a quasi-hierarchical organisation. The dotted line indicates that the light green community at t = 16.60 does not result
from a hierarchical merging. Middle: The smoothed spatial densities of neurons in each community for the different partitions
show how the communities are spatially grouped according to soma positions along a longitudinal axis normalised between 0
and 1. The merging of groups over Markov time largely retains this spatial structure. Right: The percentages of sensory (S),
inter- (I) and motor (M) neurons in each community show functional segregation in the groupings.
analysis of all single and double neuron ablations in the
connectome.
1. Single ablations: disrupting the robustness and make-up
of partitions
a. Ablations that alter the robustness of parti-
tions: To find ablations that have a strong effect on the
robustness of Partitions A–E, we detect node deletions
that induce sustained changes in the robustness (cid:104)VI(t)(cid:105) ,
i.e., they appear as outliers with respect to a Gaussian
Process fitted to the (cid:104)VI(t)(cid:105) of the ensemble of all single
node ablations (Fig. 3). For details, see Section 'IV D'.
Only seven single ablations satisfy our criterion for a
major disruption of any of the Partitions A–E (Fig. 3b).
The ablations of interneuron PVCR or of the motor neu-
ron DD3 both decrease the robustness of Partition A. In-
terestingly, PVCR (A6) and DD3 (A4) receive many in-
coming connections from their own community. Further-
more, both of these neurons are critical for motor action:
PVCR drives motion whereas DD3 coordinates it. An-
other important ablation is that of interneuron AVKL,
which links community A1 (head) with community A3
(ventral cord) and community A6 (rich club). The in-
creased robustness of the community structure upon ab-
lation of AVKL would indicate a decreased communica-
tion between these groups. The function of AVKL is un-
charted at present [45], suggesting further in vivo experi-
mental investigations to explore any behavioural changes
as a result of its ablation.
There are three important ablations in Partition B:
DD3 (again), VD2 (another D-type motor neuron yet
on the ventral side), and AIBL, an amphid interneuron.
AIBL acts as a bridge between communities A1 and A2,
which merge in Partition B (Fig. 2). The prominent role
of other amphid interneurons will become apparent in
the double ablations studied in the next section.
Partition C is rendered non robust by the ablations
of VB8 (a motor neuron responsible for forward locomo-
tion) or of interneuron DVC, with are both in community
A5. DVC has links with communities A3, A4, A5 and
A6; hence its ablation affects the subsequent merging of
these groups. Note that the ablation of DVC reduces
the robustness of both 3-way Partitions C and D, thus
blurring the spatial organisation of motor neurons. This
indicates that DVC might integrate feedback from differ-
ent parts of the body, in accordance with the fact that it
has the highest number of gap junctions in the connec-
tome, as well as substantial chemical synapses [46].
Our study of ablations that affect the robustness of
partitions can be linked to the study of 'community
roles' [47]. Using such categorisation, the neurons men-
tioned above are classified as either connector or provin-
cial hubs (e.g., DVC is a 'non-hub' connector node) [20].
b. Ablations that alter the make-up of the op-
timal partitions: To measure how much the make-up
of a partition is affected by an ablation, we use the com-
munity variation CV, defined in Eq. (14). A high value of
CV[i](P) indicates a large disruption in partition P under
the ablation of neuron i. Figure 4 shows the single ab-
lations with high CV with respect to Partitions A-E, as
t=2.88t=4.57t=5.75t=16.60t=22.91Markov time1010123456SIMSIMSIMNeuronal densitySoma Position0 1 Soma Position0 1 Neuronal densitySIMSIMSIMSIMSIMSIM20%60%100%Soma Position0 1 Neuronal densitySIMSIM20%60%100%20%60%100% 5
Certain ablations are completely destructive of Parti-
tions A and B. In particular, the ablations of DD3 or
SMDDR induce severe changes in the network flow, so
that no partition similar to A is found at any Markov
time. In general, ablations of D-type motor neurons co-
ordinating motion (e.g. DD2, DD3, VD1, VD2) have par-
ticularly severe effects for the medium resolution Parti-
tions A and B. Interestingly, D-type motor neurons have
significantly higher PageRank (median 0.0092 compared
to median of 0.0018 in the network; p = 1.7 × 10−7;
one-sided exact test), and their synapses are critically
embedded edges with few alternative routes [43]. Note
that, although robustness and make-up of partitions re-
flect different effects, the ablation of motor neurons DD3
and VD2 substantially alters both (see Figs. 3 and 4a).
In addition, the ablation of any of the command neurons
AVAR/L has important effects on Partition C. AVAR/L
are highly central neurons (with the highest in- and out-
degree in the connectome) and our method confirms that
their ablation introduces heavy distortions in the global
flow of the connectome. Finally, we observe that the
coarsest partitions D and E are strongly perturbed upon
ablation of ring motor neuron RMDVL. Experiments
have shown that ablating any of the RMD neurons di-
minishes the head-withdrawal reflex [1].
Further confirmation of the importance of inter- and
motor neurons is given in Figure 4b, where we show the
CV of single ablations averaged over the three types (S, I,
M). On average, motor neurons tend to have a stronger
effect on local organisation due to their localised con-
nectivity; this is reflected by the high CV in the finer
Partitions A and B. On the other hand, interneurons,
which are mediators of information flow from sensory to
motor neurons, can induce large changes in global flows,
as shown by larger CV for the coarser Partitions D and
E.
2. Double ablations: beyond additive effects
We have also performed an exhaustive in silico explo-
ration of all possible 38781 two-neuron ablations. Specif-
ically, we look for synergistic pairs of neurons, i.e. pairs
whose simultaneous ablation induces supra-additive dis-
ruption. To this end, we compare the CV for each double
ablation to the averaged CV of the corresponding two
single ablations, and use Quantile Regression to identify
double ablations with a combined effect significantly be-
yond the merely additive (see Section 'IV D 2').
We focus on disruptions to Partitions A and D, as
prototypical of the medium and coarse resolutions, re-
spectively (Figure 5). We select the top 1% of ablations
for each partition according to their supra-additive ef-
fect. Interestingly, 85% of the top supra-additive double
ablations for Partition A contain at least one interneu-
ron, whereas 90% of the top supra-additive double abla-
tions for Partition D contain at least one motor neuron
(Fig. 5c-d). This observation complements the results for
Figure 3. Single ablations that alter the robustness
of partitions. (a) Ensemble of (cid:104)VI(t)(cid:105) profiles of all single
node ablations (light gray lines) and the unablated connec-
tome (blue). A Gaussian process (GP) is fitted to the ensem-
ble of single ablations. (b) The GP is described by the mean
µ(t) (dark grey line) and standard deviation (grey bands).
Sustained outliers from the GP are identified using a statis-
tical criterion to find seven ablations that affect the different
partitions, as indicated by the coloured dots.
detected through a statistical criterion based on interper-
centile ranges (see Section 'IV D'). Interestingly, none is
a sensory neuron, indicating that the ablation of sensory
neurons is not influential for global flow at medium to
coarse scales, although they can have strong local effect
on the propagation of a particular stimulus.
Markov time, tbPVCRDD3AIBLVD2VB8DVCAVKLa00.080.16110110Markov time, t 6
Figure 4. Effect of single ablations on the make-up of different partitions as measured by the community
variation. (a) The disruption of every single mutation with respect to Partitions A-E is quantified through CV[i](P), as
defined in Eq. (14). The distribution of CV[i] is represented by its median (red line) and the inter-percentile range (IPR)
between the 10th and 90th percentiles (box). The whiskers correspond to the IPR for each partition, and the single ablations
detected as outliers are labelled. (b) Effect of the single ablations CV[i](P) for each partition averaged over each type: sensory
(blue), inter- (red) and motor neurons (yellow). On average, single ablations of motor neurons induce larger changes on the
finer Partitions A and B, whereas ablations of interneurons have a larger effect on the coarser Partitions D and E. The error
bars are the standard error of the mean.
single ablations in Figure 4. For Partition A, maximal
impact of a single ablation is achieved through the dele-
tion of motor neurons, but double ablations containing
interneurons are more synergistic. For the coarser Par-
tition D, the most disruptive single ablations are those
of interneurons, yet on average the most synergistically
disruptive double ablations include motor neurons. Such
joint effects underline the structured complexity of the
connectome network and reinforce the fundamental im-
portance of I and M neurons in the disruption of flows. In
particular, the relative abundances of particular neurons
in the top supra-additive pairs (Fig. 5e-f) show that in-
terneurons AIAR/L, SAAVL and PVQR and motor neu-
rons RMDL/R are overly represented for Partition A.
These neurons thus have a magnifying disruptive effect
for the medium scales of the connectome. For the coarser
Partition D, this magnifying effect on larger scales is in-
duced mostly by motor neurons DD2, VD9, VD1 and
interneuron SAAVL.
If we consider the effect on both medium and large
scales, only nine double ablations appear in the top 1%
for both partition A and D (Table I). Interestingly, none
of these pairs is linked by an edge in the connectome.
Note that eight out of these nine pairs contain interneu-
ron AIAR. The amphid interneurons AIA (along with
AIB, AIY and AIZ) have a specific position in the connec-
tome: they receive synapses from sensory neurons driv-
ing motion. Their prominent role in locomotion integra-
tion has been previously discussed and backed by in vivo
ablation experiments [6]. Our results indicate that the
deletion of pairs of neurons involving AIAR would have
a particularly magnifying effect on the disruption of the
flow organisation at all scales in the connectome. Note
that the effect of AIAL in double ablations is much less
prominent. The asymmetry observed in how the abla-
tions of AIAR and AIAL affect the flows in the connec-
tome is worth of further experimental investigation. The
full set of outcomes of both single and double ablations
are presented in SI1 as a guide for possible experimental
investigations.
Table I. Double ablations within the top 1% of supra-additive
pairs for both Partition A and D chosen according to their
quantile scores (QA > 0.9835 and QD > 0.965).
Double ablation Neuron types QA
AIAR + AQR
AIAR + AVEL
AIAR + DA2
AIAR + VA2
AIAR + VB2
AIAR + VD5
AIAR + VD6
AIAR + PVCR
SAAVL + AQR
I & S
I & I
I & M
I & M
I & M
I & M
I & M
I & I
I & S
QD
0.9975 0.9785
0.9975 0.9650
0.9875 0.9785
0.9945 0.9655
0.9970 0.9785
0.9950 0.9785
0.9950 0.9785
0.9980 0.9785
0.9955 0.9730
7
tion and consumption of flow in the network. Briefly,
RBS obtains a flow profile for each node from its in-
coming and outgoing flows at all scales. We then group
the nodes into classes ('flow roles') with similar in- and
out-flow patterns. Because they include information at
all scales, flow roles capture nuanced information about
the network, beyond pre-defined categories (e.g., sources,
sinks, hubs) or combinatorial notions based on immedi-
ate neighbourhoods (e.g., roles from Structural Equiva-
lence [49] and Regular Equivalence [50]). Details of the
RBS methodology are given in Refs. [34–37], and sum-
marised in Section 'IV E' and in Fig. S4.
In the C. elegans connectome, we identify four distinct
classes of neurons according to their flow profiles (Fig. 6).
These flow roles are distinct from the groupings into com-
munities (see an analysis of communities and their mix
of flow roles in Fig. S5). Two of the roles (R1 and R2)
have a dominant 'source' character (i.e., higher average
in-degree than out-degree) and contain most of the nodes
with high PageRank (Fig. S6). The other two roles (R3
and R4) have a dominant 'sink' character and nodes with
low PageRank. Note, however, that these roles are not
just defined by average properties, but by their global
flow patterns in the network. As seen in Fig. 6b, R1 is up-
stream from R3 and R4, whereas R2 is mostly upstream
from R4. Furthermore, R4 is an almost pure downstream
module, whereas R3 has a stronger feedback connection
with R1.
The RBS flow roles are linked to physiological proper-
ties of the neurons (Fig. 6c-d). R4 corresponds to a group
of motor neurons (mostly ventral chord motor neurons)
consistent with its downstream character, whereas R1 is a
group of mostly sensory and inter-neurons with heavy lo-
calisation in the head. R3 is a group with a balanced rep-
resentation of all three types of neurons (including some
polymodal neurons) localised in the head. Indeed, most
ring neurons in R3 are in community A1, indicating a
self-contained unit that process head-specific behaviour,
such as foraging movements and the head withdrawal re-
flex [45].
Our RBS analysis also reveals a specific flow pro-
file (R2) containing 13 neurons (mainly sensory and in-
terneurons, mostly upstream from the motor neurons
in R4), the majority of which are responsible for es-
cape reflexes triggered in the presence of noxious fac-
tors (Table II). This group can be seen as a group of
escape response neurons and include: the PVDL/R neu-
rons, which sense cold temperatures and harsh touch
along the body; FLPL/R, which perform the equivalent
task for the anterior region; PHB neurons responsible
for chemorepulsion; PHCR, which detects noxiously high
temperatures in the tail; SDQL and PQR, which medi-
ate high oxygen and CO2 avoidance, respectively; and
PLMR, a touch mechanosensor in the tail [2]. This es-
cape response group is heavily over-connected to com-
mand neurons AVAL/R, AVDL/R, DVA, PVCL/R, all
of which modulate the locomotion of the worm. (Specifi-
cally, there are 48 connections from R2 to these particular
Figure 5. Supra-additive double ablations. The com-
bined effect of each two neuron ablation is compared against
the additive effect of the corresponding two single abla-
tions. The results of Quantile Regression of CV of the pair
against the averaged CV of the two single ablations (see Sec-
tion 'IV D 2') are shown for: (a) Partition A and (b) Par-
tition D. The top 1% pairs with the largest supra-additive
effect are found above the quantile scores QA > 0.9835 and
QD > 0.965, respectively. These top 1% double ablations are
dominated by: (c) interneurons for A; (d) motor neurons
for D. Overrepresentation of neurons in the top 1% supra-
additive pairs for (e) Partition A and (f ) Partition D was
calculated using a one-sided Fisher exact test (unadjusted p-
values are reported, and also provided for all neurons in sec-
tion SI1). Neurons with p < 10−5 are listed and the names
of neurons are coloured according to their type: S (blue),
I (red), M (yellow). The word clouds are a visualisation of
these over-representations. Computing a Bayesian quantile of
higher prevalence of these neurons among the top 1% pairs
also supports these findings [48].
C.
Identifying flow profiles in the directed
connectome
A complementary analysis of the directed connectome
of C. elegans is provided by the Role Based Similar-
ity (RBS) framework [35, 36], which identifies groups of
nodes with similar flow profiles in the network without
imposing a priori the type or number of groups. Such
groups of neurons display the same character (or flow
role) in terms of their role in the generation, distribu-
abcd Partition00.040.080.12Q>0.9835Q= 0.75Q= 0.8Q= 0.85Q= 0.9Q= 0.950.00.020.040.060.08 Partition0.020.040.060.08Q= 0.75Q= 0.85Q= 0.9Q= 0.95Q>0.9650.00.010.020.030.04ef01002003001e-2923e-574e-392e-251e-178e-112e-76e-6p-value3e-3154e-231e-121e-113e-83e-82e-72e-7p-value6e-66e-6Occurrences in top 1% double ablations for Occurrences in top 1% double ablations for 389DD2VD9VD1SAAVLPHARVD6VD5VA9PDERRIMLSAAVLAIARRMDLRMDRPVQRDVCRMEDAIAL10%20%30%40%S&II&II&MM&MM&SS&S10%20%30%40%S&II&II&MM&MM&SS&S0100200300389 8
Figure 6. Flow roles for neurons in the C. elegans connectome. (a) Using RBS, we detect four flow roles in the directed
connectome. (b) The coarse-grained representation summarises the flow profiles of the roles: two upstream roles (R1, R2), with
a dominant source character and high PageRank (Fig. S6), and two downstream roles (R3, R4), with a dominant sink character
and lower PageRank. Yet each role has distinctive in- and out-flow patterns in relation to the others. (c) Spatial density of
neurons for each flow role represented as a function of the normalised soma position: R1 and R3 are localised predominantly
in the head; R2 and R4 are spread out along the body. Note how the upstream flow role R2 has noticeable localisation in the
tail. (d) The percentages of sensory (S), inter- (I) and motor neurons (M) in each role underline their functional differences.
dRole 1Role 3Role 2Role 4aR1R2R3R4cb00.20.40.60.81Soma PositionNeuronal densityRole 1Role 3Role 2Role 440%S I MS I MS I MS I M20%80%60%R1R2R3R4 command neurons in contrast to the ∼12 connections ex-
pected at random.) Note that AVDL/R and DVA are in
R1, whereas AVAL/R and PVCL/R are in R4; the R2
group thus links directly to motor locomotion neurons
across the worm. We remark that this group of neu-
rons was found exclusively through the analysis of their
all-scale in/out flow profiles, without any other extrinsic
information.
Table II. Role 2 (R2) neurons: The thirteen neurons iden-
tified in R2 constitute a group of escape response neurons
containing mostly sensory and inter-neurons linked with es-
cape reflex reactions in response to different noxious stimuli.
Neuron(s) Noxious factor
FLPL/R
PHBL/R
PHCR
PLMR
PQR
PVDL/R
SDQL
SAAVL/R No known factor
VD11
No known factor
Harsh touch, low temperature (head)
Chemicals
High temperature (tail)
Gentle touch (tail)
CO2
Harsh touch, low temperature
High O2
D.
Information propagation in the connectome:
biological input scenarios
Despite its modest size, the nervous system of C. ele-
gans can sense and react to a wide range of mechanical,
chemical and thermal factors [45]. Standard notions in
neuroscience hold that stimuli lead to motor action due
to information progressing from sensory through inter- to
motor neurons [51]. However, the underlying mechanisms
and precise signal flows are still far from understood. In
the absence of measurements probing such pathways, and
as a first approximation to more realistic nonlinear dy-
namical models, we use here simplified diffusive dynamics
(see Section 'IV B') to mimic signal propagation in the
C. elegans directed network. Such an approach, already
suggested by Varshney et al. [4], is naturally linked to MS
multiscale community detection and to the identification
of RBS flow roles, since both Markov Stability and Role
Based Similarity are intrinsically defined in terms of a
diffusive process on the graph.
To mimic the propagation of stimuli associated with
particular biological scenarios, a normalised initial flow
vector φ(0) is localised at specific input neurons and we
observe the decay towards stationarity under Eq. (5):
θ(t) = φ(t) − π.
(1)
We also define q(t), which will be used to detect over-
shooting neurons:
9
Initially, θi(0) is positive only for the input neurons where
we inject the signal, and negative for all other neurons.
Asymptotically, the vector of flows φ(t) approaches the
stationary solution π, and θi(t) → 0,∀i. However the
approach to the stationary value can be qualitatively dif-
ferent. In some cases, θi(t) can become positive, if neuron
i receives an influx of flow that drives it to 'overshoot'
above its stationary value; in other cases, neurons ap-
proach stationarity without overshooting. The different
behaviour depends on the particular initial input and the
relative location of each neuron in the network.
Motivated by several experimental studies, we have
conducted four case studies corresponding to different bi-
ological scenarios in which the input is localised on spe-
cific neurons:
(i1) Posterior (tail) mechanosensory stimulus [5, 7]:
PLML/R, PVDL/R, PDEL/R
(i2) Anterior (head) mechanosensory stimulus [5, 7]:
ADEL/R, ALML/R, AQR, AVM, BDUR/L,
FLPL/R, SIADL/R
(i3) Posterior (tail) chemosensory stimulus (also re-
ported as anus mechanosensory stimulus) [7, 38]:
PHAL/R, PHB/R
(i4) Anterior
(head)
chemosensory stimulus
[38]:
ADLL/R, ASHL/R, ASKL/R.
We exemplify the procedure in detail through the pos-
terior mechanosensory stimulus (i1), but detailed results
for the other stimuli are provided in the Fig. S8, Fig. S9,
and Fig. S10. As shown in Figure 7a, the signal proceeds
'downstream' following the expected biological informa-
tion processing sequence, S→I→M. The signal is initially
concentrated on the input neurons (mostly sensory); then
propagates out primarily to interneurons, which over-
shoot and peak at t ≈ 1.5; and is then passed on to motor
neurons, which slowly increase towards their stationary
value.
The flow roles obtained above provide further insight
into the propagation of stimuli. As seen in Figure 7c, the
input for the tail mechanosensory scenario (i1) is heavily
concentrated on R2 neurons (the escape response group),
from which the signal flows quickly towards the other up-
stream (head) group R1, followed by propagation towards
the downstream group R4. Finally, the signal spreads
more slowly to R3, the head-centric downstream unit.
This pattern of propagation carries onto the sequence of
strong response neurons (Figure 7b), and reflects the fact
that R2 contains posterior upstream units, and mirrors
the strong connectivity of R2 with motor neurons in R1
(AVDL/R and DVA) and R4 (PVCL/R), as discussed
above.
To detect key neurons comprising the specific propa-
gation pathways, we find strong response neurons, i.e.,
those with large overshoots relative to their stationary
value,
qi(t) =
φi(t)
πi
= 1 +
θi(t)
πi
.
(2)
qmax,i = max
t
qi(t) > 1 +
2
3
.
10
Figure 7. Signal propagation of posterior mechanosensory stimulus (i1). (a) As stationarity is approached (θ(t) → 0),
the input propagates from sensory to motor neurons through an intermediate stage when interneurons overshoot. (b) Signal
propagation as a cascade of strong response neurons (32 neurons with qmax,i > 1 + 2/3) with peak times concentrated around
two bursts. The number of neurons are colored according to type (top) and role (bottom). Note the overall trend S → I → M
during the propagation of strong responses, and how the sequence of strong response neurons also reflects the connectivity
between roles propagating roughly from R2 to R1 and finally to R3. (c) The input (i1), which is highly localised on R2
neurons, diffuses quickly to R1 neurons and induces an overshoot of R4 neurons followed by slower diffusion into R3 neurons.
(d) Stages of signal propagation in the network showing the strong response neurons that have peaked at each time.
See Fig. S7 for a full description of the procedure. Ac-
cording to this criterion, we obtain 26 strong response
neurons for scenario (i1). The neurons have large over-
shoots in two time windows after the inital input (Fig-
ure 7b). The details of the signal propagation (Figure 7d)
show that a first wave of peak responses (around t ≈ 1)
corresponds mostly to overshooting interneurons, includ-
ing AVDL/R and DVA, responsible for mechanosen-
sory integration, and PVCL/R, drivers of forward mo-
tion [5, 45]. The second wave of peaks (around t ≈ 3)
contains predominantly ventral B-type motor neurons,
e.g., DB2-7 and VB11.
Such B-type motor neurons
are responsible for forward motion. Hence the progres-
sion of overshooting neurons suggests a plausible bio-
logical response for a posterior mechanosensory stimu-
lus [7, 45]. The overshooting behaviour of the neurons
is not captured by other static measures of the network
(e.g., in/out degree or pagerank), as shown in S12.
1. Comparison with other biological scenarios
Detailed results of propagation under the other biolog-
ical scenarios (i2)-(i4) from the experimental literature
are presented in Fig. S8, Fig. S9, and Fig. S10. The
overall progression of the signal from S to I to M is ob-
served with small differences in all scenarios. However,
the different scenarios exhibit distinctive participation of
the flow roles. In particular, both posterior stimuli (i1)
and (i3) spread from R2 neurons quickly into R1 neu-
rons and R4 (motor) neurons, with weak propagation
into R3 neurons . On the other hand, anterior stimuli
(i2) and (i4) spread from the R1 group strongly into R3
neurons and also quickly to R2 neurons, with only weak
spreading into R4 neurons. In cases (i1)-(i3) information
flows fast out of R2 towards motor neurons, as could be
expected from neurons triggering an escape response. In-
terestingly, the (i4) scenario does not feature any strong
response neurons in the R2 group.
As shown in Fig. S8 - Fig. S10, and summarised in
Fig. 8, the signal propagation pathways have distinc-
tive characteristics for each of the scenarios. For in-
stance, although the posterior chemosensory scenario
(i3) shows strong similarities to (i1) at earlier stages
(input mostly R2 and strongly responding interneurons
PVCL/R, AVDL/R, AVJL, DVA), they show differences
in the motor neurons exhibiting a strong overshoot. In
particular, for (i3) A-type neurons (DA8, DA9, VA12)
responsible for backward motion are present in addition
to B-type neurons (DB2, DB3, DB7).
The anterior (head) scenarios (i2) and (i4) inputs show
a localised propagation mostly in head-centric groups R1
and R3. For the anterior mechanosensory scenario (i2),
command interneurons such as PVCL/R, AVDL/R re-
dSIM10-310-210-11001011020.60.40.20-0.2-0.4-0.6Markov Timea10-310-210-11001011020.50.40.30.20.10-0.1-0.2-0.3-0.4Markov TimeRole 1Role 2Role 3Role 4ct=1.778PVDRPVDLPLMRPDERPLMLPDELPVMHSNLAVKLPVRHSNRDVALUARPHCLLUALPVCRAVDRVD11PVCLt=3.5481PVDRPVDLPLMRPDERPLMLPDELPVMHSNLRIRSMBVLSMBDLAVKLPVRHSNRDVALUARPHCLLUALPVCRAVDRAVJLAVDLVD11PVCLVB11DB2DB4DB3DB7DA9DB5DB6bt=0.001PVDRPVDLPLMRPDERPLMLPDEL02.48Strong response neuronsPeak timeStrong response neurons Input neuronsR1R2R3R4 11
Figure 8. Summary of signal propagation in the four biological scenarios. The specific pathways for the signal
propagation for each of the scenarios (i1)-(i4) are shown, highlighting the input, strong response (qmax,i > 5/3) and overshooting
neurons (qmax,i > 1). The input and strong response neurons are labelled for each biological scenario.
spond strongly, together with ring interneurons, such as
RIGL/R and RIBL/R. In this case, only small excitation
of ventral cord motor neurons is attained. Instead, we
observe strong responses of polymodal ring motor neu-
rons, such as URADL/R and SIADL/R, and of sensory
neurons CEPVL/R and CEPDL/R, even though these
CEP neurons receive no external input.
Interestingly,
CEP neurons are reported [45] to be functionally re-
dundant with nose touch receptors ADE, where the in-
put signal is located. Upon anterior chemical stimula-
tion (i4), a bulk of flow is captured within the neuronal
ring and induces strong response from chemosensory neu-
rons such as PVQL, ASKL, AWAL/R, AWABL/R and
AWACL/R, as well as interneurons RICL/R, RMGL/R,
AIAL/R and AIBL/R, which are specific for integrating
chemo-sensation.
Indeed, several of these neurons also
appear in the posterior chemosensory stimulus (i3). A
summary of the strong response and overshooting neu-
rons for all scenarios is presented in Figure 8.
III. DISCUSSION
We have presented an integrated network-theoretic
analysis of the C. elegans connectome in terms of directed
flows. We exploit the connection between diffusive pro-
cesses and graph-theoretical properties, which intimately
links structure and dynamics, to elucidate dynamically
relevant features in the connectome. Although diffusive
processes are a coarse approximation of physiological sig-
nal propagation, they can be used to extract systemic dy-
namical features, specifically in the case of non-spiking
neuronal systems such as C. elegans [4].
Using the Markov Stability (MS) framework, we have
identified flow-based groupings of neurons in the C. ele-
gans connectome at different levels of granularity. Pre-
vious studies [20–22] have aimed at uncovering modules
based on structural properties of the network, usually
considering a particular scale so as to find one partition
(e.g., modularity at the standard resolution).
In sec-
tion SI2 we provide a detailed comparison of MS multi-
scale flow structures against partitions found by modular-
ity [20, 21], stochastic block models [22] and the MapE-
02.13Chemical stimulusHeadMechanical stimulusTail02.4802.31(i1)(i2)(i3)(i4)02.47Strong response neurons Input neuronsOvershooting neuronsIL2LURADLOLLLURADRBAGLCEPVLBAGRCEPVRCEPDLALACEPDRRIHAVHLRIBLSIADLAVJLRIBRAVDRAVDLSIADRFLPRAQRADELADERRIGRRIGLBDURBDULAVMALMLALMRHSNRHSNLSDQLPVCLPVCRPVRPVNRPVNLAVJRFLPLCEPVRALAASKLAFDLASKRADLLADFLAFDRAWBLAWCLADFRASGLADLRAWALASILASHLASGRAIBLASHRAWCRAWARASIRAIBRAVJRASERRIMLAVDRAVDLRICLAIARAIZRRICRAIZLAIYRAIMRAIMLAIYLRIFRADARADALRMGLRMGRHSNLPVQRPVQLAIALPVDRPVDLPLMRPDERPLMLPDELPVMHSNLRIRSMBVLSMBDLAVKLPVRHSNRDVALUARPHCLLUALPVCRAVDRAVJLAVDLVD11PVCLVB11DB2DB4DB3DB7DA9DB5DB6RIDAVALASKLAVARAVHLAVHRAVJRAVJLAIALAVDRAVDLAVFRAVFLDB02AVGDB03DB07VA12PVPLDA08DA09DVAPHALPHARPVQLPVCLPHBRPHBLPVCR quation [52]. The partitions obtained by MS at a par-
ticular scale are closer to those obtained with directed
modularity. The MS framework, however, provides a
multiscale description across all scales by sweeping the
Markov time [33], respecting and exploiting directional-
ity. In doing so, it reveals an intrinsic, quasi-hierarchical
organisation of the connectome, giving insight into rele-
vant features of signal propagation. The partitions found
by MS are in good agreement with C. elegans physiology,
and summarise previously observed features, such as the
hierarchical and spatial organisation of neuronal commu-
nities [23, 40].
The obtained flow-based organisation highlights the
prominent position of particular neurons, such as AVF
and AVH, and allows for a systematic exploration of sin-
gle and double ablations most disruptive of signal flows,
thus providing insight into candidate neurons for fur-
ther experimental investigations. Examples of such neu-
rons include, among others: the synergistic effects caused
by neuron AIAR in double ablations; the global role
of D-type motor neurons, which often appear as rele-
vant in single ablations; or the role of polymodal (I/M)
SAAVL/R head neurons [45], about which little is known
but which appear in the R2 group and are salient in
our ablations. Several other examples are discussed in
the text, and further such hypotheses may be formulated
based on the full set of ablation scores we provide in sec-
tion SI1 as a resource to experimentalists investigating
the physiology of particular neurons.
Other methods can be used to study the effect of abla-
tions using, for example, measures of centrality, efficiency
or information transfer [53, 54]. Our study of ablations
gives distinct results, as shown in Fig. S3. For instance,
because our measures focus on the disruption of the flow
community structure at different scales, our approach can
provide a structured view of the effect of ablations for dif-
ferent neuron types, as shown in Figs. 3–5.
As a complementary flow-based perspective, we have
used Role Based Similarity to identify classes of neurons
with similar patterns of flow in the C. elegans nervous
system. Rather than reflecting any measure of connect-
edness in the network, such flow roles (or flow profiles)
reflect similar roles in the generation, distribution and
consumption of flow in the directed connectome. In pre-
vious work, neurons have been assigned to roles by ex-
ploring the core-periphery structure [55], or by examining
the connections of nodes within and between communi-
ties [47, 56]. Other notions of roles have been based on
the use of centrality scores, or on combinatorial notions
of social neighbourhoods, as in regular and structural
equivalence [49, 50]. RBS takes a different approach by
grouping neurons according to their patterns of in/out
flows at multiple scales in the graph, irrespective of their
community membership and going beyond standard clas-
sifications [34, 37]. See SI3 and Fig. S6 for a comparison
of RBS flow roles, regular equivalence and community
roles.
The RBS analysis of flow profiles finds two groups of
12
mostly upstream neurons and two groups of mostly down-
stream neurons, yet with a specific inter-connectivity pat-
tern. In particular, the analysis singles out a small group
of upstream neurons (R2), which is functionally related
to escape responses from noxious factors, and could also
be the object of further experimental investigation. The
RBS roles are also informative in conjunction with signal
propagation from 'input-response' in silico biological sce-
narios (see Fig. S11). In particular, the R2 group plays an
important role in posterior biological stimuli, channelling
stronger and faster responses, whereas R3 (the down-
stream, head-centric group) constitutes a self-contained
set of neurons mainly accessible via the upstream, head-
centric R1 group. Therefore, the propagation profiles ob-
tained for different biological scenarios suggest a graded
organisation of the roles of nodes in terms of upstream-
downstream information, which could provide valuable
insight into functional circuits.
Interesting theoretical extensions of the current work
would include considering the C. elegans connectome as
a multiplex network; taking into account the different
types of synapse in a more explicit fashion; and enriching
the dynamics of the model by incorporating the effects of
inhibitory synapses and nonlinearities in the dynamics.
Furthermore, one may explore more intricate dynamics
by incorporating the memory of information flow using
higher order Markov models [57, 58].
Our computational tools could be used in conjunction
with experimental techniques, as an aid to the genera-
tion of functional hypotheses for experimental evaluation.
With the eventual aim of linking wiring properties of the
connectome with information processing and functional
behaviour, high throughput experiments (e.g., systematic
ablation of several neurons) coupled with advancements
in neuronal monitoring that can allow recordings from
thousands of neurons simultaneously [59] could deliver
time course measurements to characterise signal propa-
gation in relation to function. Another interesting area of
future work would be the evaluation of ablation and prop-
agation scenarios as related to quantitative behavioural
investigations upon more general ablational/mutational
strategies in C. elegans [10–12], as well as compara-
tive studies of the flow architecture in different nema-
tode species [60]. Such comparative analyses between
the functional and structural network of the connectome
could yield valuable information in bridging the relation
between structure and function in network neuroscience.
IV. METHODS
A. The C. elegans neuronal network
The information of the large component of the connec-
tome network is encoded into the n× n adjacency matrix
A (n = 279), where entry Aij counts the total number
of synapses (both chemical synapses and gap junctions)
connecting neuron i to neuron j [4]. Note that chemical
synapses are not necessarily reciprocal, hence A (cid:54)= AT .
Therefore the connectome is a directed, weighted network.
The network is relatively sparse, with 2990 edges: 796
edges formed by gap junctions only; 1962 containing only
chemical synapses; 232 edges with both gap junctions and
chemical synapses present. The vector of out-strengths,
which compiles the sum of all synapses for each neuron,
is d = A1 (where 1 is the n × 1 vector of ones). The
average out-strength per neuron is 29; ranging from the
maximum (256) attained by neuron AVAL to the min-
imum (0) attained by the motor neuron DD6, which is
the only sink in the network. The network is not strongly
connected.
B. Propagation dynamics in the network
Methods with different levels of complexity have been
used to study signal propagation in the C. elegans con-
nectome (see, e.g., Refs.
[4, 51, 61–63]). Here, we use
a continuous-time diffusion process as a simple proxy
for the spread of information in this neuronal network.
Note that gap junctions may be simply modelled as lin-
ear resistors and, although chemical synapses are likely
to introduce nonlinearities, their sigmoidal transfer func-
tions may be well approximated by a linearisation around
their operating point.
Indeed, as remarked by Varsh-
ney et al. [4], such an approach has additional merit in
C. elegans, where neurons do not fire action potentials
and have chemical synapses that release neurotransmit-
ters tonically [64]. Thus, linear systems analysis is in
this case an appropriate tool that can provide valuable
insights [4]. Interestingly, athough simplified, such linear
models have been successfully applied even to the anal-
ysis of spatio-temporal behaviour of strongly nonlinear
neuronal networks [65].
The signal on the nodes at time t is represented by
the 1 × n row vector φ(t) governed by the differential
equation
= φ [M − I] ,
dφ
dt
(3)
where I is the identity matrix and M is the transition
matrix defined as follows:
M = τ D†A +
[(1 − τ ) 1 + τ 1di=0] 1T .
(4)
Here, τ ∈ (0, 1) is the Google teleportation parameter
(and we take τ = 0.85 as is customary in the literature);
1di=0 is the indicator vector of sink nodes; and the diago-
nal matrix D† is the pseudo-inverse of the degree matrix:
1
n
(cid:40)
†
ii =
D
0
1/di
if di = 0
if di (cid:54)= 0.
The matrix M describes a signal diffusion along the di-
rected edges with an additional re-injection of external
'environmental noise': each node receives inputs from
13
its neighbours (which transmit flow along their outgo-
ing links according to their relative weight with proba-
bility τ ) and receives a constant external re-injection of
size (1 − τ )/n. For pure sinks, the outgoing flow is uni-
formly redistributed to all nodes so as to avoid the signal
accumulating at nodes with no out-links. Mathemati-
cally, this reinjection of probability (known as teleporta-
tion in the networks literature) guarantees the existence
of a unique stationary solution for Eq. (3), even when
the network is not strongly connected [24, 66]. Biophys-
ically, the teleportation can be understood as modelling
the random interactions with the external environment.
Let φ(0) be the input, i.e., the signal at t = 0. The
solution of Eq. (3) is then:
φ(t) = φ(0) exp (t [M − I]) ,
(5)
with stationary solution φ(t → ∞) = (φ(0) · 1) π,
where π is the dominant left eigenvector of M , known
as PageRank [66]. Therefore, under a unit-normalised
input, φ(t) · 1 = 1 ∀ t, and the stationary solution is π.
C. A dynamical perspective for community
detection in graphs: Markov Stability
The diffusive dynamics (3) can be exploited to reveal
the multiscale organisation of the C. elegans connectome
using the Markov Stability community detection frame-
work [24, 31, 32]. Markov Stability finds communities
across scales by optimising a cost function related to
this diffusion (parametrically dependent on time) over
the space of all partitions.
More formally, a partition P of the n nodes of the
network into m non-overlapping communities is encoded
as a n × m indicator matrix HP :
(cid:40)
[HP ]ic =
1 if node i belongs to community c
0 otherwise.
(6)
Given a partition matrix HP , we define the time-
dependent clustered autocovariance matrix :
R(t, HP ) = HP T(cid:2)Π exp(t[M − I]) − ππT(cid:3) HP ,
(7)
where Π = diag(π). The matrix entry [R(t, HP )]cf quan-
tifies how likely it is that a random walker starting in
community c will end in community f at time t, minus
the probability for such an event to happen by chance.
To find groups of nodes where flows are trapped more
strongly over time t than one would expect at random,
we find a partition P that maximises
r(t, HP ) = trace R(t, HP ).
(8)
We define r(t, HP ) as the Markov Stability of partition
P at time t [24, 32].
Maximising r(t, HP ) over the space of all partitions for
each time t results in the sequence of optimal partitions:
Pmax(t) = arg max
P
r(t, HP ).
(9)
Although the optimisation (9) is NP-hard, there exist
efficient heuristic algorithms that work well in practice.
In particular, it has been shown that this optimisation
can be carried out using any algorithm devised for mod-
ularity maximisation [24, 31, 32]. In this work, we use
the Louvain algorithm [67], which is known to offer high
quality solutions whilst remaining computationally effi-
cient. The code for Markov Stability can be found at
github.com/michaelschaub/PartitionStability.
As an additional improvement of the optimisation of
Pmax(t), we run the Louvain algorithm (cid:96) = 100 times
with different random initialisations for each Markov
time t, and generate an ensemble of solutions {Pi(t)}(cid:96)
i=1.
From this ensemble, we pick the best partition (cid:98)P(t) ac-
cording to our measure (8):
i=1 (cid:55)−→ (cid:98)P(t) ≈ Pmax(t).
max
{Pi(t)}(cid:96)
i
Ideally, the optimised partition from the ensemble, (cid:98)P(t),
will be close to the true optimum, Pmax(t).
To identify the important partitions across time, we
use the following two robustness criteria [33, 68]:
a. Consistency of the optimised partition: A
relevant partition should be a robust outcome of the op-
timisation,
i.e., the ensemble of (cid:96) optimised solutions
should be similar. To assess this consistency, we em-
ploy an information-theoretical distance between parti-
tions: the normalised variation of information between
two partitions P and P(cid:48) defined as [69]:
2Ω(P,P(cid:48)) − Ω(P) − Ω(P(cid:48))
VI(P,P(cid:48)) =
,
log(n)
where Ω(P) = −(cid:80)C p(C) log p(C) is a Shannon entropy,
with p(C) given by the relative frequency of finding a node
in community C in partition P; Ω(P,P(cid:48)) is the Shannon
entropy of the joint probability; and the factor log(n)
ensures that the measure is normalised between [0, 1].
(10)
To quantify the robustness to the optimisation, we
compute the average variation of information of the en-
semble of solutions obtained from the (cid:96) Louvain runs at
Markov time t:
(cid:104)VI(t)(cid:105) =
1
(cid:96)((cid:96) − 1)
VI(Pi(t),Pj(t)).
(11)
(cid:88)
i(cid:54)=j
If all runs of the optimisation return very similar parti-
tions, then (cid:104)VI(t)(cid:105) will be small, indicating robustness of
the partition to the optimisation. Hence we select parti-
tions with low values (or dips) of (cid:104)VI(t)(cid:105) .
b. Persistence of the partition over time: Rel-
evant partitions should also be optimal across stretches
of Markov time. Such persistence is indicated both by a
plateau in the number of communities over time and a
low value plateau of the cross-time variation of informa-
tion:
VI(t, t(cid:48)) = VI((cid:98)P(t),(cid:98)P(t(cid:48))).
(12)
14
Therefore, within a time-block of persistent partitions
we choose the most robust partition, i.e., that with lowest
(cid:104)VI(t)(cid:105) .
D. Quantifying the disruption of community
structure under node deletion
To mimic in silico the ablation of neuron i, we remove
the i-th row and column of the adjacency matrix A, and
analyse the change induced in the Markov Stability com-
munity structure of the reduced (n − 1) × (n − 1) matrix
A[i]. Double ablations are mimicked by simultaneously
removing two rows (and their corresponding columns) to
obtain the reduced (n − 2) × (n − 2) matrix A[i,j].
1. Detecting salient single-node deletions
We carry out a systematic study of all single node dele-
tions in the network. To detect relevant deletions, we
monitor either an induced loss of robustness or an in-
duced disruption in the make-up of particular partitions.
a. Changes induced in the robustness of par-
titions: First, we run the MS analysis on all deletions
to obtain the optimised partitions and their robustness
across all times t:(cid:110)(cid:98)P[i](t), (cid:104)VI[i](t)(cid:105)(cid:111)n
∀ t.
(13)
i=1
We then fit a Gaussian Process (GP) [70] to the ensemble
of n + 1 time series of the robustness measure (cid:104)VI[i](t)(cid:105),
plus the unablated (cid:104)VI(t)(cid:105) . The resulting GP, with mean
µ(t) and variance σ2(t), describes the average robustness
of partitions under a single-node deletion.
that disrupt the robustness of a partition over its epoch.
b. Changes induced in the make-up of parti-
tions: To detect if the deletion of node i induces a
To detect single-node deletions that induce a large
change in the robustness of a given partition we find
mal over t ∈ [t1, t2], we select node deletions i such that
(cid:104)VI[i](t)(cid:105) differs from µ(t) by at least two standard de-
viations σ(t) over a continuous time interval larger than
sustained outliers of the GP. For a partition (cid:98)P opti-
ln((cid:112)t2/t1) [68]. This criterion identifies node deletions
change in the make-up of partition (cid:98)P, we compute the
i.e., the variation of information between (cid:98)P and the most
work (cid:98)P[i](t).
CV[i]((cid:98)P) = min
similar among all optimal partitions of the ablated net-
VI((cid:98)P,(cid:98)P[i](τ )),
community variation:
(14)
τ
We detect outliers in CV for each partition using
a simple criterion based on the inter-percentile range:
the deletion of i is considered an outlier if CV[i] >
P90 + IPR90/10, where P90 is the 90th percentile, and
IPR90/10 = P90 − P10 is the interpercentile range be-
tween the 10th-90th percentiles of the ensemble of CV[i].
2. Detecting supra-additive double-node deletions
We have carried out a study of all double deletions in
the network to detect two-node deletions whose effect is
larger than the additive effect of the two corresponding
single node deletions. To this end, we first obtain the set
of MS partitions across all Markov times for all double
delections (cid:98)P[i,j](t), and compute their community varia-
tion:
We then compute the average of the individual ablations:
CV[i,j]((cid:98)P) = min
CV[i],[j]((cid:98)P) =
VI((cid:98)P,(cid:98)P[i,j](τ )).
CV[i]((cid:98)P) + CV[j]((cid:98)P)
τ
2
.
(16)
To find pairs with a supra-additive effect, we use Quan-
tile Regression (QR) [71], a method widely used in econo-
metrics, ecology, and medical statistics. Whereas least
squares regression aims to estimate the conditional mean
of the samples, QR provides a method to estimate con-
ditional quantiles of the sample distribution. Hence, QR
facilitates a more global representation of the relation-
ships between the dependent and independent variables
considered in the regression. A good introduction to QR
can be found in Ref. [72], and a more in-depth treatment
can be found in the book by Koenker [71].
For a partition (cid:98)P, we employ QR to fit quantiles
for the regression of CV[i,j]((cid:98)P) against CV[i],[j]((cid:98)P), us-
ing all 38781 two-node ablations (Fig. 5). We report
the top 1% double deletions according to their quantile
scores-this is our criterion to select double-ablations
that have a strong effect. All scores are computed
using Bayesian Quantile Regression, as implemented
in the R package BSquare (https://cran.r-project.
org/web/packages/BSquare/index.html), which fits
all quantiles simultaneously resulting in a more coherent
estimate [73]. Following Ref. [73], we fit the quantiles
to the normalised CV[i],[j]((cid:98)P) using a Gamma centering
distribution and four basis functions.
15
a weighted number of in- and out-paths of increasing
lengths beginning and ending at the node. The feature
vectors are collected in the feature matrix X:
(cid:123)
(cid:122)
(cid:2). . . (βAT )k1 . . .
(cid:125)(cid:124)
paths in
(cid:123)
(cid:122)
. . . (βA)k1 . . .(cid:3),
(cid:125)(cid:124)
paths out
(17)
=
x1
...
xn
X =
(15)
where β = α/λ1, with λ1 the spectral radius of the ad-
jacency matrix A and α ∈ (0, 1). The cosine between
feature vectors gives the similarity score between nodes:
Yij =
xixT
j
(cid:107)xi(cid:107)2 (cid:107)xj(cid:107)2
.
(18)
The n × n matrix Y quantifies how similar the directed
flow profiles between every pair of nodes are. Nodes with
identical connectivity have Yij = 1, whereas in the case
of nodes with dissimilar flow profiles (e.g., if i is a source
node with no incoming connections and j is a sink node
with no outgoing connections), then their feature vectors
are orthogonal and Yij = 0.
As outlined in Refs. [35–37], we compute the similarity
matrix Y iteratively with α = 0.95, and apply the RMST
algorithm to obtain a similarity graph, in which only the
important information of Y is retained. We then extract
flow roles in a data-driven manner without imposing the
number of roles a priori by clustering the similarity graph
(see Fig. S4). The flow roles so obtained have been shown
to capture relevant features in complex networks, where
other role classifications based on combinatorial concepts
and neighbourhoods fail [34, 37]. In particular, our flow
roles are fundamentally different from notions of roles in
social networks based on Structural Equivalence [49] and
Regular Equivalence [50]. Such equivalence measures do
not incorporate information about the large scales of the
network and are sensitive to small perturbations, mak-
ing them unsuitable for complex networks such as the
C. elegans connectome [34] (see Fig. S6 for roles based
on Regular Equivalence).
E. Finding flow roles in networks: Role-Based
Similarity
F. Acknowledgements
In directed networks, nodes can have different 'roles',
e.g., sinks, sources or hubs.
In complex directed net-
works, functional roles may not fall into such simple cat-
egories, yet nodes can still be characterised by their con-
tribution to the diffusion of in- and out-flows. Here we
use a recent method (Role-Based Similarity, RBS) to un-
cover roles in directed networks based on the patterns of
incoming and outgoing flows at all scales [35, 36]. The
main idea underpinning RBS is that nodes with a sim-
ilar in/out flow profile play a similar role, regardless of
whether they are near or far apart in the network. Each
node is associated with a feature vector xi containing
KAB acknowledges an Award from the Imperial Col-
lege Undergraduate Research Opportunities Programme
(UROP). MTS acknowledges support from the ARC and
the Belgium network DYSCO (Dynamical Systems, Con-
trol and Optimisation). YNB acknowledges support from
the G. Harold and Leila Y. Mathers Foundation. MBD
acknowledges support from the James S. McDonnell
Foundation Postdoctoral Program in Complexity Sci-
ence/Complex Systems Fellowship Award (#220020349-
CS/PD Fellow). MB acknowledges support from EPSRC
grant EP/I017267/1 under the Mathematics Underpin-
ning the Digital Economy program.
Appendix : Supplementary Information
SI1. Supplementary Data
16
Supplementary Data as XLS spreadsheet is available at http://dx.doi.org/10.1371/journal.pcbi.
1005055.
SI2. Supplementary Text 1
Comparison of MS partitions to other methods
The flow-based MS partitions are distinct from partitions obtained by several other methods. In particular, we
have compared against partitions obtained with Modularity, Stochastic Block models, and Infomap.
Modularity has been used to obtain optimised partitions in Refs. [20, 21]. The partition found in Ref. [20] is closest
to our 4-way Partition B (VI = 0.185), whereas the partition found in Ref. [21] is closest to our 3-way Partition C
(VI = 0.186). Note that optimisation of modularity at a fixed resolution imposes an intrinsic scale, so that partitions
found with modularity are well matched to a particular scale (i.e., a particular Markov time) in the Markov Stability
framework, as shown previously [24, 32]. On the other hand, as discussed in the main text, the Markov Stability
framework carries out a systematic scanning across Markov times [33] allowing the intrinsic multiscale organisation
to became apparent.
The partitions based on stochastic block models [22] and hierarchical Infomap [52] are less similar to the ones found
by MS: the partition found by stochastic block models in [22] is closest to our 3-way Partition C (but with a higher
VI =0.272), and the partition found by hierarchical Infomap in [52] is closest to our 6-way Partition A (yet with
an even higher VI =0.282). These differences in the outcomes are expected due to the contrasting methodological
approaches. In particular, Infomap is known to impose a clique-like structure to the modules leading to groupings
where strong local density is favoured [75]. We remark that, as shown in S2 Text, the MS communities are also
different from the flow roles found through RBS.
SI3. Supplementary Text 2
Comparison of RBS flow roles with other analyses of roles
Our RBS flow roles are fundamentally different from notions of roles used in social networks based on Structural
Equivalence (SE) [49], and Regular Equivalence (RE) [50] (Fig. S6 and S1 Data). Because both RE and SE consider
only one-step neighbourhoods and do not incorporate information about the long scales of the network [33], they
are less applicable to complex networks such as the C. elegans connectome [34]. In particular the roles produced by
REGE show undifferentiated PageRank and connectivity profiles.
In Refs. [20, 21], roles were assigned to neurons according to 'community roles', the technique proposed by Guimera
et al [47] which identifies certain interneurons as relevant hubs between predefined communities. It was found that
command interneurons (e.g. AVA, AVB, AVD, PVC) play the role of global hubs, whereas D-type motor neurons
play the role of provincial hubs [20]. These features are in line with our ablation results, where D-type motor neuron
ablations alter flows at finer scales and ablation of interneurons modifies flow patterns at larger scales. Indeed, the
concept of 'community role' is closer to our ablation results, in that we measure there the disruption of flow-based
communities.
Chatterjee and Sinha [55] explored the core-periphery structure of the C. elegans connectome using a k-core de-
composition based on in- and out- degree separately. The k-core of a network is the subgraph with the property that
all nodes have (in/out)degree at least k. As expected, motor neurons are overrepresented in the k-cores based on
in-degree, and sensory neurons are overrepresented in k-cores based on out-degree. This distinction between neurons
with upstream and downstream roles is also an inherent characteristic in the RBS analysis, yet from a different per-
spective, i.e., based on the global characteristics of a node with respect to the in- and out-flows in the network, rather
than based on its local connections.
To quantify the differences between the groupings into roles obtained by these different methods, we have computed
the variation of information between them. RBS roles show very low similarity to any of the other groupings with
values of VI = 0.3061, 0.3607, 0.4356 against REGE [74], Sohn et al. [21] and Pan et al. [20], respectively.
17
Figure S1. Full analysis of the C. elegans connectome with Markov Stability (MS). We show the scan across all
Markov times, from the finest possible partition (every node in its own partition) at small Markov times to the bipartition at
large Markov times. The highlighted time interval corresponds to Fig. 1 in the main text, which focusses on the medium to
coarse partitions A − E.
10-210-101011010010110210-210-110010100.050.10.150.223364ABNo. of communitiesMarkov time, t 18
Figure S2. The asymmetry in the normalised conditional entropy of the optimised MS partitions signals a
quasi-hierarchical community structure. The normalized conditional entropy Ω(P(t(cid:48))P(t))/ log(n) ∈ [0, 1] quantifies the
uncertainty in the community assignment P(t(cid:48)) given the known partition P(t). If P(t(cid:48)) can be predicted from P(t), (i.e. when
P(t(cid:48)) is a strictly hierarchical agglomeration of the communities of P(t)) then the conditional entropy will be zero. The strong
upper-triangular character of the conditional entropy of the partitions A − E indicates a quasi-hierarchical organisation.
2.8822.915.7516.604.57Markov time, t'Markov time, t00.050.10.150.20.250.30.35 19
Figure S3. The effect of ablations and other network measures. Scatter plots of the Community Variation with respect
to Partitions A and D, CV[i](A) (left column) and CV[i](D) (right column), for all single neuron ablations (i = 1, . . . , 279)
plotted against the following properties of the corresponding neuron: a, stationary flow distribution π (PageRank); b, in-
degree; c, out-degree; d, betweenness centrality; and e, local clustering coefficient. None of these quantities (which are related
to network centralities) shows a manifest correlation with the effect of the neuron ablation on community structure.
20
Figure S4. Finding role profiles with RBS. Schematic summary of the procedure to obtain flow roles using RBS analysis,
as discussed in detail in [37]. First, from the original directed network of the C. elegans connectome we create a similarity
matrix using the RBS metric, by computing a similarity score between each node in the network, based on their incoming and
outgoing weighted path profiles. Second, the similarity matrix is transformed into a similarity matrix using the RMST method,
which subsequently prunes out uninformative links (see Ref. [37] for details). Third, the resulting similarity graph is clustered
to obtain relevant groups of nodes with similar in- and out-flow profiles at all scales. Four such classes of neurons (flow roles)
are found in this case. The neurons are then colored according to their flow profile on the original connectome layout.
Figure S5. Distribution of RBS flow roles across MS communities. RBS roles in each of the six communities of partition
A. The communities and flow roles induce very different groupings in the connectome. Hence the six communities present
distinct mixes of roles: the anterior communities A1 and A2 present a dominance of roles R1 and R3, whereas the posterior
communities A3, A4 and A5 are dominated by roles R1 and R4. Community A6 has a balanced mix of roles R1, R2, and R4
giving it a distinctive information processing structure, confirming the the importance of its embedded rich-club neurons.
Role 1Role 3Role 2Role 4Markov time tt=63.1Markov time t'100101102100101102Original network(connectome)2. RBS graphNode IDNode ID501001502002505010015020025000.10.20.30.40.50.60.70.80.911. Role-based Similarity3. Clustering the RBS graphRMSTIdentified rolesNetwork rolesFinding roles through RBSNo. of communities110110210010110200.020.040.064 groupsYij 21
Figure S6. Comparison of RBS flow roles to roles obtained using Regular Equivalence. a: Roles of the nodes
according to RBS with the PageRank distribution for each role and the average in/out degree for each role. b: Same for the
roles obtained according to Regular Equivalence obtained using the REGE algorithm [74].
R1R3R4R2Roles from RBSAREGE1REGE2REGE3REGE4Roles from Regular Equivalence (REGE)B01020Role 101020Role 201020Role 300.0050.010.0150.02Pagerank01020Role 401020Role 101020Role 201020Role 300.0050.010.0150.02Pagerank01020Role 4inoutinoutinoutinoutAverage DegreeRole 1Role 2Role 3Role 40510152025303540450102030405060708090100Average DegreeRole 1Role 2Role 3Role 4inoutinoutinoutinout 22
Figure S7. Summary of the procedure for signal propagation analysis of posterior mechanosensory stimulus
scenario (i1). For all neurons, we compute φi(t), i.e., the amount of signal present at each node at Markov time t. As time
grows, the signal at each node converges to its stationary value πi. Hence θi(t) = φi(t) − πi → 0. The approach to stationarity
can happen in two ways: i) the initially negative θi(t) approaches 0 from below; ii) θi(t) 'overshoots' before decaying towards its
stationary value. We consider the signal relative to the stationary value, qi(t) = φi(t)/πi, and focus on neurons that overshoot
(i.e., those with qmax,i := maxt φi(t)/πi > 1) and we collect the times at which they reach their peak. A concise summary of
the signal propagation is given by the strong response neurons with qmax,i > 5/3. Their peak-time histogram and the particular
sequence of strong response neurons is characteristic of the different input-response biological scenarios, as well as the analyses
by neuron type and flow roles.
10-310-210-1100101024681010-310-210-110010105101520Overshooting neuronsStrong response neurons10-310-210-1100101102-0.04-0.0200.020.040.060.080.10.120.140.1610-310-210-1100101102-0.04-0.0200.020.040.060.080.10.120.140.1610-310-210-1100101102-0.04-0.0200.020.040.060.080.10.120.140.16MotorInterneuronsSensoryPosterior mechanosensory stimulusStrong response sensoryStrong response interStrong response motorTimeTimeTimeOvershooting sensoryOvershooting interOvershooting motorTimeTimeTimeStrong response neurons Input neurons02.48Overshooting neuronsInputInput10-310-210-11001010246810Time 23
Figure S8.
Signal propagation of the anterior mechanosensory stimulus (i2). Signal propagation evolving from
an initial condition localised at the mechanosensory neurons (i2). (a) As stationarity is approached (θ(t) → 0), the input
propagates from sensory to motor neurons through an intermediate stage when interneurons overshoot. (b) The propagation
seen as a cascade of strong response neurons ( qmax,i > 1 + 2/3) with peak times concentrated around two bursts. (c) The
input (i2), appears localised on R1 and to a lesser extent R2 neurons. The signal diffuses somewhat quicker out of R2 than R1
neurons, but induces not collective overshoot of R3 or R4 neurons. (d) Stages of signal propagation in the network showing
the strong response neurons that have peaked at each time.
Figure S9. Signal propagation: posterior chemosensory stimulus (i3). See caption of S8.
t=2.2387TimeRole 1Role 2Role 3Role 4SIMTimeStrong response neuronsdacb02.13Strong response neuronsInput neuronst=0.001SIADLSIADRFLPLFLPRAQRADELADERBDURBDULAVMALMLALMRIL2LURADLOLLLURADRBAGLCEPVLBAGRCEPVRCEPDLALACEPDRRIHAVHLRIBLSIADLAVJLRIBRAVDRAVDLSIADRFLPRAQRADELADERRIGRRIGLBDURBDULAVMALMLALMRHSNRHSNLSDQLPVCLPVCRPVRPVNRPVNLAVJRFLPL10-310-210-1100101102-0.6-0.4-0.200.20.40.60.81024605101520Peak time10-310-210-1100101102-0.6-0.4-0.200.20.4t=2.2387TimeStrong response neuronsdacb02.47Strong response neuronsInput neuronsSIMRole 1Role 2Role 3Role 4t= 3.5481Peak timeTimet=0.001RIDAVALASKLAVARAVHLAVHRAVJRAVJLAIALAVDRAVDLAVFRAVFLDB02AVGDB03DB07VA12PVPLDA08DA09DVAPHALPHARPVQLPVCLPHBRPHBLPVCRAVALAVARAVHLAVHRAVDRAVDLAVFLAVGVA12PVPLDA08DVAPHALPHARPVQLPVCLPHBRPHBLPVCRPHALPHARPHBRPHBL10-310-210-1100101102-0.6-0.4-0.200.20.410-310-210-1100101102-0.6-0.4-0.200.20.40.60.81024605101520 24
Figure S10. Signal propagation: anterior chemosensory stimulus (i4). See caption of S8.
Figure S11. Peak times of strong response neurons by RBS roles for each of the four input scenarios (i1)-(i4).
Histograms of peak times of the strong response neurons in the four biological scenarios from the perspective of flow roles. The
tail inputs (i1) and (i3) induce strong responses on neurons spreading from R2 to R1 and finally to R4. On the other hand,
the head inputs induce strong responses on neurons heavily based on R1 spreading downwards to R3.
t=2.2387TimeStrong response neuronsdacb02.31Strong response neuronsInput neuronst=3.9811CEPVRALAASKLAFDLASKRADLLADFLAFDRAWBLAWCLADFRASGLADLRAWALASILASHLASGRAIBLASHRAWCRAWARASIRAIBRAVJRASERAIALRIMLAVDRAVDLRICLAIARAIZRRICRAIZLAIYRAIMRAIMLAIYLRIFRADARADALRMGLRMGRHSNLPVQRPVQLCEPVRALAASKLASKRADLLADFLAWBLADFRADLRASHLAIBLASHRAWCRAWARASIRAIBRAVJRASERAIALAVDRAVDLRICLAIARRICRAIMRAIMLRIFRADARADALRMGLRMGRHSNLPVQRPVQLt=0.001ASKLASKRADLLADLRASHLASHR02460510152010-310-210-1100101102-0.6-0.4-0.200.20.410-310-210-1100101102-0.6-0.4-0.200.20.40.60.81TimeSIMRole 1Role 2Role 3Role 4Peak timeHeadTailMechChemStrong response neuronsStrong response neuronsStrong response neuronsStrong response neurons(i2)(i1)(i3)(i4) 25
Figure S12. Peak overshoots against other network measures. The maximum overshoot of each neuron qmax,i for each
of the four biological scenarios (i1)–(i4) is plotted against the following measures of the corresponding neuron: a, stationary
flow distribution π (PageRank); b, in-degree; c, out-degree; d, betweenness centrality; and e, local clustering coefficient. There
is no manifest correlation between the overshooting qmax,i and any of those centrality scores or the local clustering coefficient.
26
[1] Donald DLE. C. elegans II. Cold Spring Harbor Laboratory Press; 1997.
[2] White JG, Southgate E, Thomson JN, Brenner S. The Structure of the Nervous System of the Nematode Caenorhabditis
elegans. Philosophical Transactions of the Royal Society of London Series B, Biological Sciences. 1986;314(1165):1–340.
[3] Hall DH, Russell R. The posterior nervous system of the nematode Caenorhabditis elegans: serial reconstruction of
identified neurons and complete pattern of synaptic interactions. The Journal of neuroscience. 1991;11(1):1–22.
[4] Varshney LR, Chen BL, Paniagua E, Hall DH, Chklovskii DB. Structural Properties of the Caenorhabditis elegans Neuronal
Network. PLoS Computational Biology. 2011;7(2).
[5] Chalfie M, Sulston JE, White JG, Southgate E, Thomson JN, Brenner S. The neural circuit for touch sensitivity in
Caenorhabditis elegans. The Journal of Neuroscience. 1985;5(4):956–964.
[6] Wakabayashi T, Kitagawa I, Shingai R. Neurons regulating the duration of forward locomotion in Caenorhabditis elegans.
Neuroscience Research. 2004;50(1):103–111.
[7] Li W, Kang L, Piggott BJ, Feng Z, Xu XZS. The neural circuits and sensory channels mediating harsh touch sensation in
Caenorhabditis elegans. Nature communications. 2011;2:315.
[8] Nagel G, Brauner M, Liewald JF, Adeishvili N, Bamberg E, A AG. Light activation of channelrhodopsin-2 in excitable
cells of Caenorhabditis elegans triggers rapid behavioral responses. Current Biology. 2005;15(24):2279–2284.
[9] Ibsen S, Schutt ATC, Esener S, Chalasani SH. Sonogenetics is a non-invasive approach to activating neurons in Caenorhab-
ditis elegans. Nature Communications. 2015;6.
[10] Stephens GJ, Johnson-Kerner B, Bialek W, Ryu WS. Dimensionality and Dynamics in the Behavior of C. elegans. PLoS
Comput Biol. 2008;4(4):e1000028.
[11] Yemini E, Jucikas T, Grundy LJ, Brown AE, Schafer WR. A database of Caenorhabditis elegans behavioral phenotypes.
Nature methods. 2013;10(9):877–879.
[12] Brown AE, Yemini EI, Grundy LJ, Jucikas T, Schafer WR. A dictionary of behavioral motifs reveals clusters of genes
affecting Caenorhabditis elegans locomotion. Proceedings of the National Academy of Sciences. 2013;110(2):791–796.
[13] Chen BL, Hall DH, Chklovskii DB. Wiring optimization can relate neuronal structure and function. Proceedings of the
National Academy of Sciences of the United States of America. 2006;103(12):4723–4728. doi:10.1073/pnas.0506806103.
[14] Watts DJ, Strogatz SH. Collective dynamics of 'small-world' networks. Nature. 1998;393(6684):440–442. doi:10.1038/30918.
[15] Kim JS, Kaiser M. From Caenorhabditis elegans to the human connectome: a specific modular organization increases
metabolic, functional and developmental efficiency. Phil Trans R Soc B. 2014;369(1653).
[16] Barab´asi AL, Albert R.
Emergence of Scaling in Random Networks.
Science. 1999;286(5439):509–512.
doi:10.1126/science.286.5439.509.
[17] Towlson EK, Vertes PE, Ahnert SE, Schafer WR, Bullmore ET. The Rich Club of the C. elegans Neuronal Connectome.
The Journal of Neuroscience. 2013;33(15):6380–6387.
[18] Majewska A, Yuste R. Topology of gap junction networks in C. elegans. J Theor Biol. 2001;212(2):155–67.
[19] Arenas A, Fern´andez A, G´omez S. A complex network approach to the determination of functional groups in the neural
system of C. elegans. In: Bio-Inspired Computing and Communication. Springer; 2008. p. 9–18.
[20] Pan RK, Chatterjee N, Sinha S. Mesoscopic Organization Reveals the Constraints Governing Caenorhabditis elegans
Nervous System. PLOS ONE. 2010;5.
[21] Sohn Y, Choi MK, Ahn YY, Lee J, Jeong J. Topological Cluster Analysis Reveals the Systemic Organization of the
Caenorhabditis elegans Connectome. PLoS Comput Biol. 2011;7(5).
[22] Pavlovic DM, Vertes PE, Bullmore ET, Schafer WR, Nichols TE. Stochastic Blockmodeling of the Modules and Core of
the Caenorhabditis elegans Connectome. PLoS ONE. 2014;9(7).
[23] Sporns O, Betzel RF. Modular brain networks. Annual review of psychology. 2015;67(1).
[24] Lambiotte R, Delvenne J, Barahona M.
ganization of Complex Networks.
doi:10.1109/TNSE.2015.2391998.
Random Walks, Markov Processes and the Multiscale Modular Or-
IEEE Transactions on. 2014;1(2):76–90.
Network Science and Engineering,
[25] Jeub LG, Balachandran P, Porter MA, Mucha PJ, Mahoney MW. Think locally, act locally: Detection of small, medium-
sized, and large communities in large networks. Physical Review E. 2015;91(1).
[26] Betzel RF, Griffa A, Avena-Koenigsberger A, Goni J, Hagmann P, Thiran JP, et al. Multi-scale community organization
of the human structural connectome and its relationship with resting-state functional connectivity. Network Science.
2013;1(3):353–373.
[27] Misic B, Betzel RF, Nematzadeh A, Goni J, Griffa A, Hagmann P, et al. Cooperative and competitive spreading dynamics
on the human connectome. Neuron. 2015;86(6):1518?1529.
[28] Lizier JT, Heinzle J, Horstmann A, Haynes JD, Prokopenko M. Multivariate information-theoretic measures reveal di-
rected information structure and task relevant changes in fMRI connectivity. Journal of Computational Neuroscience.
2011;30(1):85–107.
[29] Bullmore E, Sporns O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nature
Reviews Neuroscience. 2009;10(3):186–198.
[30] Sporns O. Networks of the Brain. MIT press; 2011.
[31] Delvenne JC, Schaub MT, Yaliraki SN, Barahona M. The Stability of a Graph Partition: A Dynamics-Based Framework
for Community Detection. In: Mukherjee A, Choudhury M, Peruani F, Ganguly N, Mitra B, editors. Dynamics On and
Of Complex Networks. vol. 2. Springer New York; 2013. p. 221–242. Available from: http://arxiv.org/abs/1308.1605.
27
[32] Delvenne JC, Yaliraki SN, Barahona M. Stability of graph communities across time scales. Proceedings of the National
Academy of Sciences. 2010;107(29):12755–12760. doi:10.1073/pnas.0903215107.
[33] Schaub MT, Delvenne JC, Yaliraki SN, Barahona M. Markov Dynamics as a Zooming Lens for Multiscale Com-
PLoS ONE. 2012;7(2):e32210.
munity Detection: Non Clique-Like Communities and the Field-of-View Limit.
doi:10.1371/journal.pone.0032210.
[34] Beguerisse-D´ıaz M, Garduno Hern´andez G, Vangelov B, Yaliraki SN, Barahona M. Interest communities and flow roles in
directed networks: the Twitter network of the UK riots. J R Soc Interface. 2014;11(101). doi:10.1098/rsif.2014.0940.
[35] Cooper K, Barahona M. Role-based similarity in directed network. arXiv:10122726. 2010;.
[36] Cooper K. Complex Networks: Dynamics and Similarity. PhD Thesis, Imperial College London; 2010.
[37] Beguerisse-D´ıaz M, Vangelov B, Barahona M. Finding role communities in directed networks using Role-Based Similarity,
Markov Stability and the Relaxed Minimum Spanning Tree. In: 2013 IEEE Global Conference on Signal and Information
Processing (GlobalSIP); 2013. p. 937–940.
[38] Hilliard M, Bargmann CI, Bazzicalupo P. C. elegans responds to chemical repellents by integrating sensory inputs from
the head and the tail. Current Biology. 2002;12:730–734.
[39] Billeh YN, Schaub MT, Anastassiou CA, Barahona M, Koch C. Revealing cell assemblies at multiple levels of granularity.
Journal of neuroscience methods. 2014;236:92–106.
[40] Bassett DS, Greenfield DL, Meyer-Lindenberg A, Weinberger DR, Moore SW, Bullmore ET. Efficient physical embedding
of topologically complex information processing networks in brains and computer circuits. PLoS Comput Biol. 2010;6(4).
doi:10.1371/journal.pcbi.1000748.
[41] Gray JM, Hill JJ, Bargmann CI.
Inaugural Article: A circuit for navigation in Caenorhabditis elegans. PNAS.
2005;102(9):3184–3191. doi:10.1073/pnas.0409009101.
[42] Haspel G, O'Donovan MJ. A peri-motor framework reveals functional segmentation in the motoneuronal net-
The Journal of Neuroscience. 2011;31(41):14611–14623.
work controlling locomotion in Caenorhabditis elegans.
doi:10.1523/JNEUROSCI.2186-11.2011.
[43] Schaub MT, Lehmann J, Yaliraki SN, Barahona M. Structure of complex networks: Quantifying edge-to-edge relations by
failure-induced flow redistribution. Network Science. 2014;2(1):66–89. doi:10.1017/nws.2014.4.
[44] Hardaker LA, Singer E, Kerr R, Zhou G, Schafer WR. Serotonin modulates locomotory behavior and coordinates egg-laying
and movement in Caenorhabditis elegans. Journal of Neurobiology. 2001;49:303–313.
[45] Hall D, Altun Z, Herndon L. Worm Atlas; 2015. Available from: http://www.wormatlas.org.
[46] Altun ZF, Chen B, Wang ZW, Hall DH. High resolution map of Caenorhabditis elegans gap junction proteins. Develop-
mental Dynamics. 2009;238(8):1936–1950.
[47] Guimera R, Amaral LAN. Functional cartography of complex metabolic networks. Nature. 2005;433(7028):895–900.
doi:10.1038/nature03288.
[48] Gelman A, Carlin JB, Stern HS, Rubin DB. Bayesian data analysis. Chapman & Hall/CRC; 2014.
[49] Lorrain F, White HC. Structural equivalence of individuals in social networks. The Journal of Mathematical Sociology.
1971;1(1):49–80.
[50] Everett MG, Borgatti SP. Regular equivalence: General theory. The Journal of Mathematical Sociology. 1994;19(1):29–52.
[51] Jarrell TA, Wang Y, Bloniarz AE, Brittin CA, Xu M, Thomson JN, et al. The Connectome of a Decision-Making Neural
Network. Science. 2012;337(6093):437–444.
[52] Edler D, Rosvall M. The MapEquation software package, available online at http://www.mapequation.org;.
[53] Marinazzo D, Wu G, Pellicoro M, Angelini L, Stramaglia S. Information Flow in Networks and the Law of Diminishing
Marginal Returns: Evidence from Modeling and Human Electroencephalographic Recordings. PLoS ONE. 2012;7(9):1–9.
doi:10.1371/journal.pone.0045026.
[54] Marinazzo D, Pellicoro M, Wu G, Angelini L, Cort´es JM, Stramaglia S. Information Transfer and Criticality in the Ising
Model on the Human Connectome. PLoS ONE. 2014;9(4):1–7. doi:10.1371/journal.pone.0093616.
[55] Chatterjee N, Sinha S. In: Understanding the mind of a worm: hierarchical network structure underlying nervous system
function in C. elegans. vol. 168 of Progress in Brain Research. Elsevier; 2007. p. 145–153.
[56] Klimm F, Borge-Holthoefer J, Wessel N, Kurths J, , Zamora-L´opez G. Individual node's contribution to the mesoscale of
complex networks. New Journal of Physics. 2014;16.
[57] Rosvall M, Esquivel AV, Lancichinetti A, West JD, Lambiotte R. Memory in network flows and its effects on spreading
dynamics and community detection. Nature Communications. 2014;5.
[58] Salnikov V, Schaub MT, Lambiotte R. Using higher-order Markov models to reveal flow-based communities in networks.
Scientific Reports. 2016;6:23194–.
[59] Ahrens MB, Orger MB, Robson DN, Li JM, Keller PJ. Whole-brain functional imaging at cellular resolution using
light-sheet microscopy. Nat Meth. 2013;10(5):413–420.
[60] Bumbarger DJ, Riebesell M, Rodelsperger C, Sommer RJ. System-wide Rewiring Underlies Behavioral Differences in
Predatory and Bacterial-Feeding Nematodes. Cell. 2013;152:109–119.
[61] Zaslaver A, Liani I, Shtangel O, Ginzburg S, Yee L, Sternberg PW.
Hierarchical sparse coding in the sen-
sory system of Caenorhabditis elegans. Proceedings of the National Academy of Sciences. 2015;112(4):1185–1189.
doi:10.1073/pnas.1423656112.
[62] Koch C. Biophysics of Computation. Oxford University Press; 1999.
[63] Ferree TC, Lockery SR. Journal of Computational Neuroscience. 1999;6(3):263–277. doi:10.1023/A:1008857906763.
[64] Goodman MB, Hall DH, Avery L, Lockery SR. Active currents regulate sensitivity and dynamic range in C. elegans
neurons. Neuron. 1998;20(4):763–772.
28
[65] Schaub MT, Billeh YN, Anastassiou CA, Koch C, Barahona M. Emergence of slow-switching assemblies in structured
neuronal networks. PLoS Computational Biology. 2015;11(7):e1004196. doi:10.1371/journal.pcbi.1004196.
[66] Page L, Brin S, Motwani R, Winograd T. The PageRank Citation Ranking: Bringing Order to the Web. Stanford InfoLab;
1999. 1999-66. Available from: http://ilpubs.stanford.edu:8090/422/.
[67] Blondel VD, Guillaume JL, Lambiotte R, Lefebvre E. Fast unfolding of communities in large networks. Journal of
Statistical Mechanics: Theory and Experiment. 2008;2008(10):P10008.
[68] Amor B, Yaliraki SN, Woscholski R, Barahona M. Uncovering allosteric pathways in caspase-1 using Markov transient
analysis and multiscale community detection. Mol BioSyst. 2014;10:2247–2258. doi:10.1039/C4MB00088A.
[69] Meila M. Comparing clusterings-an information based distance. Journal of Multivariate Analysis. 2007;98(5):873–895.
doi:10.1016/j.jmva.2006.11.013.
[70] Rasmussen CE, Williams CKI. Gaussian Processes for Machine Learning (Adaptive Computation and Machine Learning).
The MIT Press; 2005.
[71] Koenker R. Quantile regression. 38. Cambridge university press; 2005.
[72] Cade BS, Noon BR. A gentle introduction to quantile regression for ecologists. Frontiers in Ecology and the Environment.
2003;1(8):412–420.
[73] Smith LB, Reich BJ. BSquare: An R package for Bayesian simultaneous quantile regression. North Carolina State
University; 2013. Available from: http://www4.stat.ncsu.edu/~reich/QR/BSquare.pdf.
[74] Borgatti SP, Everett MG. Two algorithms for computing regular equivalence. Social Networks. 1993;15(4):361–376.
[75] Schaub MT, Lambiotte R, Barahona M. Encoding dynamics for multiscale community detection: Markov time sweeping
for the map equation. Phys. Rev. E. 2012; 86(2):026112
|
1707.01484 | 1 | 1707 | 2017-07-05T17:34:14 | Cerebellar-Inspired Learning Rule for Gain Adaptation of Feedback Controllers | [
"q-bio.NC",
"cond-mat.dis-nn",
"eess.SY",
"nlin.AO",
"physics.bio-ph"
] | How does our nervous system successfully acquire feedback control strategies in spite of a wide spectrum of response dynamics from different musculo-skeletal systems? The cerebellum is a crucial brain structure in enabling precise motor control in animals. Recent advances suggest that synaptic plasticity of cerebellar Purkinje cells involves molecular mechanisms that mimic the dynamics of the efferent motor system that they control allowing them to match the timing of their learning rule to behavior. Counter-Factual Predictive Control (CFPC) is a cerebellum-based feed-forward control scheme that exploits that principle for acquiring anticipatory actions. CFPC extends the classical Widrow-Hoff/Least Mean Squares by inserting a forward model of the downstream closed-loop system in its learning rule. Here we apply that same insight to the problem of learning the gains of a feedback controller. To that end, we frame a Model-Reference Adaptive Control (MRAC) problem and derive an adaptive control scheme treating the gains of a feedback controller as if they were the weights of an adaptive linear unit. Our results demonstrate that rather than being exclusively confined to cerebellar learning, the approach of controlling plasticity with a forward model of the subsystem controlled, an approach that we term as Model-Enhanced Least Mean Squares (ME-LMS), can provide a solution to wide set of adaptive control problems. | q-bio.NC | q-bio |
Cerebellar-Inspired Learning Rule for Gain
Adaptation of Feedback Controllers
Ivan Herreros1, Xerxes D. Arsiwalla1, Cosimo Della Santina2, Jordi-Ysard
Puigbo1, Antonio Bicchi2,3, and Paul Verschure1,4
1 SPECS Lab, University Pompeu Fabra
Carrer de Roc Boronat, 138, 08018 Barcelona, Spain
2 Research Center "Enrico Piaggio", University of Pisa
Largo Lucio Lazzarino 1, 56126 Pisa, Italy
3 Department of Advanced Robotics, Istituto Italiano di Tecnologia
via Morego, 30, 16163 Genoa, Italy
4 Instituci´o Catalana de Recerca i Estudis Avan¸cats
Passeig de Llus Companys, 23, 08010 Barcelona, Spain
{[email protected]}
Abstract
How does our nervous system successfully acquire feedback control strategies in spite of a wide
spectrum of response dynamics from different musculo-skeletal systems? The cerebellum is a crucial
brain structure in enabling precise motor control in animals. Recent advances suggest that synaptic
plasticity of cerebellar Purkinje cells involves molecular mechanisms that mimic the dynamics of the
efferent motor system that they control allowing them to match the timing of their learning rule to
behavior. Counter-factual predictive control (CFPC) is a cerebellum-based feed-forward control scheme
that exploits that principle for acquiring anticipatory actions. CFPC extends the classical Widrow-
Hoff/Least Mean Squares by inserting a forward model of the downstream closed-loop system in its
learning rule. Here we apply that same insight to the problem of learning the gains of a feedback
controller. To that end, we frame a model-reference adaptive control (MRAC) problem and derive an
adaptive control scheme treating the gains of a feedback controller as if they were the weights of an
adaptive linear unit. Our results demonstrate that rather than being exclusively confined to cerebellar
learning, the approach of controlling plasticity with a forward model of the subsystem controlled, an
approach that we term as Model-enhanced least mean squares (ME-LMS), can provide a solution to
wide set of adaptive control problems.
1
Introduction
The cerebellum is arguably the brain structure whose study has had a deeper impact on the
robotics and control communities. The seminal theory of cerebellar function by Marr [17] and
Albus [1] was translated by the latter into the cerebellar model articulation controller (CMAC)
in the early seventies [2], which up until today is used both in research and applications. A
decade later, Fujita [10] advanced the adaptive filter theory of cerebellar function based on
the work by Widrow et al. [22]. Later, in the late eighties, Kawato formulated the influential
feedback error learning (FEL) model of cerebellar function [15], in which the cerebellum, im-
plemented as an adaptive filter, learned from, and supplemented, a feedback controller. Unlike
CMAC, FEL had a strong impact within the neuroscientific community as a theory of biolog-
ical motor control [24]. Within the robotics and control communities, FEL has been studied
in terms of performance and convergence properties [18]. Later, the adaptive filter theory of
cerebellar function was revived by Porrill and colleagues [6], proposing alternatives to FEL that
1
have been applied to the control of bio-mimetic actuators, like pneumatic or elastomer muscles
[16], [23].
More recently, the counterfactual predictive control (CFPC) scheme was proposed in [14],
motivated from neuro-anatomy and physiology of eye-blink conditioning, a behavior dependent
on the cerebellum. CFPC includes a reactive controller, which is an output-error feedback
controller that models brain stem or spinal reflexes actuating on peripheral muscles, and a
feed-forward adaptive component that models the cerebellum and learns to associate its own
inputs with the errors that drive the reactive controller. CFPC proposes that the learning of
adaptive terms in the linear filter should depend on the coincidence of an error signal with the
output of a forward model implemented at the synaptic level, reproducing the dynamics of the
downstream reactive closed-loop system. We refer to that learning rule as a model-enhanced
least-mean squares (ME-LMS) rule to differentiate with the basic least-mean squares (LMS)
rule that is implemented in previous models of the cerebellum, such as CMAC and FEL. In
agreement with the theoretical insight of CFPC, recent physiological evidence in [21] shown
that the timing of the plasticity rule of Purkinje cells is matched to behavioral function. That
suggests that Purkinje cells, the main cells implied in learning at the level of the cerebellum,
have plasticity rules that reflect the sensorimotor latencies and dynamics of the plants they
control.
However, in the context of CFPC, the ME-LMS rule was derived as a batch gradient-
In that sense, it can
descent rule for solving a feed-forward control task in discrete time.
be interpreted as providing a solution to a iterative-learning control scheme, an input design
technique for learning to optimize the execution of a repetitive task [5]. Hence, it remained open
the question as to whether a similar learning rule could support the acquisition of well-tuned
feedback gains. That is, whether ME-LMS could be applied in an adaptive feedback control
problem. Here we answer that using the model reference adaptive control (MRAC) frame [4].
In that frame, we first show that the biologically-inspired ME-LMS algorithm can be derived
from first principles. More concretely, we show that the ME-LMS rule emerges from deriving
the stochastic gradient descent rule for the general problem of updating the gains of linear
proportional feedback controllers actuating on a LTI system. Finally we test in simulation the
effectiveness of the proposed cerebellum-inspired architecture in controlling a damped-spring
mass system, a non-minimum phase plant and, finally, closing the loop with the biology, a
biologically-based model of a human limb.
2 Derivation of the ME-LMS Learning Rule
In the next we derive a learning rule for learning the controller gains for both state and output-
error feedback controllers. The generic architectures for a full-state-feedback and a proportional
(P) controller are shown in Fig. 1. To define the model-reference adaptive control (MRAC)
problem, we set a reference model whose output we denote by rrm. The error in following the
output of the reference model, erm = rrm − y, drives adaptation of the feedback gains. But
note that in the proportional error feedback controller, e = r − y is the signal feeding the error
feedback gain.
2
Figure 1: Adaptive architecture for the full state feedback (left) and output-error proportional
(P, right) control cases. Abbreviations: C, feedback controller, P, plant; RM, reference model;
r, reference signal; r, reference signal outputted by the reference model; y, plant's output; e,
output error; erm, output error relative to the reference model; x, state of the plant; u, control
signal; kr, reference gain; K, state feedback gains; and kp, proportional error-feedback gain.
2.1 ME-LMS for State-Feedback
For the derivation purposes we assume that the adaptive control strategy is applied to a generic
LTI dynamical system
(1)
where A ∈ RN×N , B ∈ RN×1 and C ∈ R1×N are the usual dynamics, input and output
matrices, respectively; x ∈ RN is the state vector; and u and y, both scalars, are the input and
output signals.
x = Ax + Bu
y = Cx
The control signal will be generated according to the following state-feedback law
(2)
where r is the reference signal, K ∈ R1×N is the (row) vector of state feedback gains and kr the
reference gain. Both K and kr are here time-dependent and will be updated by the learning
rule controlling adaptation.
u = Kx + krr
Substituting the control law within the dynamics equation, we obtain the closed-loop system
description
x = (A + BK)x + Bkrr
y = Cx
(3)
(4)
We set the L-2 norm of the error as the cost function to minimize
J =
1
2
e2
rm
For convenience, we write now the control
law as u = KT x with K ∈ RN +1 ≡
[k1, . . . , kN , kr]T and x ∈ RN +1 ≡ [x1, . . . , xN , r]T. To derive the gradient descent algorithm for
adjusting the vector of gains, K, we need the gradient of J with respect to K:
∇ KJ =
∂erm
∂ K
erm = − ∂y
∂ K
erm
(5)
3
closed-loop feedback systemCrRMkrPKrrmermy-xu-closed-loop feedback systemCrRMkpPrrmermuey- Now we will consider each of the gains individually, treating separately the state and the
reference gains. Let ki denote the feedback gain associated with the i-th state variable. We
have that
∂y
∂ki
= C
∂x
∂ki
(6)
We compute the partial derivative of the state vector x with respect to ki applying the
partial derivative to the differential equation that governs the closed-loop dynamics:
∂
∂ki
Using the substitution zi ≡ ∂x
∂ki
x =
((A − BK)x + Bkrr)
∂
∂ki
(7)
and applying the product rule in the derivation we obtain
zi = (A − BK)zi + Bxi
(8)
Introducing hi ≡ Czi, we get
∂J
∂ki
= hierm
Note that this has solved the problem of obtaining the partial derivative for all state feedback
gains.
In the case of the reference gain, with zr ≡ ∂x
, we obtain
∂kr
zr = (A − BK)zr + Br
(9)
And introducing hr ≡ Czr,
∂J
∂kr
= hre
We will refer to the quantities hi and hr as eligibility traces. We can write the vector of
eligibility traces as follows: h = [h1, . . . , hN , hr]T.
With this we can solve for the gradient of the cost function as follows
∇ KJ = −herm
And consequently derive a learning rule for the gains that will follow a gradient descent:
K = η herm
(10)
Note that this rule is similar to the classical Widrow-Hoff or least mean squares (LMS) rule.
However, in the standard LMS, the rate of change is obtained multiplying the error with the
input signals of the filter ( K = ηxerm) whereas in the rule we have derived the error multiplies
the quantities in h, which are obtained after passing the input signals through a forward model
of the controlled system. For this reason, we refer to the learning rule in equation 10 as model-
enhanced least mean squares (ME-LMS). Moreover, hi is the eligibility trace associated to the
input xi because, at the time that a particular error signal comes, it codes how much xi could
have contributed to canceling that error.
4
Figure 2: Schematic of the implementation of the LMS (left) and the ME-LMS (right) rule
at the level of a single adaptive weight. Note that since ki is a gain of the controller C, this
scheme is implicitly recursive: as the ki gain changes (together with the other adaptive gains
of the controller) the forward that it utilizes to drive plasticity, changes as well.
2.2 ME-LMS for Output Error Proportional Control
The control law of an output-error proportional controller is u = kpe = kpr − kpCx. Following
a derivation analogous to the previous one, we obtain the following expression for computing
eligibility trace of the proportional gain (hp).
zp = (A − BCkp)zp + Br
hp = Czp
(11)
(12)
hence, in this model, plasticity will be implemented by the rule kp = ηhpe.
2.3 Model-Enhanced Least Mean-Squares vs. Least Mean Squares
Rule
The differences between LMS and ME-LMS care summarized as follows: In LMS, the change
of the adaptive gain ki is determined based on the temporal coincidence of a global error signal
erm and the local input xi (Fig. 2 left). In ME-LMS (Fig. 2 right) the change in the gain is
controlled by the output of a gain-specific forward model F Mi, whose output hi facilitates an
eligibility trace for ki. The term eligibility trace implies that hi marks how much the input xi
could have contributed to decrease the current error. In that sense, it is a trace as long as to
be generated hi takes into account the history of xi with a time-span implicitly determined by
the dynamics of the forward model.
3 Applying ME-LMS to a Linear Damped Spring-Mass
System
We evaluate here the performance of the proposed algorithm in controlling a standard damped-
spring mass system: mq + c q + kq = u, with m = 1Kg, c = 0.5 Ns
m , k = 0.5 N
m .
5
xiermhiuikiFMiCPxiermuikiLMSME-LMS Figure 3: Damped-spring mass system with P-control. Above left. Cost as a function of kp.
Above right. Convergence of kp for three different models. Below left. Output trajectories and
reference signals. Below right. Evolution of the error in mean-square error (MSE) units. In all
panels: WH: standard LMS rule; OL: LMS with eligibility trace derived from the open-loop
system; CL: ME-LMS with eligibility trace derived from the closed-loop system.
3.1 Output Error P-Control
√
For this problem we use a reference model built as chain of two leaky integrators with identical
relaxation time constants (τ =
0.5s). The impulse response curve of this reference model
corresponds to a normalized double-exponential convolution that peaks at 0.5s. Finally, we
use as reference the following superposition two sinusoidal functions r = sin(5t/7) + sin(5t/11).
We first examine the cost function as a function of the feedback parameter, kp varying it
logarithmically from 0.01 to 100.0. Within that range, the cost function is convex and has
minimum in the near kp = 0.815 (Fig 3 above-left). At this point we check whether the
ME-LMS converges to the optimum value for kp. For comparison we also run the test with
using the standard LMS rule, and a heuristically motivated alternative algorithm wherein we
use a model of the open-loop system to generate the eligibility trace. We test two starting
values each one at a different side of the minimum and we observe that in both cases the
ME-LMS converges around the optimal kp (Fig 3 above-right) while the alternative algorithms
convergence to different equilibrium points which are non-optimal in cost terms. The difference
in performance can also be appreciated by seeing how the different algorithms track rrm at the
end of the training period (1h of simulated time) (Fig 3 below-left). Indeed, only the ME-LMS
algorithm is in-phase with rrm. Finally, in cost terms, only ME-LMS converges rapidly to the
minimum value (Fig 3 above-right).
In summary, this result shows that even for the simplest feedback learning scenario, a LMS-
like learning rule converges to the optimal gain only if it uses the eligibility trace generated by
a forward model that reproduces the dynamics of the closed-loop system.
6
10−210−11001011020.20.30.40.50.60.70.80.91kp∫J dt (cost)59.459.559.659.759.859.9−3−2−10123time (min)output (a.u.) rrmWHOLCL010203040506000.511.522.53time (minutes)kp WHOLCLbest kp510152025303540455055600.250.30.350.40.450.50.550.6time (minutes)MSE WHOLCL 4 Applying ME-LMS to a Non-Minimum Phase Plant
4.1 Full State-Feedback Control
In this section we apply ME-LMS to a non-minimum phase system, which is a system with zeros
in the right-hand side of the complex plane. Acting to reduce an error in such a system requires
foresight in that before minimizing an error one has to steer the plant in the direction that
increases it. That property of the system, namely that errors cannot be reduced instantaneously,
disqualifies the use of the standard LMS algorithm for the problem of adaptively tuning feedback
gains. On the contrary, ME-LMS, as it takes explicitly into account the dynamics of the
controlled system to control plasticity, can in principle appropriately adjust the gains of a
feedback controller even when it is applied to a non-minimum phase system.
As a non-minimum phase system, we use the following a linearized model of a balance
system (e.g., a self-balancing robot):
A =
0
0
1
m2l2g/µ −cJt/µ −γlm/µ
Mtmgl/µ −clm/µ −γMt/µ
, C =(cid:2) 0
0 (cid:3)
1
B =
0
Jt/µ
lm/µ
where µ = MtJt − m2l2. The values, chosen to mimic the ones of a custom made prototype, are
Mt = 1.58Kg, m = 1.54Kg, l = 0.035, Jt = 1.98 × 10−3, γ = 0.01 and c = 0.1. As an added
difficulty, the plant is unstable in open loop. To deal with that, we set boundary conditions
to our simulation. That is, whenever the system reaches a threshold velocity of 0.5m/s or
and inclination in absolute value above 22.5 degrees the simulation re-starts and the system is
brought back to the initial rest position. In that sense, the system is not fully autonomous but
assisted by an external agent that makes it regain the balanced position.
In practice, the control problem consisted in following a low amplitude and slow velocity
reference signal constructed as a sum of sinusoidals 0.05(sin(πt/70) + sin(πt/110)). We used
the same reference model as in the previous section.
For this system the problem of adjusting the velocity to a particular reference signal is under-
constrained as there are two possible strategies: keeping the error in the linear velocity equal to
zero while the angular position diverges or keeping that error equal to zero while maintaining
the angular position stable. In order to bias the system towards the second solution we set
the starting gains already tuned towards the right solution. However, that initial set of gains
keep the robot balanced for less than 200ms. Hence, we can divide this particular problem
in two stages: first stabilizing the plant, and next make the controlled system converge to the
dynamics of the reference model.
ME-LMS requires approximately 10 seconds for reaching a set of gains that stabilizes the
plant following fifteen falls (Fig. 4 top row ). Standard LMS fails as it converges to a solution
that controls for the linear velocity but ignores the angular position (Fig. 4 middle row ).
Indeed, by the end of the 30 seconds of training, standard LMS has reduced the errors in
velocity but the speed at which the plant loses balance remains unchanged. Regarding the
learning dynamics, we observe how the feedback gains of the ME-LMS change rapidly until the
robot maintains balance (Fig. 4 below left). After that, the change is gradual but sufficient to
achieve following closely the target velocity by the end of the 10 mins training (Fig. 4 below
right).
7
Figure 4: ME-LMS applied to a Self-Balancing system. Top row refers to the ME-LMS algo-
rithm and middle row to the standard LMS. Left panels show the velocity traces (red) and the
reference signal (black). Right panels, show the angular position traces. Bottom left: gains of
the ME-LMS system during the first 30 seconds of simulation. Bottom right: velocity trace the
last minute of 10 mins training of the ME-LMS system.
5 Applying ME-LMS to a Bio-Inspired Limb Model
In this section we test the ability of the proposed ME-LMS algorithm to control an antagonistic
pair of human muscles acting on a joint, as e.g. the biceps and triceps muscles on elbow joint
(Fig. 5).
5.1 Model Derivation
We consider the muscle model proposed in [12], as a trade of between accuracy and complexity.
According to that, it is possible to describe the force exerted by a muscle as exponential function
of its activation A, which is considered proportional to the difference between the instantaneous
muscle length l and the threshold length λ(t), assuming the role of an input. A damping term
due to proprioceptive feedback is also considered, proportional to the variation of muscle length
l. The overall activation is
A(t) = [l(t − d) − λ(t) + µ l(t − d)]+ ,
(13)
8
0102030−0.500.5vel (m/s)time (s)0102030−0.1−0.0500.050.10.15vel (m/s)time (s)0102030−20−15−10−505gainstime (s) kθkvkθkr0102030−30−20−100102030angle (deg)time (s)0102030−30−20−100102030angle (deg)time (s)99.510−0.1−0.0500.050.10.15vel (m/s)time (min) Figure 5: Agonist-antagonist actuation systems with main variables underlined. q is in both
cases the joint angle, and τ is the external torque. f1 and f2 are the forces exerted by the
biceps and triceps respectively.
where [ x ]+ is 0 when x ≤ 0, and x otherwise. In the following we neglect the reflex delay d.
Thus considering the forces exerted by biceps f1 and triceps f2 on elbow joint, and the
gravity force acting on the forearm, the overall dynamics is
I q + m L g cos(q) = R(f1 + f2),
(14)
where q is the forearm angular position w.r.t. the arm. I and m are forearm inertia and mass
respectively, and L is the distance of forearm center of mass from the joint, all considered
constant (i.e. we neglect the dependency from wrist configuration). The two forces exerted by
the muscles are
f1 = −ρ(eδA1 − 1), A1 = [ Rq − λ1 + µR q]+,
f2 = ρ(eδA2 − 1), A2 = [−Rq + λ2 − µR q]+ ,
(15)
where δ is a form parameter equal for all muscles, ρ is a magnitude parameter related to
force-generating capability. λ1 and λ1 are the length commands for each muscle. R is the
instantaneous lever arm, considered here constant [12], i.e. l = Rq.
At the equilibrium with no external torque [11] the joint angle qeq, and the stiffness σ are
qeq =
r
R
,
σ =
∂τ
∂q
= 2ρδR2eδc
(16)
where r = λ1+λ2
or co-activation [9]. We consider here r as control input, and c as fixed.
is referred in literature as r-command, and c = λ2−λ1
2
2
is referred as c-command
(cid:12)(cid:12)(cid:12)(cid:12)q= λ1+λ2
2 R
5.2 Control Derivation
System (14) in state-space form is
(cid:20) x1
(cid:21)
x2
=
(cid:20)
(cid:21)
I (R(f1(x1, x2, r) + f2(x1, x2, r)) − mLg sin(x1))
x2
1
(17)
where [x1, x2] = [q, q]. We start by linearizing the system in the equilibrium position qeq = 0
(cid:20) x1
(cid:21)
x2
(cid:20)0
=
1
κ β
(cid:21)(cid:20) x1
(cid:21)
x2
(cid:20)0
(cid:21)
b
+
r ,
(18)
9
where
κ = − 1
β = −2 ρµδR2
b = 2 ρδR
I eδc.
I
I (σ(c) + m L g)
eδc
(19)
Note that κ and β are negative for any positive choice of the system parameters, and for any
value of co-contraction c. Thus the system is always locally asymptotically stable in the origin.
In the following we consider the problem of trajectory tracking in joint space. Thus the output
function is
y = x1 .
(20)
5.3 Simulations
The average male forearm weight is m = 1.36Kg, and the average male distance of the center of
gravity from the elbow joint is L = 0.155m [19]. The moment of inertia results approximately
I = 0.0109 Kg m2. We consider the instantaneous lever arm R = 1.5cm. The gravity is
approximated to the second digit 9.81 N m
s2 . For the muscle characteristic we consider values in
[3], [12]: µ = 0.06s, δ = 0.112mm−1, ρ = 1N. Thus
κ = −(4.62 e0.112mm−1c + 189.72) 1
β = −(0.277 e0.112mm−1c) 1
b = 308 e0.112mm−1c 1
s
s2
m s2 .
(21)
We consider the following desired swing movement qd = π
4 sin(t). We consider as reference
model the same as before but 10 times faster, with a response that peaks in 50 ms.
Figure 6: Results of the control of swing movement in a human upper limb, controlled through
the proposed cerebellar inspired algorithm. On the left side trajectories are presented, on the
right the learning of parameters.
Note that the system (17) presents many aspects making the problem of controlling it
very hard, i.e. it includes exponential and trigonometric actions, the activation terms are not
derivable with continuity, and the system can not be written in an affine control form. So
the effectiveness of a linear controller in a non-local task is a result not trivial to achieve.
We rely here on the natural inspiration of the controller, and on its robust structure already
demonstrated in the previous simulations. Fig. 6 presents the simulation results. The algorithm
is able to learn the correct control action to track the reference.
10
0204060−50050angle (degrees)time (secs) rrmy050100150−0.04−0.0200.020.040.060.08time (s)gains k1k2kr 6 Discussion
The cerebellum is a crucial brain structure for accurate motor control. It is phylogenetically old
and conserved through evolution in all vertebrates [13]. Much of the interest that the cerebellum
gathered in the machine learning, robotics and adaptive control communities stemmed from its
remarkable anatomy [8], with a general connectivity layout that resembles very closely the
one of a supervised learning neural network, such as a Perceptron [20]. However, here we
have drawn inspiration for a recent insight regarding cerebellar physiology that has emerged
simultaneously at both the theoretical [14] and experimental [21] domains. That is, that in
order to solve appropriately the problem of motor control, neurons from a same type (i.e.,
the cerebellar Purkinje cells) might display different learning rules in different areas of the
cerebellum, matched to the particular behavioral needs [7]. From a control theory perspective
those behavioral needs correspond to the transport latencies and response dynamics associated
to the controlled sub-system.
Here we have shown that the model-enhanced least-mean-squares (ME-LMS) learning rule
can be easily derived for the task of learning the optimal feedback gains of a fully known plant.
Second, we have shown that the ME-LMS learning rule converges in a series of tasks in which
conventional LMS would fail, as is the case of a non-minimum phase plant.
Regarding the derivation of ME-LMS presented here, it is worth noting that although a
similar result was originally obtained in [14] using a cerebellar-based control architecture, the
derivation presented here applies to two very general control architectures; namely, proportional
full-state-feedback and proportional error-feedback control. Hence, in that sense, the current
derivation is cerebellar-independent.
7 Conclusions and Future Work
In this work we proposed a control algorithm in which a linear feedback action is combined
with an adaptation rule for the control gains inspired from cerebellar learning in animals. The
controller is also analytically derived as gradient descent estimation of feedback gains for LTI
systems.
We tested the effectiveness of the algorithm in three different simulation scenarios, including
the control of a model of human upper limb, closing the loop with the biological inspiration.
The algorithm presented better performance w.r.t. the classic LMS rule. In future work we
plan to experimentally test it on bio-inspired robotic systems.
Although the algorithm presented in practice a stable behavior, to analytically prove the
stability of the closed loop system is a challenging task. This is due not only to the non-
linearities introduced by the possibility of adapting control gains (common in the context of
adaptive control), but also to the strong interplay between the three dynamics involved: the
system, the eligibility trace, and the control gains. However we consider this step very important
for the full formalization of the proposed learning rule, and so we depute this study to future
works.
Acknowledgments
The research leading to these results has been supported by the European Commission's Hori-
zon 2020 socSMC project (socSMC-641321H2020-FETPROACT-2014), the European Research
Council's CDAC project (ERC-2013-ADG 341196), the European Commission's Grant no.
11
H2020-ICT-645599 SOMA: SOft MAnipulation and the ERC Advanced Grant no. 291166
SoftHands.
References
[1] James S Albus. A theory of cerebellar function. Mathematical Biosciences, 10(1):25–61, 1971.
[2] Js Albus. A new approach to manipulator control: The cerebellar model articulation controller
(CMAC). Journal of Dynamic Systems, Measurement, and Control, (SEPTEMBER):220–227,
1975.
[3] David G Asatryan and Anatol G Feldman. Functional tuning of the nervous system with control
of movement or maintenance of a steady posture: I. mechanographic analysis of the work of the
joint or execution of a postural task. Biophysics, 10(5):925–934, 1965.
[4] Karl Johan Astrom. Theory and applications of adaptive control??a survey. Automatica, 19(5):471–
486, 1983.
[5] Douglas A Bristow, Marina Tharayil, and Andrew G Alleyne. A survey of iterative learning
control. IEEE Control Systems, 26(3):96–114, 2006.
[6] Paul Dean, John Porrill, Carl-Fredrik Ekerot, and Henrik Jorntell. The cerebellar microcircuit
as an adaptive filter: experimental and computational evidence. Nature reviews. Neuroscience,
11(1):30–43, jan 2010.
[7] Conor Dempsey and Nathaniel B Sawtell. The timing is right for cerebellar learning. Neuron,
92(5):931–933, 2016.
[8] J.C. Eccles, M. Ito, and J. Szent´agothai. The cerebellum as a neuronal machine. Springer Berlin,
1967.
[9] Anatol G Feldman and Mindy F Levin. The origin and use of positional frames of reference in
motor control. Behavioral and brain sciences, 18(04):723–744, 1995.
[10] M Fujita. Adaptive filter model of the cerebellum. Biological cybernetics, 45(3):195–206, 1982.
[11] Manolo Garabini, Cosimo Della Santina, Matteo Bianchi, Manuel Catalano, Giorgio Grioli, and
Antonio Bicchi. Soft robots that mimic the neuromusculoskeletal system. In Converging Clinical
and Engineering Research on Neurorehabilitation II, pages 259–263. Springer, 2017.
[12] Paul L Gribble, David J Ostry, Vittorio Sanguineti, and Rafael Laboissi`ere. Are complex control
signals required for human arm movement? Journal of Neurophysiology, 79(3):1409–1424, 1998.
[13] Sten Grillner. Control of locomotion in bipeds, tetrapods, and fish. Comprehensive Physiology,
2011.
[14] Ivan Herreros, Xerxes Arsiwalla, and Paul Verschure. A forward model at purkinje cell synapses
facilitates cerebellar anticipatory control. In Advances in Neural Information Processing Systems,
pages 3828–3836, 2016.
[15] M. Kawato, Kazunori Furukawa, and R. Suzuki. A hierarchical neural-network model for control
and learning of voluntary movement. Biological Cybernetics, 57(3):169–185, 1987.
[16] Alexander Lenz, Sean R Anderson, Anthony G Pipe, Chris Melhuish, Paul Dean, and John Porrill.
Cerebellar-inspired adaptive control of a robot eye actuated by pneumatic artificial muscles. IEEE
Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 39(6):1420–1433, 2009.
[17] D Marr. A theory of cerebellar cortex. The Journal of physiology, 202(2):437–470, 1969.
[18] Jun Nakanishi and Stefan Schaal. Feedback error learning and nonlinear adaptive control. Neural
Networks, 17(10):1453–1465, 2004.
[19] Stanley Plagenhoef, F Gaynor Evans, and Thomas Abdelnour. Anatomical data for analyzing
human motion. Research quarterly for exercise and sport, 54(2):169–178, 1983.
[20] Frank Rosenblatt. The perceptron: A probabilistic model for information storage and organization
in the brain. Psychological review, 65(6):386, 1958.
12
[21] Aparna Suvrathan, Hannah L Payne, and Jennifer L Raymond. Timing rules for synaptic plasticity
matched to behavioral function. Neuron, 92(5):959–967, 2016.
[22] Bernard Widrow, Marcian E Hoff, et al. Adaptive switching circuits. In IRE WESCON convention
record, volume 4, pages 96–104. New York, 1960.
[23] Emma D Wilson, Tareq Assaf, Martin J Pearson, Jonathan M Rossiter, Sean R Anderson, John
Porrill, and Paul Dean. Cerebellar-inspired algorithm for adaptive control of nonlinear dielectric
elastomer-based artificial muscle. Journal of The Royal Society Interface, 13(122):20160547, 2016.
[24] Daniel M Wolpert, RC Miall, and Mitsuo Kawato. Internal models in the cerebellum. Trends in
cognitive sciences, 2(9):338–347, 1998.
13
|
1804.09667 | 1 | 1804 | 2018-04-25T16:31:53 | Open(G)PIAS: An open source solution for the construction of a high-precision Acoustic-Startle-Response (ASR) setup for tinnitus screening and threshold estimation in rodents | [
"q-bio.NC"
] | The acoustic startle reflex (ASR) that can be induced by a loud sound stimulus can be used as a versatile tool to, e.g., estimate hearing thresholds or identify subjective tinnitus percepts in rodents. These techniques are based on the fact that the ASR amplitude can be suppressed by a pre-stimulus of lower, non-startling intensity, an effect named pre-pulse inhibition (PPI). For hearing threshold estimation, pure tone pre-stimuli of varying amplitudes are presented before an intense noise burst serving as startle stimulus. The amount of suppression of the ASR amplitude as a function of the pre-stimulus intensity can be used as a behavioral correlate to determine the hearing ability. For tinnitus assessment, the pure-tone pre-stimulus is replaced by a gap of silence in a narrowband noise background, a paradigm termed GPIAS (gap-pre-pulse inhibition of the acoustic startle response). A proper application of these paradigms depend on a reliable measurement of the ASR amplitudes, an exact stimulus presentation in terms of frequency and intensity. Here we introduce a novel open source solution for the construction of a low-cost ASR setup for the above mentioned purpose. The complete software for data acquisition and stimulus presentation is written in Python 3.6 and is provided as an anaconda package. Furthermore, we provide a construction plan for the sensory system based on low-cost hardware components. Exemplary data show that the ratios (1-PPI) of the pre and post trauma ASR amplitudes can be well described by a lognormal distribution being in good accordance to previous studies with already established setups. Hence, the open access solution described here will help to further establish the ASR method in many laboratories and thus facilitate and standardize research in animal models of tinnitus or hearing loss. | q-bio.NC | q-bio | Open(G)PIAS: An open source solution for the
construction of a high-precision Acoustic-Startle-Response
(ASR) setup for tinnitus screening and threshold estimation
in rodents
8
1
0
2
r
p
A
5
2
]
.
C
N
o
i
b
-
q
[
1
v
7
6
6
9
0
.
4
0
8
1
:
v
i
X
r
a
Richard Gerum1∗, Hinrich Rahlfs2,3∗, Matthias Streb2,3∗, Patrick Krauss2,
Claus Metzner1, Konstantin Tziridis2, Michael Günther3,
Holger Schulze2, Walter Kellermann3, Achim Schilling2
1 Department of Physics, Center for Medical Physics and Technology, Biophysics Group,
Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany
2 Experimental Otolaryngology, ENT-Hospital, Head and Neck Surgery, Friedrich-Alexander
University Erlangen-Nürnberg (FAU), Erlangen, Germany
3 Multimedia Communications and Signal Processing, Friedrich-Alexander University Erlangen-
Nürnberg (FAU), Erlangen, Germany
∗ authors contributed equally to this work
Keywords:
celeration sensor
low-cost setup, anaconda package, tinnitus, animal model, startle, 3D ac-
Abstract
The acoustic startle reflex (ASR) that can be induced by a loud sound stimulus can be used as a
versatile tool to, e.g., estimate hearing thresholds or identify subjective tinnitus percepts in rodents.
These techniques are based on the fact that the ASR amplitude can be suppressed by a pre-stimulus
of lower, non-startling intensity, an effect named pre-pulse inhibition (PPI). For hearing threshold
estimation, pure tone pre-stimuli of varying amplitudes are presented before an intense noise burst
serving as startle stimulus. The amount of suppression of the ASR amplitude as a function of the
pre-stimulus intensity can be used as a behavioral correlate to determine the hearing ability. For
tinnitus assessment, the pure-tone pre-stimulus is replaced by a gap of silence in a narrowband noise
background, a paradigm termed GPIAS (gap-pre-pulse inhibition of the acoustic startle response).
The rationale of this approach is that the presence of a tinnitus percept leads to a masking of
that gap ("filling in" hypothesis) leading to a reduced PPI in the affected frequency range. A
proper application of these paradigms depend on a reliable measurement of the ASR amplitudes,
an exact stimulus presentation in terms of frequency and intensity. Here we introduce a novel open
source solution for the construction of a low-cost ASR setup for the above mentioned purpose. The
complete software for data acquisition and stimulus presentation is written in Python 3.6 and is
provided as an anaconda package. Furthermore, we provide a construction plan for the sensory
system based on low-cost hardware components. Exemplary data show that the ratios (1-PPI) of
the pre and post trauma ASR amplitudes can be well described by a lognormal distribution being
in good accordance to previous studies with already established setups. Hence, the open access
solution described here will help to further establish the ASR method in many laboratories and
thus facilitate and standardize research in animal models of tinnitus or hearing loss.
1
Open(G)PIAS
Introduction
A behavioral paradigm that can be used to
assess hearing abilities in animal models (i.e.
behavioral audiometry) without the necessity to
apply time-consuming conditioning paradigms is
the so called pre-pulse inhibition of the acoustic
startle reflex (PPI of ASR or PIAS). The startle
reflex is induced by an intense stimulus such as
a loud tone. This reflex, processed in the brain-
stem, is an evolutionary adaptation to prevent
subjects from harm, e.g., by sudden predators
[Koch, 1999]. The reaction to a loud tone, the
acoustic startle reflex (ASR), can be modulated
by a variety of different factors such as drug
treatment, pathological conditions, or presen-
tation of pre-stimuli [Koch, 1999]. A decrease
of the ASR amplitude caused by any kind of
pre-stimulus is called pre-pulse inhibition (PPI)
and indicates that the stimulus has actually
been perceived [Fendt et al., 2001]. Hence, e.g.
hearing thresholds can be determined by acoustic
pre-stimuli such as pure-tones of varying inten-
sity
[Tziridis et al., 2012, Walter et al., 2012].
Furthermore,
in 2006, Turner and coworkers
suggested a novel paradigm using the ASR as a
tool for tinnitus screening [Turner et al., 2006].
The paradigm is based on the fact that not
only a tone or noise pre-stimulus can reduce
the startle response but also a gap of silence
embedded in band pass filtered noise can lead
to a suppression of the ASR. This paradigm is
called "Gap-Pre-Pulse Inhibition of the Acoustic
Startle Reflex" (GPIAS). Thus, a potential
tinnitus percept leads to a masking of this gap of
silence and consequently results in a diminished
decrease of the startle amplitudes (decrease of
PPI) [Turner et al., 2006].
It is advantageous
that this paradigm does not depend on any
pre-training of the animals such as classical
conditioning (e.g.
[Jastreboff et al., 1988]).
time consuming compared to
As
conditioning
paradigm is
widely used in the tinnitus research commu-
Shore et al., 2016,
nity
Krauss et al., 2016, Pienkowski, 2018].
Thus,
[Kalappa et al., 2014,
paradigms,
this
it
is
less
Gerum et al.
are
continuously
evaluation
the GPIAS paradigm and the
advanced and
procedures
[Longenecker and Galazyuk, 2012,
improved
Schilling et al., 2017]. However, setups for the
recording of ASR responses for tinnitus screening
as well as hearing ability estimation are still quite
expensive. Our aim is to provide an open source
solution for the construction of an ASR setup
that is written in the interpreted programming
language Python and is based on commercially
available hardware components.
For clarification of the method, we also pro-
vide some testing results in our animal model,
the Mongolian gerbil (Meriones unguiculatus).
Materials and Methods
Animals and Housing
The Mongolian gerbils were housed in standard
animal racks (Bio A.S. Vent Light, Ehret Labor-
und Pharmatechnik, Emmendingen, Germany)
in groups of 3 -- 4 animals with free access to
water and food at a room temperature of 20 --
25 ◦C under a 12h/12h dark/light circle. The
care of the animals was approved by the state
of Bavaria (Regierungspräsidium Mittelfranken,
Ansbach, Germany, No. 54-2532.1-02/13). Ex-
emplary measurements were recorded using an-
imals aged ten to twelve weeks purchased from
Janvier Laboratories Inc.
Software
The complete software for the ASR setup is writ-
ten in Python 3.6 [Rossum and Drake, 1995], an
interpreted programming language optimized for
scientific purposes. For maximum efficiency sev-
eral open source libraries are used. For numerical
operations, such as matrix operations, the Numpy
library is used [van der Walt et al., 2011]. Fur-
ther complex mathematical operations, such as
signal filter functions, are implemented using the
SciPy package [Olifant, 2007]. The data are visu-
alized using the Matplotlib library [Hunter, 2007].
2
Open(G)PIAS
Gerum et al.
band noise or pure tones of low amplitudes. To
protect the broadband loudspeaker from damage
by loud stimuli, a second loudspeaker (Neo-25s,
Sinuslive) is used to present the startle stimuli.
The startle stimuli causing the animal to twitch
(ASR-amplitudes) are 20 ms noise bursts, with an
intensity of 115 dB SPL, proven to be the op-
timal stimulus to induce an ASR response (cf.
[Turner et al., 2006]). This same startle stimulus
is used for all paradigms (e.g. threshold or GPIAS
measurements). All stimuli are presented via a
soundcard (Asus Xonar STX II) connected to two
pre-amplifier (Amp 75 for pre-stimuli, Amp 74 for
startle stimuli, Thomas Wulf, Frankfurt).
The recording and digitization of the analog
signal of the acceleration sensor (ADXL 335 on
GY 61 board, Robotpark) is performed by a
data acquisition card (PCIe-6320, National In-
struments) connected to the sensor via a breakout
box (NI-BNC 2110). The synchronization of star-
tle stimulus onset and ASR response is assured by
a trigger pulse generated by the soundcard, sent
directly to the data acquisition card. The rising
edge of the trigger-pulse is used to align the mea-
sured acceleration values to the startle stimulus
onset.
The sensor platform consists of two plates (cf.
Fig. 2a1,4). The lower plate is fixed to the vibra-
tion isolated table (TMC, Peabody, MA, USA) by
four screws. The upper plate is flexibly mounted
on the lower plate by four springs, damped by
two foam rubber blocks. The upper plate has a
mount for the animal restrainer. This restrainer
consists of an acrylic tube which is closed on both
ends, with a wire mesh at the front end, facing
the speakers, and a plastic cap (custom 3D print)
at the rear end. The wire mesh causes no mea-
surable distortions in the spectra of the acoustic
stimuli and the cap prevents the animal from es-
caping the restrainer. The fixed mount ensures
a constant distance between the animal and the
loudspeakers.
As described above, ASR amplitudes are quan-
tified by a three-way acceleration sensor, fixed
underneath the upper plate of the sensor plat-
form. When the animal twitches, as a response
3
Figure 1 ASR measurement setup
Stimuli are presented via two different loudspeakers,
one for the pre-stimulus and one for the startle stimu-
lus, both placed in an anechoic chamber. Both loud-
speakers are connected via an amplifier (Amp 74 and
Amp 75) to the soundcard of the computer. The
animal is restrained in an acrylic tube, located on
a sensor platform in the anechoic chamber. ASR
amplitudes are measured via an acceleration sensor.
Sound and accelerations are recorded using a data ac-
quisition card connected to the PC over a breakout
box. Our open source ASR program (Python) con-
trols stimulus application and measurement record-
ing.
Setup
The setup is located in an anechoic cham-
ber
(Industrial Acoustics Company GmbH,
Niederkrüchten, Germany) and consists of two
main parts:
the stimulation hardware and the
recording hardware (see Fig. 1). The animal is
restrained in an acrylic tube (different inner diam-
eter: 27 mm, 37 mm, or 42 mm, depending on the
size of the animal) placed on a sensor platform
with an integrated acceleration sensor (ADXL
335 on GY 61 board). The animal is acousti-
cally stimulated using two different loudspeakers.
A broadband two-way loudspeaker (Canton Plus
XS.2) presents the pre-stimuli such as narrow-
anechoic chamberanimalsensor platformpre-stimulusamplifierAmp 74startle-stimulusamplifierAmp 75breakout-boxNI-BNC-2110soundcardcomputerNI-Daqcard Open(G)PIAS
Gerum et al.
Stimulation Software
The software consists of several parts, a configu-
ration module that allows to adapt the software
to the given hardware setting, a protocol genera-
tor that allows to prepare the stimuli that should
be presented, and a measurement module that
applies the specified stimuli to the animal and
records the responses.
Configuration
The configuration module allows to specify de-
tails of the used hardware: The soundcard and
sound driver as well as the channels for trigger
pulse, pre-stimulus, and startle-stimulus can be
specified.
The configuration module also allows to per-
form some calibration measurements, e.g.,
for
synchronizing the output channels of the sound-
card or equalizing the frequency response of the
loudspeakers.
To calibrate possible latency shifts, a TTL
pulse is presented in all selected output channels
and the results are recorded using the data ac-
quisition card. The program uses these measure-
ments to determine the relative time shifts be-
tween the channels.
Since loudspeakers often exhibit a non-flat fre-
quency response, it is desirable to correct these
deviations using an equalizing filter. To this
end, a microphone with a flat frequency response
is placed at the position of the animal in the
measurement setup and the influence of the an-
imal itself on the sound field is emulated by a
piece of rubber foam. The coefficients of the
equalizer filter are determined by first identi-
fying the loudspeaker-enclosure-microphone sys-
tem (LEMS) and subsequently inverting its trans-
fer function in the frequency domain. The sys-
tem identification task, depicted schematically in
Fig. 3a, is solved by an adaptive linear filter us-
ing the NLMS algorithm [Widrow et al., 1976],
which iteratively minimizes the power of the er-
ror signal between the adaptive filter output Dn
and the observed microphone signal Yn for a
4
Figure 2 Sensor platform for ASR quantifica-
tion
a, Sensor platform (without animal restrainer). b,
Sensor platform (with animal restrainer). c, Photo
of the sensor platform in front to the loudspeakers.
The upper plate (1) of the sensor platform holds the
mount (2) for the animal restrainer (3). The re-
strainer consists of an acrylic tube (3a) fixed in the
mount, closed by a 3D printed plastic cap (3b), fixed
with a screw (3c), at the rear end, and a wire-mesh
(3d) at the front end, facing the loudspeaker. The ac-
celeration sensor is fixed underneath the upper plate.
This plate is flexibly mounted on the lower plate (4)
via four springs (5). The lower plated is screwed to
the vibration isolated table. Rubber foam (6) to pre-
vents the sensor system from oscillating. The sensor
platform is placed in front of the two speakers: pre-
stimulus (7) and startle-stimulus loudspeaker (8).
to the startle stimulus, the upper plate moves.
This movement is quantified by the acceleration
sensor. The ASR amplitude is
A = maxt(a(t) · Θ(t) · Θ(150 ms − t)),
(1)
where Θ is the Heaviside function, t = 0 the
begin of the startle stimulus and
(cid:113)
a(t) =
(cx · ax(t))2 + (cy · ay(t))2 + (cz · az(t))2,
(2)
with ax(t), ay(t), az(t), the measured acceleration in
x, y, z direction and cx, cy, cz the calibrated val-
ues, so that same force leads to same acceleration.
b63a3c3b3da1245c78 Open(G)PIAS
Gerum et al.
known excitation signal Xn. To allow the iden-
tification of all frequencies of interest, the ex-
citation signal should exhibit a constant power
spectral density in the relevant frequency range
as provided, e.g.
by a white noise sequence
or pseudo-random maximum-length sequences
[Rife and Vanderkooy, 1989]. Finally, the time-
domain coefficients of the equalizer filter are ob-
tained by frequency bin-wise inversion of the iden-
tified transfer function over the desired frequency
range and subsequent application of the inverse
discrete Fourier transform (DFT), realized by a
Fast Fourier Transform (FFT). By convolving the
loudspeaker driving signals with the equalization
filter before playback, a flat frequency response
for the overall system, i.e.
the cascade of the
equalization filter and the loudspeaker-enclosure-
microphone system can be achieved as depicted
in Fig. 3b.
Protocols
The protocol module allows to define measure-
ment protocols for different paradigms. Currently
two types of protocols are implemented: hear-
ing threshold measurements (based on PIAS) and
GPIAS measurements, but the system is flexible
to allow the future implementation of additional
protocols.
All measurement sessions consist of multiple
trials, each ending with a startle-stimulus. At the
beginning of each measurement session, five tri-
als without pre-stimulus are presented to prevent
adaptation effects in the response amplitudes of
the animal during the subsequent trials.
For hearing threshold determination, some tri-
als have a pre-stimulus in the form of a short
(40 ms) pure tone 100 ms before the startle stim-
ulus. For the protocol, a frequency range (with
octave, 1/2 octave, or 1/4 octave steps), a sound
pressure level range, and a trial repetition count
can be specified. The number of trials with pre-
stimulus is the same as the number of trials with-
out a pre-stimulus.
For the GPIAS paradigm, each trial has a noise
presented before the startle-stimulus. The noise
Figure 3 Equalizer Scheme
a, The block diagram of the system identification
task; The output signal Xn of the soundcard is fed
into the "LEMS" (speaker) and recorded by the mi-
crophone (subject to additive measurement noise).
The same signal is fed into the adaptive filter H and
the difference signal en is used to adjust the filter
coefficients. This iterative process leads to a filter
that mimics the transfer function of the "unknown
system". b, The frequency response of the overall
system without equalization (blue) and with equal-
ization (red). The frequency response is flattened in
the range from 2 kHz to 20 kHz, where the microphone
exhibits a flat frequency response.
can be broadband or a frequency band around a
center frequency. For each measurement different
center frequencies can be specified, as well as the
number of measurement repetitions. The noise
can be interrupted by a short (50 ms, flattened
with 20 ms sin2-ramps) gap of silence. The num-
ber of trials with a gap is the same as the number
of trials without a gap.
The trials in each protocol are randomized to
prevent habituation effects of the animal.
5
01000020000frequency (Hz)604020020squared magnitude (dB)frequency responsebefore equalizationafter equalizationb HLEMSnoise-+XnXnDnYnena Open(G)PIAS
Gerum et al.
is plotted to provide direct feedback to the exper-
imenter.
The data is stored in a folder structure accord-
ing to the meta data to allow for an organized
storage. Data is saved in the form of numpy files
(.npy) for space efficient storage, but can be ex-
ported to .txt or excel files for later evaluations.
Results
In the following section, we show that exemplary
recorded data using the novel setup is consistent
with measurements from already established se-
tups.
Fig. 5 shows the ASR amplitudes for one exem-
plary animal stimulated according to the GPIAS
paradigm (30 trials, center frequency 2,000 Hz ±
1/2 octave).
The data is in good agreement with results from
a previous study conducted with another setup
[Schilling et al., 2017, Tziridis et al., 2012], show-
ing that ASR amplitudes are not normally dis-
tributed, but the startle responses can be well
described by a lognormal distribution (Fig. 5).
Also, the ratios of gap and no gap amplitudes
are lognormally distributed (Fig. 5c), i.e. the log-
arithmic ratios are almost perfectly normally dis-
tributed (cf. Fig. 5d).
The exemplary data clearly show that the here
described measure for the ASR amplitudes based
on the 3D acceleration vector (Equ. 1) is a valid
choice and leads to the same quantitative results
as the 1D-peak-to-peak amplitudes used before.
The average of the logarithm of the ratios is nega-
tive (Fig. 5d) which means that the gap of silence
leads to a pre-pulse inhibition, thus demonstrat-
ing the validity of the method.
Discussion
In this study, we present an open source setup
for the measurement of acoustic startle reflexes
evoked by a loud noise burst with preceding pure-
tone and/or gap pre-stimuli for the estimation of
6
Figure 4 Measurement Interface
The interface of the measurement program. The dif-
ferent tabs provide access to the different modules.
The measurement module provides inputs for the
meta data, a status of the components, a log of the
measurement, and a plot of the data from the last
measurement.
Measurement
The main part of the software is the measure-
ment module. This module checks whether both
the soundcard and data acquisition card are con-
figured properly and are ready to play or record,
respectively. Furthermore, it checks the selected
protocol file to be valid.
Before starting the measurement, the user has
to supply some meta data, the experimenter
name, the animal name, and optionally a treat-
ment, to allow for consistent storage of the data.
When everything is ready and the animal has
been placed in the restrainer, the measurement
can be started. Each trial is presented to the an-
imal using three channels of the sound card (trig-
ger pulse, pre-stimulus, and startle-stimulus) and
recorded using the data acquisition card (three
input channels from the acceleration sensor and
three input channels directly from the sound-
card). Data is saved after each trial and raw-data
Open(G)PIAS
Gerum et al.
Figure 5 Distribution of ASR amplitudes.
Distributions of the ASR amplitudes measured as a response to GPIAS paradigm stimuli, with a center
frequency of 2 kHz and a spectral width of ±1/2 octaves. a, ASR amplitudes without a gap of silence (blue).
b, ASR amplitudes with gap stimuli (red). Both data sets are fitted with a lognormal distribution (solid line)
using the maximum likelihood estimator. Histogram of the full combinatorial of the ratios of gap (Agap) and
no gap (Anogap) ASR-amplitudes. c, Ratio histogram (green bars) can be described well with a lognormal
distribution (solid line) as we have shown in a previous study [Schilling et al., 2017]. d, The logarithmized
ratios (green bars) are therefore Gaussian distributed (solid line). The mean of the distribution is negative,
indicating that on average the pre-pulse inhibited the response.
hearing thresholds and identification of possible
tinnitus percepts in rodents.
The pre-stimuli and startle stimuli
(noise
bursts) are presented using two separate loud-
speakers, controlled by a commercial soundcard.
The response amplitudes of the animals are cap-
tured by an 3D-acceleration sensor, whose out-
put is digitized using a data acquisition card con-
trolled via a Python program.
We have shown that the described setup can be
used to present well defined pure-tone and band
noise stimuli and to quantify the ASR amplitudes.
Additionally, we were able to show that the pre-
stimuli lead to a clear PPI (cf. Fig. 5c) and that
the ASR amplitudes (cf.
[Csomor et al., 2008])
as well as the ratios of gap and no gap ampli-
tudes (cf. [Schilling et al., 2017]) are lognormally
distributed.
In contrast to existing, more elaborate and
costly systems, our setup requires only low-cost
hardware components and, with open-source soft-
ware, allows for more flexible adoptions to differ-
ent measurement protocols.
All our hardware components are listed in de-
tail and the software is provided as open source
(hardware and software documentation: http:
//open-gpias.readthedocs.io).
References
[Csomor et al., 2008] Csomor, P. A., Yee, B. K.,
Vollenweider, F. X., Feldon, J., Nicolet, T., and
Quednow, B. B. (2008). On the Influence of
Baseline Startle Reactivity on the Indexation of
Prepulse Inhibition. Behavioral Neuroscience,
122(4):885 -- 900.
[Fendt et al., 2001] Fendt, M., Li, L., and Yeo-
mans, J. S. (2001). Brain stem circuits medi-
7
051015probability densityfitdatano gapgapratio0.00.20.40.6ASR-Amplitude (m/s2)051015probability density012345Agap/AnoGap0.00.20.40.60.81.0probability density21012log(Agap/AnoGap)0.00.20.40.6probability densityabcd Open(G)PIAS
Gerum et al.
ating prepulse inhibition of the startle reflex.
Psychopharmacology, 156(2-3):216 -- 224.
[Hunter, 2007] Hunter, J. D. (2007). Matplotlib
: A 2D Graphics Environment. Computing in
science & engineering, 9(3):90 -- 95.
[Jastreboff et al., 1988] Jastreboff, P. J., Bren-
nan, J. F., Coleman, J. K., and Sasaki, C. T.
(1988). Phantom Auditory Sensation in Rats:
An Animal Model for Tinnitus. Behavioral
Neuroscience, 102(6):811 -- 822.
[Kalappa et al., 2014] Kalappa, B. I., Brozoski,
T. J., Turner, J. G., and Caspary, D. M. (2014).
Single unit hyperactivity and bursting in the
auditory thalamus of awake rats directly cor-
relates with behavioural evidence of tinnitus.
Journal of Physiology, 592(22):5065 -- 5078.
[Koch, 1999] Koch, M. (1999). The neurobiology
of startle. Progress in Neurobiology, 59(2):107 --
128.
[Krauss et al., 2016] Krauss, P., Tziridis, K.,
Buerbank, S., Schilling, A., and Schulze, H.
(2016). Therapeutic Value of Ginkgo biloba
Extract EGb 761 R(cid:13) in an Animal Model
(Meriones unguiculatus) for Noise Trauma In-
duced Hearing Loss and Tinnitus. Plos One,
11(6):e0157574.
[Longenecker and Galazyuk, 2012] Longenecker,
R. J. and Galazyuk, A. V. (2012). Method-
ological optimization of tinnitus assessment
using prepulse inhibition of the acoustic startle
reflex. Brain Research, 1485:54 -- 62.
[Olifant, 2007] Olifant, T. E. (2007). Python for
Scientific Computing. Computing in science &
engineering, 9(3):10 -- 20.
[Pienkowski, 2018] Pienkowski, M. (2018). Can
long-term exposure to non-damaging noise lead
to hyperacusis or tinnitus ? In Proceedings of
the International Symposium on Auditory and
Audiological Research, volume 6, pages 83 -- 94.
[Rife and Vanderkooy, 1989] Rife, D. D. and
Vanderkooy, J.
Transfer-function
measurement with maximum-length sequences.
the Audio Engineering Society,
Journal of
37(6):419 -- 444.
(1989).
[Rossum and Drake, 1995] Rossum, G. V. and
Drake, F. L. (1995). Python Reference Man-
ual.
[Schilling et al., 2017] Schilling, A., Krauss, P.,
Gerum, R., Metzner, C., Tziridis, K., and
Schulze, H. (2017). A New Statistical Approach
for the Evaluation of Gap-prepulse Inhibition
of the Acoustic Startle Reflex (GPIAS) for Tin-
nitus Assessment. Frontiers in Behavioral Neu-
roscience, 11(October):1 -- 12.
[Shore et al., 2016] Shore, S. E., Roberts, L. E.,
Langguth, B., Physiology, I., and Clinic, I. T.
(2016). Maladaptive plasticity in tinnitus-
triggers, mechanisms and treatment Susan. Na-
ture Reviews Neurology, 12(3):150 -- 160.
[Turner et al., 2006] Turner, J. G., Brozoski,
T. J., Bauer, C. a., Parrish, J. L., Myers, K.,
Hughes, L. F., and Caspary, D. M. (2006).
Gap Detection Deficits in Rats With Tinnitus:
A Potential Novel Screening Tool. Behavioral
Neuroscience, 120(1):188 -- 195.
[Tziridis et al., 2012] Tziridis, K., Ahlf, S., and
Schulze, H. (2012). A low cost setup for behav-
ioral audiometry in rodents. JoVE, 68.
[van der Walt et al., 2011] van der Walt, S., Col-
bert, S. C., and Varoquaux, G. (2011). The
NumPy Array: A Struture for Efficient Numer-
ical Computation. Computing in Science {&}
Engeneering, 13(2):22 -- 30.
[Walter et al., 2012] Walter, M., Tziridis, K.,
Ahlf, S., and Schulze, H. (2012). Context De-
pendent Auditory Thresholds Determined by
Brainstem Audiometry and Prepulse Inhibition
in Mongolian Gerbils. Open Journal of Acous-
tics, 2:34 -- 49.
8
Open(G)PIAS
Gerum et al.
[Widrow et al., 1976] Widrow, B., Mc Cool,
J. M., Larimore, M. G., and Johnson, R. C.
(1976). Stationary and Nonstationary Learn-
ing Characteristics of the LMS Adaptive Filter.
PROCEEDINGS OF THE IEEE, 64(8):1151 --
1162.
9
|
1603.04687 | 1 | 1603 | 2016-03-14T15:00:12 | Recurrent Network Models Of Sequence Generation And Memory | [
"q-bio.NC",
"cond-mat.dis-nn",
"physics.bio-ph"
] | Sequential activation of neurons is a common feature of network activity during a variety of behaviors, including working memory and decision making. Previous network models for sequences and memory emphasized specialized architectures in which a principled mechanism is pre-wired into their connectivity. Here we demonstrate that, starting from random connectivity and modifying a small fraction of connections, a largely disordered recur- rent network can produce sequences and implement working memory efficiently. We use this process, called Partial In-Network Training (PINning), to model and match cellular resolution imaging data from the posterior parietal cortex during a virtual memory- guided two-alternative forced-choice task. Analysis of the connectivity reveals that sequences propagate by the cooperation between recurrent synaptic interactions and external inputs, rather than through feedforward or asymmetric connections. Together our results suggest that neural sequences may emerge through learning from largely unstructured network architectures. | q-bio.NC | q-bio | TITLE
Recurrent Network Models Of Sequence Generation And Memory
AUTHORS
Kanaka Rajan, Christopher D Harvey and David W Tank
AFFILIATIONS
Joseph Henry Laboratories of Physics and Lewis–Sigler Institute for Integrative Genomics,
Princeton University, Princeton, NJ.
Department of Neurobiology, Harvard Medical School, Boston, MA.
Department of Molecular Biology and Princeton Neuroscience Institute, Princeton University,
Princeton, NJ.
CORRESPONDENCE
KR: [email protected]
CDH: [email protected]
DWT: [email protected]
The sequential activation of neurons is a common feature of network activity during a variety
SUMMARY
of behaviors and has been proposed as a mechanism for cortical computation, including short
term memory. Previous modeling approaches for sequences and memory networks have
emphasized highly specialized architectures in which a principled mechanism is pre-wired into
the connectivity of the network. Here, we demonstrate that starting from random synaptic
connectivity and allowing a small fraction of connections to undergo modification, a largely
disordered recurrent network can produce sequences and short-term memory. We use this
process, which we call Partial In-Network training (PINning), to model and match data from
cellular-resolution imaging of neural activity in the mouse posterior parietal cortex (PPC) during
a memory-guided two-alternative forced choice task in a virtual environment [Harvey, Coen &
Tank, 2012]. In the model, as in the PPC data, individual neurons exhibit transient activations
that are staggered relative to one another in time to form sequences spanning the duration of
the task, and different sequences are activated on trials with different cues and choices.
Analysis of the connectivity matrices of the minimally structured model networks revealed that
the time-ordered neural activity is produced by the cooperation between recurrent synaptic
interactions and external inputs, rather than feedforward connections, or the asymmetric
connections of ring attractor models for sequences. In addition, our model showed that
sequential activation across a population of neurons is an efficient mechanism for implementing
short-term memory with comparable memory capabilities to previously proposed fixed point
mechanisms. Together our results develop a new modeling framework based on generic,
minimally modified networks and suggest that neural activity sequences may emerge through
learning from largely unstructured network architectures.
!1
Sequential firing has emerged as a prominent motif of population activity in several
INTRODUCTION
experiments involving temporally structured behaviors, such as short-term memory and decision
making. Neural sequences have been observed in many brain regions including the cortex
[Luczak et al, 2007; Schwartz & Moran, 1999; Andersen et al, 2004; Pulvermutter & Shtyrov,
2009; Buonomano, 2003; Ikegaya et al, 2004; Tang et al, 2008; Seidemann et al, 1996;
Fujisawa et al, 2008; Crowe et al, 2010; Harvey, Coen & Tank, 2012], hippocampus [Nadasdy et
al, 1999; Louie & Wilson, 2001; Pastalkova et al, 2008; Davidson, Kloosterman & Wilson, 2009],
basal ganglia [Barnes et al, 2005; Jin, Fuji & Graybiel, 2009], cerebellum [Mauk & Buonomano,
2004], and area HVC of the songbird [Hahnloser, Kozhevnikov & Fee, 2002; Kozhevnikov &
Fee, 2007]. In all these cases, the observed sequences span a wide range of time durations,
but individual neurons fire transiently only during a small portion of the full sequence. The
ubiquity of neural sequences in different brain regions suggests that they are of widespread
functional use, and that they may be produced by general circuit-level mechanisms.
more generic circuits adapted through the learning of a specific task. Highly structured circuits of
this type have a long history [Kleinfeld & Sompolinsky, 1989; Goldman, 2009], for example, as
synfire chain models [Hertz & Prugel-Bennett,1996; Levy et al, 2001; Hermann, Hertz & Prugel-
Bennet, 1995; Fiete et al, 2010], in which excitation flows unidirectionally from one active
neuron to the next along a chain of connected neurons, or as ring attractor models [Yishai, Bar-
Or & Sompolinsky, 1995; Zhang, 1996], in which increased ("central") excitation between nearby
neurons surrounded by long-range inhibition and asymmetric connectivity are responsible for
time-ordered neural activity. Constructing these models typically involves implementing a task-
specific mechanism (for producing sequences, for instance) into the synaptic connectivity,
producing highly specialized networks. Neural circuits are highly adaptive and involved in a wide
variety of tasks, and furthermore, sequential neural activity often emerges through the learning
of a specific task and retains significant variability. It is therefore unlikely for highly structured
approaches to produce models with flexible circuitry or to generate dynamics with the temporal
complexity needed to make a connection with experimental recordings of neural sequences.
Broadly speaking, sequences can be produced by highly structured neural circuits or by
In contrast, random networks of model neurons interconnected with excitatory and inhibitory
connections in a balanced state [Sompolinsky, Crisanti & Sommers, 1988], rather than being
specifically designed for one single task, have been modified by training to perform a variety of
tasks [Buonomano & Merzenich, 1995; Buonomano, 2005; Williams & Zipser, 1989; Pearlmutter,
1989; Jaeger & Haas, 2004; Maass, Joshi & Sontag, 2007; Sussillo & Abbott, 2009; Maass,
Natschlager & Markram, 2002; Jaeger, 2003]. Here, we built on these lines of research and
asked whether a general implementation using relatively unstructured random networks could
!2
create sequential neural dynamics resembling experimental data. We used data from
sequences observed in the posterior parietal cortex (PPC) of mice trained to perform a two-
alternative forced-choice (2AFC) task in a virtual reality environment [Harvey, Coen & Tank,
2012], and also constructed models that extrapolated beyond these experimental data.
To address how much network structure is required to support sequences like those
observed in the neural recordings, we introduced a novel modeling framework called Partial In-
Network Training or PINning. In this scheme, any desired fraction of the initially random
connections within the networks we construct can be modified by a synaptic change algorithm,
enabling us to explore the full range of networks between completely random and fully
structured. Using network models constructed by PINning, we first demonstrated that
sequences resembling the PPC data are most consistent with minimally structured neural
circuitry, with small amounts of structured connectivity that supports sequential activity patterns
and a much larger fraction of unstructured connections. Next, we investigated the circuit-level
mechanism of sequence generation in largely random networks containing some learned
structure. Finally, we determined the role modeled sequences play in short-term memory, for
instance, by storing information during the delay period about whether a left or right turn was
indicated early in the 2AFC task [Harvey, Coen & Tank, 2012]. Going beyond models meant to
reproduce the experimental data, we also analyzed multiple sequences initiated by different
sensory cues and computed the capacity of this form of short-term memory.
RESULTS
1. Sequences from highly structured or random networks do not match PPC data
Networks of rate-based model neurons, in which the outputs of individual neurons are
characterized by firing rates and units are interconnected through excitatory and inhibitory
synapses of various strengths (Experimental Procedures 1), form the basis of our studies. To
interpret the outputs of the rate networks we construct in terms of experimental data, we extract
firing rates from the calcium fluorescence signals recorded in the PPC using two complementary
deconvolution methods (Figures 1C and D show calcium data from [Harvey, Coen & Tank, 2012]
and the rates extracted from these data, respectively; see also Supplemental Figure 1 and
Experimental Procedures 3). We define two measures to compare the rates from the model to
the rates extracted from data. The first, bVar, measures the stereotypy of the data or the
network output by quantifying the variance that is explained by the translation along the
sequence of an activity profile with an invariant shape (Figure 1G and Experimental Procedures
6). The second metric, pVar, quantifies the percent variance of the experimental data from the
PPC [Harvey, Coen & Tank, 2012] that is captured by the outputs of the different networks we
!3
build and is therefore useful for tracking their performance as a function of network parameters
(Figure 1H and Experimental Procedures 7). bVar and pVar are used throughout this paper (for
the results in Figures 1–6).
Many models have suggested that highly structured synaptic connectivity, i.e., containing
ring-like or chain-like interactions, in networks is responsible for neural sequences [Yishai, Bar-
Or & Sompolinsky, 1995; Zhang, 1996; Hertz & Prugel-Bennett,1996; Levy et al, 2001;
Hermann, Hertz & Prugel-Bennet, 1995; Fiete et al, 2010]. We therefore first asked how much
of the variance in long-duration neural sequences seen experimentally, for instance, in the PPC
[Harvey, Coen & Tank, 2012], was consistent with a bump of activity moving across the network,
as expected from such highly structured connectivity. To quantify this, we used bVar. The
stereotypy of PPC sequences was found to be quite small – bVar = 40% for Figures 1D and 2A,
which were averaged over hundreds of trials and pooled across different animals; in fact, bVar
was lower in both single trial data (10–15% for data in Supplemental Figure 13 from [Harvey,
Coen & Tank, 2012]) and trial-averaged data from a single mouse (15% for the data in Figure 2c
from [Harvey, Coen & Tank, 2012]). The relatively small fraction of the variance explained by a
moving bump (low bVar), combined with a weak relationship between the activity pattern of a
neuron and anatomical location in the PPC data (Figure 5d in [Harvey, Coen & Tank, 2012]),
motivated us to consider network architectures with disordered connectivity composed of a
balanced set of excitatory and inhibitory weights drawn independently from a random
distribution (Experimental Procedures 1).
ongoing dynamics have been shown to be chaotic [Sompolinsky, Crisanti & Sommers, 1988];
however, the presence of external stimuli can channel the ongoing dynamics in random network
models by suppressing their chaos [Molgedey, Schuchhardt & Schuster, 1992; Bertschinger &
Natschlager, 2004; Rajan, Abbott & Sompolinsky, 2010; Rajan, Abbott & Sompolinsky, 2011].
Furthermore, in experiments, strong inputs have been shown to reduce the Fano factor and
trial-to-trial variability associated with spontaneous activity [Churchland et al, 2010; White,
Abbott & Fiser, 2011]. Thus we asked whether PPC-like sequences could be constructed either
from the spontaneous activity or the input-driven dynamics of such random networks. To test
this, we first simulated a random network of firing rate-based model neurons operating in a
chaotic regime (schematized in Figure 1A, described in Experimental Procedures 1, N = 437,
network size chosen to match the size of the dataset under consideration). The individual firing
rates were normalized by the maximum over the duration of each trial (10.5s here, consistent
with Figure 1C) and sorted in ascending order of their time of center-of-mass (tCOM), matching
the procedures applied to the real data [Harvey, Coen & Tank, 2012]. Although the resulting
ordered chaotic spontaneous activity was sequential (not shown, but similar to Figure 1B), the
In random network models, when excitation and inhibition are balanced on average, the
!4
level of extra-sequential background activity (bVar = 5 + 2% and pVar = 0.15 + 0.1%) was
higher than data. Sparsifying this background activity by increasing the threshold of the
sigmoidal activation function increased bVar to a maximum of 22%, still considerably smaller
than the data value of 40% (Supplemental Figure 8).
Next, we introduced external inputs to the random network that were time-varying to
represent the effects of the visual stimuli in the virtual environment (a few example inputs are
shown in the right panel of Figure 1A, see also Experimental Procedures 2). Another sequence
was obtained by normalizing and sorting the firing rates from this input-driven random network
(as with the spontaneous activity and with the data [Harvey, Coen & Tank, 2012]), but this too
did not match the PPC data, bVar = 10 + 2% and pVar = 0.2 + 0.1% (Figure 1B, see also left
panel of Figure 1H and Supplemental Figure 8).
resembling the data (Figure 1B, see also Supplemental Figure 3), which are more structured
and temporally constrained (Figures 1C–D and 2A, compared to Figure 1B). Therefore,
sequences like those observed during timing and memory experiments [Harvey, Coen & Tank,
2012] are unlikely to be an inherent property of completely random networks. Furthermore,
since real neural sequences arise during the learning of various experimental tasks, we asked
whether initially disordered networks could also be modified by training to produce realistic
sequences.
External inputs and completely disordered connectivity are insufficient to evoke sequences
2. Temporally Constrained Neural Sequences Emerge With Synaptic Modification
To construct networks that match the activity seen in the PPC, we developed a training
scheme, Partial In-Network Training (PINning), in which different sized subsets of synapses
were modified (Experimental Procedures 4). In our synaptic modification scheme, the inputs to
individual model neurons in the network were compared directly with target functions or
templates derived from real experimental data [Fisher et al, 2013], on both left and right correct-
choice outcome trials from 437 trial averaged neurons, pooled across 6 mice, during a 2AFC
task (Figure 2A and Figure 5D). During training, the internal synaptic weights in the connectivity
matrix of the recurrent network were modified using a variant of the recursive least-squares
(RLS) or or first-order reduced and controlled error (FORCE) learning rule [Haykins, 2002;
Sussillo & Abbott, 2009] until the network rates matched the target functions (Experimental
Procedures 4). Crucially, the learning rule was applied only to all the synapses of a randomly
selected and often small fraction of neurons. Only a fraction p (pN2 << N2) of the total number of
synapses in the network were modified (plastic synapses are depicted in orange in Figures 2B,
5A and 6A). In particular, while every neuron in the network had a target function, only the
!5
outgoing synapses from a subset of neurons (i.e., only the outgoing synaptic weights from pN
chosen neurons) were subject to the learning rule (schematized in Figure 2B, see also
Experimental Procedures 4). The remaining elements of the synaptic matrix remained
unmodified and in their randomly initialized state (random synapses are depicted in gray in
Figures 1A, 2B, 5A and 6A). The PINning method for building networks therefore provides a way
to span the entire spectrum of possible neural architectures – from disordered networks with
random connections (p = 0 for Figures 1A and B), through networks with partially structured
connectivity, in which only a small subset p of connections are task-specifically wired (p < 25%
for the examples in Figures 2, 5 and 6), while the majority of connections remain disordered, to
networks containing entirely trained connections (p = 100%, Figure 3D–F).
We first applied PINning to a network with templates obtained from a single PPC-like
sequence (Figure 2A–C) using a range of values for the fraction of plastic synapses, p.
Modification of only a small percentage (p = 12%) of the synaptic connections in an initially
disordered network (Figure 1F) was sufficient for its sequential outputs to become more
temporally constrained (bVar = 40%) and to match the PPC data with high fidelity (pVar = 85%,
Figure 2C, see also Supplemental Data 12 for results of a cross-validation analysis). Although
the networks shown here are typically as large as the size of the experimental dataset (N = 437
for Figures 1–2 and 569 for Figure 5), our results are consistent both for larger N networks and
for networks in which non-targeted neurons are included to simulate the effect of unobserved
but active neurons present in the experimental data (Supplemental Figure 7). The dependence
of pVar on p is shown in Figure 2D.
the relative fraction of synapses modified from their initially random values, but what is the
overall magnitude of synaptic change required to produce these sequences? As shown in
Figure 2E, we found that although the individual synapses changed more in sparsely PINned
(small p) networks, the total amount of change across the synaptic connectivity matrix was
smaller. We return to other implications of this issue in Section 3.
neurons in PINned networks, we used principal component analysis (PCA) (see for example,
[Rajan, Abbott & Sompolinsky, 2011; Sussillo, 2014], and references therein). For an untrained
random network (here, with N = 437 and p = 0) operating in a spontaneously active regime
(Experimental Procedures 1), the top 38 principal components accounted for 95% of the total
variance (therefore, the effective dimensionality, Qeff = 38) (gray circles in Figure 2F). In
comparison, Qeff of the data in Figure 1D is 24. In the network with p = 12% with outputs
matching the PPC data, the dimensionality was lower (Qeff = 14) (orange circles in Figure 2F).
Qeff asymptoted around 12 dimensions for higher values of p (inset of Figure 2F). The circuit
Figure 2D quantifies the amount of structure required for generating sequences in terms of
To uncover how the sequential dynamics are distributed across the population of active
!6
Two features are critical for the production of sequential activity in a neural circuit. The first is
To develop a simplified prototype for further investigation of the mechanisms of sequential
dynamics are higher dimensional for the output of the PINned network than for a sinusoidal
bump attractor but lower than for the data.
3. Circuit Mechanism For Sequential Activation Through PINning
activation, we created a synthetic sequence "idealized" from data, lasting as long as the
duration of the PPC sequence (10.5 seconds, Figure 2A). We first generated a Gaussian curve,
f(t) (green curve in the inset of Figure 1F and in the left panel of Supplemental Figure 4A) that
best fit the average extracted from the tCOM-aligned PPC data (Rave, red curve in the inset of
Figure 1F and in the left panel of Supplemental Figure 4A). The curve f(t) was then translated
uniformly over a 10.5s time period to derive a set of target functions (N = 500, right panel of
Supplemental Figure 4A, bVar = 100%). Increasing fractions of the initially random connectivity
matrix, p, were trained using PINning with the target functions of the idealized sequence. As
before, pVar increases as a function of p, reaches pVar = 92% at p = 8% (highlighted by a red
circle in Supplemental Figure 4B), asymptoting at p ~ 10%. The plasticity required for producing
the idealized sequence was therefore smaller than the p = 12% required for the PPC-like
sequence (Figure 2D), due to the lack of the idiosyncrasies present in the experimental data, for
example, irregularities in the temporal ordering of the individual transients and background
activity away from the sequence.
the formation of a subpopulation of active neurons ("bump"), maintained by excitation between
co-active neurons and restricted by inhibition. The second is an asymmetry in the synaptic
inputs from neurons ahead in the sequence and those behind, needed to make the bump move.
We therefore looked for these two features in PINned networks by examining both the structure
of their connectivity matrices and the synaptic currents into individual neurons.
In a classic moving ring/bump attractor network, the synaptic connectivity is only a function
of the "distance" between pairs of network neurons in the sequence, i – j, assuming the neurons
are labelled in order of their appearance in the sequence (for our analysis of the connectivity in
PINned networks, we follow the same approach and order the neurons in a similar manner).
Furthermore, the connectivity that sustains, constrains and moves the bump is all contained in
the connectivity matrix, which is localized in i – j and asymmetric. We looked for similar
structure in the trained network models.
We considered 3 connectivity matrices interconnecting a population of model neurons, N =
500 – before PINning (the randomly initialized matrix denoted by JRand), after sparse PINning
(JPINned, 8%), and after full PINning (JPINned, 100%, built as a useful comparison). To analyze how
the synaptic strengths in these 3 matrices varied with i – j, we first computed the means and the
!7
In the p = 100% model, the band-averages of JPINned, 100% formed a localized and
standard deviations of the diagonals, and the means and the standard deviations of successive
off-diagonal "bands" moving away from the diagonals, i.e., i – j = constant (Experimental
Procedures 11). These band-averages and the fluctuations around them were plotted as a
function of the interneuronal distance, i – j (Figures 3B and 3E). Next, we generated "synthetic"
interaction matrices, in which all the elements along each diagonal were replaced by their band-
averages and fluctuations, respectively. Finally, these synthetic matrices were used in networks
of rate-based model neurons driven by the same external inputs as the original PINned
networks (Experimental Procedures 2).
asymmetric profile (orange circles in Figure 3E) and produce a moving "Gaussian bump" of
activity (right panel of Figure 3G) that is qualitatively similar to moving ring attractor dynamics
[Yishai, Bar-Or & Sompolinsky, 1995; Zhang, 1996]. On the other hand, the band-averages for
JPINned, 8% (orange circles in Figure 3B) exhibited a localized zone of excitation for small values
of i – j that was symmetric, a significant inhibitory self-interactive or autaptic feature at i – j = 0,
and relatively diffuse flanking inhibition for larger values of i – j. This is reminiscent of the
features expected in the interaction matrices of stationary bump models [Yishai, Bar-Or &
Sompolinsky, 1995]. Furthermore, neither the band-averages of JPINned, 8% by themselves
(shown in Figure 3C) nor the fluctuations by themselves (not shown), were sufficient to produce
moving sequences similar to the output of a network containing the full matrix JPINned, 8%.
Instead, the outputs of the synthetic networks built from the components of the band-averaged
JPINned, 8% were stationary bumps (right panel of Figure 3C). In this case, what causes the bump
to move?
To address this, we considered the fluctuations around the band averages of the sparsely
PINned connectivity matrices, JPINned, 8%. As expected (from the color-bars in Figures 3A, the
lines in Figures 3B, and Figure 3F), the fluctuations around the band-averages of JPINned, 8% (red
lines in Figure 3B) were much larger and more structured than those of JRand (small blue lines in
Figure 3B). To uncover the mechanistic role of these fluctuations, we examined the input to
each neuron produced by the sum of the fluctuations of JPINned, 8% around the band-averages
and the external input. We realigned the sums of the fluctuations and the external inputs for all
the neurons in the network by their tCOM and then averaged over neurons (see Experimental
Procedures 6 for an example of a similar procedure). This yielded an aligned population
average (bottom right panel in Figure 4) that clearly revealed the asymmetry responsible for the
movement of the bump across the network. Therefore, in the presence of external inputs that
are constantly changing in time, the mean synaptic interactions do not have to be asymmetric,
as we observed for JPINned, 8% (Figure 3B). Instead, the variations in the fluctuations of JPINned, 8%
(i.e., after the mean has been subtracted) and the external inputs create the asymmetry that
!8
moves the bump along. It is difficult to visualize this asymmetry at an individual neuron level
because of fluctuation, necessitating this type of population-level measure. While the mean
synaptic interactions in sparsely PINned networks cause the formation of the localized bump of
excitation, it is the non-trivial interaction of the fluctuations in these synaptic interactions with the
external inputs that causes the bump to move across the network. Therefore, this is a novel
circuit mechanism for non-autonomous sequence propagation in a network.
Additionally, we looked at other PINning-induced trends in the elements of JPINned, 8% more
directly by plotting the synaptic weight values directly (Supplemental Figure 5A). The changed
elements in JPINned, 8% (orange dots) were scattered away from the identity line and skewed
= 40,000 of these, corresponding to the 8%
toward more negative values. There were !
pN 2
1/ N
, Experimental Procedures 1), 0 skewness and 0
plastic synaptic weights, while the other 92% of the weights remained unchanged from their
initial random values (the eigenvalue spectra are shown in Supplemental Figure 5B). For all the
3 connectivity matrices of sequential networks, there was a substantial increase in the
magnitudes spanned by the synaptic weights as p decreased (color-bars in Figures 3A and 3D),
These increases in magnitude were manifested in PINned networks built with different p values
in different ways (Figure 3F). The initial weight matrix, JRand had 0 mean (by construction),
0.005 variance (generally of order !
kurtosis. The partially structured matrix, JPINned, 8%, on the other hand, had a negative mean of –
0.1, variance of 2.2, skewness at –2, and kurtosis of 30, all of which were indicative of a
probability distribution that was asymmetric about 0 and had heavy tails from a small number of
strong weights. This corresponds to a network in which the large sequence-facilitating synaptic
changes come from a small fraction of the weights, as suggested in experimental
measurements [Song et al, 2005]. In JPINned, 8%, the ratio of the size of the largest synaptic
weight to the size of the "typical" is ~20. If we assume the typical synapse corresponded to a
post-synaptic potentiation (PSP) of 0.05mV, then the "large" synapses had a 1mV PSP. This is
within the range in which existing experimental data support the plausibility of the network [Song
et al, 2005]. For comparison purposes, the connectivity matrix for a fully structured network,
JPINned, 100%, had a mean = 0, variance = 0.7, skewness = –0.02, and kurtosis = 0.2,
corresponding to a network in which the synaptic changes responsible for sequences were
numerous and distributed throughout the network.
sensitive to small amounts of structural noise, i.e., perturbations in the matrix JPINned, 8%.
However, when stochastic noise (described in Experimental Procedures 2) is used during
training, slightly more robust networks are obtained (Supplemental Figure 11).
Finally, we determined that synaptic connectivity matrices obtained by PINning are fairly
!9
4. Delayed Paired Association And Working Memory Can Be Implemented Through
Sequences In PINned Networks
Delayed Paired Association (DPA) tasks, such as the two-alternative forced-choice (2AFC)
task from [Harvey, Coen & Tank, 2012], engage working memory during each trial because the
mouse must remember the identity of the cue, or cue-associated motor response, as it runs
through the T-maze. DPA tasks also engage declarative memory across trials, because the
mouse must remember whether each cue stimulus is associated with a left or a right turn.
Therefore, in addition to being behavioral paradigms that produce sequential neural activity
[Harvey, Coen & Tank, 2012], DPA tasks are useful for exploring the different neural correlates
of short-term memory [Gold and Shadlen, 2007; Brunton et. al., 2013; Hanks et. al., 2015; Amit,
1995; Amit & Brunel,1995; Hansel & Mato, 2001; Hopfield & Tank, 1985; Shadlen & Newsome,
2001; Harvey, Coen & Tank, 2012]. By showing that the partially structured network we
constructed by PINning could accomplish a 2AFC task, we argued that sequences can mediate
an alternative form of short term memory.
During the first third of the 2AFC experiment [Harvey, Coen & Tank, 2012], the mouse
received either a left or a right visual cue, and during the last third, it had two different
experiences depending on whether it made a left turn or a right turn. Therefore, we modeled the
cue period and the turn period of the maze by two different time-varying inputs. In the middle
third, when the left- or right-specific visual cues were off (blue and red traces in Figure 5C), the
mouse ran through a section of the maze that was visually identical for both types of trials for a
duration equal to the delay period. In our simulation of the 2AFC task, the inputs to individual
network neurons coalesced into the same time-varying waveform during the delay periods of
both the left and the right trials (purple traces in Figure 5C). The correct execution of this type of
task therefore depended on the network generating more than one sequence – in this case, a
left sequence or a right sequence, which maintained the memory of the identity of the visual cue
during a period in the task when the sensory inputs were identical.
A network with only p = 16% plastic synapses generated outputs that were consistent with
experimental data (Figures 5D and E, pVar = 85%, bVar = 40%, also compare with Figure 2c in
[Harvey, Coen & Tank, 2012], here, N = 569, 211 network neurons selected at random to
activate in the left trial condition, schematized in blue in Figure 5A; 226 to fire in the right
sequence, red in Figure 5A; and the remaining 132 to fire in the same order in both left and right
sequences, non-choice-specific neurons, depicted in green in Figure 5A). This network retained
the memory of cue identity by silencing the left preferring network neurons during the delay
period of a right trial, and the right preferring network neurons, during a left trial, and generating
sequences with the active neurons. Non-choice-specific neurons, on the other hand, were
sequentially active in the same order in trials of both types, like real no-preference PPC neurons
!10
observed experimentally (outputs shown in Figure 5E, see also Supplemental Figure 7b in
[Harvey, Coen & Tank, 2012]).
To compare the task performance of the three types of memory networks under
Are sequences a comparable alternative to fixed point models commonly used for storing
5. Comparison With Fixed Point Memory Networks
memories? To test this idea, we compared two types of sequential-memory networks with a
fixed point memory network (Figure 5F, see also Supplemental Figure 6). As before, different
fractions of plastic synapses, controlled by p, were embedded by PINning against different
target functions (Experimental Procedures 3 and 4). Here, these targets represented the values
of a variable being stored and were chosen based on the dynamical mechanism by which
memory is implemented – idealized for the sequential memory network (orange in Figure 5F,
based on Supplemental Figure 4A), firing rates extracted from PPC data [Harvey, Coen & Tank,
2012] for the PPC-like delayed paired association network (green in Figure 5F, outputs from the
PPC-like DPA network with p = 16% plasticity are shown in Figure 5E), and constant valued
targets for the fixed point-based memory network (blue in Figure 5F, see also Supplemental
Figure 6).
consideration here, we computed a selectivity index (Experimental Procedures 9, similar to the
one used in Figure 4 in [Harvey, Coen & Tank, 2012]). We found that the network exhibiting
long-duration population dynamics and memory activity through idealized sequences (orange
triangles in Figure 5F) had a selectivity = 0.91, when only 10% of its synapses were modified by
PINning (p = 10%). In comparison, the PPC-like DPA network (whose outputs are shown in
Figure 5E) needed p = 16% of its synapses to be structured to match data [Harvey, Coen &
Tank, 2012] and to achieve a selectivity = 0.85 (shown in the green triangles in Figure 5F). We
compared the performance of fixed point memory networks relative to both these sequential
memory networks, and found that both sequential memory networks performed comparably with
the fixed point network in terms of their selectivity-p relationship. The fixed point memory
network achieved an asymptotic selectivity of 0.81 for p = 23%. The magnitude of synaptic
change required (Experimental Procedures 8) was also comparable between sequential and
fixed point memory networks, suggesting therefore that sequences may be a viable alternative
to fixed points as a mechanism for storing memories in neural circuits.
During experiments, the fraction of trials on which the mouse makes a mistake is about
15-20% (accuracy of the performance of mice at the 2AFC task was found to 83 ± 9% correct
[Harvey, Coen & Tank, 2012]). We interpret errors as arising from trials in which PPC delay
period activity failed to retain the identify of the cue leading to chance performance at the time
that the animal makes a turn. Given a 50% probability of turning in the correct direction by
!11
chance, this implies that the cue identity is forgotten on 30-40% of trials. Adding noise to our
model, we can reproduce this level of delay period forgetting with a noise amplitude of
ηc = 0.4 − 0.5
of noise levels where the model shows a fairly abrupt decrease in performance.
(Supplemental Figure 10B). It should be noted that this noise value is in a region
Thus far, we have discussed one specific instantiation of delayed paired association through
6. Capacity Of Sequential Memory Networks
a sequence-based memory mechanism – the 2AFC task. Next, we extended the same basic
PINning framework for constructing networks, to generate and maintain several sequential
patterns of population activity. We asked whether PINned networks could accomplish memory-
based tasks that required the activation of multiple, i.e., >2, non-interfering sequences.
Additionally, we computed the capacity of such multi-sequential networks (denoted by Ns) as a
function of the parameters of the networks we model, specifically, network size (N), PINning
fraction (p), fraction of non-choice-specific neurons (NNon-choice-specific/N), and temporal
sparseness (fraction of sequential neurons active at any instant in the task, characterized by
NActive/N, Experimental Procedures 10).
We adjusted the PIN parameters to simulate a memory task mediated by multiple non-
interfering sequences (Figure 6). Different fractions of synaptic weights in an initially random
network were trained by PINning to match different target functions (schematized in Figure 6A),
non-overlapping sets of Gaussian curves (similar to Supplemental Figure 4A), each of which
was evenly spaced so that, collectively, they spanned a duration of 8s. The width of these
waveforms, defined in terms of NActive/N, was varied as a parameter that controlled the
sparseness of the sequence (Experimental Procedures 10). Because the turn period is omitted
here in the interest of clarity, the total duration of multi-sequential memory tasks modeled here is
8s-long. Similar to the DPA task modeled before (Figure 5), each network neuron received a
different filtered white noise input for each cue during the cue period (0–4s), but during the delay
period (4–8s), these inputs coalesced to a common cue-invariant waveform, albeit a different
one for each neuron. In the most general case, we assigned N/Ns neurons to each sequence
that we wanted the network to produce. Here, Ns is the number of memories (capacity), which
also equals the number of "trial types" or the number of "cue preferences". Once again, only p%
of the synapses in the network were plastic.
8s), network neurons fire in a sequence only on trials of the same type as their cue preference,
and are silent during other trials. For example, the set of neurons selective for Cue 1 fires in a
sequence only during Trial Type 1 and not during Trial Types numbering 2 – Ns, neurons
A correctly executed multi-sequential task is one in which during the delay period (here 4–
!12
!
selective for Cue 2 activate only during Trial Type 2, and so on (schematized in Figure 6A). A
network of 500 neurons with p = 25% plastic synapses performed such a task easily, generating
Ns = 5 sequences with delay period memory of the appropriate cue identity with temporal
sparseness of NActive/N = 3% of the total number of network neurons (Figure 6B).
Temporal sparseness of the sequences was found to be a crucial factor that determined how
well a network performed a multi-sequential memory task (Experimental Procedures 10). When
sequences were forced to be sparser than a certain minimum (by using narrower target
waveforms, for this particular 8s-long task, this occurred when NActive/N < 1.6%) the network
failed. Although there was sequential activation, the memory of the identity of the 5 separately
memorable cues was not maintained across the delay period. Surprisingly, we found that this
failure could be rescued by adding a small number of non-choice-specific neurons (NNon-choice-
specific/N = 4%) that fired in the same temporal order in all 5 trial types (Figure 6C). In other
words, the remaining N – NNon-choice-specific network neurons (< N) now successfully executed the
memory-based multi-sequential task using 5 sparse choice-specific sequences and 1 non-
choice-specific sequence. Thus, the capacity of the network for producing multiple temporally
sparse sequences increased when non-choice-specific neurons were present to stabilize them,
without requiring a concomitant increase in the amount of synaptic modification. Therefore, we
predict that non-choice-specific neurons, also seen in the PPC [Harvey, Coen & Tank, 2012],
may function as a "conveyor belt" of working memory, providing recurrent synaptic current to
sequences that may be too sparse to sustain themselves otherwise in a non-overlapping
scheme.
presence of non-choice-specific neurons increased the capacity of networks to store memories
if they were implemented through sequences (Figure 6C). Up to a constant factor, the memory
capacity, the maximum of the fraction Ns/N, scaled in proportion to the fraction of plastic
synapses, p and the network size, N, and inversely with the sparseness, NActive/N (Figure 6D).
The slope of this capacity-to-network-size relationship increased when non-choice-specific
neurons were included because they enabled the network to carry sparser sequences.
Furthermore, we found that networks with a bigger fraction of plastic synaptic connections and
networks containing non-choice-specific neurons were more stable against stochastic
perturbations (Figure 6E). Once the amplitude of added stochastic noise exceeded the
maximum tolerance of a particular network (denoted by !
Finally, we focused on two of these aspects – memory capacity and noise tolerance. The
ηc
, Experimental Procedures 2),
however, non-choice-specific neurons were no longer effective at repairing the memory capacity
(not shown); non-choice-specific neurons could not rescue inadequately PINned multi-
sequential schemes either (small p, not shown).
!13
In this paper, we used and extended earlier work based on liquid-state [Maass, Natschlager
DISCUSSION
& Markram, 2002] and echo-state machines [Jaeger, 2003], which has shown that a basic
balanced-state network with random recurrent connections can act as a general purpose
dynamical reservoir, the modes of which can be harnessed for performing different tasks by
means of feedback loops. These models typically compute the output of a network through
weighted readout vectors [Buonomano & Merzenich, 1995; Maass, Natschlager & Markram,
2002; Jaeger, 2003; Jaeger & Haass, 2004] and feed the output back into the network as an
additional current, leaving the synaptic weights within the dynamics-producing network
unchanged in their randomly initialized configuration. The result was that networks generating
chaotic spontaneous activity prior to learning [Sompolinsky, Crisanti & Sommers, 1988],
produced a variety of regular non-chaotic outputs that match the imposed target functions after
learning [Sussillo & Abbott, 2009]. The key to making these ideas work is that the feedback
carrying the error during the learning process forces the network into a state in which it
produces less variable responses [Molgedey, Schuchhardt & Schuster, 1992; Bertschinger &
Natschlager, 2004; Rajan, Abbott & Sompolinsky, 2010; Rajan, Abbott & Sompolinsky, 2011]. A
compelling example, FORCE learning [Sussillo & Abbott, 2009], has been implemented as part
of several modeling studies [Laje & Buonomano, 2013; Sussillo, 2014], and references therein),
but it had a few limitations. Specifically, the algorithm in [Sussillo & Abbott, 2009] included a
feedback network separate from the network generating the actual trajectories, which contained
aplastic connections in a black-box, and learning was restricted only to a set of readout weights.
The approach we developed here, called Partial In-Network Training or PINning, avoids these
issues, while propagating neural sequences resembling data (Sections 1 and 2). During
PINning, only a fraction of the recurrent synapses in an initially random network are modified in
a task-specific manner and by a biologically reasonable amount, leaving the majority of the
network heterogeneously wired. Furthermore, the fraction of plastic synaptic connections in
these networks is a tunable parameter that controls the contribution of the structured portion of
the network relative to the random part, allowing these models to interpolate smoothly between
highly structured and completely random architectures until we find the point that best matches
the relevant experimental data (a similar point is made in Barak et al, 2013). PINning is not the
most general way to restrict learning to a limited number of synapses, but it allows us to do so
without losing the efficiency of the learning algorithm.
To illustrate the applicability of this framework for constructing partially structured networks,
we used experimental data recorded from the posterior parietal cortex (PPC) of mice performing
a working memory-based decision making task in a virtual reality environment [Harvey, Coen &
Tank, 2012]. It is unlikely that a high-order association cortex such as the PPC evolved
!14
specialized circuitry solely for sequence generation, especially considering that the PPC
appears to retain its ability to mediate other complex temporal tasks, from working memory and
decision making to evidence accumulation and navigation [Shadlen & Newsome, 2001; Gold &
Shadlen, 2007; Freedman & Assad, 2011; Snyder, Batista & Anderson, 1997; Anderson & Cui,
2009; Bisley & Goldberg, 2003; McNaughton, 1994; Nitz, 2006; Whitlock et al, 2008; Carlton &
Taube, 2009; Hanks et al, 2015]. We also computed a measure of stereotypy from PPC data
(bVar, Figure 1C–G) and found it to be much lower than if the PPC were to generate sequences
based on highly specialized intrinsic connectivity (for example, with chain-like or ring-like
connections) or if it merely read out sequential activity from a highly structured upstream region.
We therefore started with random recurrent networks, imposed a small amount of structure in
their connectivity through PINning and duplicated many features of the experimental data
[Harvey, Coen & Tank, 2012], primarily among them, choice-specific neural sequences and
retention of the memory of cue identity during the delay period.
We analyzed the the structural features in the synaptic connectivity matrix of sparsely
PINned networks, concluding that the probability distribution of the synaptic strengths is heavy
tailed due to the presence of a small percentage of strong interaction terms (Figure 3F). There
is experimental evidence that there might be a small fraction of very strong synapses embedded
within a milieu of a large number of relatively weak synapses in the cortex [Song et al, 2005],
most recently, from the primary visual cortex [Cossell & Mrsc-Flogel, 2015]. Synaptic
distributions measured in slices have been shown to have long tails [Song & Nelson, 2005], and
the experimental result in [Cossell & Mrsc-Flogel, 2015] has demonstrated that rather than
occurring at random, these strong synapses significantly contribute to network tuning by
preferentially interconnecting neurons with similar orientation preferences. These strong
synapses may be the plastic synapses that are induced by PINning in our scheme. The model
exhibits some structural noise sensitivity (Supplemental Figure 11), and this is not completely
removed by training in the presence of noise. It is possible that dynamic mechanisms of ongoing
plasticity could enhance stability to structural fluctuations.
excitation in the connectivity, and the manner in which the bump of excitation propagates across
the network, were elucidated. For the first, we quantified the influence of network neurons that
are away from the ones active in the sequence (Figure 3). We found that the mechanism for
bump formation in the networks constructed by sparse PINning is consistent with "center
surround"-like features in the mean synaptic interactions (Figure 3B, see also Figure 3C), as
expected [Yishai, Bar-Or & Sompolinsky, 1995; Zhang, 1996]. For the second, we analyzed the
collective impact of recurrent synaptic currents and external inputs onto an example neuron
active at one specific time point in the sequence (Figure 4). From these results, we concluded
In this paper, the circuit mechanisms underlying both the formation of a localized bump of
!15
that the circuit mechanism for the propagation of the sequence is non-autonomous, relying on a
complex interplay between the connections in these networks and the external inputs. The
mechanism for bump propagation is therefore distinct from the standard moving bump/ring
attractor model [Yishai, Bar-Or & Sompolinsky, 1995; Zhang, 1996], but has similarities to the
models developed in [Fiete et al, 2010; Hopfield, 2015] (explored in Supplemental Figure 9). In
particular, the model in [Fiete et al, 2010] successfully generates highly stereotyped and noise-
free sequences, similar to those observed in area HVC experimentally. This model is initialized
as a recurrently connected network similar to ours and subsequently uses spike-timing
dependent plasticity, heterosynaptic competition and correlated external inputs to learn
sequential activity. A critical difference between [Fiete et al, 2010] and the approach presented
here lies in the underlying network connectivity responsible for sequences – their learning rule
results in synaptic chain-like connectivity.
We suggest an alternative hypothesis that sequences might be a more general and effective
dynamical form of working memory (Sections 4 and 5), making the prediction that sequences
may be observed in many experiments involving diverse tasks that require working memory
(also suggested in the Discussion section of [Harvey, Coen & Tank, 2012]). This contrasts with
previous models of working memory that relied on fixed point attractors to retain information and
exhibited sustained activity [Amit, 1992; Amit, 1995; Amit & Brunel,1995; Hansel & Mato, 2001;
Hopfield & Tank, 1985]. Finally, we computed the capacity of sequential memory networks for
storing more than two memories (Section 6) by extending the same sparse PINning approach
developed for matching the data from a 2AFC experiment [Harvey, Coen & Tank, 2012] in
Section 4. Experiments on freely behaving animals performing more complex tasks, e.g., with
more than two contingencies, will likely show that more than two memories can be stored
through sequences. Furthermore, ongoing experiments in several laboratories are testing the
effect of switching the contextual meaning of multiple cues and cue-combinations, including
olfactory, visual, and multi-sensory cues, in basic memory-guided decision-making tasks and in
complex variants of such tasks. Analyzing population data from such studies will soon allow
experimental tests of the multi sequential network models constructed by PINning.
the "computational bandwidth" of a general-purpose neural circuit to perform different timing-
based computations. Here, we showed that a small amount of structured connections
embedded in a much larger skeleton of disordered connections is sufficient for sequential timing
signals of the order of 10s. We can also ask, what else the rest of the network can do. A network
could be channeled by sparse training to first perform two (or more) different temporally
structured tasks, for instance, accumulation of evidence ([Brunton, Botvinick & Brody, 2013;
Hanks et al, 2015], and references therein) and delayed paired association ([Harvey, Coen &
The capacity of the sequential memory networks constructed by PINning can be thought of
!16
The term "pre-wired" is used in this paper to mean a scheme in which a principled
Tank, 2012], and references therein), and then switch between these tasks in a context-
dependent manner. In the future, using the same analyses developed in this paper on such
"multi-purpose" networks could make biological predictions about circuit-level mechanisms
operating in areas such as the PPC that are implicated in both these, as well as in other tasks
[Shadlen & Newsome, 2001; Gold & Shadlen, 2007; Freedman & Assad, 2011; Snyder, Batista
& Anderson, 1997; Anderson & Cui, 2009; Bisley & Goldberg, 2003; McNaughton, 1994; Nitz,
2006; Whitlock et al, 2008; Carlton & Taube, 2009; Hanks et al, 2015].
mechanism for executing a certain task is first assumed, and then incorporated into the network
circuitry, for example, a moving bump architecture [Yishai, Bar-Or & Sompolinsky, 1995; Zhang,
1996], or a synfire chain [Hertz & Prugel-Bennett,1996; Levy et al, 2001; Hermann, Hertz &
Prugel-Bennet, 1995; Fiete et al, 2010]. In contrast, the models built and described in this paper
are constructed without bias or assumptions. If a moving bump architecture [Yishai, Bar-Or &
Sompolinsky, 1995; Zhang, 1996] had been assumed at the beginning and the network pre-
wired accordingly, we would of course have uncovered it through the analysis of the synaptic
connectivity matrix (similar to Section 3, Figures 3 and 4 and Supplemental Figure 9, see also
Experimental Procedures 11). However, by starting with an initially random configuration and
learning a small amount of structure, we found an alternative mechanism for input-dependent
sequence propagation (Figure 4). We would not have encountered this mechanism for the non-
autonomous movement of the bump by pre-wiring a different mechanism into the connectivity of
the model network. While the models constructed here are indeed trained to perform the task,
the fact that they are unbiased means that the opportunity was present to uncover mechanisms
that were not thought of a priori.
!17
EXPERIMENTAL PROCEDURES
13.1. Network elements
We consider a network of N fully interconnected neurons described by a standard firing rate model.
Each model neuron is characterized by an activation variable, !
for i = 1, 2, ... N, where, N = 437 for the
xi
PPC-like sequence in Figures 1D and 2A, N = 500 for the single idealized sequence in Supplemental
Figure 4 and the multi-sequential memory task in Figure 6, and N = 569 for the 2AFC task in Figure 5 (we
generally build networks of the same size as the experimental dataset we are trying to model, however
the results obtained remain applicable to larger networks, see for example, Supplemental Figure 7), and a
nonlinear response function, !
φ(x) =
. This function ensures that the firing rates, !
]
ri =φ(xi)
,
1
[
1+ e −(x−θ)
go from a minimum of 0 to a maximum at 1. Adjusting !
allows us to set the firing rate at rest, x = 0 , to
some convenient and biologically realistic background firing rate, while retaining a maximum gradient at
x =θ
We introduce a recurrent synaptic weight matrix J with element !
(but see also Supplemental Figure 8).
representing the strength of the
. We use !
θ= 0
θ
Jij
connection from presynaptic neuron j to postsynaptic neuron i (schematic in Figure 1A) The individual
synaptic weights are initially chosen independently and randomly from a Gaussian distribution with mean
and variance given by !
〈 Jij 〉J = 0
and !
〈Jij 〉J
2 = g2 / N
, and are either held fixed or modifiable,
depending on the fraction of plastic synapses p that can change by applying a learning algorithm
(Experimental Procedures 4).
The activation variable for each network neuron !
xi
is determined by,
!
dxi
dt = –xi +
τ
∑ φ(xj) + hi
Jij
N
j
.
τ
In the above equation, !
= 10ms is the time constant of each unit in the network and the control
parameter g determines whether (g > 1) or not (g < 1) the network produces spontaneous activity with
non-trivial dynamics [Sompolinsky, Crisanti & Sommers, 1988; Rajan, Abbott & Sompolinsky, 2010;
Rajan, Abbott & Sompolinsky, 2011]. We use g values between 1.2 and 1.5 for the networks in this paper,
so that the randomly initialized network generates chaotic spontaneous activity prior to activity-dependent
modification (gray trace in the schematic in Figure 2B), but produces a variety of regular non-chaotic
outputs that match the imposed target functions afterward (red trace in the schematic in Figure 2B, see
also results in Figures 2, 5 and 6, see also Experimental Procedures 4 later). The network equations are
is the external input to the unit i.
integrated using Euler method with an integration time step, dt = 1ms. !
hi
!18
!
13.2. Design of External Inputs
During the course of the real two-alternative forced-choice (2AFC) experiment [Harvey, Coen & Tank,
2012], as the mouse runs through the virtual environment, the different patterns projected onto the walls
of the maze (colored dots, stripes, pillars, hatches, etc.) translate into time-dependent visual inputs
arriving at the PPC. Therefore, to represent sensory (visual and proprioceptive) stimuli innervating the
PPC neurons, the external inputs to the neurons in the network, denoted by !
, are made from filtered
h(t)
and spatially delocalized white noise that is frozen (repeated from trial to trial), using the equation,
τWN
dh
dt = −h(t) + h0η(t),
where !
η
is a random variable drawn from a Gaussian distribution with 0 mean
and unit variance, and the parameters !
time, respectively. We use !
h0 =1
and !
h0
and !
τWN =1
control the scale of these inputs and their correlation
τWN
s. There are as many different inputs as there are model
neurons in the network, with individual model neurons receiving the same input on every simulated trial. A
few example inputs are shown in the right panel of Figure 1A.
In addition to the frozen noise h, which acts as external inputs to these networks, described above,
we also test the resilience of the memory networks we built (Figure 6E) to injected stochastic noise. This
stochastic injected noise varies randomly (i.e., is a Gaussian random variable between 0 and 1, drawn
from a zero mean and unit variance distribution) and independently at every time step. The diffusion
constant of the white noise is given by !
2 /2τ
Aη
, where the amplitude is !
Aη
2
and !
τ
is the time constant of
the network units (we use 10ms, as detailed in Experimental Procedures 1). We define "Resilience" or
, at which the delay
"Noise Tolerance" as the critical amplitude of this stochastic noise, denoted by !
ηc
period memory fails and the Selectivity Index of the memory network drops to 0 (Figure 6E, see also
Supplemental Figure 10).
To derive the target functions for our activity-dependent synaptic modification scheme termed Partial
13.3. Extracting Target Functions From Calcium Imaging Data
In-Network Training or PINning, we convert the calcium fluorescence traces from PPC recordings
[Harvey, Coen & Tank, 2012] into firing rates using two complementary methods. We find that for this
dataset, the firing rates estimated by the two methods agree quite well (Supplemental Figure 1).
The first method is based on the assumption that the calcium impulse response function, which is a
K ∝ e−t/384 − e−t/52
difference of exponentials (!
[Tian et al, 2009; Harvey, Coen & Tank, 2012]), is approximated by an alpha function of the form
K ∝ te−t/τCa
According to this assumption, the scaled firing rate s and calcium concentration, [Ca2+] are related by,
, where there is only a single (approximate) time constant for the filter, !
, with a rise time of 52ms and a decay time of 384ms
= 200ms.
τCa
!
τCa
dCai
dt = −Cai(t) + xi(t)
and !
τ
dxi
dt = −xi(t) + si(t),
!19
!
!
where, x(t) is an auxiliary variable.The inverse of the above model is obtained by taking a derivative of the
calcium data, writing,
!
xi(t) = Cai(t) +τCa
dCai
dt
and !
si(t) = xi(t) +τ
dxi
dt .
Once we have s(t), we rectify it and choose a smoothing time constant !
τR
for the firing rate we need to
compute. Finally, integrating the equation, !
τR
dRi
dt = −Ri(t) + si(t)
, and normalizing by the maximum
The second method is a fast Bayesian deconvolution algorithm [Pnevmatikakis et al, 2014; Vogelstein
gives us an estimate for the firing rates extracted from the calcium data, denoted by R (Supplemental
Figure 1).
et al, 2010, available online at https://github.com/epnev/continuous_time_ca_sampler] that infers spike
trains from calcium fluorescence data. The inputs to this algorithm are the rise time (52ms) and the decay
time (384ms) of the calcium impulse response function [Tian et al, 2009; Harvey, Coen & Tank, 2012] and
a noise parameter [Pnevmatikakis et al, 2014; Vogelstein et al, 2010]. Typically, if the frame rate for
acquiring the calcium images is low enough (the data in [Harvey, Coen & Tank, 2012] are imaged at 64ms
per frame), the outputs from this algorithm can be interpreted as a normalized firing rate. To verify the
accuracy of the firing rate outputs obtained from trial-averaged calcium data (for example, from Figure 2c
in [Harvey, Coen & Tank, 2012]), we smoothed the spike trains we got from the above method for each
trial separately through a Gaussian of the form !
−(t−ti )
2
2τR
e
∑
i
, normalized by!
2π×τR
, and then averaged
over single trials to get trial-averaged firing rates (this smoothing and renormalization procedure has also
been recommended for faster imaging times [Pnevmatikakis et al, 2014; Vogelstein et al, 2010]).
Once the values of !
τR
and !
are determined that make the results obtained by both deconvolution
τCa
τR
= 100ms and !
methods consistent (we used !
τCa
= 384ms), we used the firing rates extracted as target
functions for PINning through the transform,!
fi(t) = ln
Ri(t)
1− Ri(t)
⎡
⎢
⎣
⎤
⎥
⎦
. The above expression is obtained by
solving the activation function relating input current to firing rate of model neurons, !
Ri(t) =
1
1+ e− fi(t) ,
since the goal of PINning is to match the input to neuron i, say, denoted by !
zi(t)
to its target function,
denoted by!
fi(t)
. Finally, to verify our estimates, we re-convolved (Supplemental Figure 1) the output
firing rates from the network neurons with a difference of exponentials using a rise time of 52ms and a
decay time of 384ms.
!20
During PINning, the inputs of individual network neurons are compared directly with the target
13.4. Synaptic Modification Rule For PINning
functions to compute a set of error functions, i.e., !
inputs are expressed as
zi(t) =
∑ rj(t)
Jij
j
ei(t) = zi(t) − fi(t)
rj(t)
, for i = 1, 2, ... N. Individual neuron
, where
is the firing rate of the jth or the presynaptic
During learning, the subset of plastic internal weights in the connectivity matrix J of the random
neuron.
recurrent network, denoted by the fraction p, undergo modification at a rate proportional to the error term,
the presynaptic firing rate of each neuron, !
and a pN x pN matrix, P (with elements !
) that keeps
Pij
rj
track of the rate fluctuations across the network at every time step. Here, p is the fraction of neurons
whose outgoing synaptic weights are plastic; since this is a fully connected network, this is also the
fraction of plastic synapses in the network. Mathematically, !
Pij = < rirj >−1
, the inverse cross-correlation
matrix of the firing rates of the network neurons (Pij is computed for all i but is restricted to j = 1, 2, 3,
…,pN). The basic algorithm is schematized in Figure 2B. At time t, for i = 1, 2, ... N neurons, the learning
rule is simply that the elements of the matrix J are moved from their values at a time step !
through !
earlier
. Here, the synaptic update term, according to the RLS/FORCE
Jij(t) = Jij(t −1) + ΔJij(t)
Δt
procedure [Haykins, 2002; Sussillo & Abbott, 2009] (since other methods for training recurrent networks,
such as backpropagation would be too laborious for our purposes) follows,
ΔJij(t) = c[zi(t) − fi(t)] Pjk
where the above update term is restricted to the p% of plastic
∑ (t)rk(t),
k
synapses in the network, which are indexed by j and k in the above expression. While c can be thought of
as an effective learning rate, it is given by the formula, !
c =
. The only free parameter in
1
1+ ′r (t)P(t)r(t)
the learning rule is P(0) (but the value to which it is set is not critical [Sussillo & Abbott, 2009]). When
there are multiple sequences (such as in Figures 5 and 6), we choose pN synapses that are plastic and
we use those same synapses for all the sequences.
P(t) = P(t −1) −
The matrix P is generally not explicitly calculated but rather updated according to the rule,
in matrix notation, which includes a regularizer [Haykins,
P(t −1)r(t) ′r (t)P(t −1)
1+ ′r (t)P(t −1)r(t)
2002]. In our scheme, all indices in the above expression are restricted to the neurons with plastic
synapses in the network. The algorithm requires the matrix P to be initialized to the identity matrix times a
factor that controls the overall learning rate, i.e., !
, and in practice, values from 1 to 10 times
P(0) =α× I
the overall amplitude of the external inputs (denoted by h0 in Experimental Procedures 2) driving the
network are effective (other values are explored in [Sussillo & Abbott, 2009]).
!21
!
!
For numerically simulating the PINned networks whose sequential outputs are shown in Figures 2C, 5
and 6, the integration time step used is dt = 1ms (as described in Experimental Procedures 1 and 2, we
use Euler method for integration). The learning occurs at every time step for the p% of pre-synaptic
neurons with plastic outgoing synaptic weights. Starting from a random initial state (Experimental
Procedures 1), we first run the program for 500 learning steps, which include both the network dynamics
and the PINning algorithm, and then an additional 50 steps with only the network dynamics after the
learning has been terminated (convergence metrics below, see also Supplemental Figure 2). A "step" is
defined as one run of the program for the duration of the relevant trial, denoted by T. Each step is
equivalent to T = 10500 time points (10.5s) in Figures 2C and 5E, and 8000 time points (8s) in Figure 6B.
On a standard laptop computer, the first 500 such steps for a 500-unit rate-based network producing
a 10s-long sequence (such as in Supplemental Figure 4) take approximately 8 minutes to complete in
realtime and the following 50, about 30 seconds in realtime (about 1s/step for the 10s-long single
sequence example, scaling linearly with network size N, total duration or length of the trial T, and number
of sequences produced.)
outputs with the data (as in the case of Figures 2A–C and 5D–E) or the set of target functions used
The convergence of the PINning algorithm was assayed as follows: (a) By directly comparing the
(Supplemental Figure 4) and (b) By calculating and following the !
χ2
-squared error between the network
rates and the targets, both during PINning (inset of Supplemental Figure 2) and at the end of the
simulation (Supplemental Figure 2). The performance of the PINning algorithm was assayed by
computing the percent variance of the data or the targets captured by the sequential outputs of different
networks (pVar, see Experimental Procedures 7, see also Figure 1H) or by computing the Selectivity
Index of the memory network (Selectivity, see Experimental Procedures 9, see also Figure 5F).
13.5. Dimensionality Of Network Activity (Qeff)
We use state space analysis based on PCA (see for example, [Rajan, Abbott & Sompolinsky, 2011;
Sussillo, 2014] and references therein) to describe the instantaneous network state by diagonalizing the
Qij = 〈(ri(t)− < ri >)(rj(t)− < rj >)〉
equal-time cross-correlation matrix of network firing rates given by, !
,
where <> denotes a time average. The eigenvalues of this matrix expressed as a fraction of their sum
indicate the distribution of variances across different orthogonal directions in the activity trajectory. We
define the effective dimensionality of the activity, Qeff, as the number of principal components that capture
95% of the variance in the dynamics (Figure 2F).
13.6. Stereotypy Of Sequence (bVar), %
bVar quantifies the variance of the data or the network output that is explained by the translation
along the sequence of an activity profile with an invariant shape. For example, for Figure 1C–G, we
extracted an aggregate waveform, denoted by !
(red trace in Figures 1D and Supplemental Figure
Rave
4A), by averaging the tCOM-realigned firing rates extracted from trial-averaged PPC data collected during a
!22
2AFC task [Harvey, Coen & Tank, 2012]. Undoing the tCOM shift, we can write this function for the
aggregate or "typical" bump-like waveform as !
The amount of variability in the data that is
Rave(t − ti).
explained by the moving bump, !
Rave(t − ti)
is given by a measure we call bVar.
bVar = 1− 〈Ri(t) − Rave(t − ti)〉2
〈Ri(t) − R(t)〉2
⎡
⎢
⎣
⎤
⎥,
⎦
where !
<> =
N
∑
i
T
∑
t
.
In the above expressions, R's denote the firing
rates extracted from calcium data (Experimental Procedures 3);
Ri(t)
is the firing rate of the ith PPC
neuron at time t and !
R(t)
is the average over neurons. The total duration,T is 10.5s in Figures 1, 2 and
In some ways, bVar is similar to the ridge-to-background ratio computed during the analysis of
5, and 8s in Figure 6.
experimental data for measuring the level of background activity (see for example, Supplemental Figure
14 in [Harvey, Coen & Tank, 2012]); however, bVar additionally quantifies the stereotypy of the shape of
the transient produced by individual neurons.
13.7. Percent Variance Of Data Explained By Model (pVar), %
We quantify the match between the experimental data or the set of target functions, and the outputs
of the model by the amount of variance of the data that is captured by the model,
pVar = 1− 〈Di(t) − ri(t)〉2
〈Di(t) − D(t)〉2
⎡
⎢
⎣
⎤
⎥,
⎦
which is one minus the ratio of the Frobenius norm of the difference
between the data and the outputs of the network, and the variance of the data. The data referred to here,
denoted by D, is trial-averaged data, such as from Figures 1D, 2A and 5D.
In Figure 2E and in Figure 5G, we compute the magnitude of the synaptic change required to
13.8. Magnitude Of Synaptic Change
implement a single PPC-like sequence, an idealized sequence and three memory tasks, respectively. In
combination with the fraction of plastic synapses in the PINned network, p, this metric characterizes the
amount of structure that needs to be imposed in an initially random network to produce the desired
temporally structured dynamics. This is calculated as, normalized mean synaptic change =
∑ Jij
Rand
where, following the same general notation as in the main text, JPINned, p%
ij
PINned, p% − Jij
∑ Jij
Rand
ij
,
denotes the connectivity matrix of the PINned network constructed with p% plastic synapses and JRand
denotes the initial random connectivity matrix (p = 0).
!23
!
!
In Figures 5F, a Selectivity Index is computed (similar to Figure 4 in [Harvey, Coen & Tank, 2012]) to
13.9. Selectivity Index For Memory Task
assess the performance of different PINned networks at maintaining cue-specific memories during the
delay period of delayed paired association tasks. This metric is based on the ratio of the difference and
the sum of the mean activities of preferred neurons at the end of the delay period during preferred trials,
and the mean activities of preferred neurons during opposite trials. We compute Selectivity Index as,
1
2
where the notation is as follows:
r.n + < r >l.t
< r >l.t
l.n − < r >r.t
l.n
l.n + < r >l.t
l.n
< r >r.t
< r >r.t
⎡
⎢
⎣
⎤
⎥,
⎦
right pref neurons
∑ ri
i
and !
< r >r.t
l.n=
1
Nleft pref neurons
left pref neurons
∑ ri
i
are the average firing
r.n − < r >l.t
r.n
r.n + < r >l.t
1
< r >r.t
!
r.n=
Nright pref neurons
rates of right-preferring and left-preferring model neurons on right trials;
< r >l.t
r.n=
1
Nright pref neurons
right pref neurons
∑ ri
i
and !
< r >l.t
l.n=
1
Nleft pref neurons
left pref neurons
∑ ri
i
are the average firing
rates of right-preferring and left-preferring model neurons on left trials of the simulated task. The end of
the delay period is at approximately 10s for the network in Figure 5 (after [Harvey, Coen & Tank, 2012])
and at ~7s time point for the network in Figure 6.
13.10. Temporal Sparseness Of Sequences (NActive/N)
The temporal sparseness of a sequence is defined as the fraction of neurons active at any instant
during the sequence, the fraction, NActive/N. To compute this, first, the normalized firing rate from each
model neuron in the network or from data [Harvey, Coen & Tank, 2012], denoted by Ri(t), is realigned by
the center-of-mass, iCOM(t), given by,
iCOM(t) = ∑i i*R(i,t)/∑i R(i,t),
where i is the neuron number
after using a circular shift rule to undo the iCOM-shift. The standard
and t is time. The realigned rates are then averaged over time to obtain <R>time, i.e.,
〈R(i)〉time = 〈R(i–iCOM(t), t)〉time
deviation of the best Gaussian that fits this curve <R>time is the number, NActive, and the ratio of NActive to
the network size, N, yields the temporal sparseness of the sequence, NActive/N. For the data in Figure 2A,
for example, NActive/N = 3% (i.e., 16 neurons out of a total of 437). Practically speaking, decreasing this
fraction makes the sequence narrower and at a critical value of sparseness (NActive/N = 1.6% in Figure 6),
there is not enough current in the network to propagate the sequence. However, up to a point, making the
sequences sparser increases the capacity of a network for carrying multiple non-interfering sequences
(i.e., sequences with delay period memory of cue identity), without demanding an increase in either the
fraction of plastic synapses, p, or in the overall magnitude of synaptic change.
!24
!
!
This is pertinent to Section 3 (Figures 3 and 4), in which we quantify how the synaptic strength varies
13.11. Analyzing the Structure of PINned Synaptic Connectivity Matrices
with the "distance" between pairs of network neurons in connectivity space, i – j, in PINned sequential
networks. We first compute the means and the standard deviations of the principal diagonals, i.e.,
Jij
N
∑ →
Ni= j
, in the 3 connectivity matrices under consideration here – JPINned, 8%, JRand, and just for
N
∑ Jii
N
i=1
comparison purposes, JPINned, 100%. Then, we compute the means and the standard deviations of
successive off-diagonal "stripes" moving away from the principal diagonal, i.e., !
Jij
N
∑ →
i− j=a
N
∑ Ji(i−a)
N
i=a+1
, for
the same three matrices. These are plotted in Figures 3B for the sparsely PINned matrix, JPINned, 8%,
relative to the randomly initialized matrix, JRand, and in Figure 3E, for the fully PINned matrix constructed
for comparison purposes, JPINned, 100%. The same analysis is also used to compare different partially
structured matrices in Supplemental Figure 9.
!25
FIGURES AND CAPTIONS
Output from input-driven random network
A Random network + external inputs
B Driven random network output
h1
h2
h3
h4
.
.
.
hN
Random
synapses
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
.
.
.
2s
437
0
Time, s
10.5
Stereotypy of PPC sequence
C Calcium fluorescence from PPC
1
1
D Firing rates extracted from data
1
1
0
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
437
0
Time, s
10.5
E Firing rates realigned by tCOM
1
s
n
o
r
u
e
N
d
e
t
r
o
s
,
s
n
o
r
u
e
N
437
0
0
10.5
Time, s
F Average waveform and residual
1
s
n
o
r
u
e
N
0.7
Rave
1
0
f(t)
1s
0
1
437
–5.2
0
Time relative to tCOM, s
0
5.2
437
–5.2
0
Time relative to tCOM, s
–0.5
5.2
G Variance explained by moving bump dynamics bVar
bVar = 40%
f(t)
PPC
neuron
1
0
bVar = 100%
Idealized
sequence target
1
0
s
e
t
a
r
g
n
i
r
i
f
n
o
r
u
e
n
e
p
m
a
x
E
l
s
e
t
a
r
g
n
i
r
i
f
n
o
r
u
e
n
e
p
m
a
x
E
l
0
Time, s
H Variance of PPC data explained by network outputs pVar
Time, s
10.5
0
10.5
s
e
t
a
r
g
n
i
r
i
f
n
o
r
u
e
n
e
p
m
a
x
E
l
pVar = 0.2%
Random network
neuron, p = 0
1
0
pVar = 85%
PINned network
neuron, p = 8%
1
0
s
e
t
a
r
g
n
i
r
i
f
n
o
r
u
e
n
e
p
m
a
x
E
l
0
Time, s
10.5
0
Time, s
10.5
Figure 1: Stereotypy Of PPC Sequences
And Random Network Output
A. Schematic of a randomly connected
model network of firing rate neurons
(depicted in blue, Experimental Procedures
1) operating in a spontaneously active
regime and a few examples of the
temporally irregular external inputs to the
network (orange traces, Experimental
Procedures 2) are shown here. Random
synapses are depicted in gray. Note that the
networks we build contain as many rate-
based model neurons as the size of the PPC
dataset [Harvey, Coen & Tank 2012] under
consideration, and are typically all-to-all
connected, but only a fraction of these
model neurons and their interconnections
are depicted in these schematics.
B. When the individual firing rates of the
random untrained network driven by external
inputs in (A) are normalized by the
maximum per neuron and sorted by their
tCOM, the activity across the unstructured
network appears time ordered. However,
this sequence contains large amounts of
extra-sequential or background activity,
compared to the PPC data (bVar = 12%,
pVar = 0.2%).
C.Calcium fluorescence (i.e., normalized
ΔF / F
neurons during an 10.5 second-long 2AFC
experiment, corresponding to both left and
right correct-choice outcomes, from 437 trial
averaged neurons and pooled across 6
) data collected from 437 PPC
!26
!
mice.
D. Normalized firing rates extracted from (C) using deconvolution methods (Experimental
Procedures 3, see also Supplemental Figure 1) are shown here.
E. The firing rates from the 437 neurons shown in (D) are realigned by their time of center-of-
mass (abbreviated tCOM) and plotted here.
F. Inset: A "typical" waveform (Rave in red) is obtained by averaging the realigned rates from
(C) over neurons, and a Gaussian curve with mean = 0 and variance = 0.3 (f(t), green trace),
that best fits the neuron-averaged waveform, are both plotted here. Main panel: Residual
activity level that is not explained by translations of the best fit to Rave, f(t), is shown here.
G. The variance in the population activity that is explained by translations of the best fit to Rave,
f(t), is a measure of the stereotypy in the activity (abbreviated bVar in the main text,
Experimental Procedures 6). Left panel shows the normalized firing rates from 4 example
PPC neurons (red) from (D), and the curves f(t) for each example neuron (green). bVar =
40% for these data. Right panel shows the normalized firing rates from 4 model neurons (red)
from a network generating an idealized sequence (such as in Supplemental Figure 4A) and
the corresponding curves f(t) for each (green). bVar = 100% for this idealized sequential
network.
H. The variance of the PPC data that is explained by the outputs of the different networks
constructed by PINning is given by pVar (Experimental Procedures 7) and illustrated with 4
example neurons here. Left panel shows the normalized rates from 4 example PPC neurons
(red) picked from (D) and 4 model neuron outputs from a random network driven by time-
varying external inputs (gray, network schematized in (A)) with no training (p = 0). For this
example, pVar = 0.2%. Right panel shows the same PPC neurons as in the left panel in red,
along with 4 model neuron outputs from a PINned network with p = 12% plastic synapses.
For this example, pVar = 85%.
!27
A Rates from correct trials of 2AFC task
1
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
437
0
Time, s
0
10.5
B PINning all outgoing synapses from a few network neurons
Plastic
synapses
Input to network
neuron i, zi(t)
Firing rate of neuron i
i
1s
BEFORE AFTER
PINning
1s
Target derived from
Ca2+ data, fi(t)
Error, zi(t) – fi(t)
Synaptic update, ∆Jij(t)
j = 1 ... pN
C Output of PINned network (p = 12%)
1
pVar = 85%
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
437
0
Time, s
0
10.5
D Variance of data explained increases with p
E Magnitude of synaptic change required
F Dimensionality of sequential activity
%
i
l
,
r
a
V
p
d
e
n
a
p
x
e
a
t
a
d
f
o
e
c
n
a
i
r
a
V
100
85
80
60
40
20
0
0
10
8
6
4
2
e
g
n
a
h
c
c
i
t
p
a
n
y
s
n
a
e
m
d
e
z
i
l
a
m
r
o
N
100
0
0
20
40
60
80
100
Fraction of plastic synapses p, %
20
12
Fraction of plastic synapses p, %
40
60
80
%
,
C
P
y
b
d
e
n
a
p
x
e
l
i
e
c
n
a
i
r
a
v
e
v
l
i
t
a
u
m
u
C
100
95
Qeff = 14
80
60
40
20
0
0
Qeff = 38
40
30
20
10
0
0
5
12
p, %
20
25
f
f
e
Q
y
t
i
l
i
a
n
o
s
n
e
m
D
i
PINned network, p = 12%
Random network
10
20
30
40
50
Principal component
Figure 2: Partial In-Network Training (PINning) Matches PPC-Like Sequences
A. Identical to Figure 1C.
B. Schematic of the activity-based modification scheme we call Partial in-Network Training or
PINning is shown here. Only the synaptic weights carrying the recurrent inputs from a small
and randomly selected subset, controlled by the fraction p of the N = 437 firing rate neurons
(blue) in the network are modified (plastic synapses depicted in orange) at every time step by
an amount proportional to the difference between the input to the respective neuron at that
time step (zi(t), plotted in gray), and a target waveform (fi(t), plotted in green), the presynaptic
firing rate, rj, and the inverse cross-correlation matrix of the firing rates of the PINned neurons
(denoted by the matrix P, Experimental Procedures 4). Each neuron in the PINned network
fires a bump like the one in the red trace with its peak at a time point staggered relative to the
other neurons in the network, and together, the population spans the entire duration of the
task. Here, the target functions we use for PINning are extracted from the firing rates shown
in (A).
C. Normalized activity from the network with p = 12% plastic synapses (also indicated by the red
circle in (D)). The activity of the PINned network is temporally constrained, has a relatively small
!28
amount of extra-sequential activity (bVar = 40% for (C) and (A)), and shows an excellent match
with data (pVar = 85%).
D. Effect of increasing the PINning fraction, p, in a network producing a single PPC-like
sequence like (A), is shown here. pVar increases from 0 for random networks with no plastic
connections (i.e., p = 0) to pVar = 50% for p = 8% (not shown), and asymptotes at pVar =
85% as p > 12% (highlighted by the red circle and outputs shown in (C)).
E. The total magnitude of synaptic change to the connectivity matrix as a function of p is shown
here (computed as described in Experimental Procedures 8 for the matrix, JPINned, 12%). The
normalized mean synaptic change grows from a factor of ~7 for sparsely PINned networks (p
= 12%) to ~9 for fully PINned networks (p = 100%) producing the PPC-like sequence. This
means that although the individual synapses change more in small-p networks, the total
amount of change across the synaptic connectivity matrix is smaller.
F. Dimensionality of the sequential activity is computed (Experimental Procedures 5) by plotting
the cumulative variance explained by the different principal components (PCs) of the 437-
neuron PINned network generating the PPC-like sequence (orange circles) with p = 12% and
pVar = 85%, relative to those of the random network (p = 0, gray circles). Of the 437 possible
PCs that can capture the total variability in the activity of this 437-neuron network, 14 PCs
account for over 95% of the variance when p = 12%. This effective dimensionality, denoted
by Qeff, is smaller than the 38 accounting for >95% of the variability of the random untrained
network. Inset shows this effective dimensionality, Qeff (depicted in red circles), of the
manifold of the overall activity of the network, as p increases. Qeff drops from about 38
dimensions for p = 0 to about 14 dimensions by p > 12%, indicating therefore that even a
fully PINned sequence network does not construct a low-dimensional dynamical solution. In
comparison, Qeff of the data in Figure 1D is 24.
!29
1
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
1
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
500
1
1
0
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
A Sparsely PINned matrix B Distance-dependent synaptic interactions
JPINned, 8%
p = 8%
15
mean JPINned, 8% +/– std. dev.
mean JRand +/– std. dev.
2
1
0
J
f
o
s
t
n
e
m
e
E
l
0
–15
500
0.25
p = 0
500
1
JRand
1
Presynaptic neuron
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
500
1
0
–1
–2
–0.25
500
–200
Presynaptic neuron
0
i – j
C Dynamics from the mean of JPINned, 8%
Mean connectivity
Rates from the mean
y
t
i
s
n
e
d
y
t
i
l
i
b
a
b
o
r
p
g
o
L
–2
–4
–6
–15
200
1
100% PINned matrix
JPINned, 100%
p = 100%
5
0
D
1
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
–5
F
500
Presynaptic neuron
500
1
Distribution of weights
JPINned, 100%
0
JPINned, 8%
JRand
^
0
1
,
E Distance-dependent synaptic interactions
mean JPINned, 100% +/– std. dev.
mean JRand +/– std. dev.
2
1
0
J
f
o
s
t
n
e
m
e
E
l
–1
–2
–5
0
5
Synaptic strength
15
–200
0
i – j
G Dynamics from the mean of JPINned, 100%
Mean connectivity
Rates from the mean
1
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
0.5
1
0
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
Presynaptic neuron
–1
500
500
0
Time, s
0
10.5
500
1
Presynaptic neuron
–2
500
500
0
Time, s
0
10.5
Figure 3: Properties Of PINned Connectivity Matrices
A. Synaptic connectivity matrix of a 500-neuron network with p = 8% that produces an idealized
sequence with pVar = 92% (highlighted by the red circle in Supplemental Figure 4B), denoted
by JPINned, 8%, and that of the randomly initialized network with p = 0, denoted by JRand, are
shown here. Colorbars on the panels indicate, in pseudocolor, the magnitudes of synaptic
strengths after PINning.
B. Influence of neurons away from the sequentially active neurons is estimated by computing
the mean (circles) and the standard deviation (lines) of the elements of JRand (in blue) and
JPINned, 8% (in orange) in successive off-diagonal "stripes" away from the principal diagonal (as
described in Experimental Procedures 11). These quantities are plotted as a function of the
"inter-neuron distance", i - j. In units of i – j, 0 corresponds to the principal diagonal or self-
interactions, and the positive and the negative terms are the successive interaction
magnitudes of neurons a distance i – j away from the primary sequential neurons.
Fluctuations around the mean interaction are much larger and much more structured for
JPINned, 8% (red lines) compared to those for JRand (blue lines).
C. Dynamics from the band-averages of JPINned, 8% are shown here. Left panel is a synthetic
matrix generated by replacing the elements of JPINned, 8% by their respective means (orange
circles in (B)). The normalized activity from a network with this synthetic connectivity is shown
!30
on the right. Although there is a localized "bump" of excitation around i – j = 0 and long range
inhibition, these features are not responsible for sequences, but rather lead to a fixed point.
D. Same as panel (A), except for the synaptic connectivity matrix from a fully PINned (p =
100%) network, denoted by JPINned, 100% and shown here for comparison purposes. The initial
random network with p = 0, JRand, is identical to the one in (A) and is omitted here.
E. Same as (B), except comparing JPINned, 100% and JRand.. Band-averages (orange circles) are
bigger and more asymmetric compared to those for JPINned, 8%. Notably, these band-averages
are also negative for i – j = 0 and in the neighborhood of 0. Fluctuations around the band-
averages (red lines) for JPINned, 100% are smaller than those for JPINned, 8%.
F. Log of the probability density of the elements of JRand (gray squares), JPINned, 8% (yellow
squares) and for comparison, JPINned, 100% (red squares) are shown here. JRand is normally
distributed with zero mean, variance = 0.005, and zero skewness and zero kurtosis. JPINned,
8% is skewed toward negative or inhibitory weights (mean = –0.1, variance = 2.2, and
skewness = –2) with heavy tails (kurtosis = 30). JPINned, 100% has mean = 0, variance = 0.7,
skewness = –0.02, and kurtosis = 0.2.
G. Same as panel (C), except showing the firing rates from the mean JPINned, 100%. In contrast
with JPINned, 8%, the band averages of the fully PINned network can be sufficient to evoke a
"Gaussian bump" that is qualitatively similar to the moving bump of activity evoked by a ring
attractor model, since the movement of the bump is driven by the asymmetry in the mean
connectivity.
!31
Formation and propagation mechanisms for an idealized sequence
Current from full JPINned, 8%
4
0
−6
Current from mean JPINned, 8%
5
0
–5
Time, s
10.5
0
Time, s
10.5
0
0
0
Current from fluctuations
2
0
−2
Time, s
Input to neuron i, hi
1
0
−1
Time, s
Aligned population average of
fluctuations + inputs h
1
0.6
0
0.6
–1
–5
Time relative to tCOM , s
5
N
>
+ <
10.5
10.5
Figure 4: Mechanism For Formation And Propagation Of Sequence
The currents from the matrix JPINned, 8% (red trace in the top left panel) and from its components
are examined to uncover the mechanism for the formation and the propagation of a single
idealized sequence. The band-averages of JPINned, 8% cause the bump to form, as shown in the
plot of the current from mean JPINned, 8% to one neuron in the network (red trace in the top panel
on the right, see also Figure 3C). On the other hand, the cooperation of the fluctuations around
the means (whose currents are plotted in red in the middle panel on the left) with the external
inputs (in yellow in the bottom left) causes the bump to move. We demonstrate this by
considering the currents from the fluctuations around mean JPINned, 8% for all the neurons in the
network combined with the currents from the external inputs. The summed currents are
realigned to the tCOM of the bump (see Experimental Procedures 6 for a similar realignment) and
then averaged over neurons. The resulting curve, an aligned population average of the sum of
the fluctuations and external inputs to the network, is plotted in the bottom panel on the right,
and reveals the asymmetry that is responsible for the movement of the bump across the
network.
!32
A Schematic of PINned network of delayed paired association
B Example firing rates
C Task-relevant inputs
Plastic
synapses
Left sequential
neuron
Left Trial Right Trial
Left/Right
inputs
Common
inputs
Left/Right
"inputs"
Right sequential
neuron
Non-choice-specific
neuron
i
1s
h1
h2
h3
.
.
.
hN
.
.
.
Random
synapses
BEFORE
AFTER PINning
Synaptic update, ∆Jij(t)
j = 1 ... pN
0
4
7
Time, s
10.5
D Normalized firing rates extracted from PPC data
Left trial, Right preferring
Right trial, Right preferring
E Normalized firing rates from PINned network (p = 16%)
Left trial, Right preferring
Right trial, Right preferring
1
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
226
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
211
1
1
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
226
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
211
1
Left trial, Left preferring
Right trial, Left preferring
Left trial, Non-choice-specific Right trial, Non-choice-specific
Left trial, Left preferring
Right trial, Left preferring
Left trial, Non-choice-specific Right trial, Non-choice-specific
132
0
4
7
Time, s
10.5
0
4
7
Time, s
0
132
0
10.5
4
7
Time, s
10.5
0
4
7
Time, s
0
10.5
F Selectivity of PINned networks of working memory
G Magnitude of synaptic change required
1
0.91
0.85
0.81
0.8
0.6
0.4
0.2
x
e
d
n
i
y
t
i
v
i
t
c
e
l
e
S
0
0
10
16
20
23
Idealized sequential memory network
PPC-match delayed
paired association network
Fixed point memory network
40
80
Fraction of plastic synapses p, %
60
100
e
g
n
a
h
c
c
i
t
p
a
n
y
s
n
a
e
m
d
e
z
i
l
a
m
r
o
N
10
8
6
4
2
0
0
Idealized sequential memory network
Fixed point memory network
20
40
60
80
100
Fraction of plastic synapses p, %
Figure 5: Delayed Paired Association In PINned Networks Of Working Memory
A. Schematic of a sparsely trained network constructed by PINning to implement a delayed
paired association task, a two alternative forced-choice task [Harvey, Coen & Tank, 2012].
Only the outgoing synapses (plastic synapses are depicted in orange) from small subsets of
!33
randomly interconnected (random synapses are depicted in gray) network neurons are
plastic, using as targets, the firing rates extracted from Ca2+ imaging data from left preferring
PPC cells (schematized in blue), right preferring cells (schematized in red) and cells with no
choice preference (in green, called non-choice-specific neurons). As schematized in Figure
2B (see also Experimental Procedures 4), the learning rule is applied only to p% of the
synapses in the network.
B. Example single neuron firing rates, normalized to the maximum, before (gray) and after
PINning (blue trace for left preferring, red trace for right preferring and green trace for non-
choice-specific neurons). After PINning, left preferring neurons (blue trace) are active only
during a left trial and are silent during a right trial, right preferring neurons (red trace) are
active only during a right trial, and non-choice-specific neurons (green trace) are active in
trials of both types.
C. A few example task-specific inputs (!
hi
for i = 1, 2, 3, …, N) are shown here. Each network
neuron gets a different one of these irregular, spatially delocalized, filtered white-noise inputs
(Experimental Procedures 2), but receives the same one on every simulated trial of the task.
Inputs for the left trial are plotted in blue and those for the right trial are in red. During the first
4s of each 10.5s-long trial of the task, the cue period, each network neuron receives a
different input, depending on whether that specific trial is a left trial or a right trial. During the
middle 3s, the delay period, left- and right-specific inputs coalesce into a single time-varying
pattern to simulate the fact that during the real delay period in the experimental task, visual
inputs seen by the mouse are not choice-specific. During the final 3.5s of the trial, the turn
period, left and right specific "inputs" are again separate and different, to simulate the distinct
experiences the mouse might have while executing a left turn versus a right turn.
D. Normalized firing rates extracted from trial-averaged Ca2+ imaging data collected in the PPC
during a 2AFC task are shown here. Spike trains are extracted by deconvolution
(Experimental Procedures 3) from mean calcium fluorescence traces for the 437 choice-
specific and 132 non-choice-specific, task-modulated cells (one cell per row) imaged on
preferred and opposite trials [Harvey, Coen & Tank, 2012]. These firing rates are used to
extract the target functions for PINning (schematized in (A), Experimental Procedures 3 and
4). Traces are normalized to the peak of the mean firing rate of each neuron on preferred
trials and sorted by the time of center-of-mass (tCOM). Vertical gray lines indicate the time
points corresponding to the different epochs of the task – the cue period ending at 4s, the
delay period ending at 7s, and the turn period concluding at the end of the trial, at 10.5s.
!34
E. The outputs of the 569-neuron recurrent network with p = 16% plastic synapses, sorted by
tCOM and normalized by the peak of the output of each neuron, showing a match with the
experimental data (D). For this network, pVar = 85% (Experimental Procedures 7).
F. Selectivity index, shown here, is computed as the ratio of the difference and the sum of the
mean activities of preferred neurons at the 10s time-point during preferred trials and the
mean activities of preferred neurons during opposite trials (Experimental Procedures 9). Task
performance of 3 different PINned networks of working memory as a function of the fraction
of plastic synapses, p – an idealized sequential memory model (orange triangles, N = 500), a
model that exhibits PPC-like dynamics (green triangles, N = 569, using rates from (D)) and a
fixed point memory network (blue triangles, N = 500, see also Supplemental Figure 5) – are
shown here. The network that exhibits long-duration population dynamics and memory
activity through idealized sequences (orange triangles) has selectivity = 0.91, when p = 10%
of its synapses are plastic; the PPC-like network needs p = 16% plastic synapses for
selectivity = 0.85, and the fixed point network needs p = 23% of its synapses to be plastic for
selectivity = 0.81.
G. Magnitude of synaptic change (computed as for Figure 2E, Experimental Procedures 8) for
the 3 networks shown in (F) here. The idealized sequential memory network (orange
squares) and the fixed point memory network (blue squares) require comparable amounts of
mean synaptic change to execute the DPA task, growing from a factor of ~3 to ~9 for p
values ranging from 10% through 100%. This indicates that across the connectivity matrix,
the total amount of synaptic change is smaller, even though individual synapses change
more in sparsely PINned (small p) networks. The PPC-like memory network is omitted for
clarity.
!35
A Schematic for multi-sequential memory networks
Plastic
synapses
Cue 1
B Network output, 5 choice-specific sequences
N = 500, NActive/N = 3%, p = 25%,
1
1
Cue 2
Cue 3
.
.
.
Cue Ns
d
e
t
r
o
s
,
s
n
o
r
u
e
N
100
200
300
400
Random
synapses
8
Time, s
C Temporally sparse sequences rescued by inclusion of non-choice-specific neurons
Non-choice-specific neuron
8 0
8 0
4
500
0
4
4
0
4
8
0
4
0
8
N = 500, NActive/N = 1.6%, p = 25%,
1
1
NActive/N = 1.6%, p = 25% + NNon-choice-specific/N = 4%
1
20
1
100
200
300
400
d
e
t
r
o
s
,
s
n
o
r
u
e
N
d
e
t
r
o
s
,
s
n
o
r
u
e
N
116
212
308
404
500
0
4
8 0
4
8 0
4
8
Time, s
0
4
8
0
4
0
8
500
0
4
8 0
4
8 0
4
8
Time, s
0
D Capacity of multi-sequential memory networks
E Resilience to stochastic noise
4
8
0
4
0
8
p = 100% + NNon-choice-specific/N = 4%
p = 100%
p = 25% + NNon-choice-specific/N = 4%
p = 25%
25
20
15
10
5
s
N
s
e
c
n
e
u
q
e
s
f
o
r
e
b
m
u
N
0
100
300
700
500
Network size N
900
1100
c
η
,
e
c
n
a
r
e
o
t
l
i
e
s
o
N
0.5
0.4
0.3
0.2
0.1
0
0
p = 100% + NNon-choice-specific/N = 4%
p = 100%
p = 25% + NNon-choice-specific/N = 4%
p = 25%
0.04
0.02
0.08
Capacity : Network size Ns/N
0.06
0.1
Figure 6: Capacity Of Multi-Sequential Memory Networks
A. Schematic of a multi-sequential memory network of rate-based neurons (N = 500). N/Ns
neurons are assigned to each one of the multiple sequences that we want the network to
simultaneously produce. Here, Ns is the network capacity, which is equal to the number of
types of "trials", as well as the number of "cue preferences". As described in the text, only p%
of the synapses in the network are plastic. The target functions are identical to those used in
!36
Supplemental Figure 4A, however, their widths, denoted by the ratio, NActive/N, can vary as a
task parameter that controls how many network neurons are active at any instant, the
temporal sparseness of the sequence (Experimental Procedures 10).
B. The normalized firing rates of the 500-neuron network with p = 25% and NActive/N = 3% are
shown here. The memory task (the cue periods and the delay periods only) is correctly
executed through 5 choice-specific sequences; during the delay period (4–8s), neurons fire in
a sequence only on trials of the same type as their cue preference and are silent during other
types of trials.
C. Left panel shows the same network as (B) failing to perform the task correctly when the
widths of the targets was halved (NActive/N = 1.6%), sparsifying the sequences in time. While
there is sequential activation, the memory of the cue identity is not maintained during the
delay period. Right panel shows the result of including a small number of non-choice-specific
neurons (NNon-choice-specific/N = 4%) that fire in the same order in trials of all 5 types. Including
non-choice-specific neurons restores the memory of cue identity during the delay period,
without requiring an increase in p.
D. Capacity of multi-sequential memory networks, Ns, as a function of network size, N, for
different values of p is shown here. Mean values are indicated by orange circles for p = 25%
PINned networks, red circles for p = 25% + NNon-choice-specific/N = 4%, light blue squares for p =
100% and dark blue squares for p = 100% + NNon-choice-specific/N = 4% neurons. PINned
networks containing additional non-choice-specific neurons have temporally sparse target
functions with NActive/N = 1.6%, the rest have NActive/N = 3%. Error bars are calculated over 5
different random instantiations each and decrease under the following conditions: as network
size is increased, as fraction of plastic synapses p is increased, and moderately with the
inclusion of non-choice-specific neurons. The green square highlights the 500-neuron
network whose normalized outputs are plotted in (B). The maximum sequence-carrying
capacity, Ns, is directly proportional to pN and inversely proportional to NActive/N.
E. Resilience of different multi-sequential networks is computed as the critical amount of
stochastic noise (denoted here by !
ηc
, Experimental Procedures 2) tolerated before the
memory task fails, is shown here. Tolerance, plotted here as a function of the ratio of multi-
sequential capacity and network size, Ns/N, decreases as p is lowered and as capacity
increases, although it falls slower with the inclusion of non-choice-specific neurons. Only
mean values are shown here for clarity.
!37
SUPPLEMENTAL DATA
A
1
0
1
0
1
0
1
0
B
1
0
1
0
1
0
1
0
Deconvolving Ca2+ into rates
Data Ca2+
Rate = P[spike]
C
Extracting firing rates from data
Ca2+ data from 437 PPC neurons
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
437
Firing rates extracted from data
1
0
Time, s
10.5
Inferred Ca2+ from rates
Data Ca2+
Inferred Ca2+
d
e
t
r
o
s
,
s
n
o
r
u
e
N
437
Inferred Ca2+ from rates
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
1
0
1
0
1
0
Time, s
10.5
437
0
4
7
Time, s
0
10.5
Figure S1: Extracting Target Functions From Data By Deconvolution
A. The use of deconvolution algorithms for the extraction of firing rates is illustrated here. A few
example PPC neurons showing the Ca2+ fluorescence signals in red (replotted from [Harvey,
Coen & Tank, 2012]) and their extracted firing rates (normalized to a maximum of 1) in green.
These rates were obtained using both methods described in Experimental Procedures 3. The
implementation of the Bayesian algorithm (method #2 from Experimental Procedures 3)
yields spike times along with the statistical confidence of a spike arriving at that particular
time point (P[spike]). However, since the frame rate of the imaging experiments in [Harvey,
Coen & Tank, 2012] was relatively slow (64ms per frame), we used these "probabilistic" spike
trains as a normalized firing rate estimate. To verify, we first deconvolved single trial Ca2+
data, and then performed an average over all the single trial P[spike] estimates to obtain the
smooth firing rates for each of the 437 sequential neurons that we show in the middle panel
of (C).
!38
B. To see how accurate the firing rate estimates extracted from the data were, we re-convolved
the extracted firing rates of the example units in (A) through a difference of exponentials with
a rise time of 52ms and decay time of 384ms. The PPC Ca2+ data is plotted in red and the
inferred Ca2+ is plotted in blue here.
C. Top panel shows the original Ca2+ fluorescence signals from the PPC (similar to Figure 2A of
main text, adapted from Figure 2c in [Harvey, Coen & Tank, 2012]). Middle panel shows the
firing rates (from 0 to a maximum of 6Hz, normalized by the peak to 1) extracted from these
data by using deconvolution methods. Bottom panel shows the inferred Ca2+ signals obtained
by convolving these extracted firing rates by a Ca2+ impulse response function as explained
for (B) (see also Experimental Procedures 3).
!39
Convergence of algorithm for multiple networks
2
χ
g
n
n
r
a
e
i
l
r
e
t
f
a
l
s
k
r
o
w
t
e
n
e
p
i
t
l
u
m
s
s
o
r
c
a
r
o
r
r
E
0.06
0.05
0.04
0.03
0.02
0.01
0
0
Convergence criteria for one p = 8% network
0.45
0.35
0.25
0.15
0.05
2
χ
g
n
n
r
a
e
i
l
g
n
i
r
u
d
r
o
r
r
E
0
0
Run 5
Run 4
Run 3
Run 2
Run 1
200
100
400
Learning steps per run
300
500
p = 8%
p = 100%
500
1000
2000
3000
4000
5000
Maximum number of learning steps
Figure S2: Convergence Of PINning Algorithm
The convergence of our PINning algorithm (Experimental Procedures 4) is shown here for multiple
sequential networks generating sequential outputs (similar to Supplemental Figure 4). We plot the
χ2
error between the target functions and the outputs of several PINned networks as a function of
the number of learning steps for two values of p – p = 8% plastic synapses (mean values in the red
circles, means computed over 5 instantiations each) and p = 100% plastic synapses (mean values in
the yellow circles, means computed over 5 instantiations each) for comparison purposes. When the
χ2
500th learning step, we terminate the learning and simulate the network with the PINned connectivity
matrix (denoted as JPINned, p% in general) for an additional 50 steps before the program graphs the
network outputs (firing rates, inferred calcium to compare with data, statistics of JPINned, p%, etc.). The
point highlighted in the gray square corresponds to a PINned network with p = 8% that ran for 500
error drops below 0.02, which for both sparsely and fully PINned networks occurs before the
learning steps, at the end of which, the !
error was 0.018. Additionally, this network had a pVar
(Experimental Procedures 7) of 92%. Inset shows the speed of convergence for runs of different
lengths for the p = 8% PINned network. Run #5 corresponds to the example network highlighted in
the gray square.
χ2
!40
!
!
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
218
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
Left/Right
"inputs"
i
h
,
i
n
o
r
u
e
n
o
t
t
u
p
n
I
.
.
.
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
218
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
A Random recurrent network
B Network output (N = 436, g = 1.5)
Left Trial, Right Preferring
Right Trial, Right Preferring
1
+ Task-specific inputs
Common
Left/Right
specific inputs
input
0
Left Trial, Left Preferring
7
4
10.5
0
10.5
Right Trial, Left Preferring
7
4
0
Time, s
10.5
218
0
7
10.5
218
0
4
Time, s
0
7
10.5
4
Time, s
Figure S3: No Delay Period Memory In Random Network Driven By Choice-Specific
Inputs
A. Schematic of a 436-neuron random recurrent network operating in a spontaneously active
regime (g = 1.5, top panel) driven by filtered white noise inputs (Experimental Procedures 1
and 2) for a two alternative forced-choice task (bottom panel, same as Figure 5C). We assign
218 of the neurons in this network to be left preferring and the other 218 to be right
preferring.
B. Normalized firing rates from the 436 random network neurons are sorted by their tCOM and
shown here. Both left preferring and right preferring neurons are active during the delay
periods of both types of trials (between the 4s and 7s time points). The memory of the identity
of the cue disappears during the delay period in this network because the inputs coalesce
between 4–7s and become cue-invariant.
!41
Sequential target functions for PINning
1
1
d
e
t
r
o
s
,
s
n
o
i
t
c
n
u
f
t
e
g
r
a
T
e
t
a
r
g
n
i
r
i
f
d
e
z
i
l
a
m
r
o
N
0
–5
A Deriving target functions for an idealized sequential PINned network
Gaussian fit to averaged waveform
1
Rave
Gaussian fit, f(t)
0
Time, s
500
0
5
0
10.5
Time, s
B PINning an idealized sequence
C Dimensionality of sequential activity
%
i
l
,
r
a
V
p
d
e
n
a
p
x
e
e
c
n
a
i
r
a
V
100
92
80
60
40
20
0
0
Magnitude of synaptic change
e
g
n
a
h
c
c
i
t
p
a
n
y
s
n
a
e
m
d
e
z
i
l
a
m
r
o
N
10
8
6
4
2
0
20
0
80 100
Fraction of plastic synapses p, %
60
40
20
8
Fraction of plastic synapses p, %
60
80
40
100
%
,
C
P
y
b
d
e
n
i
a
l
p
x
e
e
c
n
a
i
r
a
v
e
v
l
i
t
a
u
m
u
C
100
95
80
60
40
20
0
0
10
Qeff = 10
PINned network, p = 8%
20
40
Principal component
30
50
Figure S4: Idealized Sequence-Generating PINned Network
A. Left panel shows a Gaussian with mean = 0 and variance = 0.3, denoted by f(t) (green trace),
that best fits the neuron-averaged waveform (red trace, Rave, identical to Figure 1D). This
waveform f(t) is used to generate the target functions (right panel) for a network of 500 rate-
based model neurons PINned to produce an idealized sequence of population activity (bVar
= 100%).
B. Effect of increasing the PINning fraction, p, in a network producing the single idealized
sequence is shown here. pVar (Experimental Procedures 7) plotted as a function of p,
increases from 0 for a random unmodified network (p = 0) network and plateaus at pVar =
~92% for and above p = 8%. The sequence-facilitating properties of the connectivity matrix in
the p = 8% network highlighted in red are analyzed in Figure 4 of the main text. Inset shows
the magnitude of synaptic change required to generate an idealized sequence as a function
of p, computed as in Figure 2E (Experimental Procedures 8). The overall magnitude grows
from a factor of ~3 for sparsely PINned networks (p = 8%) to between 8 and 9 for fully
PINned networks (p = 100%) producing a single idealized sequence that match the targets in
(A). As explained in the main text, although the individual synapses change more in sparsely
PINned (small p) networks, the total amount of change across the synaptic connectivity
matrix is smaller.
!42
C. Dimensionality of sequential activity is computed (as in Figure 2F, Experimental Procedures
5) for the 500-neuron PINned network generating an idealized sequence (orange circles) with
p = 8% and pVar = 92%. Of the 500 possible PCs that can capture the total variability in the
activity of this 500-neuron network, 10 PCs account for over 95% of the variance when p =
8%, i.e., Qeff = 10.
!43
Elements of JPINned, 8% and JRand
25
PINned synapses, 8%
Unchanged, random, 92%
B
Eigenvalue spectra of JPINned, 8% and JRand
0.08
0.06
0.04
0.02
0
0.02
s
,
d
o
i
r
e
P
–0.04
–0.06
–0.08
–3000 –2000 –1000
0
Decay/Growth rate, s
1000
A
%
8
,
d
e
n
N
P
I
J
f
o
s
t
n
e
m
e
l
E
15
5
0
–5
–15
–25
–0.3
0.1
–0.1 0
Elements of JRand
0.3
Figure S5: Elements And Eigenvalues Of JPINned, 8% Relative To JRand
A. Elements of the matrices JRand and JPINned, 8% (red dots) are plotted against each other here.
The largest changes in JPINned, 8% are in the scattered 40,000 light red dots (8% of the total
number of weights, pN2) and there is a negative bias to their spread. Identity line is plotted in
dark red.
B. The eigenvalue spectra of JRand – I (gray circles) and JPINned, 8% – I (orange circles) showing
is the real part of the eigenvalues and!
τ
the time period (computed as !
Re(λ)
Re(λ)/τ
where !
is the time constant for network units) and decay/growth rates (computed as !
Im(λ)/2πτ
where !
Im(λ)
is the imaginary part of the eigenvalues) corresponding to the different modes of
the two networks. The gray line at 0 is the line of stability. As expected, the eigenvalues of
JRand lie uniformly within a circle, however, those of JPINned, 8% are distributed non-uniformly
with several large non-zero eigenmodes, some positive, others, more negative.
!44
A Working memory implemented through fixed points
(N = 500, p = 23%, Selectivity = 0.81)
B Incorrect delayed paired association through fixed points
(N = 500, p = 16%, Selectivity = 0)
Right trial, Right preferring
Left trial, Right preferring
1
1
Left trial, Right preferring
0
Left trial, Left preferring
4
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
8
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
0
Right trial, Left preferring
4
8
0
8
Right trial, Overlap neurons
4
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
0
Left trial, Left preferring
4
Right trial, Right preferring
1
0
Right trial, Left preferring
4
8
0
8
Right trial, Overlap neurons
4
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
8
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
0
Left trial, Overlap neurons
4
8
0
Left trial, Overlap neurons
4
8
100
0
4
Time, s
100
0
8
4
Time, s
0
8
100
0
4
Time, s
100
0
8
4
Time, s
0
8
Figure S6: Sequential Memory As Alternative Memory Mechanism To Fixed Points
A. Working memory implemented through fixed points in a PINned network with N = 500 and p
= 23% plastic synapses is shown here. This network has an asymptotic selectivity of 0.81, as
shown in Figure 5F of the main text.
B. Inadequately PINned fixed point network (p = 16%) fails to maintain cue memory during the
delay period of the task (selectivity = 0).
!45
A Effect of introducing non-targeted neurons
100
%
p = 10%
Idealized sequence
PPC-like sequence
p = 15%
80
60
40
20
,
r
a
V
p
s
n
o
r
u
e
n
d
e
t
e
g
r
a
t
y
b
d
e
r
u
t
p
a
c
e
c
n
a
i
r
a
V
0
0
20
30
40
10
80
Non-targeted : targeted neurons Nnt /N, %
B Idealized sequential network outputs, N = 1000
Nnt /N = 75%
Nnt /N = 5%
70
50
60
90
1
1
pVar = 90%
Nnt /N = 50%
pVar = 95%
pVar = 98%
d
e
t
r
o
s
,
s
n
o
r
u
e
N
1000
0
0
C PPC-like sequential network, N = 437
Time, s
Time, s
10.5
10.5
0
0
10.5
Time, s
Nnt /N = 25%
pVar = 80%
1
Nnt /N = 50%
pVar = 85%
Nnt /N = 75%
pVar = 88%
1
,
s
n
o
r
u
e
N
d
e
t
r
o
s
437
0
Time, s
10.5
0
Time, s
10.5
0
Time, s
0
10.5
Figure S7: Simulating Unobserved Neurons By Including Non-Targeted Neurons In
PINned Networks
A. Variance of the target functions captured by the outputs of the network neurons that have
been PINned, pVar (evaluated as described in Experimental Procedures 7) is plotted here as
a function of the ratio of non-targeted to targeted neurons in the network, denoted by Nnt/N.
Overall, pVar does not decrease appreciably when the relative number of untrained neurons
introduced into the PINned network is increased. Interestingly however, pVar improves
!46
slightly for sparsely PINned (p = 10% networks in the red circles for an idealized sequence-
generating network and p = 15% networks in green for a PPC-like sequence) when untrained
neurons are introduced. This improvement gets smaller as p increases (not shown). The
inclusion of non-targeted neurons in the networks constructed by PINning simulates the
effect of unobserved but active neurons that may exist in the experimental data and might
influence neural activity. It should be noted, however, that these additional neurons, however,
might add irregularity to the sorted outputs of the full network (including both targeted and
non-targeted neurons) and reduce the stereotypy of the overall outputs (indicated by a
decrease in the bVar computed over the full network, not shown). This effect is independent
of p.
B. Example network outputs are shown here for 3 values of Nnt/N. for a network generating an
idealized sequence similar to Supplemental Figure 4. pVar = 90% for Nnt/N = 5%, 95% for
Nnt/N = 50% and 98% for Nnt/N = 75%. The sequences become noisier overall as more
randomly fluctuating untrained neurons are introduced, but the percent variance of the
targets captured by the PINned neurons remains largely unaffected, even showing a slight
improvement.
C. Same as panel (B), except for a 437-neuron network constructed by PINning to generate a
PPC-like sequence similar to Figure 2C with different fractions of neurons left untrained. pVar
= 80% for Nnt/N = 25%, 85% for Nnt/N = 50% and 88% for Nnt/N = 75%, confirming the same
general trend as above.
!47
A Effect of sparsifying random network output
%
,
r
a
V
b
k
r
o
w
t
e
n
m
o
d
n
a
r
f
o
y
p
y
t
o
e
r
e
t
S
25
20
15
10
5
0
–10
Response function r =
1
[1 + e – (x – θ) ]
20
–5
15
Threshold of response function θ
10
5
0
B Example outputs with low & high thresholds
θ = 0, bVar = 12%
θ = 5, bVar = 22%
1
Time, s
10.5
0
Time, s
0
10.5
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
500
0
Figure S8: Changing The Threshold Of The Response Function And Sparsifying The
Random Network Output
A. Effect of sparsifying the outputs of random networks of model neurons constructed with
θ
, is shown here. Stereotypy of the sequences made by
different thresholds, denoted by !
sorting the outputs of these random networks, bVar, increases from 0 to 12% for !
= 0
networks used as initial configurations throughout the paper, and saturates at 22% for
networks with threshold values of !
> 2.
θ
θ
B. Example outputs from two random networks, one with !
identical to Figure 1B) and one with !
rates are normalized by the tCOM and sorted to yield the sequences shown here.
θ
= 0 (left, with bVar = 12%, this is
= 5 (right, with bVar = 22%) are shown here. Firing
θ
!48
A Sparsely PINned matrix B Distance-dependent synaptic interactions
JPINned, 12%
mean JPINned, 12% +/– std. dev.
p = 12%
15
0
2
1
%
2
1
,
1
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
1
437
1
500
Rates from JPINned, 12%
Presynaptic neuron
–25
d
e
n
N
P
J
I
0
f
o
s
t
n
e
m
e
E
l
–1
1
D Distance-dependent synaptic interactions
2
0
2
1
mean JHybrid +/– std. dev.
C Hybird matrix, JHybrid
JMoving bump + JRand
1
n
o
r
u
e
n
c
i
t
p
a
n
y
s
t
s
o
P
500
1
Presynaptic neuron
Rates from JHybrid
1
–2.5
500
1
d
i
r
b
y
H
0
J
f
o
s
t
n
e
m
e
E
l
–1
437
1
Time, s
–2
–437
0
10.5
0
i – j
500
1
437
Time, s
–2
–500
0
10.5
0
i – j
500
Figure S9: Comparison Of Sparsely Pinned Matrix, JPinned, 12% And An Additive Hybrid
Connectivity Matrix Of The Form, JHybrid = JMoving Bump + JRand
A. Synaptic connectivity matrix of a 437-neuron network with p = 12% that produces a PPC-like
sequence with pVar = 85%, denoted by JPINned, 12%, and the firing rates obtained (identical to
Figure 2C, also highlighted by the red circle in Figure 2D) are shown here.
B. Influence of neurons away from the sequentially active neurons is estimated by computing
the mean (circles) and the standard deviation (lines) of the elements of JRand (in blue) and
JPINned, 8% (in orange) in successive off-diagonal "stripes" away from the principal diagonal (as
described in Experimental Procedures 11 and analogous to Figure 3 in the main text). These
quantities are plotted as a function of the "inter-neuron distance", i – j. In units of i – j, 0
corresponds to the principal diagonal or self-interactions, and the positive and the negative
terms are the successive interaction magnitudes of neurons a distance i – j away from the
primary sequential neurons.
C. Same as panel (A), except for the synaptic connectivity matrix from an additive hybrid of the
form, JHybrid = JMoving bump + JRand is shown here, where, JRand. is a random matrix similar to
the one shown in the lower panel of Figure 3A and JMoving bump is the connectivity for a
moving bump model [Yishai, Bar-Or & Sompolinsky, 1995]. The hybrid matrix contains a
structured and a random part, and is constructed by the addition of a moving bump
connectivity matrix [Yishai, Bar-Or & Sompolinsky. 1995],
JMoving bump = –J0 + J2[cos(φi–φj)]+ 0.06 ×[sin(φi–φj)] φ= 0,...,π
JRand, similar to the one used to initialize PINning (lower panel of Figure 3A). Mathematically,
) and a random matrix,
, !
!49
!
the hybrid is of the form,
JHybrid =
1
N [ABJMoving bump]+
1
N [ARJRand]
, where AB is the relative
amplitude of the structured or moving bump part, scaled by network size, N, and AR is the
relative amplitude of the random part of the N-neuron hybrid network, scaled by !
lower panel shows the firing rates from the additive hybrid network whose connectivity is
given by JHybrid. pVar for the output of this hybrid network is only 1%, however, its stereotypy,
bVar = 92%.
. The
N
D. Same as (B), except for JHybrid. Band-averages (orange circles) are bigger and more
asymmetric compared to those for JPINned, 12%. Notably, these band-averages are positive for i
– j = 0 and in the neighborhood of 0. Fluctuations around the band-averages (red lines) for
JHybrid are less structured than those for JPINned, 12% and result from JRand.
!50
A Sensitivity of single sequences to stochastic noise
Idealized sequence network
PPC-like sequential network
1
0.8
0.6
0.4
0.2
%
,
r
a
V
p
e
c
n
a
i
r
a
v
t
n
e
c
r
e
P
0
0
Tolerance = 1
0.5
1
Noise amplitude
1.5
2
B Resilience of sequential memory tasks to noise
x
e
d
n
i
y
t
i
v
i
t
c
e
l
e
S
1
0.8
0.6
0.4
0.2
0
0
Sequential memory network
PPC-match delayed
paired association network
Tolerance
0.5
0.4
0.5
1
Noise amplitude
1.5
C Correctly performed task in sequential PINned network
(N = 500, p = 9%, Selectivity = 0.91)
Left trial, Right preferring
1
Right trial, Right preferring
1
1
D "Error" caused by stochastic noise in inputs
(N = 500, p = 9%, ηc = 0.6, Selectivity = 0)
Left trial, Right preferring
1
Right trial, Right preferring
1
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
Left trial, Left preferring
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
Right trial, Left preferring
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
Left trial, Left preferring
d
e
t
r
o
s
,
s
n
o
r
u
e
N
200
1
d
e
t
r
o
s
,
s
n
o
r
u
e
N
Right trial, Left preferring
200
0
Time, s
200
0
8
Time, s
8
0
200
0
Time, s
200
0
8
Time, s
8
0
Figure S10: Robustness Of PINned Networks To Stochastic Noise In Inputs
A. Stochastic noise, !
η
is injected as an additional current to test whether, and how much, the
neural sequences learned through PINning are stable against perturbations (as described in
Experimental Procedures 2), and the results are shown here. Percent variance of the target
functions explained by the network outputs, pVar, drops as amplitude of the injected noise is
increased. The maximum tolerance is indicated by the gray line and for single sequences, we
define it as the amplitude of noise at which pVar drops below 50% and denote it by !
. For
ηc
the idealized sequence and for the PPC-like sequence, !
ηc =1
.
B. Noise tolerance of memory networks that implement working memory through an idealized
sequence (orange triangles) and through PPC-like sequences (green triangles, identical to
the network shown in Figure 5) is plotted here. Selectivity of both networks drops at different
values of !
–
, indicating the maximum resilience or noise-tolerance of each, denoted by !
η
ηc
!51
the sequential memory network has a maximum tolerance of !
ηc = 0.4
memory network has a maximum noise tolerance of !
ηc = 0.5
.
while the PPC-match
C. Correctly performed delayed paired association task with delay period memory of the cue
identity implemented through two idealized sequences (each similar to Supplemental Figure
4A) in a network of 500 neurons with p = 9% plastic synapses. For the outputs shown here,
the turn period is omitted for clarity, and the delay period starts at the 5s time-point in the
10.5s-long task. The performance of this network, quantified by the selectivity index
(Experimental Procedures 9) is 0.91.
D. The same network from panel (D) fails to perform the task (selectivity = 0) because the high
levels of stochastic noise present in the inputs (!
memory.
η= 0.6
, here) quenches delay period
!52
Sensitivity to structural noise
1
0.8
%
,
r
a
V
p
e
c
n
a
i
r
a
v
t
n
e
c
r
e
P
0.6
0.4
0.2
0
–0.2
0
Trained without noise
Trained with stochastic noise
0.4
0.2
0.8
Structural noise level
0.6
1
Figure S11: Structural Noise Sensitivity Of PINned Networks
Structural noise (described by !
level × N (0,1)
N
, where N = 500 here) is added to the connections
in the sparsely PINned connectivity matrix, JPINned, 8% to test whether, and by how much, the
synaptic connections are finely tuned. pVar, computed as in Experimental Procedures 9, is
plotted as a function of structural noise amplitude, for sequential networks obtained by PINning
in the presence of stochastic noise in the inputs (red squares) and for networks trained without
any noise (blue squares). Gray line is at pVar = 50%, the noise-free-PINned network has a
tolerance (noise amplitude at which pVar drops below 50%) of 0.6, and the network trained with
noise, 0.8. Training in the presence of stochastic noise therefore leads to slightly more robust
networks, although the drop in pVar with the addition of structural noise does not fully recover
with training noise.
!53
Supplemental Data S12: Cross-Validation Analysis
We first divided the data from 436 neurons in Figure 1D into two separate "synthetic" sequences
by assigning the even-numbered cells to one (let's call this Sequence A, containing 218
neurons) and the odd-numbered cells to another sequence (say, Sequence B, also with 218
neurons). Then we constructed a PINning-based network with 218 model neurons, exactly like
we described in the main text, using data from Sequence A as target functions for PINning.
Next, we computed the percent variance of the data in Sequence A and the data in Sequence B,
denoted by pVarTest A, Train A and pVarTest B, Train A, respectively, that are captured by the outputs of
the PINned network (Experimental Procedures 7). Following a similar procedure, we also
computed pVarTest A, Seq B and pVarTest B, Seq B, after PINning a second network against target
functions derived from Sequence B.
We obtained the following estimates for all fractions of plastic synapses, p :
pVar
Test A, Train A = 91 + 2%
pVar
Test A, Train B = 46 + 2%
pVarTest B, Train A = 45 + 2%
pVarTest B, Train B = 90 + 2%
For comparison purposes, a random network such as the one in Figure 1B only captures a tiny
amount of the variability of data (in this notation, pVarTest Data, Random Network = 0.2%, see also, right
panel of Figure 1H). Additionally, the data from one set only accounts for 49% of the variance of
the data of the other set, i.e., pVar = 49%. Thus the model does almost as well as it possibly
could.
!54
ACKNOWLEDGEMENTS
The authors thank Larry Abbott for providing guidance and critiques throughout this project;
Eftychios Pnevmatikakis and Liam Paninski for the deconvolution algorithm [Pnevmatikakis et
al, 2014; Vogelstein et al, 2010]; and Dmitriy Aronov, Bill Bialek, Selmaan Chettih, Cristina
Domnisoru, Tim Hanks, and Matthias Minderer for comments. This work was supported by the
NIH (DWT: R01-MH083686; RC1-NS068148; 1U01NS090541-01 and CDH: R01-MH107620;
R01-NS089521), a grant from the Simons Collaboration on the Global Brain (DWT), a Young
Investigator Award from NARSAD/Brain & Behavior Foundation (KR), a fellowship from the
Helen Hay Whitney Foundation (CDH), and a Burroughs Wellcome Fund Career Award at the
Scientific Interface (CDH). CDH is a New York Stem Cell Foundation-Robertson Investigator
and a Searle Scholar.
REFERENCES
1.
Seidemann E, Meilijson I, Abeles M, Bergman H, Vaadia E (1996) Simultaneously
recorded single units in the frontal cortex go through sequences of discrete and stable
states in monkeys performing a delayed localization task. J Neurosci 16:752–768.
Jin DZ, Fuji N, Graybiel AM (2009) Neural representation of time in cortico-basal ganglia
circuits. Proc Natl Acad Sci USA 106:19156–19561.
Shadlen MN, Newsome WT (2001) Neural basis of a perceptual decision in the parietal
cortex (area LIP) of the rhesus monkey. J Neurophysiol 86, 1916–1936.
Gold JI, Shadlen MN (2007) The neural basis of decision making. Annu Rev Neurosci 30,
535–574.
Freedman DJ, Assad JA (2011) A proposed common neural mechanism for categorization
and perceptual decisions. Nat Neurosci 14, 143–146.
Snyder LH, Batista AP, Andersen RA (1997) Coding of intention in the posterior parietal
cortex. Nature 386, 167–170.
Andersen RA, Cui H (2009) Intention, action planning, and decision making in parietal-
frontal circuits. Neuron 63, 568–583.
Bisley JW, Goldberg ME (2003) Neuronal activity in the lateral intraparietal area and
spatial attention. Science 299, 81–86.
Bisley JW, Goldberg ME (2010) Attention, intention, and priority in the parietal lobe. Annu
Rev Neurosci 33, 1–21.
2.
3.
4.
5.
6.
7.
8.
9.
!55
10. McNaughton BL, Mizumori SJ, Barnes CA, Leonard BJ, Marquis M, Green EJ (1994)
Cortical representation of motion during unrestrained spatial navigation in the rat. Cereb
Cortex 4, 27–39.
11. Nitz DA (2006) Tracking route progression in the posterior parietal cortex. Neuron 49,
747–756.
12. Whitlock JR, Sutherland RJ, Witter MP, Moser MB, Moser EI (2008) Navigating from
hippocampus to parietal cortex. Proc Natl Acad Sci USA 105, 14755–14762.
13. Calton JL, Taube JS (2009) Where am I and how will I get there from here? A role for
posterior parietal cortex in the integration of spatial information and route planning.
Neurobiol Learn Mem 91, 186–196.
Luczak A, Bartho P, Marguet SL, Buzsaki G, Harris KD (2007) Sequential structure of
neocortical spontaneous activity in vivo. Proc Nat Acad Sci USA 104(1): 347–352.
14.
15. Harvey CD, Coen P, Tank DW (2012) Choice-specific sequences in parietal cortex during
a virtual-navigation decision task. Nature 14,484(7392): 62–68.
16. Schwartz AB, and Moran DW (1999) Motor cortical activity during drawing movements:
population representation during lemniscate tracing. J Neurophysiol 82: 2705–2718.
17. Andersen RA, Burdick JW, Musallam S, Pesaran B, Cham JG (2004) Cognitive neural
prosthetics. Trends Cogn Sci 8: 486–493.
18. Pulvermutter F, Shtyrov Y (2009) Spatiotemporal signatures of large-scale synfire chains
for speech processing as revealed by MEG. Cereb. Cortex 19: 79–88.
19. Buonomano DV (2003) Timing of neural responses in cortical organotypic slices. Proc Natl
20.
21.
Acad Sci USA 100: 4897–4902.
Ikegaya Y, Aaron G, Cossart R, Aronov D, Lampl I, Ferster D, Yuste R (2004) Synfire
chains and cortical songs: temporal modules of cortical activity. Science 304: 559–564.
Tang A, Jackson D, Hobbs J, Chen W, Smith JL, Patel H, Prieto A, Petrusca D, Grivich MI,
Sher A, Hottowy P, Dabrowski W, Litke AM, Beggs JM (2008) A maximum entropy model
applied to spatial and temporal correlations from cortical networks in vitro. J Neurosci 28:
505–518.
22. Nadasdy Z, Hirase H, Czurko A, Csicsvari J, Buzsaki G (1999) Replay and time
compression of recurring spike sequences in the hippo- campus. J Neurosci 19: 9497–
9507.
Louie K, Wilson MA (2001) Temporally structured replay of awake hippocampal ensemble
activity during rapid eye movement sleep. Neuron 29: 145–156.
23.
!56
24. Pastalkova E, Itskov V, Amarasingham A, and Buzsaki G (2008) Internally generated cell
assembly sequences in the rat hippocampus. Science 321: 1322–1327.
25. Davidson TJ, Kloosterman F, Wilson MA (2009) Hippocampal replay of extended
experience. Neuron 63: 497–507.
26. Barnes TD, Kubota Y, Hu D, Jin DZ, Graybiel AM (2005) Activity of striatal neurons reflects
dynamic encoding and recoding of procedural memories. Nature 437: 1158–1161.
27. Mauk MD, Buonomano DV (2004) The neural basis of temporal processing. Annu Rev
Neurosci 27: 307–340.
28. Hahnloser RHR, Kozhevnikov AA, Fee MS (2002) An ultra-sparse code underlies the
generation of neural sequences in a songbird. Nature 419: 65–70.
29. Kozhevnikov AA, Fee MS (2007) Singing-related activity of identified HVC neurons in the
zebra finch. J Neurophysiol 97: 4271–4283.
30. Hertz J, Prügel-Bennett A (1996) Learning short synfire chains by self-organization.
31.
Network 7(2): 357–363.
Levy N, Horn D, Meilijson I, Ruppin E (2001) Distributed synchrony in a cell assembly of
spiking neurons. Neural Netw 14(6-7): 815–824.
32. Hermann M, Hertz J, Prugel-Bennet A (1995) Analysis of synfire chains. In Network:
Computation Neural Systems, vol 6, pp 403–414.
33. Kleinfeld D, Sompolinsky H (1989) Associative network models for central pattern
generators In Methods in Neuronal Modeling, Koch and Segev, Editors, Cambridge, MA:
MIT Press, pp. 195–246.
Zhang K (1996) Representation of spatial orientation by the intrinsic dynamics of the
head-direction cell ensemble: A Theory. Journal of Neuroscience 16: 2112–2126.
34.
35. Ben-Yishai R, Bar-Or RL, Sompolinsky H (1995) Theory of orientation tuning in visual
cortex. Proc Natl Acad Sci, 92: 3844–3848.
36. Goldman MS (2009) Memory without feedback in a neural network. Neuron 61:621-634.
37. Sompolinsky H, Crisanti A, Sommers HJ (1988) Chaos in Random Neural Networks. Phys
Rev Lett 61: 259–262.
38. Buonomano DV, Merzenich MM (1995) Temporal information transformed into a spatial
code by a neural network with realistic properties. Science 267: 1028–1030.
39. Buonomano DV. (2005) A learning rule for the emergence of stable dynamics and timing
in recurrent networks. J Neurophysiol 94(4): 2275–2283.
!57
40. Williams RJ, Zipser D (1989) A learning algorithm for continuously running fully recurrent
neural networks. Neural Comput 1:270–280.
41. Pearlmutter B (1989) Learning state space trajectories in recurrent neural networks.
42.
Neural Comput 1: 263–269.
Jaeger H, Haas H (2004) Harnessing nonlinearity: predicting chaotic systems and saving
energy in wireless communication. Science 304: 78–80.
43. Maass W, Joshi P, Sontag ED (2007) Computational aspects of feedback in neural
circuits. PLoS Comput Biol 3: e165.
44. Sussillo D, Abbott LF (2009) Generating coherent patterns of activity from chaotic neural
networks. Neuron 63: 544–557.
45. Maass W, Natschlager T, Markram H (2002) Real-time computing without stable states:a
new framework for neural computation based on perturbations. Neural Comput 14: 2531–
2560.
Jaeger H (2003) Adaptive nonlinear system identification with echo state networks. In
Advances in Neural Information Processing Systems 15, Becker S, Thrun S, Obermayer
K, eds. MIT Press, Cambridge MA, 593–600.
46.
47. Molgedey L, Schuchhardt J, Schuster HG (1992) Suppressing chaos in neural networks
by noise. Phys Rev Lett 69: 3717–3719.
48. Bertschinger N, Natschläger TN (2004) Real-time computation at the edge of chaos in
recurrent neural networks. Neural Computation, 16(7): 1413–1436.
49. Churchland MM, Yu BM, Cunningham JP, Sugrue LP, Cohen MR, Corrado GS, Newsome
WT, Clark AM, Hosseini P, Scott BB, Bradley DC, Smith MA, Kohn A, Movshon JA,
Armstrong KM, Moore T, Chang SW, Snyder LH, Lisberger SG, Priebe NJ, Finn IM,
Ferster D, Ryu SI, Santhanam G, Sahani M, Shenoy KV (2010) Stimulus onset quenches
neural variability: a widespread cortical phenomenon. Nat Neurosci 13(3): 369–78.
50. White, B., Abbott, L.F. and Fiser, J. (2012) Suppression of cortical neural variability is
stimulus- and state-dependent. J. Neurophysiol. 108: 2383–2392.
51. Rajan K, Abbott LF, Sompolinsky H (2011) Inferring Stimulus Selectivity from the Spatial
Structure of Neural Network Dynamics, In Lafferty, Williams, Shawne-Taylor, Zemel and
Culotta. Advances in Neural Information Processing Systems 23.
52. Rajan K, Abbott LF, Sompolinsky H (2010) Stimulus-dependent Suppression of Chaos in
Recurrent Neural Networks. Phys Rev E 82: 01193.
!58
53. Pnevmatikakis E, Merel J, Pakman A & Paninski L (2014) Bayesian spike inference from
calcium imaging data. Asilomar Conf. on Signals, Systems, and Computers. http://
arxiv.org/abs/1311.6864
54. Vogelstein J, Packer A, Machado T, Sippy T, Babadi B, Yuste R & Paninski L (2010) Fast
non-negative deconvolution for spike train inference from calcium imaging. J.
Neurophysiol 104(6): 3691–3704.
55. Amit DJ (1992) Modeling Brain Function: The World of Attractor Neural Networks.
Cambridge University Press
56. Amit DJ (1995) The Hebbian paradigm reintegrated: local reverberations as internal
representations. Behav Brain Sci 18: 617.
57. Amit DJ, Brunel N (1995) Learning internal representations in an attractor neural network
with analogue neurons. Network 6:359–388.
58. Hansel D, Mato G (2001) Existence and stability of persistent states in large neuronal
networks. Phys Rev Lett 10: 4175–4178.
59. Hopfield JJ, Tank D W (1985) "Neural" computation of decisions in optimization problems.
Biol Cyber 55: 141–146.
60. Laje R, Buonomano DV (2013) Robust timing and motor patterns by taming chaos in
recurrent neural networks. Nat. Neurosci, 16: 925-933.
61. Cossell L, Iacaruso MF, Muir DR, Houlton R, Sader EN, Ko H, Hofer SB, Mrsic-Flogel TD
(2015) Functional organization of excitatory synaptic strength in primary visual cortex. Nature
518: 399–405.
62. Sussillo D (2014) Neural circuits as computational dynamical systems. Curr Op Neurobiol
25: 156-163.
63. Fisher D, Olasagasti I, Tank DW, Aksay ER, Goldman MS (2013) A modeling framework for
deriving the structural and functional architecture of a short-term memory microcircuit.
Neuron 79(5): 987–1000.
64. Tian L, Hires SA, Mao T, Huber D, Chiappe ME, Chalasani SH, Petreanu L, Akerboom J,
McKinney SA, Schreiter ER, Bargmann CI, Jayaraman V, Svoboda K, Looger LL (2009)
Imaging neural activity in worms, flies and mice with improved GCaMP calcium indicators.
Nat Methods. 6(12): 875-881.
65. Haykin S (2002) Adaptive Filter Theory, Upper Saddle River, NJ: Prentice Hall.
66. Brunton BW, Botvinick MM, Brody CD (2013) Rats and Humans can Optimally Accumulate
Evidence for Decision-making. Science, 340: 95-98.
!59
67. Hanks TD, Kopec CD, Brunton BW, Duan CA, Erlich JC, Brody CD (2015) Distinct
relationships of parietal and prefrontal cortices to evidence accumulation. Nature 520: 220–
223.
68. Song S, Sjöström PJ, Reigl M, Nelson S, Chklovskii DB (2005) Highly Nonrandom Features
of Synaptic Connectivity in Local Cortical Circuits. PLoS Biol 3(10): e350.
69. Fiete IR, Senn W, Wang C, Hahnloser RHR (2010) Spike time-dependent plasticity and
heterosynaptic competition organize networks to produce long scale-free sequences of
neural activity. Neuron 65(4): 563-576.
70. Fujisawa S, Amarasingham A, Harrison MT, Buzsaki G (2008) Behavior-dependent short-
term assembly dynamics in the medial prefrontal cortex. Nature Neurosci 11, 823–833.
71. Crowe DA, Averbeck BB, Chafee MV (2010) Rapid sequences of population activity patterns
dynamically encode task-critical spatial information in parietal cortex. J Neurosci 30, 11640–
11653.
72. Hopfield JJ (2015) Understanding emergent dynamics: Using a collective activity coordinate
of a neural network to recognize time-varying patterns. Neural Comp, 27(10): 2011-2038.
73. Barak O, Sussillo D, Romo R, Tsodyks M, Abbott LF (2013) From fixed points to chaos: three
models of delayed discrimination, Progress in neurobiology 103, 214-222.
!60
|
1905.10281 | 2 | 1905 | 2019-05-28T19:00:57 | Information parity in complex networks | [
"q-bio.NC",
"physics.data-an",
"physics.soc-ph"
] | A growing interest in complex networks theory results in an ongoing demand for new analytical tools. We propose a novel measure based on information theory that provides a new perspective for a better understanding of networked systems: Termed "information parity," it quantifies the consonance of influence among nodes with respect to the whole network architecture. Considering the statistics of geodesic distances, information parity detects how similar a pair of nodes can influence and be influenced by the network. This allows us to quantify the quality of information gathered by the nodes. To demonstrate the method's potential, we evaluate a social network and human brain networks. Our results indicate that emerging phenomena like an ideological orientation of nodes in a social network is severely influenced by their information parities. We also show that anatomical brain networks have a greater information parity in inter-hemispheric correspondent regions placed near the sagittal plane. Finally, functional networks have, on average, greater information parity for inter-hemispheric correspondent regions in comparison to the whole network. We find that a pair of regions with high information parity exhibits higher correlation, suggesting that the functional correlations between cortical regions can be partially explained by the symmetry of their overall influences of the whole brain. | q-bio.NC | q-bio |
Information parity in complex networks
Aline Viol,1, 2, ∗ Vesna Vuksanovi´c,3 and Philipp Hovel1, 2, 4, †
1Institute of Theoretical Physics, Technische Universitat Berlin, Hardenbergstrasse 36, 10623 Berlin, Germany
2Bernstein Center for Computational Neuroscience Berlin,
Humboldt-Universitat zu Berlin, Philippstrasse 13, 10115 Berlin, Germany
3Aberdeen Biomedical Imaging Centre, University of Aberdeen, Foresterhill, Aberdeen AB25 2ZD, UK
4School of Mathematical Sciences, University College Cork, Western Road T12 XF62, Cork, Ireland
(Dated: May 30, 2019)
A growing interest in complex networks theory results in an ongoing demand for new analytical tools.
We propose a novel measure based on information theory that provides a new perspective for a better
understanding of networked systems: Termed "information parity," it quantifies the consonance of
influence among nodes with respect to the whole network architecture. Considering the statistics
of geodesic distances, information parity detects how similarly a pair of nodes can influence and
be influenced by the network.This allows us to quantify the quality of information gathered by
the nodes. To demonstrate the method's potential, we evaluate a social network and human brain
networks. Our results indicate that emerging phenomena like an ideological orientation of nodes in
social network is severely influenced by their information parities. We also show that anatomical
brain networks have a greater information parity in inter-hemispheric homologous regions placed
near the midsagittal plane. Finally, functional networks have, on average, greater information parity
for inter-hemispheric homologous regions in comparison to the whole network. We find that a pair
of regions with high information parity exhibits higher correlation, suggesting that the functional
correlations between cortical regions can be partially explained by the symmetry of their overall
influences of the whole brain.
I.
INTRODUCTION
Within the last few decades, complex network ap-
proaches have pervaded a number of scientific fields. This
interest expanded, in part, by virtue of technological ad-
vances that acquire novel datasets such as brain imag-
ing techniques and online communication networks [1, 2].
Besides much progress towards a general understanding
of complex systems, the popularity of network model-
ing has resulted in a growing demand for new analytical
methods. We use notions of the information theory [3]
to create a new method for analyzing the influence of the
network topology on nodes' behavior. The study of net-
works using concepts of information theory has increased
in literature [4 -- 7]. The main assets of the method pro-
posed in this paper are (i) the simplicity - it is based on
an established metric of the network topology; (ii) the
clear and interpretable meaning of the output and the
convenience for experimental studies; (iii) the novelty in
regards to an understanding of the overall influence of
network topology on its nodes.
We are driven by the following questions: How are the
nodes of a network influenced or influencing the over-
all structure? Does the overall network topology affect
the nodes' role and local behaviors? We tackle these
questions quantifying the similarity of influence patterns
between pairs of nodes. In a networked system, the ele-
ments influence each other directly via links connecting
them and indirectly via nodes that are two or more links
∗ [email protected]
† [email protected]
away. The diversity of influences can be assessed by the
geodesic distance matrix. In that matrix, the rows record
the relative position of a node to the others [8]. Hence,
the geodesic distances matrix is a key object to evaluate
the symmetry of influences in networks [9]. Inspired by
mutual information [3], the measure proposed in this pa-
per, termed "information parity," uses the distribution
of geodesic distances to quantify the similarity of influ-
ence of pairs of nodes. In other words, information parity
quantifies to what extent one can infer the influences of
a node given the knowledge about influences of another.
We use the general term influences because the meaning
of relations depends on the nature of the links. For ex-
ample, if the links represent a channel of communication,
the information parity reflects how similar is the infor-
mation received for two nodes in the network and how
similar is their impact on the whole network.
We define the information parity formula in section II.
In section III, we illustrate the potential of this measure
by evaluating empirical networks. There have been many
efforts to understand the phenomena of social ideologi-
cal polarization [10 -- 12], that is, when a community splits
in two groups with different opinions about one or more
issues. Our results suggest that individual opinion in a
polarized network are strongly influenced by the whole
networks topology. Subjects -- represented as nodes --
with high information parity tend to choose the same
side of polarization. Besides, we demonstrate that infor-
mation parity can also bring valuable insight into studies
of human brain networks unveiling patterns on anatom-
ical and functional brain networks. Decoding network
patters of brain networks is one of the major challenges
of contemporary neuroscience.
II. METHODS
Let us define a unweighted and undirected network
G(V, E), where V is a set of N nodes and E is the set
of their links. The adjacency matrix A is defined here
as Aij = 1 if a pair of nodes are connected and Aij = 0
otherwise. Links describe relations of different natures
depending on the problem. Since it is often impossible to
find a spatial embedding of networks and corresponding
Cartesian coordinates, the geodesic distance is defined by
the shortest number of links along a path leading from a
specific node to another [8]. The geodesic distance matrix
reflects the network topology. Each row of the geodesic
distance matrix comprises the relationship of a node with
all nodes of the network; it shows how the whole network
is seen from that node's perspective [9]. We define pi(r)
as the probability of finding a node in the network at a
distance r from the node i and pij(r) as the probability
of finding a node at a distance r from a pair of nodes i
and j:
pi(r) =
1
N − 1
δDik,r
(1)
pij(r) =
1
N − 2
δDik,r δDjk,r ,
(2)
(cid:88)
k∈V
k(cid:54)=i
(cid:88)
k∈V
k(cid:54)=i,j
where δ·,· denotes the Kronecker delta, {Dij}i,j=1,··· ,N
is the geodesic distance matrix, and the parameter r as-
sumes integer values in the interval 1 ≤ r ≤ rmax with
rmax being the nodes' maximum neighborhood radius [9].
The information parity is defined as:
Iij =
pij(r) log
pij(r)
pi(r)pj(r)
.
(3)
rmax(cid:88)
r=1
This formula differs from mutual information [3] because
it does not consider a joint probably of two random vari-
ables. Instead, it considers the probability of two nodes
being equidistant to the same node. Unlike mutual in-
formation, the information parity may assume negative
values. In that case, the pair of nodes evaluated does not
have a considerable amount of common influences corre-
sponding to a high geodesic entropy [9]. The negative
values can be considered as a "information disparity."
Information parity quantifies the similarity of influ-
ences for a pair of nodes taking into account their relative
positions in the network structure. One can also consider
ri and rj in the definition of the probability pij(ri, rj) as
two independent parameters. Then, Iij quantifies how
2
the influences of i is constrained to the influences j, and
vice versa. See supplementary material for further infor-
mation. In this paper, we focus on the special case where
the distances are constrained ri = rj = r, that is, we con-
sider the network neighborhoods equidistant to nodes i
and j. This simplified case has an intuitive meaning: the
higher the information parity of a pair of nodes, the more
similar are their interaction patterns in the network.
In order to illustrate the capability of the influence
symmetry, we address two relevant problems in networks
science:
ideological polarization in social networks and
bilateral symmetry in human brain networks. The ap-
proach can be easily transferred for other contexts as
well as generic classes of networks.
III. RESULTS
A. Social network with ideological polarization
We evaluate the information parity of a small and well-
documented social network known as Karate-club net-
work, firstly studied by Zachary [13]. This network have
been extensively explored in the literature, becoming a
classical example of an ideologically polarized social net-
work. In short, a disagreement between the president of
the club and one of the instructors led to an ideological
polarization resulting later in a rupture of the club. One
part of the members left along with the instructor, Mr.
Hi, and the other part stayed in the original club [13].
The topology of the Karate-club network was defined by
Zachary according to the personal relationships between
the karate-club's members. He proposed a model to de-
tect the polarization considering the structure including
the strength of connections between members [13]. Later,
Girvan and Newman demonstrated that the polarization
could be detected considering only the structure of con-
nections [14]. Using the unweighted and undirected net-
work, we show that the individual decision, about which
group to join, is reflected by the information parity.
Figure 1 (a) shows the information parity matrix of the
Karate-club network. The matrix elements refer to the
information parity between pairs of members and are re-
ordered according to the two groups for better visualiza-
tion. Note that the information parity between members
of the same group is greater than between members of
the opposed group. These relations are depicted by col-
ored connections on top of the original network (dashed
links) in panel (b).
Figure 2 compares the distribution of information par-
ity within the same group (blue) and between the two
groups (red) for members of Mr. Hi group and Officers
group in panels (a) and (b), respectively. Note that, for
the subjects where the information parity is not greater
inside its group (members/nodes 17 and 32), the differ-
ence between inter- and intra-group parity is not signif-
icant. Similarly the difference for subject 9, which is a
subject miss-classified by Zachary [13], is not significant
3
(a)
(b)
Figure 1: Information parity of the Karate-club network. (a) Information parity matrix sorted according to the
group that each of the 34 members belong after the fission keeping the labels from the original study. (b) Schematic
illustration showing the information parity on the network. Connections with Iij < 0.3 are shaded to improve the
visualization. Members from Mr. Hi and the Officers group are represented by pink and green nodes, respectively.
The network links are represented in black-dashed line.
either. This example shows that the measure of informa-
tion parity can help to understand the functioning and
limitations of communities detection algorithm [15].
The distribution of neighbors of a node in dependence
on the neighborhood radius reflects the diversity of influ-
ences concerning this node [9? ]. If two members share
exactly the same structure of influences, that is, they
are equidistant to the same individuals, they receive the
same quality of information and tend to have similar per-
ception and opinions. This explains why members with
high information parity chose the same side after the fis-
sion, although they might not have a direct link; they
are exposed to similar information which impacted their
decision. Therefore, information parity is a valuable re-
source for creating strategies to analyze the influence in
social groups, e.g., with respect to ideological polariza-
tion. One can use the knowledge of information parity to
control the information flow on a network, for instance,
by introducing strategic connections to reduce network
polarization via a few local interventions.
B. Bilateral symmetry on anatomical brain
networks
As a second example, we evaluate the information
parity between human brain cortical regions with inter-
hemispheric spatial correspondence regarding the struc-
ture of the anatomical connections. The data evaluated
here was firstly studied by Kahn et al.
[16]. See sup-
plementary material for further information. Anatomi-
cal networks are generated from structural connectivity
maps built by tracking the white matter fibers linking
cortical regions [17]. High information parity between
cortical regions indicates that they potentially have sim-
ilar influences taking into account the whole structure
of the whole brain. Figure 3 (a) shows the information
parity matrix of one subject with the regions reordered
according to right and left hemispheres. The illustration
in panel (b) depicts the magnitude of the information
parity between inter-hemispheric homologous cortical re-
gions expressed by the color and size of the nodes. Note
that the information parity of cortical regions situated
on the mid-sagittal plane is significantly higher than the
others.
In panel (c) we compare the average informa-
tion parity (over 21 subjects) of all pairs of homologous
inter-hemispheric cortical regions. Several regions show
negative information parity, or information disparity, in-
dicating a strong lack of overlapping relations. The aver-
age information parity over all possible pairs of nodes in
the network is I ≈ 0.07, considerably lower than the av-
erage for the regions on near the mid-sagittal plane. The
data presented here refer to networks with mean degree
(cid:104)k(cid:105) ≈ 20, but we find that the qualitative behavior does
not change varying this parameter. We describe on the
supplementary material the method we use to generate
unweighted networks. The interpretation of these finding
requires further analysis and is out of the scope of this
article. One possible hypothesis is that the greater infor-
mation parity near the mid-sagittal plane could be due
the proximity to the corpus callosum, the largest white
matter structure connecting the hemispheres [18]. An-
other possibility is that the divergence between these re-
gions and the other regions results from the limitation of
diffusion tensor imaging (DTI) image to mapping prop-
erly all anatomical connections, in particular between the
two hemispheres [17, 19].
C. Bilateral symmetry on functional brain
networks
Functional brain networks are derived from tempo-
ral correlation between regional activities [17]. A link
in these networks denotes a statistical dependence be-
tween cortical regions signals. High correlation reflects
a functional relationship according the leading paradigm
on neuroscience, i.e., regions that show correlated sig-
nals support similar function [20 -- 22]. Information par-
ity evaluated in functional networks quantifies the overall
statistical dependencies of one cortical region of the brain
having information of the statistical dependencies of an-
other.
We evaluate information parity in functional human
brain networks that are acquired from functional mag-
netic resonance images of subjects in resting state with
open eyes; no frequency filter were applied. The func-
tional matrices are the same used in the study of refer-
ences [23, 24]. Our focus is again in the inter-hemispheric
homologous cortical regions, that is, regions spatially cor-
respondents placed on opposite hemispheres. We bina-
rized the functional maps to define unweighted networks
using a threshold method [25, 26]. See supplementary
material.
Figure 4 (a) shows the correlation matrix of one of
the subjects on left and the information parity matrix on
right. The matrices are reordered to split right-left hemi-
spheres. Panel (b) compares the average of information
parity of homologous inter-hemispheric regions with the
average over all pairs of regions for each of 26 subjects.
The bars represent the standard deviation considering
different network densities. The average information par-
ity between homologous regions are significantly greater
when compared with the average mutual information be-
tween all cortical regions. Panel (c) shows these two in-
formation parity values -- evaluated over the average of all
subjects -- in dependence on the average degree, that is,
the network density. The values of information parity are
different for each subject due to the natural variation on
functional activity among distinct subjects [27]. Unlike
anatomical networks, there is no significant discrepancy
among the pairs regions. In particular, all evaluated pairs
of bilaterally homologous regions show relative high in-
formation parity. We argue that high information parity
between cortical regions could contribute to an increase
of correlation of their signals. In other words, the symme-
try on the overall statistical dependencies influences the
local functional connectivity. Panel (d) relates the Pear-
son correlation and the information parity of all pairs of
regions. Note that nodes with high information parity
tends to have high correlation. The gray plot depicts
the information parity considering only the first neigh-
bors, that is, only the direct influences. A linear rela-
tion is also observed in this case with a smaller slope.
4
The influence of common first neighbors on the correla-
tion between cortical regions signals has previously been
suggested in synchronization models of brain networks
[23, 28]. Our analysis corroborates with these results.
Moreover, we extend them; the correlation is not only
influenced by the common first neighbors but by the sym-
metry of neighbors considering the overall structure.
IV. DISCUSSION
Aiming to characterize nontrivial symmetries on com-
plex networks, we have proposed a novel method based
on statistics of the geodesic distances distributions. The
probability distribution of the geodesic distances reflects
the diversity of influences that each node is subjected
to [9], taking into account direct and indirect influences.
On the other hand, the probability distribution to find
equidistant neighbors reflects the diversity of common in-
fluences for a pair of nodes. Information parity quantifies
the congruity of influences between nodes imposed by the
whole network structure, that is, how similar is the infor-
mation that a pair of nodes share considering the whole
network topology. We have selected three problems of
network sciences to illustrate the potential of the infor-
mation parity. Our analyses have indicated that individ-
ual ideological orientations in social networks are strongly
influenced by the information parity between members of
a community. Information parity reveals the role of each
individual in the network.
We have also characterized bilateral symmetries eval-
uating information parity of homologous cortical regions
in human brain networks. We have detected that on
anatomical networks regions near the midsagittal line
have a significant greater information parity. This can
be explained by the brain anatomy or limitations of the
data acquisition technique. We also have shown that
on functional brain networks, information party is high
for all inter-hemispheric homologous regions. Our results
have indicated that the correlation between cortical sig-
nals is influenced by the nontrivial symmetries quantified
by information parity.
V. CONCLUSION
Information parity is an insightful tool for the analysis
of complex networks. The few problems explored here
can be largely extended. One can explore, for example,
functional correlations of the brain under task-evoked
conditions or disorders and underlying brain plasticity
rules. Similarly, sophisticated strategies to promote or
avoid ideological polarization in social networks can be
developed on the basis of the information parity. Con-
sider, for instance, adding or removing links to a social
networks based on the information parity. The concept
of information parity has the potential to bring valuable
insights for many other fields of network science as well.
5
(a)
(b)
Figure 2: Comparing information parity of members inside and outside their groups. The blue boxplot depicts the
information parity of each member with the members of the same group and the red boxplot with the members of
the opposite group for (a) Mr. Hi group members and (b) officers group members. The dots inside the boxplots
show the pairwise relations. The asterisk * marks the significance p < 0.05 of the difference measure by the
Student's t-test.
6
(a)
(b)
(c)
Figure 3: Information parity on anatomical brain networks. (a) Exemplary information parity matrix of one of the
subjects. The mirrored inter-hemispheric regions correspond to the diagonal in the lower left quadrant (black
square). (b) Information parity of the inter-hemispheric homologous regions. The size and color representing the
magnitude of the information parity (visualized with the BrainNet Viewer [29]). (c) Average over 21 subjects of the
inter-hemispheric homologous regions with the standard deviation represented by the green bar. The blue dashed
lines represent the global standard deviation σ. See supplementary material for the brain regions description.
7
(a)
(b)
(c)
(d)
Figure 4: Information parity of functional brain networks. (a) Exemplary Pearson correlation {Cij} and information
parity {Iij} matrices of one subject. The homologous inter-hemispheric regions correspond to the diagonal in the
lower left quadrant (black square). (b) Information parity averaged over the corresponding inter-hemispheric regions
(green) and the whole brain (purple) for 26 subjects considering different average degrees, that is, different network
densities. (c) Information parity as shown in panel (b), but in dependence on the average degree and averaged over
all subjects. (d) Scatter of all links of one sample showing the relationship between information parity and Pearson
correlation in purple; the inter-hemispheric homologous regions are highlighted in green. The information parity
calculated considering only the first neighbors is shown in gray.
AUTHOR CONTRIBUTIONS
ACKNOWLEDGMENTS
AV designed the present study and prepared all figures.
VV built the functional brain networks data. All authors
discussed the results and wrote the manuscript.
The authors acknowledge
support by Deutsche
Forschungsgemeinschaft under Grant No. HO4695/3-1
and within the framework of Collaborative Research Cen-
ter 910. The authors thank Dr. Danielle Bassett for pro-
viding anatomical networks data. AV thank Dr. Gandhi
Viswanathan and Jorge Ruiz for discussions.
8
[1] M. Rubinov and O. Sporns, NeuroImage 52, 1059 (2010).
[2] A. Vespignani, Science 325, 425 (2009).
[3] C. E. Shannon, Bell Syst. Tech. J. 27, 379 (1948).
[4] A. Viol, F. Palhano-Fontes, H. Onias, D. B. de Araujo,
and G. M. Viswanathan, Sci. Rep. 7, 7388 (2017).
[5] K. Anand and G. Bianconi, Phys. Rev. E 80, 045102
(2009).
[6] F. Tostevin and P. R. ten Wolde, Phys. Rev. Lett. 102,
218101 (2009).
[7] C. J. Honey and O. Sporns, Hum. Brain Mapp. 29, 802
[27] E. S. Finn, X. Shen, D. Scheinost, M. D. Rosenberg,
J. Huang, M. M. Chun, X. Papademetris, and R. T.
Constable, Nat. Neurosci. 18, 1664 (2015), article.
[28] V. Vuksanovi´c and P. Hovel, NeuroImage 97, 1 (2014).
[29] M. Xia, J. Wang, and Y. He, PLOS ONE 8, 1 (2013).
[30] A. E. Kahn, M. G. Mattar, J. M. Vettel, N. F. Wymbs,
S. T. Grafton, and D. S. Bassett, Cereb. Cortex 27, 173
(2017).
[31] V. Vlasov and A. Bifone, Sci. Rep. 7, 10403 (2017).
[32] B. B. Biswal et. al., Proc. Natl. Acad. Sci. 107, 4734
(2008).
(2010).
[33] Y. Liu, M. Liang, Y. Zhou, Y. He, Y. Hao, M. Song,
C. Yu, H. Liu, Z. Liu, and T. Jiang, Brain 131, 945
(2008).
[8] M. E. J. Newman, Networks: an introduction (OUP, Inc.,
New York, 2010).
[9] A. Viol, F. Palhano-Fontes, H. Onias, D. B. de Araujo,
and G. M. Viswanathan, Entropy 21, 128
P. Hovel,
(2019).
[10] R. Interian and C. C. Ribeiro, Applied Mathematics and
Computation 339, 651 (2018).
[11] M. Del Vicario, A. Scala, G. Caldarelli, H. E. Stanley,
and W. Quattrociocchi, Sci. Rep. 7, 40391 (2017).
[12] C. A. Bail, L. P. Argyle, T. W. Brown, J. P. Bumpus,
H. Chen, M. B. F. Hunzaker, J. Lee, M. Mann, F. Mer-
hout, and A. Volfovsky, Proc. Natl. Acad. Sci. 115, 9216
(2018).
[13] W. W. Zachary, J. Anthropol. Res. 33, 452 (1977).
[14] M. Girvan and M. E. J. Newman, Proc. Natl. Acad. Sci.
99, 7821 (2002).
[15] M. E. J. Newman, Eur. Phys. J. B 38, 321 (2004).
[16] A. E. Kahn, J. M. Vettel, D. S. Bassett, M. G. Mattar,
S. T. Grafton, and N. F. Wymbs, Cereb. Cortex 27, 173
(2016).
[17] O. Sporns, G. Tononi, and R. Kotter, PLoS Comput
Biol. 1 (2005), 10.1371/journal.pcbi.0010042.
[18] S. Hofer and J. Frahm, NeuroImage 32, 989 (2006).
[19] R. F. Dougherty, M. Ben-Shachar, R. Bammer, A. A.
Brewer, and B. A. Wandell, Proc. Natl. Acad. Sci. 102,
7350 (2005).
[20] B. Horwitz, NeuroImage 19, 466 (2003).
[21] C. D. Frith, K. J. Friston, P. Fletcher, P. F. Liddle, and
R. S. J. Frackowiak, Cereb. Cortex 6, 156 (1996).
[22] D. S. Bassett and O. Sporns, Nat. Neurosci. 20, 353 -- 364
(2017).
[23] P. Hovel, A. Viol, P. Loske, L. Merfort, and V. Vuk-
sanovi´c, J Nonlinear Sci. (2018).
[24] V. Vuksanovi´c and P. Hovel, Cogn. Neurodyn. 10, 361
(2016).
[25] H. Onias, A. Viol, F. Palhano-Fontes, K. C. Andrade,
M. Sturzbecher, G. Viswanathan, and D. B. de Araujo,
Epilepsy Behav. 38, 71 (2014).
[26] S. Achard and E. Bullmore, PLoS Comput Biol. 3, 1
(2007).
SUPPLEMENTARY INFORMATION
A. Generalized formula for information parity
9
A more general formula of information parity considers different radii ri and rj, which can be interpreted as
(4)
(5)
independent parameters:
pij(ri, rj) =
1
N − 2
δDik,ri δDjk,rj ,
(cid:88)
k∈V
k(cid:54)=i,j
which yields a general formula :
rmax(cid:88)
rmax(cid:88)
ri=1
rj =1
Iij =
pij(ri, rj) log
pij(ri, rj)
pi(ri)pj(rj)
.
This formula can be used to evaluate constraints and nontrivial dependencies in a network regarding to all relations.
B. Properties of the Information parity
• The probabilities are nonzero for distance smaller than rmax and zero otherwise.
p(r < rmax) (cid:54)= 0
p(r ≥ rmax) = 0.
• Defining the contribution ξij(r) = pij(r) log
pi(r) = 0 or pj(r) = 0 =⇒ ξij(r) ≡ 0;
pij(r) = 0 =⇒ ξij(r) = 0;
pij(r) > pi(r)pj(r) =⇒ ξij(r) > 0;
pij(r) < pi(r)pj(r) =⇒ ξij(r) < 0.
• The maximum number of possibilities (P n) to distribute N − 1 nodes in rmax neighborhood radius is given by
the binomial coefficients
pi(r)pj (r) to the mutual information Iij =(cid:80) ξij(r), we defined:
pij (r)
P n =
(N − 1)!
rmax!(N − 1 − rmax)!
.
C. Aditional information for the Karate club network
10
Figure 1: Information parity of all Karate-club members. The pink bars depicts the average information parity of
each member with Mr. Hi group and the green bars with the officers group. The error bars shows the standard
deviations. The group of each member belonged after the fission are indicated by the dots following the same colors
of the legend.
D. Brain datasets
Anatomical brain networks
11
The anatomical networks we evaluated here are the same of reference [30]; we consider only the first MRI scan section
of 21 subjects the AAL atlas for 90 regions. The data are available on the website https://complexsystemsupenn.
com/. We normalized the weighted matrix {Wij}i,j=1,··· ,N created from diffusion imaging tractography to force it
variate from 0 to 1; the adjacency matrix was designed by considering Aij = 1 when Wij is greater than a defined
threshold. Because the qualitative results does not change according the threshold we choose the ones that yield
network with average degree (cid:104)k(cid:105) = 20.
Functional brain networks
The functional networks explored in this paper was obteined from fMRI data of 26 subjects in resting state with
open eyes. The data are same studied in the study of the reference [28, 31]. The raw fMRI data are available on the
1000 Functional Connectome Project website (http://fcon_1000.projects.nitrc.org/; data "Berlin Margulies"
[32]). The functional matrices were created considering the whole frequency spectrum, without any filter.
We create the unweighted networks thresholding the correlation matrix {Cij}i,j=1,··· ,N . The adjacency matrix
elements are define as Aij = 1 if the Cij is greater than a defined correlation threshold and Aij = 0, otherwise. To
ensure that the threshold does not influence in the results, we evaluate a range of thresholds that yield networks with
different densities. The influence of the network density can be observed in the figure 4 (c). The method used here is
described in the references [4, 25, 33].
Table I: Brain regions according to the automated anatomic labelling (AAL) template. Indexes from 1-45/46-90
E. List of brain regions
12
indicate right (R)/left (L) hemisphere.
Index R/L Anatomical Description
1/46
2/47
3/48
4/49
5/50
6/51
7/52
8/53
9/54
10/55
11/56
12/57
13/58
14/59
15/60
16/61
17/62
18/63
19/64
20/65
21/66
22/67
23/68
24/69
25/70
26/71
27/72
28/73
29/74
30/75
31/76
32/77
33/78
34/79
35/80
36/81
37/82
38/83
39/84
40/85
41/86
42/87
43/88
44/89
45/90
Precentral
Frontal Sup
Frontal Sup Orb
Frontal Mid
Frontal Mid Orb
Frontal Inf Oper
Frontal Inf Tri
Frontal Inf Orb
Rolandic Oper
Supp Motor Area
Olflactory
Frontal Sup Medial
Frontal Mid Orb
Gyrus Rectus
Insula
Cingulum Ant
Cingulum Mid
Cingulum Post
Hippocampus
ParaHippocampal
Amygdala
Calcarine
Cuneus
Lingual
Occipital Sup
Occipital Mid
Occipital Inf
Fusiform
Postcentral
Parietal Sup
Parietal Inf
Supra Marginal Gyrus
Angular
Precuneus
Paracentral Lobule
Caudate
Putamen
Pallidum
Thalamus
Heschi
Temporal Sup
Temporal Pole sup
Temporal Mid
Temporal Pole Mid
Temporal Inf
Label
PreCG
SFGdor
ORBsup
MFG
ORBmid
IFGoperc
IFtriang
ORBinf
ROL
SMA
OLF
SFGmed
ORBsupmed
REC
INS
ACG
DCG
PCG
HIP
PHG
AMYG
CAL
CUN
LING
SOC
MOG
IOG
FFG
PoCG
SPG
IPL
SMG
ANG
PCUN
PCL
CAU
PUT
PAL
THA
HES
STG
TPOsup
MTG
TPOmid
ITG
|
1910.03349 | 1 | 1910 | 2019-10-08T11:54:43 | Analysis of an Automated Machine Learning Approach in Brain Predictive Modelling: A data-driven approach to Predict Brain Age from Cortical Anatomical Measures | [
"q-bio.NC",
"stat.ML"
] | The use of machine learning (ML) algorithms has significantly increased in neuroscience. However, from the vast extent of possible ML algorithms, which one is the optimal model to predict the target variable? What are the hyperparameters for such a model? Given the plethora of possible answers to these questions, in the last years, automated machine learning (autoML) has been gaining attention. Here, we apply an autoML library called TPOT which uses a tree-based representation of machine learning pipelines and conducts a genetic-programming based approach to find the model and its hyperparameters that more closely predicts the subject's true age. To explore autoML and evaluate its efficacy within neuroimaging datasets, we chose a problem that has been the focus of previous extensive study: brain age prediction. Without any prior knowledge, TPOT was able to scan through the model space and create pipelines that outperformed the state-of-the-art accuracy for Freesurfer-based models using only thickness and volume information for anatomical structure. In particular, we compared the performance of TPOT (mean accuracy error (MAE): $4.612 \pm .124$ years) and a Relevance Vector Regression (MAE $5.474 \pm .140$ years). TPOT also suggested interesting combinations of models that do not match the current most used models for brain prediction but generalise well to unseen data. AutoML showed promising results as a data-driven approach to find optimal models for neuroimaging applications. | q-bio.NC | q-bio |
Analysis of an Automated Machine Learning
Approach in Brain Predictive Modelling: A
data-driven approach to Predict Brain Age from
Cortical Anatomical Measures
Jessica Dafflon∗1, Walter H. L. Pinaya2,3, Federico Turkheimer1, James H. Cole1,
Robert Leech1, Mathew A. Harris4, Simon R. Cox5,6, Heather C. Whalley4, Andrew
M. McIntosh4 and Peter J. Hellyer†1
1Department of Neuroimaging, Institute of Psychiatry, Psychology and Neuroscience,
2Department of Psychosis Studies, Institute of Psychiatry, Psychology and Neuroscience,
King's College London, UK
King's College London, UK
3Center of Mathematics, Computation and Cognition, Universidade Federal do ABC, Brazil
4Division of Psychiatry, University of Edinburgh, UK
5Lothian Birth Cohorts group, Department of Psychology, University of Edinburgh, UK
6Scottish Imaging Network, A Platform for Scientific Excellence (SINAPSE) Collaboration, Edinburgh,
UK
October 9, 2019
Abstract
The use of machine learning (ML) algorithms has significantly increased in neuroscience.
However, from the vast extent of possible ML algorithms, which one is the optimal model to pre-
dict the target variable? What are the hyperparameters for such a model? Given the plethora
of possible answers to these questions, in the last years, automated machine learning (autoML)
has been gaining attention. Here, we apply an autoML library called TPOT which uses a tree-
based representation of machine learning pipelines and conducts a genetic-programming based
approach to find the model and its hyperparameters that more closely predicts the subject's
true age. To explore autoML and evaluate its efficacy within neuroimaging datasets, we chose a
problem that has been the focus of previous extensive study: brain age prediction. Without any
prior knowledge, TPOT was able to scan through the model space and create pipelines that
outperformed the state-of-the-art accuracy for Freesurfer-based models using only thickness
and volume information for anatomical structure. In particular, we compared the performance
of TPOT (mean accuracy error (MAE): 4.612± .124 years) and a Relevance Vector Regression
(MAE 5.474 ± .140 years). TPOT also suggested interesting combinations of models that do
not match the current most used models for brain prediction but generalise well to unseen
data. AutoML showed promising results as a data-driven approach to find optimal models for
neuroimaging applications.
K eywords predictive modelling · automated machine learning · age prediction · neuroimaging
∗[email protected]
†[email protected]
1
A preprint - October 9, 2019
1
Introduction
The last few decades have seen significant progress in neuroimaging methodologies and techniques
focused on identifying structural and functional features of the brain associated with behaviour.
These methods, have been widely applied to assess differences at a group level between, for example,
clinical groups. However, group-level statistics are limited and fail to make inferences that are
applicable to the individual. With the advance of machine learning (ML) algorithms and their
increased application in neuroimaging, the field is rapidly becoming more focused on exploring
relationships between individual difference and behaviour, as well as, developing clinically relevant
biomarkers of disorders (Pereira et al., 2009; Liem et al., 2017; Glaser et al., 2019; Yarkoni and
Westfall, 2017; Bzdok and Ioannidis, 2019).
This recent shift was mainly due to the use of predictive modelling approaches, consisting of
using ML algorithms to learn patterns from features in a dataset and to build an accurate model
to predict an independent variable of interest in unseen data. However, choosing a model which
is unsuitable for the statistical distribution the underlying data leads to significant problems with
over -estimation of the model and loss of generalisation. Secondly, the sheer mass of learning
approaches that are available with a vast array of different properties provides a bewildering set of
choices for the practitioner; each with advantages and disadvantages both in terms of generalisation
and computational complexity. This issue results in the occurrence of both type I and II errors,
simply as a result of picking an inappropriate analysis technique for the underlying data. This is
particularly problematic as new fields adopt machine learning approaches, and the choice of the
methodology is often based on applications in other fields where data may have quite different
statistical properties - or indeed simply be the product of whichever technique is currently in the
zeitgeist.
The no free lunch principle (Wolpert and Macready, 1997) applied to ML, suggests that there
are no single estimator and parameter combinations that will always perform well on every dataset.
The selection of preprocessing steps, the choice of the algorithm, the selection of features and
the model's hyperparameters are crucial and will vary with the task and data. Hence, the optimal
application of ML technology requires the answer to at least three questions: What are the necessary
preprocessing steps that should be performed to prepare the data? Is there a way of reducing the
feature space to only the relevant features? Among the many available ML algorithms which one
is the most appropriate for the data under analysis? That these choices are often arbitrary and
defined only on prior -wisdom, is a challenge for neuroimaging which continues to face a significant
replication crisis (Open Science Collaboration, 2015).
ML algorithms vary greatly in both their properties, complexity and the assumptions they make
about the data they are applied to. They can be linear, non-linear and optimise different functions to
predict continuous (regression) or categorical (classification) variables. Moreover, the performance
of all ML algorithms depends on the fine-tuning of its hyperparameters (Jordan and Mitchell, 2015).
In addition, feature extraction and feature selection methods are often used in series to reduce or
enhance data complexity during the preprocessing stages of analysis. The consequence is that there
are potentially infinite combinations of approaches that can be taken to identify relationships out
of data. To cut through this complexity requires the development of tools that can automatically
select the appropriate (combination of) preprocessing and ML techniques to apply to a dataset to
2
A preprint - October 9, 2019
highlight relationships that are both generalisable and computationally efficient.
In recent years, automated machine learning (autoML) has been gaining attention. The aim of
autoML is to take advantage of complexity in the underlying dataset to help guide and identify the
most appropriate model (and their associated hyperparameters), optimising performance, whilst
simultaneously attempting to maximise the reliability of resulting predictions.
In this context,
many different autoML libraries have been developed. Auto-WEKA (Thornton et al., 2013), Auto-
Sklearn (Feurer et al., 2015) and Tree-based Pipeline Optimisation Tool (TPOT) (Olson et al.,
2016a) are just a few examples. While the first two implement a hierarchical Bayesian method,
the latter uses a tree-based genetic programming algorithm. Due to its user-friendly interface and
the pipeline flexibility offered by the optimisation of a tree-based approach (Hutter et al., 2019),
we have chosen to evaluate TPOT's performance on this problem. The main idea behind the tree-
based genetic programming is to explore different pipelines (i.e. combination of different operators
that perform features selection, feature generation and model analysis) for solving a classification
or regression problem. This is done through a multi-generation approach, starting from a collection
of random models. Based on the performance and reliability of predictions at each generation
those with the highest performance will be bred (i.e. combined or crossed-over), whilst random
mutations of these models are also introduced. Therefore combinations of models that maximise
both performance and have lower complexity survive and the 'best' candidate pipeline yielded by
TPOT will consist of a combination of models and preprocessing methods that are best suited to
the relationship being probed. Figure 1 presents a high-level schematics of our approach.
In this paper, we explore the application of TPOT as an autoML approach to structural neu-
roimaging data. As a test-case, we evaluated its efficacy to predict chronological age using structural
brain data. Ageing is one factor inducing major variability in brain structure. Grey matter atrophy,
increase in the ventricle sizes, cortical thinning are a few examples of structures that alter while we
age (Hogstrom et al., 2013; Cole and Franke, 2017). As age-related changes can be detected with
structural magnetic resonance imaging (MRI) different machine learning models, have been trained
to learn the relationship between age and brain structure (Franke et al., 2010; Aycheh et al., 2018;
Cole et al., 2015; Liem et al., 2017; Valizadeh et al., 2017; Madan and Kensinger, 2018; Becker et al.,
2018). The main idea behind brain age studies is to find discrepancies between the predicted and
chronological age, which might be used as biomarkers (Cole and Franke, 2017). As brain-age pre-
diction has been extensively studied and its accuracy can be evaluated against the reported model
accuracies the existing brain-age corpus (Aycheh et al., 2018; Cole et al., 2017; Franke et al., 2010;
Valizadeh et al., 2017), we used this problem to test the settings, validity and limitations of au-
toML for imaging applications in using a regression approach. In this study, we demonstrate that:
(1) the model's performance is highly dependent on the initial model population defined by the
initial model pool passed as a configuration and the population size; (2) there is no single analysis
model that predicts age with the highest performance from the underlying structural imaging data;
(3) models suggested by TPOT outperforms relevance vector regressor (RVR), a state-of-the-art
model used to predict brain age. Therefore, TPOT can be used as a data-driven approach to learn
patterns in the data, to automatically select the best hyperparameters and models in a researcher
unbiased fashion to avoid common pitfalls from ML algorithms such as overfitting.
3
A preprint - October 9, 2019
2 Methods
2.1 Subjects and Datasets
In this analysis, T1-weighted MRI scans from N=10,307 healthy subjects (age range 18-89 years,
mean age = 59.40) were obtained from 13 publicly available datasets where each dataset used one
or more scanners to acquire the data. A summary of the demographics and imaging information
can be found in Table S2 (for more details about the BANC dataset see (Cole et al., 2017)) and for
the UK Biobank (Sudlow et al., 2015; Alfaro-Almagro et al., 2018) ((https://biobank.ctsu.ox.
ac.uk/crystal/crystal/docs/brain_mri.pdf). From the original n=2001 subjects present on
the BANC dataset, we only used 1227 subjects and excluded all subjects from the WUSL Cohort
after performing Freesurfer quality control checks.
2.2 MRI Preprocessing
Using the recon-all pipeline in Freesurver version v6.0 (Dale et al., 1999), individual T1-weighted
MRI images were preprocessed and parcelled into 116 thickness and volume information for anatom-
ical structures (for the full list of features see Table S3), according to the Desikan-Killiany atlas
and ASEG Freesurfer atlas (Desikan et al., 2006). From these segmented regions, we extracted the
cortical thickness and volume to be the input data for our further analysis.
2.3 TPOT Automated Analysis
TPOT (Olson et al., 2016b,a) uses genetic programming to search through different operators
(i.e. preprocessing approaches, machine learning models, and their associated hyperparameters)
to iteratively evolve the most suitable pipeline with high accuracy. It does so by 1) generating a
pool of random analysis models sampled from a dictionary of preprocessing approaches and analysis
models (See Table S1 for a list of the models used); 2) evaluating these models using 10-fold cross-
validation, to identify the most accurate pipeline with the lowest amount of operators; 3) breeding
the top 20 selected pipelines and applying local perturbations (e.g. mutation and cross-over); 4)
re-evaluating the pipeline in the next generation. This process is repeated for a specified number of
generations before settling on a final optimal pipeline that has high accuracy and low complexity
(i.e., lowest number of pipeline operators). To make sure that the operators are combined in a
flexible way, TPOT uses a tree-based approach. That means that every pipeline is represented as a
tree where the nodes represented by the different operators. Every tree-based pipeline starts with
one or more copies of the dataset and every time the data is passed through a node the resulting
prediction is saved as a new feature. In particular, TPOT uses a genetic programming algorithm
as implemented in the Python package DEAP (Fortin et al. (2012); for a more detailed description
of the TPOT implementation see Olson et al. (2016a)).
4
2.3 TPOT Automated Analysis
5
Figure 1: Overview of experimental design: The subject's structural MRI is used to create
a parcellation of cortical and subcortical regions. The dataset was split into 2 independent sets:
TPOT training set and evaluation set. The TPOT training set was passed to TPOT, which de-
pending on the specified configuration performed feature selection, feature transformation, feature
generation, or a combination of those and evaluated the model's performance. For each gener-
ation, a 10-fold cross-validation was performed and the best models for that specific generation
were identified, crossed-over/mutated, and passed to the next generation. At the last generation,
the pipeline with the lowest mean accuracy error was identified and returned by TPOT. We then
retrained the optimised pipeline on the independent evaluation set and tested its performance using
a 10-fold cross-validation. Finally, we compared the MAEs between different TPOT configurations
and between TPOT and RVR.
Training set (orange)Test set (pink)TPOT training setFeature SelectionFeature ConstructionFeature TransformationModel Selection and CrossvalidationOptimisedPipelineCross-over/ Mutation of best modelsT1 weighted StructuralImagei) Segmentation X Subjectsii) Extract cortical/sub-cortical Informationiii) Split into TPOT training / evaluation set++X=iv) Automated TPOT Analysisv) Retrain best model10 repetitionsRVRModel10K-foldvi) Compare perfomanceof different TPOT settingsor between TPOT and RVR A preprint - October 9, 2019
2.3.1 Regression
TPOT hyperparameters exploration We used TPOT to find the 'best' pipelines to predict
brain age, where the fitness of the pipeline is defined by a low MAE between the predicted and the
subject's chronological age. To do this, we randomly selected 1546 subjects from the dataset (TPOT
training set), and we applied TPOT on them for 10 generations to find the most fitted ML pipeline
- the pipelines with the highest accuracy. The optimal pipeline suggested by TPOT was then used
to train an independent (n=8761) dataset and its performance was evaluated using a 10-fold cross-
validation. The TPOT analysis and the evaluation of the model in an independent training set
was repeated 10 times. As a result, we obtained 100 performance scores for each configuration that
were used to evaluate the impact of manipulating a) the types of model preprocessing, b) number
of models tested on the first generation, and c) mutation and cross-over rate.
Comparison between TPOT and RVR We also performed a 10 times repetition with 10-fold
cross-validation (as described above) to assess the difference in performance between the 'best'
pipelines yielded by TPOT and the RVR, a standard model used in brain-age prediction (Franke
et al., 2010; Madan and Kensinger, 2018; Kondo et al., 2015; Wang et al., 2014).
In addition,
to check if the underlying age distribution would have an effect on the models yielded by TPOT,
we repeated the analysis using 784 subjects whose age was uniformly distributed between 18-77
years old. In this case, we used n=117 subjects to train TPOT and obtain the best pipeline. The
remaining subjects (n=667) were used to train the best pipeline using a 10-fold cross-validation.
Similarly to the other analyses, this evaluation process was also repeated 10 times resulting in 100
MAE values for each condition.
While a Student's t-test is often used to check the difference in performance between two mod-
els; Student's test assumes that samples are independent, an assumption that is violated when
performing a k-fold cross-validation. As part of the k-fold cross-validation procedure, one subject
will be used in the training set k-1 times. Therefore, the estimated scores will be dependent on
each other, and there is a higher risk of type I error. For this reason, we used a corrected version
of the t-test that accounts for this dependency (Nadeau and Bengio, 2000) when comparing the
performance of TPOT and RVR and the Friedman test when comparing different hyperparameters
from TPOT (Demsar, 2006).
3 Results
We firstly investigated which models survived thought the different generations. Figure 2 shows the
counts of the different models in one of the repetitions. Random Forests and Extra Trees Regressors
are the most popular models followed by Elastic Nets. Decision Trees and K-Nearest Neighbours
also have a high popularity for the feature selection configuration.
6
A preprint - October 9, 2019
Figure 2: Overview of the models count for each generation from one repetition for the
different configurations experiments: A: Models with a darker colour were more popular then
models with lighter colour. Across the four experiments, Random Forest, K-Nearest Neighbours,
Linear Regression and Extra Trees Regressors are the models with the highest count per generation.
To make sure that all models were represented we increased the number of times the models were
evaluated in the first generation.
3.1 TPOT parameter exploration
We then explored if the changes in the TPOT configuration are associated with a different perfor-
mance (Figure 3B). We observed that independent of the preprocessing we chose the performance
varied between 4.3 and 4.9 years. In addition to that, for every repetition TPOT found a different
pipeline which was considered to be most accurate (Figure 3A). Similarly, we analysed the change
in performance when varying the initial population of pipelines (Figure 3C). If a model was not se-
lected on the initial population it will never be present in future generations, therefore we expected
that a larger initial population would lead to a more diverse pool and therefore be associated with
higher performances. We also explored the effect of mutation and crossover rate on the performance
of the derived pipelines. For a combination of high (0.9), low (0.1), mid-ranges (0.5) mutation and
cross-over rates. (Figure 3D). For all configurations, the performance of the best models yielded by
TPOT oscillated between 4.3 and 4.9 years. These suggest that there is not one single model that
best describes the dataset but a combination of many models leads to a higher performance and
independent of the of the underlying data structure TPOT was able to a pipeline that yielded high
performance.
3.2 Comparison between TPOT and RVR
To assess the efficacy of the TPOT approach applied to neuroimaging data, we compared the
performance of the TPOT's pipelines using the Full Analysis configuration with Relevance Vector
Regression. When using the entire dataset TPOT had a lower MAE and higher Pearson's Correla-
7
Gaussian Process RegressionRelevance Vector RegressionSupport Vector RegressionRandom Forest RegressionK-Nearest Neighbours RegressionLinear RegressionRidge RegressionElastic NetsExtra Trees RegressorLasso Lars Decision Tree Regression7550250125100Full AnalysisNo PreprocessingGenerationsFeature SelectionFeature Generation 3.2 Comparison between TPOT and RVR
8
Figure 3: Overview of the ensembles for the different analysis configurations at each
repetition and their performance: A: Schematic overview of the models composing the 'best'
ensembles yielded by TPOT at each repetition. A darker colour represents models with higher
counts. Random Forest Regression, Extra Trees Regressors, Lasso Lars and Linear Regression,
were the most frequently represented. Despite the different models combinations among the different
preprocessing analysis (B:), initial population size (C:), and mutation/cross-over rate (D:), there
was no difference in the yielded performance.
012Feature GenerationLasso Lars Elastic NetsGaussian Process RegressionRelevance Vector RegressionSupport Vector RegressionRandom Forest RegressionK-Nearest Neighbours RegressionLinear RegressionRidge RegressionExtra Trees RegressorDecision Tree RegressionFull AnalysisNo PreprocessingRandom SeedsFeature SelectionA: B: 4.34.44.54.64.74.84.9C: 10 individuals100 individuals1000 individuals4.34.44.54.64.74.84.90.1 mutation0.9 cross-over0.5 mutation0.5 cross-over0.9 mutation0.1 cross-overD: 4.34.44.54.64.74.84.9Mean Absolute ErrorFeature SelectionFeature GenerationFullAnalysisNoPreprocessing A preprint - October 9, 2019
TPOT
RVR
TPOT
(uniform distribution)
RVR
(uniform distribution)
MAE
4.612 ± .124
5.474 ± 0.140
5.594 ± .0706
5.975 ± .525
p-value
t
< .01
-6.441
> .5
-.616
Pearson's Correlation
.874 ± .012
.813 ± .0102
.917 ± .027
.919 ± .013
p-value
t
< .01
3.745
> .5
.007
Table 1: Comparison between TPOT and RVR: While TPOT has a significant higher accuracy
and Pearson's Correlation when using the original data distribution, when using the uniformly
distributed dataset both models had a similar performance. (The values represent ±SD).
tion between true and predicted age (Figure 4). However, when we applied TPOT to a uniformly
distributed dataset there was no significant difference between the models yielded by TPOT and
RVR (Table 1). Nevertheless, the models suggested by TPOT using both datasets with the different
age distribution were similar (Figure S1).
4 Discussion
The successful choice of an ML pipeline to predict variables of interest (such as age) from neuroimag-
ing data is driven by the statistical characteristics and distribution of the dataset under analysis.
In most cases, the choice of machine learning model applied in multivariate analysis of neuroimag-
ing data is rather arbitrary - based on prior models that 'have worked', or by selecting whichever
model is most novel in the eyes of the analysis community. To explore an alternative approach to
model selection for a relatively simple problem, in this work, we investigated the application of an
automated analysis technique: TPOT. The TPOT approach is a data-driven methodology which
is agnostic to statistical model and prepossessing of the dataset - aiming to find the best pipeline
available to fit the statistical properties of the underlying dataset, whilst simultaneously controlling
for overfit and reliability. We showed that: (1) the performance of the models suggested by TPOT
is highly dependent on the specified model pool (i.e. algorithms and hyperparameters) that TPOT
has available to use. However, feature selection, feature generation, initial population size the mu-
tation rate and cross-values rate do not have a substantial effect on the TPOT's performance. (2)
There is not one single machine learning algorithm that performs the best, but good performance is
achieved by a combination of models. (3) The pipelines suggested by TPOT performed significantly
better than commonly used methods when performing a brain age regression from brain MRI scans.
Commonly used algorithms to predict brain age include a combination of linear and non-linear
ML algorithms such as: Multiple linear regression (Valizadeh et al., 2017), Gaussian Process Regres-
sors (Cole et al., 2015; Becker et al., 2018), K-nearest neighbours (Valizadeh et al., 2017), Relevance
Vector Regression (Wang et al., 2014; Valizadeh et al., 2017; Franke et al., 2010), Random Forests
(Valizadeh et al., 2017) and Neural Networks (Cole et al., 2017; Valizadeh et al., 2017). In this
study, we used an autoML approach that searched for the most accurate pipeline over a pool of the
9
10
Figure 4: Comparison of model's performance between TPOT and RVR: We compared
the MAE (top panel left) and Pearson's Correlation (top panel right) between True and Predicted
age of the optimised model suggested by TPOT with and RVR on the test set. The lower panels
show the Predicted vs the True age for one of the optimal pipelines suggested by TPOT (left) and
RVR (right). Note that although both models use the same subject's to make prediction, the scales
of the TPOT and RVR predictions are different, the RVR model predicts young subject's to be
younger and old as older. Asterisks show differences that are statistically significant at p < 0.01
(t-test corrected); Error bars indicate ±1SD.
A preprint - October 9, 2019
commonly used algorithms and compared its performance to RVR. We observed that the variance
in the predicted accuracy is very low on the test dataset for the pipelines suggested by TPOT but
also for the RVR model. This suggests that the models are not fitting to noise but are finding
interesting patterns in the data. Nevertheless, it is interesting to note that for every analysis's rep-
etition, a different pipeline was yielded by TPOT which had the lowest MAE (i.e. 'best' pipeline;
Figure 3). This is likely because there exists no single model that always performs better for this
type of regression problem. Similarly, when analysing age prediction using voxel-wise data Varikuti
et al. (2018) previous work showed that the pattern of 'important' voxels is different across differ-
ent training sets. Given the strength of the association between brain structure and age, and high
levels of correlation between different brain regions, it seems that multiple different approaches can
achieve high levels of prediction accuracy. This cautions against the over-interpretation of specific
sets of model weights or coefficients as being those specifically important for brain ageing, as it
seems that a different weighting on the brain could reach similar levels of performance. Inference
on which brain regions are most associated with ageing is better conducted using a longitudinal
within-subjects study design, rather than a multivariate predictive model such as those used in
TPOT. Our results also highlight that all models yielded a similar MAE and were composed by a
combination of linear and non-linear models (Random Forest Regression, Extra Tree Regression, K-
Nearest Neighbours and Ridge or Lasso Regression; Figure 3 and Figure 2). In accordance with our
results, Valizadeh et al. (2017) also reported similar brain-age prediction accuracy when comparing
Random Forest and multiple linear regression. One of the main advantages of Random Forests
is that it can deal with correlated predictors, while in a linear regression correlated predictors
might bias the results. Therefore, by combining both algorithms in an ensemble, TPOT combines
the strengths of both algorithms. Random forests have also been used by Liem et al. (2017) to
combine multi-modal brain imaging data and generate brain-age prediction. In particular, Liem
et al. (2017) used a Linear Support Vector Regression to predict age and stacked these models
with Random Forests. This combined approach was able to improve brain-age prediction. Our
interpretation of these observations is that the use of Random Forests and the hyperparameters
found by TPOT 'better fit' the non-trival non-linearities present in the dataset, transforming them
within a n-dimensional manifold which can then be fed trivially into a linear classifier. A similar
observation has been described by Aycheh et al. (2018), where a combination of Sparse Group Lasso
and Gaussian process regression was used to predict brain age. On the other hand, whilst stable,
and able to generalise, this non-linear transformation and combinations of different models into a
pipeline makes interpretation of important features within the dataset impossible.
We also noted that when using a subsample of the dataset that has a uniform distribution,
similar models were used by TPOT to build ensembles, nevertheless the difference in performance
between TPOT and RVR was not significant (Table 1). We hypothesise that by using a uniform
distribution we make the problem of age regression easier and therefore obtained similar performance
between the TPOT and RVR approach, or that the reduced sample used to pre-train TPOT was
not sufficient to obtain an accurate fit. It would be interesting for future research to explore these
hypotheses further.
In the context of other literature, it is important to note that more accurate brain-age prediction
models (Cole et al., 2017), do exist. As shown by Cole et al. (2017), Convolutional Neural Networks
can predict brain age with a MAE of 4.16 years using a similar age range (18-90 years, mean
age=36.95). As developing neural networks requires in-depth knowledge of architecture engineering,
11
A preprint - October 9, 2019
it would be interesting to use autoML approaches to explore and select the most appropriate network
architecture. However, the approach in the present study does not make any assumptions about
the underlying statistics of the dataset and does not require any fine-tuning of the model of choice
but still achieve state-of-the-art accuracy. When comparing the accuracy of different studies, it is
important to take into account the age range of the analysed sample, as age prediction in a small
range has less variability than in a large range. In fact, using a sample with subjects aged 45-91
Aycheh et al. (2018) obtained a MAE of 4.02 years. While Valizadeh et al. (2017) had a similar age
range as that described in our project, they do not report the MAE for the entire sample and use
instead 3 age groups (8-18 years, 18-65 and 65-96 years) to test the accuracy of different models.
In general, Valizadeh et al. (2017) reported lower accuracy for the older group with MAE ranging
between 4.90 and 14.23 years, when using only the thickness information. On the other hand, Liem
et al. (2017) using only the cortical thickness reported a MAE of 5.95 years (analysed age range 18
- 89 years, mean = 58.68).
In the specific case of Deep-Neural Network approaches to the brain age problem, whilst im-
provements can be made on the accuracy of the model, often this is at the cost of reliability. As
TPOT can accommodate a wider set of models, it would be interesting to include Neural Networks
on the model pool and compare its performance against the range of selected models or to use
other autoML toolboxes like autokeras (Jin et al., 2019) or Efficient Neural Architecture Search via
Parameter Sharing (Pham et al., 2018). This automated approach will allow an extensive search
of models and parameters and might also shed light into the question if deep learning is beneficial
neuroimaging analysis. Recently, Schulz et al. (2019) showed that linear, kernels and deep learning
models show very similar performance in brain-imaging datasets. Combining the potential power
of deep-learning with a model-agnostic technique such as employed by TPOT, offers an potentially
interesting route for further research.
5 Conclusion
Overall, our results show that the TPOT approach can be used as a data-driven approach to find
ML models that accurately predict brain age. The models yielded by TPOT were able to generalise
to unseen dataset and had a significantly better performance then RVR. This suggests that the
autoML approach is able to adapt efficiently to the statistical distribution of the data. Although
more accurate brain-age prediction models have been reported (Cole et al., 2017), the approach in
the present study uses a wide age range (18-89 years old), uses only cortical anatomical measures,
but most of all, it does not make any assumptions about the underlying statistics of the dataset and
does not require any fine-tuning of the model of choice. By extensively testing different models and
its hyperparameters, TPOT will suggest the optimal model for the training dataset. This approach
removes possible introduced bias out of the loop and allows decisions about the model to be made
in an automated, data-driven and reliable way.
12
A preprint - October 9, 2019
6 Acknowledgements
We thank Sebastian Popescu for his help in carrying out the Freesurfer analysis on the BANC
dataset, and Pedro F. da Costa for enlightening discussions and feedback on the analysis.
JD is funded by the Kings College London & Imperial College London EPSRC Centre for Doc-
toral Training in Medical Imaging (EP/L015226/1). WHLP was supported by Wellcome Trust
(208519/Z/17/Z). FET is funded by the PET Methodology Program Grant (Ref G1100809/1) and
the project grant Development of quantitative CNS PET imaging probes for the glutamate and
GABA systems from the Medical Research Council UK (MR/K022733/1). SRC was supported
by the Medical Research Council (MR/M013111/1 and MR/R024065/1), the Age UK-funded Dis-
connected Mind project (http://www.disconnectedmind.ed.ac.uk), and by a National Institutes of
Health (NIH) research grant (R01AG054628). PJH is supported by a Sir Henry Wellcome Postdoc-
toral Fellowship from the Wellcome Trust (WT/106092/Z/14/Z)
Author contribution : JD, WHLP, FT, JHC, RL, PJH designed the study. MAE, SRC, HCW,
AMM, JHC preprocessed the dataset. JD performed the experiments. JD, WHLP and PJH anal-
ysed the data. JD, PJH, WHLP, FT, JHC, SRC, HCW wrote and edited the manuscript.
References
F. Alfaro-Almagro, M. Jenkinson, N. K. Bangerter, J. L. R. Andersson, L. Griffanti, G. Douaud,
S. N. Sotiropoulos, S. Jbabdi, M. Hernandez-Fernandez, E. Vallee, D. Vidaurre, M. Webster,
P. McCarthy, C. Rorden, A. Daducci, D. C. Alexander, H. Zhang, I. Dragonu, P. M. Matthews,
K. L. Miller, and S. M. Smith. Image processing and quality control for the first 10,000 brain
imaging datasets from uk biobank. Neuroimage, 166:400 -- 424, 02 2018. doi: 10.1016/j.neuroimage.
2017.10.034.
H. M. Aycheh, J.-K. Seong, J.-H. Shin, D. L. Na, B. Kang, S. W. Seo, and K.-A. Sohn. Biological
brain age prediction using cortical thickness data: A large scale cohort study. Front Aging
Neurosci, 10:252, 2018. doi: 10.3389/fnagi.2018.00252.
B. G. Becker, T. Klein, C. Wachinger, A. D. N. Initiative, et al. Gaussian process uncertainty in
age estimation as a measure of brain abnormality. NeuroImage, 175:246 -- 258, 2018.
D. Bzdok and J. P. A. Ioannidis. Exploration, inference, and prediction in neuroscience and
biomedicine. Trends Neurosci, 42(4):251 -- 262, Apr 2019. doi: 10.1016/j.tins.2019.02.001.
J. H. Cole and K. Franke. Predicting age using neuroimaging: Innovative brain ageing biomarkers.
Trends Neurosci, 40(12):681 -- 690, 12 2017. doi: 10.1016/j.tins.2017.10.001.
J. H. Cole, R. Leech, D. J. Sharp, and A. D. N. Initiative. Prediction of brain age suggests
accelerated atrophy after traumatic brain injury. Annals of neurology, 77(4):571 -- 581, 2015.
13
A preprint - October 9, 2019
J. H. Cole, R. P. Poudel, D. Tsagkrasoulis, M. W. Caan, C. Steves, T. D. Spector, and G. Montana.
Predicting brain age with deep learning from raw imaging data results in a reliable and heritable
biomarker. NeuroImage, 163:115 -- 124, 2017.
A. M. Dale, B. Fischl, and M. I. Sereno. Cortical surface-based analysis: I. segmentation and
surface reconstruction. Neuroimage, 9(2):179 -- 194, 1999.
J. Demsar. Statistical comparisons of classifiers over multiple data sets. Journal of Machine learning
research, 7(Jan):1 -- 30, 2006.
R. S. Desikan, F. S´egonne, B. Fischl, B. T. Quinn, B. C. Dickerson, D. Blacker, R. L. Buckner,
A. M. Dale, R. P. Maguire, B. T. Hyman, M. S. Albert, and R. J. Killiany. An automated
labeling system for subdividing the human cerebral cortex on mri scans into gyral based regions
of interest. Neuroimage, 31(3):968 -- 80, Jul 2006. doi: 10.1016/j.neuroimage.2006.01.021.
M. Feurer, A. Klein, K. Eggensperger, J. Springenberg, M. Blum, and F. Hutter. Efficient and
robust automated machine learning. In Advances in neural information processing systems, pages
2962 -- 2970, 2015.
F.-A. Fortin, F.-M. D. Rainville, M.-A. Gardner, M. Parizeau, and C. Gagn´e. Deap: Evolutionary
algorithms made easy. Journal of Machine Learning Research, 13(Jul):2171 -- 2175, 2012.
K. Franke, G. Ziegler, S. Kloppel, C. Gaser, and Alzheimer's Disease Neuroimaging Initia-
tive. Estimating the age of healthy subjects from t1-weighted mri scans using kernel meth-
ods: exploring the influence of various parameters. Neuroimage, 50(3):883 -- 92, Apr 2010. doi:
10.1016/j.neuroimage.2010.01.005.
J. I. Glaser, A. S. Benjamin, R. Farhoodi, and K. P. Kording. The roles of supervised machine learn-
ing in systems neuroscience. Prog Neurobiol, 175:126 -- 137, Apr 2019. doi: 10.1016/j.pneurobio.
2019.01.008.
L. J. Hogstrom, L. T. Westlye, K. B. Walhovd, and A. M. Fjell. The structure of the cerebral
cortex across adult life: age-related patterns of surface area, thickness, and gyrification. Cereb
Cortex, 23(11):2521 -- 30, Nov 2013. doi: 10.1093/cercor/bhs231.
F. Hutter, L. Kotthoff, and J. Vanschoren. Automated machine learning. The Springer Series on
Challenges in Machine Learning, 2019. ISSN 2520-1328. doi: 10.1007/978-3-030-05318-5. URL
http://dx.doi.org/10.1007/978-3-030-05318-5.
H. Jin, Q. Song, and X. Hu. Auto-keras: An efficient neural architecture search system.
In
Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery &
Data Mining, pages 1946 -- 1956. ACM, 2019.
M. I. Jordan and T. M. Mitchell. Machine learning: Trends, perspectives, and prospects. Science,
349(6245):255 -- 260, 2015.
C. Kondo, K. Ito, K. Wu, K. Sato, Y. Taki, H. Fukuda, and T. Aoki. An age estimation method
using brain local features for t1-weighted images. Conf Proc IEEE Eng Med Biol Soc, 2015:666 -- 9,
08 2015. doi: 10.1109/EMBC.2015.7318450.
14
A preprint - October 9, 2019
F. Liem, G. Varoquaux, J. Kynast, F. Beyer, S. K. Masouleh, J. M. Huntenburg, L. Lampe,
M. Rahim, A. Abraham, R. C. Craddock, et al. Predicting brain-age from multimodal imaging
data captures cognitive impairment. Neuroimage, 148:179 -- 188, 2017.
C. R. Madan and E. A. Kensinger. Predicting age from cortical structure across the lifespan. Eur
J Neurosci, 47(5):399 -- 416, Mar 2018. doi: 10.1111/ejn.13835.
C. Nadeau and Y. Bengio. Inference for the generalization error. In Advances in neural information
processing systems, pages 307 -- 313, 2000.
R. S. Olson, N. Bartley, R. J. Urbanowicz, and J. H. Moore. Evaluation of a tree-based pipeline
optimization tool for automating data science. In Proceedings of the Genetic and Evolutionary
Computation Conference 2016, pages 485 -- 492. ACM, 2016a.
R. S. Olson, R. J. Urbanowicz, P. C. Andrews, N. A. Lavender, J. H. Moore, et al. Automating
biomedical data science through tree-based pipeline optimization. In European Conference on
the Applications of Evolutionary Computation, pages 123 -- 137. Springer, 2016b.
Open Science Collaboration. Psychology. estimating the reproducibility of psychological science.
Science, 349(6251):aac4716, Aug 2015. doi: 10.1126/science.aac4716.
F. Pereira, T. Mitchell, and M. Botvinick. Machine learning classifiers and fmri: a tutorial overview.
Neuroimage, 45(1):S199 -- S209, 2009.
H. Pham, M. Y. Guan, B. Zoph, Q. V. Le, and J. Dean. Efficient neural architecture search via
parameter sharing. arXiv preprint arXiv:1802.03268, 2018.
M.-A. Schulz, B. Yeo, J. Vogelstein, J. Mourao-Miranada, J. Kather, K. P. Kording, A. Richard,
and D. Bzdok. Deep learning for brains?: Different linear and nonlinear scaling in uk biobank
brain images vs. machine- learning datasets. hal-02276649, 2019.
C. Sudlow, J. Gallacher, N. Allen, V. Beral, P. Burton, J. Danesh, P. Downey, P. Elliott, J. Green,
M. Landray, B. Liu, P. Matthews, G. Ong, J. Pell, A. Silman, A. Young, T. Sprosen, T. Peakman,
and R. Collins. Uk biobank: an open access resource for identifying the causes of a wide range
of complex diseases of middle and old age. PLoS Med, 12(3):e1001779, Mar 2015. doi: 10.1371/
journal.pmed.1001779.
C. Thornton, F. Hutter, H. H. Hoos, and K. Leyton-Brown. Auto-weka: Combined selection and hy-
perparameter optimization of classification algorithms. In Proceedings of the 19th ACM SIGKDD
international conference on Knowledge discovery and data mining, pages 847 -- 855. ACM, 2013.
S. A. Valizadeh, J. Hanggi, S. M´erillat, and L. Jancke. Age prediction on the basis of brain
anatomical measures. Hum Brain Mapp, 38(2):997 -- 1008, 02 2017. doi: 10.1002/hbm.23434.
D. P. Varikuti, S. Genon, A. Sotiras, H. Schwender, F. Hoffstaedter, K. R. Patil, C. Jockwitz,
S. Caspers, S. Moebus, K. Amunts, C. Davatzikos, and S. B. Eickhoff. Evaluation of non-negative
matrix factorization of grey matter in age prediction. Neuroimage, 173:394 -- 410, 06 2018. doi:
10.1016/j.neuroimage.2018.03.007.
J. Wang, W. Li, W. Miao, D. Dai, J. Hua, and H. He. Age estimation using cortical surface
pattern combining thickness with curvatures. Med Biol Eng Comput, 52(4):331 -- 41, Apr 2014.
doi: 10.1007/s11517-013-1131-9.
15
A preprint - October 9, 2019
D. H. Wolpert and W. G. Macready. No free lunch theorems for optimization. IEEE transactions
on evolutionary computation, 1(1):67 -- 82, 1997.
T. Yarkoni and J. Westfall. Choosing prediction over explanation in psychology: Lessons from
machine learning. Perspectives on Psychological Science, 12(6):1100 -- 1122, 2017.
16
7 Supplementary Info
Sklearn Implementation
PCA
FastICA
SelectFwe
SelectPercentile
VarianceThreshold
Algorithms
Feature Selection
Principle Component Analysis
Fast algorithm for Independent Component Analysis
Select the p-values corresponding to Family-wise error rate
Select features according to a percentile of the highest scores
Remove low-variance Features
Feature Generation
Agglomerate features
Regression
Elastic Net model with iterative fitting along a regularisation path ElasticNetCV
Randomised Decision Trees on sub-samples of the dataset
k-Nearest Neighbours Regression
Cross-validated Lasso using the LARS algorithm
Linear Support Vector Regression
Linear Least squares with l2 regularisation
Random Forest Regressor
Ordinary Least Squares Linear Regression
Decision Tree Regressor
Gaussian process regression
Relevance Vector Regression
FeatureAgglomeration
ExtraTreesRegressor
KNeighborsRegressor
LassoLarsCV
LinearSVR
Ridge
RandomForrestRegressor
LinearRegression
DecisionTreeRegressor
GaussianProcessRegressors
RVR
Table S1: List of used Feature Selection, Feature Generation and Regression Algorithms
.
s
t
e
s
a
t
a
d
t
n
e
r
e
ff
d
i
e
h
t
r
o
f
s
r
e
t
e
m
a
r
a
p
g
n
i
g
a
m
i
d
n
a
s
c
i
h
p
a
r
g
o
m
e
d
e
h
t
f
o
w
e
i
v
r
e
v
O
:
2
S
e
l
b
a
T
18
Cohort N Age mean (SD) Age range Sex male/female Repository details Scanner (Field strength) Scan Voxel dimensions ABIDE (Autism Brain Imaging Data Exchange) 147 24.43 (4.89) 18-40 130/17 INDI Various (all 3T) MPRAGE Various Beijing Normal University 151 21.36 (1.95) 18-28 63/88 INDI Siemens (3T) MPRAGE 1.33x1.0x1.0 Berlin School of Brain & Mind 49 30.99 (7.08) 20-60 24/25 INDI Siemens Tim Trio (3T) MPRAGE 1.0x1.0x1.0 CADDementia 12 62.33 (6.26) 58-79 9/3 http://caddementia.grand-challenge.org GE Signa (3T) 3D IR-FSPGR 0.9x0.9x1.0 Cleveland Clinic 31 43.55 (11.14) 24-60 11/20 INDI Siemens Tim Trio (3T) MPRAGE 2.0x1.0x1.2 ICBM (International Consortium for Brain Mapping) 42 27.71 (5.75) 24-60 14/28 LONI IDA Siemens Magnetom (1.5T) MPRAGE 1.0x1.0x1.0 IXI (Information eXtraction from Images) 394 46.21 (16.11) 20-86 159/235 http://biomedic.doc.ic.ac.uk/brain-development Philips Intera (3T); Philips Gyroscan Intera (1.5T); GE Signa (1.5T) T1-FFE; MPRAGE 0.9375x0.93751x1.2 MCIC (MIND Clinical Imaging Consortium) 92 32.33 (11.92) 18-60 63/29 COINS Siemens Sonata/Trio (1.5/3T); GE Signa (1.5T) MPRAGE; SPGR 0.625x0.625x1.5 MIRIAD (Minimal Interval Resonance Imaging in Alzheimer's Disease) 23 69.66 (7.18) 58-85 12/11 https://www.ucl.ac.uk/drc/research/miriad-scan-database GE Signa (1.5T) 3D IR-FSPGR 0.9375x0.93751x1.5 NEO2012 (Adelstein, 2011) 39 29.59 (8.38) 20-49 18/21 INDI Siemens Allegra (3T) MPRAGE 1.0x1.0x1.0 Nathan Kline Institute (NKI) / Rockland 151 41.92 (18.24) 18-85 94/57 INDI Siemens Tim Trio (3T) MPRAGE 1.0x1.0x1.0 OASIS (Open Access Series of Imaging Studies) 61 42.82 (20.42) 18-89 20/41 http://www.oasis-brains.org/ Siemens Vision (1.5T)* MPRAGE 1.0x1.0x1.25 TRAIN-39 36 22.77 (2.52) 18-28 10/25 INDI Siemens Allegra (3T) MPRAGE 1.33x1.33x1.3 UK BIOBANK 9080 62.45 (7.48) 45-79 4334/4746 https://biobank.ctsu.ox.ac.uk/crystal/crystal/docs/brain_mri.pdf Siemens Skyra (3T) MPRAGE 1.0x1.0x1.0 Training set total 10307 59.40 (12.33) 18-89 4961/5346 - - - - INDI = International Neuroimaging Data-sharing Initiative (http://fcon_1000.projects.nitrc.org) COINS = Collaborative Informatics and Neuroimaging Suite (http://coins.mrn.org) LONI = Laboratory of Neuro Imaging Image & Data Archive (https://ida.loni.usc.edu) ABIDE consortiums comprising data from various sites with different scanners/parameters *OASIS scans were acquired four times and then averaged to increase signal-to-noise ratio. A preprint - October 9, 2019
Table S3: List of used Freesurfer features
lh bankssts thickness
lh cuneus thickness
lh inferiorparietal thickness
lh lateraloccipital thickness
lh medialorbitofrontal thickness
lh paracentral thickness
lh parstriangularis thickness
lh posteriorcingulate thickness
lh rostralanteriorcingulate thickness
lh superiorparietal thickness
lh frontalpole thickness
lh insula thickness
rh caudalanteriorcingulate thickness
rh entorhinal thickness
rh inferiortemporal thickness
rh lateralorbitofrontal thickness
rh middletemporal thickness
rh parsopercularis thickness
rh pericalcarine thickness
rh precentral thickness
rh rostralmiddlefrontal thickness
rh superiortemporal thickness
rh temporalpole thickness
Left-Cerebellum-White-Matter
Left-Thalamus-Proper
Left-Pallidum
Brain-Stem
CSF
Left-vessel
Right-Thalamus-Proper
Right-Pallidum
Right-Accumbens-area
CC Posterior
CC Mid Anterior
CortexVol
CerebralWhiteMatterVol
SupraTentorialVol
MaskVol
EstimatedTotalIntraCranialVol
lh caudalanteriorcingulate thickness
lh entorhinal thickness
lh inferiortemporal thickness
lh lateralorbitofrontal thickness
lh middletemporal thickness
lh parsopercularis thickness
lh pericalcarine thickness
lh precentral thickness
lh rostralmiddlefrontal thickness
lh superiortemporal thickness
lh temporalpole thickness
lh MeanThickness thickness
rh caudalmiddlefrontal thickness
rh fusiform thickness
rh isthmuscingulate thickness
rh lingual thickness
rh parahippocampal thickness
rh parsorbitalis thickness
rh postcentral thickness
rh precuneus thickness
rh superiorfrontal thickness
rh supramarginal thickness
rh transversetemporal thickness
Left-Cerebellum-Cortex
Left-Caudate
3rd-Ventricle
Left-Hippocampus
Left-Accumbens-area
Right-Cerebellum-White-Matter
Right-Caudate
Right-Hippocampus
Right-VentralDC
CC Mid Posterior
CC Anterior
lhCerebralWhiteMatterVol
SubCortGrayVol
SupraTentorialVolNotVent
BrainSegVol-to-eTIV
lh caudalmiddlefrontal thickness
lh fusiform thickness
lh isthmuscingulate thickness
lh lingual thickness
lh parahippocampal thickness
lh parsorbitalis thickness
lh postcentral thickness
lh precuneus thickness
lh superiorfrontal thickness
lh supramarginal thickness
lh transversetemporal thickness
rh bankssts thickness
rh cuneus thickness
rh inferiorparietal thickness
rh lateraloccipital thickness
rh medialorbitofrontal thickness
rh paracentral thickness
rh parstriangularis thickness
rh posteriorcingulate thickness
rh rostralanteriorcingulate thickness
rh superiorparietal thickness
rh frontalpole thickness
rh insula thickness
rh MeanThickness thickness
Left-Putamen
4th-Ventricle
Left-Amygdala
Left-VentralDC
Right-Cerebellum-Cortex
Right-Putamen
Right-Amygdala
Right-vessel
CC Central
rhCortexVol
rhCerebralWhiteMatterVol
TotalGrayVol
SupraTentorialVolNotVentVox
MaskVol-to-eTIV
19
20
Figure S1: Model counts for the different distributions for one single repetition over the different
generations. While Left shows the model count for the normal data distribution, Right illustrates
the model counts for the uniform distribution. Using both distributions, the most common models
explored by TPOT are Random Forrest Regression, Extra Tree Regressors and Elastic Nets.
Gaussian Process RegressionRelevance Vector RegressionSupport Vector RegressionRandom Forest RegressionK-Nearest Neighbours RegressionLinear RegressionRidge RegressionElastic NetsExtra Trees RegressorLasso Lars Decision Tree Regression7550250125100Original DistributionUniform Distribution |
1810.00786 | 2 | 1810 | 2019-12-02T00:19:07 | Caulking the Leakage Effect in MEEG Source Connectivity Analysis | [
"q-bio.NC",
"stat.ME"
] | Simplistic estimation of neural connectivity in MEEG sensor space is impossible due to volume conduction. The only viable alternative is to carry out connectivity estimation in source space. Among the neuroscience community this is claimed to be impossible or misleading due to Leakage: linear mixing of the reconstructed sources. To address this problematic we propose a novel solution method that caulks the Leakage in MEEG source activity and connectivity estimates: BC-VARETA. It is based on a joint estimation of source activity and connectivity in the frequency domain representation of MEEG time series. To achieve this, we go beyond current methods that assume a fixed gaussian graphical model for source connectivity. In contrast we estimate this graphical model in a Bayesian framework by placing priors on it, which allows for highly optimized computations of the connectivity, via a new procedure based on the local quadratic approximation under quite general prior models. A further contribution of this paper is the rigorous definition of leakage via the Spatial Dispersion Measure and Earth Movers Distance based on the geodesic distances over the cortical manifold. Both measures are extended for the first time to quantify Connectivity Leakage by defining them on the cartesian product of cortical manifolds. Using these measures, we show that BC-VARETA outperforms most state of the art inverse solvers by several orders of magnitude. | q-bio.NC | q-bio | Caulking the "Leakage Effect" in MEEG Source Connectivity Analysis
Eduardo Gonzalez-Moreira1,3,4#, Deirel Paz-Linares1,2#, Ariosky Areces-Gonzalez1,5, Rigel Wang1, Jorge Bosch-
Bayard4, Maria Luisa Bringas-Vega1,2 and Pedro A. Valdés-Sosa1,2*
# contributed equally as first authors
[email protected]
[email protected]
* corresponding author
[email protected]
Affiliations:
(1) The Clinical Hospital of Chengdu Brain Science Institute, MOE Key Lab for Neuroinformation, University of
Electronic Science and Technology of China, Chengdu, China
(2) Cuban Neuroscience Center, La Habana, Cuba
(3) Centro de Investigaciones de la Informática, Universidad Central "Marta Abreu" de Las Villas
(4) Departamento de Neurobiología Conductual y Cognitiva, Instituto de Neurobiología, Universidad Nacional
Autónoma de México. Boulevard Juriquilla 3001, Querétaro, 76230, México
(5) Departamento de Informatica, Universidad de Pinar del Rio, Pinar del Rio, Cuba.
Summary
Simplistic estimation of neural connectivity in MEEG sensor space is impossible due to head volume
distortion. The only viable alternative is to carry out connectivity estimation in source space. Among the
neuroscience community this is claimed to be impossible or misleading due to "Leakage": linear mixing of
the reconstructed sources. To address this problematic we propose a novel solution method that "caulks"
the "Leakage" in MEEG source activity and connectivity estimates: Brain Connectivity Variable Resolution
Tomographic Analysis (BC-VARETA). It is based on a joint estimation of source activity and connectivity in
the frequency domain representation of MEEG time series. To achieve this, we go beyond current
methods that assume a fixed gaussian graphical model for source connectivity. In contrast we estimate
this graphical model in a Bayesian framework by placing priors on it, which allows for highly optimized
computations of the connectivity, via a new procedure based on the local quadratic approximation under
quite general prior models. A further contribution of this paper is the rigorous definition of leakage via
the Spatial Dispersion Measure and Earth Movers Distance, based on the geodesic distances over the
cortical manifold. Both measures are extended for the first time to quantify "Connectivity Leakage" by
defining them on the cartesian product of cortical manifolds. Using these measures, we show that BC-
VARETA outperforms most state of the art inverse solvers by several orders of magnitude.
Highlights
• Nonlinear method BC-VARETA allows to caulk the leakage in MEEG source activity and
connectivity estimates.
• Optimized solution of gaussian graphical models with general penalization function via local
quadratic approximation.
• Quantification of the connectivity leakage by the extension of spatial dispersion measure and
earth movers distance to cartesian product of cortical manifold spaces.
• Demonstrated the inefficacy of MEEG sensors space connectivity analysis by comparison
against source space analysis with BC-VARETA
1 Introduction
The estimation of neural connectivity from EEG or MEG data is at the crossroad today. The essential
debate is whether these estimates should be obtained in sensor space or source space and the limitations
of each of these approaches, see a discussion on this topic in (Palva, et al., 2018). Equating neural
connectivities in brain networks to the statistical dependencies at the sensors space is a common fallacy
(Blinowska, 2011). It is rendered invalid by two combined effects: head volume distortive inhomogeneities
and mixture of source activity from the whole brain (Brunner et al., 2016). Alternatively, there has been a
quest for measures that somehow 'ameliorate' the volume distortive effect (Kaminski and Blinowska,
2014; Kaminski and Blinowska, 2017). This is also an unlikely enterprise. None of these procedures use
explicitly knowledge about the Forward Model or Lead Field to cancel its effect (Van de Steen et al., 2016).
It would thus seem that the only sensible procedure to estimate neural connectivity would be to analyze
interactions between estimated sources given the Forward Model inversion (Inverse Problem or
Electrophysiological Sources Imaging). While attempting to counteract the volume distortive effect, with
Electrophysiological Sources Imaging methods, also suffers from two difficulties. First, for different
reasons, there are neural generators whose activity is not reflected at the sensors. Second, any of the
methods to estimate sources suffers from the "Leakage Effect".
The first problem refers to that invisible sources can only be encountered by prior knowledge encoded
into the source activity and connectivity estimation procedure (Krishnaswamy et al., 2017). The second,
Leakage (Schoffelen and Gross, 2009), refers to a blurred reconstruction of point sources that entails
spillover of activity between them thus distorting the estimates of their inter-connections. Leakage is a
well-known problem in all medical imaging techniques but is much more severe for MEEG source
reconstruction methods. It is not surprising that there are many attempts to modify MEEG inverse
methods to ameliorate or avoid source Leakage (Freeman, 1980; Brookes et al., 2012; McCoy and Troop,
2013; Colclough et al., 2015; Wens et al., 2015; Colclough et al., 2016; Silva Pereira et al., 2017; Hedrich
et al., 2017, Farahibozorg et al., 2018).
Unfortunately, a difficulty in evaluating "Leakage Correctors" is the lack of a direct metric of the distortions
in connectivity. Rather, what exists are measures of Leakage distortion of source estimators--not
connectivity. One such measure is the dispersion of the "Point Spread Function" (PSF) -- the reconstruction
of a point sources. There is no doubt that reducing the distortions in activation, i.e. Type I Leakage, will be
a good thing for connectivity estimates, but much better would be a direct measure of the distortion in
connectivity, i.e. Type II Leakage. Another difficulty towards Leakage correction, is that the most
stablished methods are based on connectivity postprocessing of estimated source activity given by a
source localization procedure. Thus, they do not make use of more consistent models of sources activity
and connectivity, i.e. system identification. In this sense the state of the art of sophisticated Non-linear
source activity and connectivity estimators has been overlooked (Patterson and Thompson, 1971, Harville,
1977; Friston et al., 2007; Wipf et al., 2009; Belardinelli et al., 2012; Wu et al., 2016; Valdes-Sosa, 1996,
Bosch-Bayard, et al., 2001). These are precisely the points of this paper:
1- Presenting a family of both Source Activity and Connectivity Non-linear estimators, denominated
here Brain Connectivity Variable Resolution Tomographic Analysis (BC-VARETA).
2- Introducing a highly optimized method for the connectivity estimation based on the local
quadratic approximation of hermitic gaussian graphical models with penalization function of the
LASSO family.
3- Proposing measures of the Type I and II Leakage distortion in the context of BC-VARETA, that will
be generalizable to other Non-linear MEEG methods.
2 Methods
2.1 Bayesian Model of MEEG Sources Activity and Connectivity
For the MEEG recorded signals , in the Fourier space, the Forward Model at a single Frequency Component
is expressed by the general equation below. Check Appendix for the mathematical notation and definition
of variables across this manuscript.
𝒗𝓂 = 𝐋𝒗𝜾𝜾𝓂 + 𝝃𝓂; 𝓂 ∈ 𝕄
[2.1.1]
The vectors of MEEG measurements 𝒗𝓂 and signal noise 𝝃𝓂, are independent Random Samples 𝓂 ∈ 𝕄,
defined on the p-size Scalp Sensors (Electrodes) Space 𝔼, meanwhile the sources activity random vector
𝜾𝓂 is defined on the q-size discretized Gray Matter Space 𝔾. The p × q-size Source to Data Transfer
Operator (SDTO) 𝐋 (transformation of spaces 𝔾 → 𝔼) builds on a discretization of the Lead Field from a
head conductivity model (Riera and Fuentes, 1998; Valdés-Hernández et al. 2009).
Construing a Model of MEEG source localization and connectivity, upon the equation [2.1.1], can be
tackled in general by the Bayesian formalism (MacKay, 2003), which involves categorizing as random
variables the MEEG Measurements (Data) 𝒗𝓂 and Source Activity (Parameters) 𝜾𝓂. The model builds on
a Parametric representation of the signal noise 𝝃𝓂 and sources activity 𝜾𝓂 Probability Density Functions
(pdf). It is commonly given by hierarchically conditioned Gaussian models, i.e. Multivariate Circularly
Symmetric Complex Gaussians 𝑁ℂ, of the Data Likelihood and Parameters Prior. The parametrization
within these distributions introduces an additional category of random variables denominated
(Hyperparameters) 𝚵. See its schematic representation by More-Penrose diagrams in Figure 1.
Likelihood
𝒗𝓂𝜾𝓂, 𝚵~𝑁p
ℂ(𝒗𝓂𝐋𝒗𝜾𝜾𝓂, 𝚺𝝃𝝃); 𝓂 ∈ 𝕄
𝚺𝝃𝝃 = 𝜎𝝃
2𝐑𝝃𝝃; 𝜎𝝃
2 ∈ 𝚵
Parameters Prior
𝜾𝓂𝚵~𝑁q
ℂ(𝜾𝓂0, 𝚺𝜾𝜾); 𝓂 ∈ 𝕄
𝚯𝜾𝜾 = 𝚺𝜾𝜾
−1; 𝚯𝜾𝜾 ∈ 𝚵
Hyperparameters Priors
𝚯𝜾𝜾~ 𝑒𝑥𝑝(Π(𝚯𝜾𝜾, 𝐀)α)
2~ 𝑒𝑥𝑝(1 𝜎𝝃
2⁄
𝜎𝝃
𝑏)
[2.1.2]
[2.1.3]
[2.1.4]
[2.1.5]
2}.represents the Model Parametrization (Hyperparameters). The Noise Covariance or
Above, 𝚵 = {𝚯𝜾𝜾, 𝜎𝝃
Data GGM conditional Covariance matrix 𝚺𝝃𝝃 of formula [2.1.2], is assumed to be composed by a scalar
2 , representing the unknown nuisance Variance, and a known matrix 𝐑𝝃𝝃 (in the
random variable 𝜎𝝃
Cartesian space product 𝔼 × 𝔼) of the noise Covariance structure. The noise Covariance structure encodes
information about the sensors correlated activity. These correlations are given either by shorting currents
to
inputs
the scalp conductivity or common
from
between adjacent electrodes' due
instrumentation/environmental noisy sources. The noise Precision (Variance) Exponential (Jeffry
Improper) Gibbs pdf set up on, see formula [2.1.5], aims to bypass the nuisance level that could be
assimilated into the Parameters. This is possible due to the monotonically increasing values of the noise
Variance probability density assigned by the Jeffry Improper Prior, which allows for encoding the
information about the noise inferior threshold into the parameter 𝑏.
The inverse of the Covariance matrix 𝚺𝜾𝜾, Precision matrix 𝚯𝜾𝜾 (in the Cartesian space product 𝔾 × 𝔾), of
the Parameters GGM, represents the source connectivity, see formula [2.1.3]. The general penalization
function Π at the argument of the exponential Prior in formula [2.1.4], imposes certain Structured Sparsity
pattern on the connectivity. The Structured Sparsity can be encoded given information from the Gray
Matter anatomical segmentation, by penalizing the groups of variables corresponding to the Gray Matter
areas Intra/Inter-connections. The matrix 𝐀 (in the Cartesian space product 𝔾 × 𝔾) within the General
Penalization function, represents a probability mask of the anatomically plausible connections. The
probability mask in case of the dense short-range connections, e.g. Intra-Cortical Connections, is defined
as a deterministic spatially invariant empirical Kernel of the connections strength decay with distance. For
the long-range connections, e.g. Inter-Cortical connection, it is given by probabilistic maps of the White
Matter tracks connectivity strength from Diffusion Tensor Imaging (DTI). The global influence in the
Parameters GGM of the connectivity Structured Sparsity penalization Π(𝚯𝜾𝜾, 𝐀) is controlled by the Scale
Parameter (Regularization Parameter) α, which can be fitted to the Data by means of some statistical
criteria of goodness.
In general, the construction of the Model corresponds to the ubiquitous Bayesian representation of Linear
State Space Models (LSSM), in both Time (Real) and Frequency (Complex) domain. The LSSM Data
(Observation Equation) and Parameters (Autoregressive State Equation) are modeled by Multivariate
Gaussian pdfs, whereas the Connectivity (Autoregression Coefficients Matrix) is represented by the
Parameters' Precision Matrix, (Wills et al., 2009; Faes et al., 2012; Galka et al., 2004; Pascual-Marqui et
al., 2014; Valdes-Sosa 2004; Lopes da Silva et al., 1980; Baccalá and Sameshima, 2001; Babiloni et al.,
2005).
Figure 1: More-Penrose diagram of the MEEG Source Activity and Connectivity Bayesian Model and its Priors. The
model Variables and Prior knowledge are represented with gray circles and squares respectively. The filled arrows
represent Random Variables generation by a specific pdf and the unfilled arrows the corresponding pdf
parametrization.
2.2 Type I Leakage and the Brain Connectivity Variable Resolution Tomographic Analysis
The Bayesian Model depicted in Section 2.1 revendicates a large family of Linear, or Non-Linear/Hybrid
iterated Sources Activity estimators, see Table 1. Along this family, the formulation of the source activity
estimator, denominated First Level of Inference into the Bayesian Formalism, is also common to BC-
VARETA. The estimators of this Model are given independently for each frequency component, since it
does not consider Priors that could link the analysis along Frequency Domain. The First Level of Inference
consists on maximizing the Multivariate Gaussian pdf derived from Parameters' Posterior Analysis. This is
given upon fixed values of the Hyperparameters 𝚵(𝑘) , within an outer cycle indexed (𝑘) of the
Parameters and Hyperparameters iterated computation, see Section C of (Paz-Linares and Gonzalez-
Moreira et al, 2018):
𝜾𝓂𝒗𝓂, 𝚵(𝑘)~𝑁q
(𝑘))
ℂ (𝜾𝓂𝜾𝓂
(𝑘), 𝚺𝜾𝜾
[2.2.1]
(𝑘) , given the Data 𝒗𝓂 , is
The Parameters' Posterior Mean (Sources Activity iterated estimator) 𝜾𝓂
expressed through the iterated auxiliary quantities of the Data to Sources 'Transference Operator' 𝐓 (𝑘)
(𝑘) (in the space 𝔾 × 𝔾).
(transformation of spaces 𝔼 → 𝔾) and the Parameters' Posterior Covariance 𝚺𝜾𝜾
Both depending on the Hyperparameters iterated estimators 𝚵(𝑘).
(𝑘) ← 𝐓 (𝑘)𝒗𝓂
𝜾𝓂
𝐓 (𝑘) = 𝚺𝜾𝜾
(𝑘)𝐋𝒯 (𝜎𝝃
2(𝑘)
−1
𝐑)
(𝑘) = (𝐋𝒯 (𝜎𝝃
𝚺𝜾𝜾
2(𝑘)
−1
𝐑)
𝐋 + 𝚯 𝜾𝜾
(𝑘))
−1
Table 1: Family of Linear/Non-Linear/Hybrid Source Activity estimators
Parameters
Covariance
Data
Conditional
Covariance
Algorithm
[2.2.2]
[2.2.3]
[2.2.4]
Linear
Non-
Linear
Hybrid
Linear
Minimum Norm Estimator
(MME) (Hämäläinen and
Ilmoniemi, 1994)
Low Resolution
Electromagnetic
Tomography (LORETA)
(Pascual-Marqui et al.,
1994).
Exact Low Resolution
Electromagnetic
Tomography (eLORETA)
(Pascual-Marqui et al.,
2006).
Constrained to
Scaled Identity
Matrix. Fixed
value.
Connectivity-
Postprocessing.
Fixed Laplacian
Operator.
Connectivity-
Postprocessing.
Constrained to a
Diagonal Matrix.
Prior Free.
Connectivity-
Postprocessing.
Constrained
Explicit Parameters Posterior
Mean estimator.
to Scaled
Identity
Matrix.
Fixed value.
Full Matrix.
Prior Free.
Iterated Explicit Parameters
Posterior Mean estimator and
Hybrid
empirical formulas of the
Conditional Data Covariance.
Full Matrix.
Prior Free.
Iterated Explicit Parameters
Posterior Mean estimator and
Hybrid
empirical formulas of the
Conditional Data Covariance
and Parameters Variances.
Provides Zero Localization
Error in case of a single Source
reconstruction.
Standardized Low
Full Matrix
Resolution Electromagnetic
(Connectivity).
Full Matrix.
Prior Free.
Iterated Explicit Parameters
Posterior Mean estimator and
Hybrid
Tomography (eLORETA)
(Pascual-Marqui, 2002).
Prior Free.
Variable Resolution
Tomographic Analysis
(VARETA) (Valdes-Sosa,
1996, Bosch-Bayard, et al.,
2001).
Full Matrix
Constrained
(Connectivity).
Prior Free.
to Scaled
Identity
Matrix. Prior
(EM) formulas of the
free.
empirical formulas of the
Conditional Data Covariance
and Parameters Covariances.
Iterated Explicit Parameters
Posterior Mean estimator and
Expectation Maximization
Conditional Data Variance
and Parameters Covariance.
Iterated Explicit Parameters
Posterior Mean estimator and
Empirical Bayes (EB) formulas
of the Conditional Data
Variance and Parameters
Variances. Induces Sparsity by
pruning the Parameters
Variances estimates.
Non-
linear
Non-
linear
Automatic Relevance
Determination (ARD) (Neal,
1998; Tipping, 2001; Sato et
al., 2004; Wipf et al., 2006;
Constrained to a
Diagonal Matrix.
Jeffrey Improper
Constrained
to Scaled
Identity
Priors.
Matrix. Prior
Wipf and Rao, 2007;
Daunizeau and Friston,
Connectivity-
Postprocessing.
free.
2007).
Full Matrix
Constrained
to Scaled
Identity
Matrix. Prior
free.
Iterated Explicit Parameters
Posterior Mean estimator and
Non-
linear
Maximum Likelihood plus
Restrictions formulas of the
Conditional Data Variance
and Function Basis
Hyperparameters.
Restricted Likelihood
Maximization (ReLM)
(Patterson and Thompson,
1971, Harville, 1977; Friston
et al., 2007; Wipf et al.,
2009; Belardinelli et al.,
2012; Wu et al., 2016)
Structured Sparse Bayesian
Learning (SSBL) (Wipf et al.,
2010; Zhang and Rao, 2011;
Babacan et al., 2012; Wan et
al., 2014; Balkan et al. 2014;
Paz-Linares et al., 2017;
Zhang et al., 2016; Zhang et
al., 2017).
(Connectivity)
hyper-
parametrized on
Function Basis.
Sparse Priors on
the Function Basis
Hyperparameters
.
Constrained to a
Diagonal Matrix
(or Block
Diagonal). Sparse
Gamma Priors.
Connectivity-
Postprocessing.
Constrained
to Scaled
Identity
Matrix.
Jeffrey
Improper
Prior.
Iterated Explicit Parameters
Posterior Mean estimator and
Empirical Bayes (EB) formulas
of the Conditional Data
Variance and Parameters
Variances.
Non-
linear
2.2.1 Measures of Type I (Activity) Leakage
Evaluating the Type I Leakage given to Volume Conduction distortion itself should be done in an isolated
point Source scenario. In the general context of Signal Processing this situation is represented by the
concept of 'Point Spread Function' (PSF). In this case it is defined as the estimated Sources Activity from
synthetic Data due to a single Unitary Source, at an arbitrary point 𝑗0 of the Gray Matter space 𝔾. The
Activity of a single Unitary Active Source is mathematically represented by the Kronecker Delta, denoted
by the vector 𝛿(𝑗0):
𝛿𝑗(𝑗0) = {
0, 𝑗 ≠ 𝑗0
1, 𝑗 = 𝑗0
[2.2.5]
Computing the PSF involves the projection 𝔼 → 𝔾, by effectuating formulas [2.2.2], [2.2.3] and [2.2.4] till
convergence, given inputted Data obtained from the projected Kronecker Delta 𝔾 → 𝔼 in [2.2.5], i.e.
𝑙𝑖𝑚𝑘→∞ (𝐓 (𝑘)𝐋𝛿(𝑗0)). Expressing the PSF by this formula alone constitutes an idealization of the real
MEEG Source Localization scenario. A fair approach should consider avoiding the Inverse Crime (Kaipio &
Somersalo 2004), which consists on modifying the Lead Field of Data generation from the one used within
the Inference framework, and corrupting the Data with Noise, from all possible sources including
biological, environmental and instrumentation, at acceptable levels.
In a situation as such, a suitable definition of PSF, will be given by the Variances of the iterated auxiliary
(𝑘) (in the Cartesian space product 𝔾 × 𝔾). An
quantity of Sources Activity Empirical Covariance (SEC) 𝐒
𝜾𝜾
explicit and compact expression of the SEC can be attained by the projection 𝔼 × 𝔼 → 𝔾 × 𝔾 of the Data
Empirical Covariance 𝐒𝒗𝒗 (in the Cartesian space product 𝔼 × 𝔼) through the Transference Operator 𝐓 (𝑘),
see Section D of (Paz-Linares and Gonzalez-Moreira et al, 2018).
𝐒
(𝑘) ← 𝐓 (𝑘)𝐒𝒗𝒗𝐓 (𝑘)†
𝜾𝜾
𝐒𝒗𝒗 =
1
m
∑
m
𝓂=1
𝒗𝓂𝒗𝓂
†
[2.2.6]
[2.2.7]
𝑠𝑖𝑚, obtained from the projection 𝔾 × 𝔾 → 𝔼 × 𝔼,
The computation of PSF is effectuated on a synthetic 𝐒𝒗𝒗
through an independent Lead Field 𝐋𝑠𝑖𝑚, of the Kronecker Delta 𝛿(𝑗0) plus Noise samples, represented
by the Real/Complex vectors 𝝃𝓂:
𝑠𝑖𝑚 = 𝐋𝑠𝑖𝑚𝛿(𝑗0) + 𝝃𝓂, 𝓂 = 1 … m
𝒗𝓂
[2.2.8]
The explicit formula of the PSF, denoted here as 𝓟, is given by the diagonal values (Variances) of the SEC
iterated estimator, in formula [2.2.6], after convergence of the outer cycle:
𝓟(𝑗0) ← 𝑙𝑖𝑚𝑘→∞ (𝑑𝑖𝑎𝑔 (𝐓 (𝑘)𝐒𝒗𝒗
𝑠𝑖𝑚𝐓 (𝑘)†
))
[2.2.9]
The PSF distortion can be evaluated in general by any Measure of its difference to Ground Truth,
Kronecker delta in [2.2.5]. Particularly, for single point spreading like scenarios and when the Gray Matter
space 𝔾 is collapsed to a bidimensional Manifold, i.e. surfaces of different Brain structures, a measure
universally adopted is the Spatial Dispersion (SD) according to the Geodesic Distance. In such scenario the
Type I Leakage due to Volume Conduction distortion is expressed through the Spatial Dispersion of the
Point Spread Function (SD-PSF). It is defined as the Standard Deviation of the Geodesic Distance 𝑑𝑗𝑗0
between pairs of points indexed (𝑗, 𝑗0), for 𝑗 = 1 … q, in the Gray Matter space 𝔾, with probability mass
given by the absolute values of the PSF, denoted mathematically as 𝑺𝑫𝛿(𝑗0)(𝓟(𝑗0)), see formula below:
∑
𝑺𝑫𝛿(𝑗0)(𝓟(𝑗0)) = √
q
𝑗=1
𝑞
∑
𝑗=1
2 𝓟𝑗(𝑗0)
𝑑𝑗𝑗0
𝓟𝑗(𝑗0)
[2.2.10]
In a more general scenario, where the Data is given by a composition of multiple Unitary Active Sources
𝑠𝑖𝑚 = 𝐋𝑠𝑖𝑚𝛿(𝑗0, 𝑗1, ⋯ ) + 𝝃𝓂, 𝓂 = 1 … m, the concept of PSF requires to be extended, i.e. 'Generalized
𝒗𝓂
Spread Function' GSF, denoted mathematically as 𝓟(𝑗0, 𝑗1, ⋯ ). In this case the Type I Leakage is given by
the composition of two distortive effects, i.e. the Volume Conduction and superposed Scalp projection of
multiple Sources, which cannot be measured by simply using the SD. The Earth Movers' Distance (EMD)
between the GSF and the Ground Truth (EMD-GSF) would suit as a measure representative of the
distortion in this general scenario, denoted mathematically as 𝑬𝑴𝑫𝛿(𝑗0,𝑗1,⋯ )(𝓟(𝑗0, 𝑗1, ⋯ )). In the State of
the Art of Inverse Solution the EMD has been stablished as the most sensitive when compared to other
quality measures, i.e. typical Dipole Localization Error or Binary Classification, i.e. Receiving Operating
Characteristic, Precision, Recall and F1. See its definition in (Molins et al. 2008, Grova et al., 2006, Haufe
et al., 2008). This concept is applicable in general to any definition of the Active sources, e.g. Vector made
of patches with random extensions and random elements.
2.2.2 Second Level of Inference of the Brain Connectivity Variable Resolution Tomographic
analysis and its influence on variable selection (Leakage)
Meanwhile, the First Level of Inference of the Methods described in Table 1 constitutes an invariant, a
distinct aspect was its Parametrization structure and Priors defined. This is definitory at the denominated
Second Level of inference or estimation of Hyperparameters 𝚵(𝑘), which biases the variables selection into
the iterated estimation scheme, and thus the amount of Leakage carried by the Parameters and
Hyperparameters. This effect is determined, at the First Level of Inference, by the Resolution (sparsity) in
Variable Selection of the Transference Operator 𝐓 (𝑘), which is influenced by the scale and/or degree of
2(𝑘) estimators. Those transitively influence the
sparsity of the Precision matrix 𝚯 𝜾𝜾
(𝑘) and Data Nuisance 𝜎𝝃
(𝑘), see formulas [2.2.3],
Resolution through its balance into the Parameters' Posterior Covariance Matrix 𝚺𝜾𝜾
[2.2.4], [2.2.6] and [2.2.7]. The Resolution in the estimation will be globally determined by two interacting
elements. First: The choices of Hyperparameters Posterior analysis strategies, i.e. EB, EM, ARD, ReLM, etc.
Second: The bias introduced by the Hyperparameters Priors, i.e. ad hoc structure of the Data Conditional
2.
Covariance 𝐑, Prior pdfs on the Sources' Activity Precisions matrix 𝚯𝜾𝜾 and Data Nuisance Variance 𝜎𝝃
Thus, the essential constituent of the BC-VARETA methodology is the definition of the Priors and Inference
strategy at the Second Level along with adequate statistical guarantees.
The EM algorithm (Dempster et al., 1977; Liu and Rubin, 1994; McLachlan and Krishnan, 2007) constitutes
an explicit way to tackle the Hyperparameters estimation. This is done by iteratively maximizing its
m 𝚵), by the so-called
approximated representation of the intractable Data Type II likelihood 𝑝({𝒗𝓂}𝓂=1
Data Expected Log-Likelihood 𝑄(𝚵, 𝚵(𝑘)) . The Data Expected Log-Likelihood is construed by the
m 𝚵)), given
marginalization (Expectation) of the Data and Parameters Joint pdf 𝑙𝑜𝑔(𝑝({𝒗𝓂}𝓂=1
through formulas [2.1.2] and [2.1.3], by the parameters Posterior pdf 𝑝(𝜾𝓂𝒗𝓂, 𝚵(𝑘)), see formula [2.2.1].
The BC-VARETA methodology, meshed to the EM scheme at the Second Level of Inference, constitutes a
special case of Hyperparameters Penalized Posterior Analysis. It is defined as the maximization of the
approximated Hyperparameter's Posterior pdf, given by the combination the Data Expected Likelihood
and Hyperparameter's Priors, see Section D of (Paz-Linares and Gonzalez-Moreira et al, 2018):
m , {𝜾𝓂}𝓂=1
𝚵{𝒗𝓂}𝓂=1
m , 𝚵(𝑘)~𝒆𝑄(𝚵,𝚵(𝑘))𝑝(𝚵)
[2.2.11]
The Data Expected Log-Likelihood has a close form expression on the Hyperparameters, given the Data
Empirical Covariance and the iterated estimators of the Data to Sources Transference Operator, Sources
Posterior Covariance, and an iterative auxiliary quantity denominated Effective Sources Empirical
Covariance (ESEC) 𝚿𝜾𝜾
(𝑘):
𝑄(𝚵, 𝚵(𝑘)) = −mp 𝑙𝑜𝑔(𝜎𝑒
2) − (m 𝜎𝑒
2⁄
) 𝑡𝑟 ((𝐈p − 𝐋𝐓(𝑘)
)
†
𝐑−1 (𝐈p − 𝐋𝐓(𝑘)
) 𝐒𝒗𝒗) ⋯
2⁄
−(m 𝜎𝑒
(𝑘)
(𝑘) = 𝚺𝜾𝜾
𝚿𝜾𝜾
(𝑘)
)𝑡𝑟 (𝐋𝒯𝐑−1𝐋𝚺𝜾𝜾
+ 𝐒
(𝑘)
𝜾𝜾
) + m 𝑙𝑜𝑔𝚯𝜾𝜾 − m 𝑡𝑟 (𝚯𝜾𝜾𝚿 𝜾𝜾
(𝑘))
[2.2.12]
[2.2.13]
The Data Expected Log-Likelihood in formula [2.2.12] is a Concave function on 𝚵, but the Concavity of its
associated Posterior pdf in formula [2.2.11] can be only ensured with a pertinent definition of the Priors.
In addition, the approximated Hyperparameter's Posterior analysis of EM algorithm only guarantees
m ) ∝
reaching a
m 𝚵)𝑝(𝚵). The proximity of this local maximum to the global maximum is determined also by
𝑝({𝒗𝓂}𝓂=1
the selection of the Priors. The Gibbs Priors of formulas [2.1.5] and [2.1.6] guarantee the Concavity
whenever the exponent arguments redress the mathematical definition of norm, i.e. a non-negative scalar
function that satisfies 1) triangle inequality, 2) absolutely-scalable and 3) positive-definite.
local maximum of the actual Hyperparameters' Posterior,
i.e. 𝑝(𝚵{𝒗𝓂}𝓂=1
The Precision Matrix estimation, given by Posterior Analysis of equation [2.2.11], is expressed through the
minimization of a Target Function that resembles the structure of an equivalent Sources GGM with
(𝑘). Under the convention α = λ𝑚, where the λ
Effective Sources Empirical Covariance (ESEC) Matrix 𝚿𝜾𝜾
can be interpreted as the GGM effective Regularization Parameter, the Precision matrix estimator is given
by the following formula, see Section E of (Paz-Linares and Gonzalez-Moreira et al, 2018):
(𝑘+1) ← 𝑎𝑟𝑔𝑚𝑖𝑛𝚯𝜾𝜾
𝚯
𝜾𝜾
{− log𝚯𝜾𝜾 + 𝑡𝑟 (𝚯𝜾𝜾𝚿𝜾𝜾
(𝑘)) + λΠ(𝚯𝜾𝜾, 𝐀)}
[2.2.14]
Setting up Sparse Models as General Penalty Function has been stablished in similar scenarios of Variable
Selection, i.e. Graphical Models estimation (Jordan, 1998; Attias, 2000; Friedman et al, 2008; Mazumder
et al. 2012; Wang, 2012; Wang, 2014; Schmidt, 2010; Hsieh, 2014; Danaher et al., 2014; Zhang and Zou,
2014; Yuan and Zheng, 2017; Drton and Maathuis, 2017). Some of the most common Penalty Functions
are referred into the family of Graphical LASSO Models, see Table 2 below.
Table 2: Graphical LASSO family Penalty Functions Models
Graphical LASSO (GLASSO)
Graphical Elastic Net (GENET)
Penalty function Π(𝚯𝜾𝜾, 𝐀)
‖𝚯𝜾𝜾‖1,𝐀
‖𝚯𝜾𝜾‖1,𝐀1 + ‖𝚯𝜾𝜾‖2,𝐀2
2
Graphical Group Lasso (GGLASSO)
∑
𝕢
𝒿=1
‖𝚯𝜾𝜾(𝕂𝒿)‖
2,𝐀
(𝕂𝒿)
; 𝕂𝒿 ⊂ 𝔾 × 𝔾; 𝒿 = 1 ⋯ n
Nevertheless, as for choosing the Penalty Function and Regularization Parameter there is not ubiquitous
rule. It is usually assumed that, for a given Penalty Function, fitting the Regularization Parameter by some
Statistical Criteria by would suffice to rule out the ambiguity on the Variable Selection sparsity level
(Resolution). This approach does not provide a Statistical guarantee, as discussed in (Jankova and Van De
Geer, 2015, 2017), due the biasing introduced in the estimation by the Sparse Penalty in any case. For the
typical Graphical LASSO, a solution was recently presented in (Jankova and Van De Geer, 2018) through
an unbiased Precision Matrix estimator (𝚯 𝜾𝜾)
(𝑘+1)
.
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
(𝚯 𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
= 2𝚯 𝜾𝜾
(𝑘+1) − 𝚯 𝜾𝜾
(𝑘+1)𝚿𝜾𝜾
(𝑘)𝚯 𝜾𝜾
(𝑘+1)
[2.2.15]
For the conditions Π(𝚯𝜾𝜾, 𝐀) = ‖𝚯𝜾𝜾‖1 and λ = √𝑙𝑜𝑔(q) m⁄ , it is demonstrated, for the elements into the
unbiased estimator of formula [2.2.5] ((𝚯 𝜾𝜾)
, a tendency to the Model Precision Matrix
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
)
𝑖𝑗
elements (𝚯𝜾𝜾)𝑖𝑗
(𝚯 𝜾𝜾
(𝑘+1))
𝑖𝑖
(𝚯 𝜾𝜾
(𝑘+1))
+ (𝚯 𝜾𝜾
(𝑘+1))
𝑖𝑗
rated by √m:
𝑗𝑗
with Complex Normal pdf of consistent variances 𝜎𝑖𝑗 ((𝚯 𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
) =
((𝚯 𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
)
𝑖𝑗
~𝑁1
ℂ (((𝚯 𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
(𝚯𝜾𝜾)𝑖𝑗,
)
𝑖𝑗
𝜎𝑖𝑗((𝚯𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
√m
)
)
[2.2.16]
For the Nuisance Hyperparameter a close form estimator can be obtained, by the unique zero of the
resulting equation given the derivative
2 of formula [2.2.11]. It is expressed in below, by reformulating
the Jeffry Improper Prior pdf Rate Parameter as 𝑏 = mp𝜖, where 𝜖 is the Data Nuisance effective inferior
threshold, see Section E of (Paz-Linares and Gonzalez-Moreira et al, 2018):
𝜕
𝜕𝜎𝑒
2(𝑘+1)
𝜎𝝃
←
†
𝑡𝑟((𝐈p−𝐋 𝐓 (𝑘))
𝐑−1(𝐈p−𝐋 𝐓(𝑘))𝐒𝒗𝒗)
p
+
𝑡𝑟(𝐋𝒯𝐑−1𝐋𝚺
(𝑘)
𝜾𝜾
)
p
+ 𝜖
[2.2.17]
With equation [2.2.17] we attain a formulation in which the biasing given the Nuisance Hyperparameter
Prior relies on the Data Nuisance effective inferior limit 𝜖 and the Noise Covariance Structure 𝐑, which are
quantities that can be experimentally informed. The interpretability of this Prior structure would thus rule
out any ambiguity on its choice.
2.3 Estimation of the MEEG Sources Gaussian Graphical Model
Despite the growing interest on the GGM's given its applicability in several fields, drawbacks of the State
of Art methodologies prevent of utilizing them in the scenario of Electrophysiological Sources Localization
and Connectivity, we mention some of them in Table 3 below.
Table 3: Drawbacks of the Gaussian Graphical Models methodologies
Complex
Variable
Stability
The stablished algorithms do not deal with Complex Variables, limiting their implementation on
the Frequency Domain Connectivity Analysis.
They are based on very unstable strategies, such as Coordinate Updates of the Target Function
Descend Direction or Alternating Direction Methods of Multipliers.
Dimensions The High Dimensionality (common scenario in Brain Connectivity) along with the combined effect
of instability and Inverses computation, at every iteration of the Coordinate Updates cycle,
constitutes an important reason for these algorithms frequent crashing.
Addressing this problem from the perspective of Machine Learning or Optimization Theory is the
common trend to most of Data Analysis groups, while a complete Bayesian insight to the structure
and properties of the GGM and its Precision matrix Priors is still missing in State of the Art.
Bayesian
analysis
Here we propose a revindication of the MEEG SGGM from the Bayesian perspective, that allows for
obtaining a more general class of explicit Precision matrix (Connectivity) estimators. This is done by
considering invariance properties of the GGM Wishart Likelihood and the hierarchical representation of
the GGM Gibbs Priors'. The analogous SGGM representation of the Precision matrix Expected Posterior
(𝑘) of m degrees
pdf can be expressed by a Wishart Likelihood 𝑊ℂ (Real/Complex) on the ESEC matrix 𝚿𝜾𝜾
of freedom and scale matrix (m𝚯𝜾𝜾)−1, combined with the Gibbs Prior pdf of formula [2.1.5].
𝚿𝜾𝜾
(𝑘) 𝚯𝜾𝜾~𝑊q
ℂ (𝚿𝜾𝜾
(𝑘)(m𝚯𝜾𝜾)−1, m)
𝚯𝜾𝜾~𝑒−αΠ(𝚯𝜾𝜾)
𝑊q
ℂ (𝚿𝜾𝜾
(𝑘)(m𝚯𝜾𝜾)−1, m) = 𝚿𝜾𝜾
(𝑘)
(m−q)
𝚯𝜾𝜾𝑚𝒆−m 𝑡𝑟(𝚯𝜾𝜾𝚿
(𝑘)
𝜾𝜾
)
[2.3.1]
[2.3.2]
[2.3.3]
The rigorous theoretical derivation of the strategy for the minimization of the Sources GGM Target
Function is presented in Section G of (Paz-Linares and Gonzalez-Moreira et al, 2018). It is done by a
modified Model through the hierarchical representation (in Complex Variable) of the exponential Priors
by mixtures of Gaussian and Gamma pdf 's (Andrews and Mallows, 1974; Tipping, 2001; Schmolck and
Everson, 2007; Faul and Tipping, 2002; Li and Lin, 2010; Kyung et al., 2010). With this hierarchical model
it is derived a concave Local Quadratic Approximation (LQA) estimation strategy of the SGGM (Fan and Li,
2001; Valdés-Sosa et al., 2006; Sánchez-Bornot et al., 2008). We reformulate the LQA Target Function of
the Precision Matrix into simple Quadratic Model, due to the Standardization GGM's Wishart Likelihood
(Srivastava, 1965; Drton et al., 2008). The explicit Connectivity estimator of the Standard Quadratic Model
is expressed as the unique solution of a Matrix Riccati equation (Lim, 2006; Honorio and Jaakkola, 2013).
(𝑘+1)
2.4 Type II Leakage and the Brain Connectivity Variable Resolution Tomographic Analysis
(𝑘) with unbiased Precision Matrix
The whole estimation strategy consists on the computation ESEC 𝚿𝜾𝜾
(𝚯
computed from its SGGM LQA estimator, after the convergence of the 𝚯
𝜾𝜾
inner cycle indexed 𝑙-th, see Section H of (Paz-Linares and Gonzalez-Moreira et al, 2018):
, at the outer cycle indexed 𝑘-th. The unbiased Precision Matrix, given in formula [2.2.15], is
(𝑘,𝑙) and 𝚪 (𝑘,𝑙), given within an
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
𝜾𝜾)
(𝑘,𝑙+1) ←
𝚯
𝜾𝜾
1
2𝜆
(𝑘))
𝚪 (𝑘,𝑙) ⊙ (√((𝚿
𝜾𝜾
−1
⊘ 𝚪 (𝑘,𝑙))
−2
+ 4𝜆𝐈q − (𝚿𝜾𝜾
(𝑘)−1
−1
⊘ 𝚪 (𝑘,𝑙))
)
𝚪 (𝑘,𝑙) ← (−𝟏q + (𝟏q + 4(λm)2𝐀.2 ⊙ 𝑎𝑏𝑠 (𝚯
.2
(𝑘,𝑙))
𝜾𝜾
.
1
2
)
⊘ (2
1
2
.
)
1
1
2(λm)
2𝐀)
[2.4.1]
[2.4.2]
As in Section 2.3, the elementwise matrix operations in the notation introduce: ⊙ as Hadamard product,
).2 (elementwise matrix
⊘ as Hadamard division, 𝑎𝑏𝑠(
1
2 (elementwise matrix Square Root), 𝐈q (q × q identity matrix), 𝟏q (q × q
) as elementwise matrix absolute value, (
).
Square exponentiation), (
matrix of ones).
2.4.1 Measures of Type II (Connectivity) Leakage
Evaluating the Type II Leakage given to Volume Conduction distortion itself should be done in a scenario
where a single connection (pair connected sources) is present. For the Connectivity the PSF definition
admits an extension to the Cartesian spaces product of Cortical Manifolds 𝔾 × 𝔾, i.e. 'Cartesian Point
Spread Function' (CPSF). It is defined as the Precision matrix estimator from synthetic Data due to a pair
of cortical sources at arbitrary points 𝑗0 and 𝑗1 with Unitary Connectivity (Precision matrix). The Unitary
Precision matrix of the pair of connected Sources is denoted here with the matrix 𝚫(𝑗0, 𝑗1) on the space
𝔾 × 𝔾, given by an outer product of Kronecker Delta functions:
𝚫(𝑗0, 𝑗1) = (𝛿(𝑗0) + 𝛿(𝑗1))(𝛿(𝑗0) + 𝛿(𝑗1))
𝒯
[2.4.3]
In analogy to the construction of the PSF in Section 2.2.1, the computation of the CPSF is effectuated on
𝑠𝑖𝑚, obtained from the projection 𝔾 × 𝔾 → 𝔼 × 𝔼, through an independent Lead Field 𝐋𝑠𝑖𝑚
a synthetic 𝐒𝒗𝒗
(avoiding inverse Crime) of the simulated Cortical Sources Activity plus Noise samples (corruption of Data),
represented by the Real/Complex vectors 𝝃𝓂 . The simulated Sources Activity, represented by the
Real/Complex vectors 𝜾𝓂, is taken from random samples of Gaussian Random Generator with Covariance
+
matrix (𝚫(𝑗0, 𝑗1))
, given by the Pseudoinverse operation over the nonzero block of 𝚫(𝑗0, 𝑗1):
𝑠𝑖𝑚 =
𝐒𝒗𝒗
1
m
∑
m
𝓂=1
𝑠𝑖𝑚𝒗𝓂
𝒗𝓂
𝑠𝑖𝑚†
𝑠𝑖𝑚 = 𝐋𝑠𝑖𝑚𝜾𝓂 + 𝝃𝓂, 𝓂 = 1 … m
𝒗𝓂
+
𝜾𝓂~𝑁q (𝜾𝓂𝟎, (𝚫(𝑗0, 𝑗1))
), 𝓂 = 1 … m
[2.4.4]
[2.4.5]
[2.4.6]
(𝑘) , expressed by
The Precision matrix iterated estimator of formula [2.4.1] comprises the ESEC 𝚿𝜾𝜾
substituting into formula [2.2.13] the SEC formula [2.2.6] computed for synthetic Data Empirical
Covariance from [2.4.5]. Expressing 𝚺𝜾𝜾
(𝚯 𝜾𝜾)
we obtain:
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
(𝑘) and 𝐓 (𝑘) into [2.2.13] through the unbiased Precision Matrix
𝚿𝜾𝜾
(𝑘) = (𝐋𝒯 (𝜎𝝃
2(𝑘)
−1
𝐑)
𝐋 + (𝚯 𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
−1
)
⋯
× (𝐈q +
𝟏
2(𝑘)
(𝜎𝝃
)
2 𝐋𝒯𝐑−1𝐒𝒗𝒗
𝑠𝑖𝑚𝐑−1†
𝐋 ((𝐋𝒯 (𝜎𝝃
2(𝑘)
−1
𝐑)
𝐋 + (𝚯 𝜾𝜾)
(𝑘+1)
𝑢𝑛𝑏𝑖𝑎𝑠𝑒𝑑
−1
)
†
)
)
[2.4.7]
The explicit formula of the CPSF is given by the Precision matrix iterated estimator of formula [2.4.1] with
the ESEC defined in formula [2.4.7] after convergence of the inner cycle (indexed 𝑙) and the outer cycle
(indexed 𝑘):
𝓒(𝑗0, 𝑗1) ←
1
𝜆
𝑙𝑖𝑚 𝑙→∞
𝑘→∞
(𝑘)−1
(𝚪 (𝑘,𝑙) ⊙ (√(𝚿
𝜾𝜾
−2
⊘ 𝚪 (𝑘,𝑙))
+ 4𝜆𝐈q − (𝚿𝜾𝜾
(𝑘)−1
−1
⊘ 𝚪 (𝑘,𝑙))
)) …
−
1
4𝜆2 𝑙𝑖𝑚 𝑙→∞
𝑘→∞
(𝑘)−1
((𝚪 (𝑘,𝑙) ⊙ (√(𝚿
𝜾𝜾
−2
⊘ 𝚪 (𝑘,𝑙))
+ 4𝜆𝐈q − (𝚿𝜾𝜾
(𝑘)−1
−1
⊘ 𝚪 (𝑘,𝑙))
)) ⋯
× 𝚿𝜾𝜾
(𝑘)−1
(𝑘) (𝚪 (𝑘,𝑙) ⊙ (√(𝚿
𝜾𝜾
−2
⊘ 𝚪 (𝑘,𝑙))
+ 4𝜆𝐈q − (𝚿𝜾𝜾
(𝑘)−1
−1
⊘ 𝚪 (𝑘,𝑙))
)))
[2.4.8]
The CPSF distortion (difference to the Ground Truth of the Precision matrix in [2.4.3]) can be evaluated by
considering a generalization of the SD measure to Connectivity. Its expression can be directly deduced
from the Natural extension of Geodesic Distance Natural to pairs of points belonging to the Cartesian
product of Cortical Manifolds 𝔾 × 𝔾. Analogously to the SD-PSF formulation of Section 2.2.1, the Spatial
Dispersion of the Cartesian Point Spread Function (SD-CPSF), denoted mathematically as 𝑺𝑫𝓒(𝑗0,𝑗1), is
defined as follows:
"The Standard Deviation of the Cartesian Geodesic Distance 𝐷(𝑗,𝑗′)(𝑗0,𝑗1) between pairs of points indexed
{(𝑗, 𝑗′), (𝑗0, 𝑗1)}, for 𝑗, 𝑗′ = 1 ⋯ q in the Cartesian product 𝔾 × 𝔾 of Cortical Manifolds, with probability
mass built on the superior triangle of the CSF" (see its schematic representation in Figure 2).
∑
𝑺𝑫𝚫(𝑗0,𝑗1)(𝓒(𝑗0, 𝑗1)) = √
q
𝑗=1
∑
2
(𝑗,𝑗′)(𝑗0,𝑗1)
𝐷
q
𝑗′=𝑗+1
q
q
𝑗′=𝑗+1
𝑗=1
∑
∑
𝓒𝑗𝑗′(𝑗0,𝑗1)
𝓒𝑗𝑗′(𝑗0,𝑗1)
𝐷(𝑗,𝑗′)(𝑗0,𝑗1) = √𝑑𝑗𝑗0
2 + 𝑑𝑗′𝑗1
2
[2.4.9]
[2.4.10]
Figure 2: Schematic representation of the SD-CPSF at the product of Cortical Manifold spaces given two elements.
1. The Cartesian Geodesic Distance 𝐷(𝑗,𝑗′)(𝑗0,𝑗1) between the pair of Connected generators (𝑗, 𝑗′), among the
generators in the Non-zero elements of the CPSF superior triangle (Yellow Circles), and the pair of generators
(𝑗0, 𝑗1) (Red Circles), at the centers of the of the Kronecker Deltas 𝛿(𝑗0) and 𝛿(𝑗1) . 2. The corresponding
contribution to the SD of the Cartesian Geodesic Distance weighted by the CPSF 𝐷(𝑗,𝑗′)(𝑗0,𝑗1)
𝓒𝑗𝑗′(𝑗0, 𝑗1).
2
In similitude to what was discussed on the PSF and GSF, also the CPSF requires an extended representation
for a general scenario in which the Data is given by a Unitary Precision Matrix given the composition of
multiple Active Sources, i.e. 𝚫(𝑗0, 𝑗1, ⋯ ) = (𝛿(𝑗0) + 𝛿(𝑗1) + ⋯ )(𝛿(𝑗0) + 𝛿(𝑗1) + ⋯ )𝒯. We denominate
this representation "Cartesian Generalized Spread Function" CGSF, denoted mathematically as
𝓒(𝑗0, 𝑗1, ⋯ ). Consequently, the CGSP distortion cannot be measured by using the SD. For this we consider
the generalization to the Cartesian spaces product of Cortical Manifolds of the EMD measure (EMD-CGSF),
denoted mathematically 𝑬𝑴𝑫𝚫(𝑗0,𝑗1,⋯ )(𝓒(𝑗0, 𝑗1, ⋯ )), defined as follows:
"The sum of the typical EMD between all rows (columns) of the CGSF and Gold Standard, i.e. the
projections in the cortical Manifold space 𝔾 𝓒𝑗,:(𝑗0, 𝑗1, ⋯ ) ( 𝓒:,𝑗′(𝑗0, 𝑗1, ⋯ ) ) and 𝚫𝑗,:(𝑗0, 𝑗1, ⋯ )
( 𝚫:,𝑗′(𝑗0, 𝑗1, ⋯ ) ) for all 𝑗 = 1 ⋯ q ( 𝑗′ = 1 ⋯ q ) in the Cortical Manifold space 𝔾 . The Complex
rows/columns EMD is given by the sum of the EMD between its corresponding Real and Imaginary part."
𝑞
𝑬𝑴𝑫𝚫(𝑗0,𝑗1,⋯ )(𝓒(𝑗0, 𝑗1, ⋯ )) = ∑
𝑗=1
𝑬𝑴𝑫𝚫𝑗,:(𝑗0,𝑗1,⋯ ) (𝓒𝑗,:(𝑗0, 𝑗1, ⋯ ))
𝑞
𝑬𝑴𝑫𝚫(𝑗0,𝑗1,⋯ )(𝓒(𝑗0, 𝑗1, ⋯ )) = ∑
𝑗′=1
𝑬𝑴𝑫𝚫:,𝑗′(𝑗0,𝑗1,⋯ ) (𝓒:,𝑗′(𝑗0, 𝑗1, ⋯ ))
[2.4.11]
[2.4.12]
This concept is applicable in general to any definition of the sources Precision Matrix, e.g. Hermitian
Matrix made of blocks with random extensions and random Complex elements.
3 Results
3.1 Simulation substrate
We evaluate the proposed estimators of Sources Localization and Connectivity on simulated EEG data.
The simulation substrate was set up on a Cortical Manifold space 𝔾, defined as 15K points Surface of the
Gray Matter, with coordinates on the MNI Brain template (http://www.bic.mni.mcgill.ca). The Scalp
Sensors space 𝔼 was built on 343 electrodes, within 10-5 EEG Sensors system (Oostenveld and Praamstra,
2001). The Lead Fields, for both Simulations 𝐋𝑠𝑖𝑚 and for Reconstruction 𝐋, were computed by BEM
integration method accounting for a model of 5 head compartments (gray matter, cerebrospinal fluid,
inner skull, outer skull, scalp) (Fuchs et al., 2002; Valdés-Hernández et al., 2009).
To avoid the Inverse Crime two individual subject Head Models were extracted from the corresponding
T1 MRI images. The Electrophysiological Noise was defined by a composition of Sensors Noise and 500
Noisy Cortical Sources (approximating a 3% of the 15K points of the Cortical Manifold space 𝔾), projected
to the Scalp Sensors space 𝔼. For practical computational limitations the Reconstruction was compute on
a 6K points of Cortical Manifold space 𝔾, obtained as a reduction from the original 15K points Surface of
the Gray Matter. We design three kinds of characteristic simulations, to evaluate under different
conditions the Leakage Effect in Source Localization (Type I) and Connectivity (Type II), see Table 4.
Table 4: Description of Simulations
Simulation 1
Configuration A Unitary Source was placed at 500 random locations (following the procedure described in
Section 2.2.1), with amplitude defined by the Kronecker Delta of formula [2.2.5]. 400 EEG Data
trials were created at the Scalp Sensors space, by adding Noise Samples at 5dB Level to each
projected Unitary Source.
Evaluate the Type I Leakage given to Volume Conduction distortion in an isolated point Source
scenario, by the SD-PSF and EMD-PSF.
Aims
Simulation 2
Configuration Two Sources were placed in 500 random configurations (following the procedure described in
Section 2.4.1). 400 EEG Data trials were created, at the Scalp Sensors space, by projecting
Samples from a Gaussian Random Generator given for each configuration. The Gaussian Random
Generator Covariance structure was defined as the Inverse of the Unitary Precision matrix
(Connectivity), given in formula [2.4.3]. Also, 5dB Level Noise were added to the Data of each
projected Sample.
Evaluate the Type I Leakage given the composition the distortive effects of Volume Conduction
and two Sources superposed at the Scalp projection, by the EMD-GSF.
Evaluate the Type II Leakage given to Volume Conduction distortion in a scenario where a single
connection is present, by the SD-CPSF and EMD-CPSF.
Aims
Simulation 3
Configuration Four Sources were placed in 500 random configurations where only thwo of them were
connected. 400 EEG Data trials were created, at the Scalp Sensors space following the same
procedure as in Simulation 2.
Evaluate the Type I Leakage given the composition the distortive effects of Volume Conduction
and multiple Sources superposed at the Scalp projection, by the EMD-GSF.
Evaluate the Type II Leakage given to Volume Conduction distortion in a scenario where multiple
connections are present, by the EMD-CGSF.
Aims
3.3 Validation Methods and software platforms
For reproducibility of the entire methodology proposed in this work results we provide a complementary
routines package, i.e. Sources Activity and Connectivity estimators along with the Type I and II Leakage
measures SD-PSF, EMD-GSF, SD-CPSF and EMD-CGSP, publicly available in MATLAB format at the GitHub
link: https://github.com/dpazlinares/BC-VARETA. The evaluation of SD-PSF, EMD-GSF, SD-CPSF and EMD-
CGSP measures was also extended to different Methods within the State-of-the-Art Electrophysiological
Source analysis,
the FIELDTRIP software package publicly available at:
https://github.com/fieldtrip/fieldtrip/blob/master/ft_sourceanalysis.m.
implemented
into
For comparison purpose we selected the Exact version of LORETA (eLORETA) (Pascual, 2002), which
constitutes the most stablished and robust (under a wide range of conditions), among the of family Source
Activity estimators described in Table 1. The eLORETA SD-PSF and EMD-GSF were computed from the SEC
at the Fist Level of Inference after convergence, according to the theory in Subsection 2.2.1. The SD-CPSF
and EMD-CGSF was taken from the Inverse of the Source Covariance ma Enpirical Formula matrix at the
Second Level of Inference after convergence.
We also consider the Linearly Constrained Minimum Variance (LCMV) (Van Veen et al. 1997), well
stablished among the family of Beam Former methods. The LCMV constitutes a qualitative different
approach in comparison with the family of iterated Source Activity estimators described in Section 2.2,
that enriches our validation with a higher contrast of the results. Roughly, it consists on the Spatial
Filtering of the Forward Equation [2.1.1], under similar assumptions of the Noise and Sources Activity pdf's
and in Subsection 2.1. The LCMV focuses only in the Sources' Variances, ruling out the Covariance
structure from the First Level of Inference, but it also provides a Connectivity analysis through empirical
formulas of the Sources' Covariance. For the computation of the SD-PSF, EMD-GSF, SD-CPSF and EMD-
CGSF we follow analogous procedure as it was applied before to eLORETA solution.
3.4 Study of the Type I and II Leakage in simulations
Figure 4 below shows the results for a typical trial of Simulation 1 and Simulation 2 described in Subsection
3.1, as tridimensional colormaps of the Activity within an interval between 0 and maximum value 1. First,
𝑠𝑖𝑚) at the Sensors space 𝔼 (343 points),
we present the Data Empirical Covariance diagonal values 𝑑𝑖𝑎𝑔(𝐒𝒗𝒗
see Figure 4 a) c), as portraying of the Volume Conduction effect in the Gold Standard Scalp projection.
Second, the estimated PSF and GSF and the corresponding simulation Gold Standard are presented at the
reduction of the Cortical Manifold space 𝔾 (6K points), for the Methods eLORETA, LCMV and BC-VARETA,
see Figure 4 b) d).
Correspondingly, Figure 5 below shows the results of the Connectivity for the typical trial of Simulation 2,
as bidimensional colormaps within an interval between 0 and maximum value 1. First, we present the
𝑠𝑖𝑚) at the Cartesian product of Sensor spaces 𝔼 × 𝔼
Data Empirical Covariance Matrix Inverse 𝑖𝑛𝑣(𝐒𝒗𝒗
(343×343 points), see Figure 5 a), as portraying the Volume conduction effect in the Connectivity of the
Gold Standard Scalp projection. Second, the estimated CPSF and CGSF and the corresponding simulation
Gold Standard are presented at the Cartesian product of reduced Cortical Manifold spaces 𝔾 × 𝔾, (6K×6K
points), for the Methods eLORETA, LCMV and BC-VARETA, see Figure 5 b). For an easier visualization we
don't show the full space, but a subspace defined by 20 neighbors of each Active Source in the actual
configuration.
Simulation 1
Simulation 2
Scalp activity
a)
b)
c)
d)
eLORETA cortical activity
LCMV cortical activity
BC-VARETA cortical activity
Ground truth cortical activity
Figure 4: A typical trial of the two configurations (one active dipole, two active dipoles) to evaluate the Volume Conduction
effect and Localization performance. Projected Scalp Activity given one active generator a), two active generators c). Activity
estimated with the methods eLORETA, LCVM and BC-VARETA given one active generator b), two active generators d). The
tridimensional colormaps are shown within an interval between 0 and maximum value 1.
Simulation 2
a)
Scalp connectivity
eLORETA cortical connectivity
LCMV cortical connectivity
b)
BC-VARETA cortical connectivity
Ground truth cortical connectivity
Figure 5: A typical trial of the simulated configuration (two active dipoles) to evaluate the Volume Conduction
effect and Connectivity performance. Projected Scalp Connectivity given two active generators a), four active
generators c). Connectivity estimated with the methods eLORETA, LCVM and BC-VARETA given two active
generators b), four active generators d). The bidimensional colormaps are shown within an interval between 0
and maximum value 1.
We performed further analysis in Simulation 3, regarding the distance between Active Sources at each of
the 500 configurations. This was done stablishing three classifications of distance. Short Range: The
maximum distance between Sources was smaller than 5 cm. Middle Range: The minimum distance
between Sources was greater than 5cm and the maximum smaller than 8 cm. Long Range: The minimum
distance between Sources was greater than 8 cm. We show the results for typical trials corresponding to
each classification of distance analogously to Figure 4 and Figure 5. See Figure 6 for the tridimensional
colormaps of Activity and in Figure 7 for the bidimensional colormaps of the Connectivity.
Long Range
a)
Middle Range
c)
Scalp activity
Short Range
e)
b)
d)
eLORETA cortical activity
f)
LCVM cortical activity
BC-VARETA cortical activity
Ground truth activity
Figure 6: A typical trial of each classification (long range, middle range and short range) of Simulation 3 to evaluate
the Volume Conduction effect and Localization performance. Projected Scalp Activity given the long range a),
middle range c), short range e) configurations. Activity estimated with the methods eLORETA, LCVM and BC-
VARETA given the long range b), middle range d), short range f) configurations. The tridimensional colormaps are
shown within an interval between 0 and maximum value 1.
Long Range
a)
Middle Range
c)
Scalp connectivity
Short Range
e)
b)
d)
f)
eLORETA cortical connectivity
LCMV cortical connectivity
BC-VARETA cortical connectivity
Ground truth cortical connectivity
Figure 7: A typical trial of each classification (long range, middle range and short range) of Simulation 3 to evaluate
the Volume Conduction effect and Connectivity performance. Projected Scalp connectivity given the long range a),
middle range c), short range e) configurations. Connectivity estimated with the methods eLORETA, LCVM and BC-
VARETA given the long range b), middle range d), short range f) configurations. The bidimensional colormaps are
shown within an interval between 0 and maximum value 1.
We report the Mean and Standard Deviation of the SD-PSF, EMD-GSF, SD-CPSF and EMD-CGSF measures
for the Methods eLORETA, LCMV and BC-VARETA, computed for the 500 configurations of Simulation 1
and Simulation 2, see Table 5 below. The corresponding results for Simulation 3 of the EMD-GSF and EMD-
CGSF measures Mean and Standard Deviation are reported separately for each classification (long range,
middle range and short range), see Table 6.
Table 5: Mean and Standard Deviation of the SD-PSF, EMD-GSF, SD-CPSF and EMD-CGSF measures
for the methods eLORETA, LCMV and BC-VARETA in Simulation 1 and 2.
Methods
Simulation 1
eLORETA
LCMV
BC-VARETA
EMD-GSF
26.9305 ±
3.9276
10.2282 ±
11.4917
0.4632 ±
0.7121
SD-PSF
0.1093 ±
0.0166
0.1592 ±
0.0447
0.0105 ±
0.0128
EMD-GSF
30.4747 ±
6.5958
37.1807 ±
10.2112
0.9712 ±
0.5589
Simulation 2
SD-CPSF
107.8553 ±
25.3005
EMD-CGSF
107.8553 ±
25.3005
132.7289 ±
132.7289 ±
32.6940
2.8131 ±
0.4494
32.6940
2.8131 ±
0.4494
Table 6: Mean and Standard Deviation of the SD-PSF, EMD-GSF, SD-CPSF and EMD-CGSF measures
for the methods eLORETA, LCMV and BC-VARETA in each distance classification of Simulation 3.
Methods
Long Range
Middle Range
Short Range
eLORETA
EMD-GSF
11.4388 ±
2.9366
LCMV
14.2701 ±
BC-VARETA
1.8652
2.2695 ±
1.7096
EMD-CGSF
231.0420
± 60.7255
201.7472
± 18.8072
43.7651 ±
4.5070
EMD-GSF
7.9462 ±
1.5958
20.3405 ±
10.2112
1.8992 ±
0.6810
EMD-CGSF
216.6924 ±
25.3005
EMD-GSF
16.0355 ±
0.2562
EMD-CGSF
139.9369 ±
10.1142
274.7283 ±
18.1182 ±
163.6560 ±
32.6940
21.4402 ±
9.8305
1.9750
4.4229 ±
1.5213
29.1482
16.4168 ±
5.6970
4 Discussion
4.1 Methodology of Brain Connectivity Variable Resolution Tomographic Analysis
The State of Art of Electrophysiological Brain Source Localization and Connectivity is quite diverse. Despite
this fact, the structure of the proposed Bayesian Model, underlying the BC-VARETA framework, is common
for a large family of Methods, see Subsection 2.1 and Table 1 in Subsection 2.2. Furthermore, we intended
enough generality of our Model set up, within the Square of acceptable Physiological and Mathematical
assumptions, by avoiding the use of constraints that could prevent for a fair analysis of the Parameters
and Hyperparameters. It is based on a Hierarchically Conditioned and fully Multivariate Two Levels
Gaussian Graphical Model of Data and Parameters, see Subsection 2.1, in consistency with the Bayesian
representation of the LSSM, in both Time and Frequency Domain (Wills et al., 2009; Faes et al., 2012;
Galka et al., 2004; Pascual-Marqui et al., 2014; Valdes-Sosa 2004; Lopes da Silva et al., 1980; Baccalá and
Sameshima, 2001; Babiloni et al., 2005).
The ad-hoc Conditional Data Covariance structure was regarded as an observable property of
Instrumental/Environmental/Biological Noisy processes (Waldorp et al., 2001; De Munck et al., 2002;
Huizenga et al., 2002). In general, this strategy improves the algorithm convergence by ruling out adverse
effects, given the Non-linear interaction between the estimation of multiple Hyperparameters within the
Data Conditional Covariance, and the Parameters Covariance, as it happens with LORETA, eLORETA and
eLORETA (Table 1). In similitude to previous works we regarded the Noise Variance as a Hyperparameter,
see VARETA, ARD, ReLM and SSBL in Table 1. Distinctively, a Noise Variance Jeffrey Improper Gibbs Prior
allowed us to set a Noise inferior limit, that constitutes also an observable property of Noisy processes
(Van Hoey et al., 2000; Lemm et al., 2006; Bigdely-Shamlo et al., 2015; Phillips et al., 2002; Phillips et al.,
2005). Setting a Noise inferior limit also prevents for adverse effects like assimilating higher amount of
Noise at the Sources Level, which can lead to the overestimation of the Parameters Covariance. This also
contributes to the stability of the algorithms in general.
We formulated general Sparse Gibbs Priors on the Parameters Precision Matrix and not on the Covariance,
see Subsection 2.1 and Table 2 of Subsection 2.3. To regard the Covariance as informative of the
Connectivity instead, is a wrong approach in sight of the LSSM theory (Wills et al., 2009; Faes et al., 2012;
Galka et al., 2004; Pascual-Marqui et al., 2014; Valdes-Sosa 2004; Lopes da Silva et al., 1980; Baccalá and
Sameshima, 2001; Babiloni et al., 2005). This constitutes a common misdeed of previous works that
pursue the Connectivity analysis through Sparse Sources Covariance estimation, like ReLM, or Covariance
extraction as a Postprocessing of Sources Activity estimates, like MNE, LORETA, eLORETA, eLORETA, ARD,
VARETA, SSBL. In addition, our formulation attains to incorporate information on the Gray Matter areas
Intra/Inter-connections. The Model Priors were rather general, allowing to include Structured Sparsity
given an anatomical segmentation and probability mask of Intra-Cortical connections strength decay with
distance and probabilistic maps of the White Matter tracks connectivity.
With the BC-VARETA framework we attained interpretable formulations of the First Level of Inference.
The Maximum Posterior analysis (First Level of Inference) leads to an estimator of the Parameters that
has identical expression to those of previous methodologies, see Table 1 in Subsection 2.2. Thus,
formulating the Type I Leakage measures upon BC-VARETA is representative and extendable to the whole
State of the Art in Electrophysiological Brain Source Localization. The Second Level of Inference
constituted the most particular aspect in setting up the BC-VARETA. We proposed a unification of the
State of Art in Electrophysiological Source Localization and Connectivity with the theory GGM's,
denominated here as SGGM. This was done by applying the EM strategy (Dempster et al., 1977; Liu and
Rubin, 1994; McLachlan and Krishnan, 2007) and expressing the Expected Log-Likelihood as a function of
the ESEC Matrix, see Subsection 2.2.2. The Posterior analysis with Gibbs Priors on the Precision Matrix
Hyperparameters lead to Target Function analogous to that of a typical GGM's.
The current GGM's theory however was not able to serve the mathematical semblance of the
Electrophysiological SGGM (Friedman et al, 2008; Mazumder et al. 2012; Schmidt, 2010; Hsieh, 2014;
Danaher et al., 2014; Drton and Maathuis, 2017), see Table 2. In consequence we explored the GGM's
from the Bayesian perspective and proposed a new solution that better suits the Electrophysiological
SGGM scenario, see Subsection 2.3. To achieve this, we used a generalization of Andrews and Mallows
Lemma to the Gibbs Priors of the Precision Matrix with Penalization functions in the LASSO family
(Andrews and Mallows, 1974; Tipping, 2001; Schmolck and Everson, 2007; Faul and Tipping, 2002; Li and
Lin, 2010; Kyung et al., 2010). This generalization allowed for Concave Quadratic reformulation of the
SGGM Target Function in resembling the strategy of LQA algorithms (Fan and Li, 2001; Valdés-Sosa et al.,
2006; Sánchez-Bornot et al., 2008). In the context of the SGGM LQA we implemented a Standardization
technique that simplifies the Target Function minimization problem, given the Scale Invariance properties
of the Wishart Likelihood (Srivastava, 1965; Drton et al., 2008). The Connectivity estimator was derived,
as consequence of applying the LQA and Standardization to the original SGGM, from the direct solution
of a Matrix Riccati equation (Lim, 2006; Honorio and Jaakkola, 2013). To prevent the biasing of the
Precision Matrix estimator given the selection of specific SGGM Penalty function and Regularization
parameters (Jankova and Van De Geer, 2015, 2017), we use a debiasing operation proposed in (Jankova
and Van De Geer, 2018).
The Posterior analysis of the Noise Variance Hyperparameter lead to an interpretable formula composed
of the also typical EM Naïve estimator plus the inferior limit (Dempster et al., 1977; Liu and Rubin, 1994;
McLachlan and Krishnan, 2007; Valdes-Sosa, 1996, Bosch-Bayard, et al., 2001), in correspondence to what
was discussed before about the Noise Variance Jeffrey Improper Gibbs Prior. The bias of this formula lies
on quantities that can be interpreted experimentally, thus there is not ambiguity on the definition such
estimator.
4.2 Theoretical analysis of the Type I and II Leakage effect measures
In this work we provided a unified description of the Leakage by a Model that regards the Source Activity
and Connectivity as elements to be estimated into a System Identification approach. The two different
scenarios of Leakage were correspondingly represented by the two Levels of Inference of the BC-VARETA
methodology. The PSF and GSF formulated at the Firs Level of Inference, were representative of the Type
I Leakage effect given Volume Conduction in multiple scenarios as proposed by (Schoffelen and Gross,
2009; Wens et al., 2015; Brunner et al., 2016; Silva Pereira et al., 2017; Hincapié et al., 2017). The original
PSF concept was too ideal to cover all the aspects of realistic Electrophysiological Sources Localization,
thus it was generalized to consider multiple Active cortical sources, the Inverse Crime (Kaipio & Somersalo
2004) and the Noise from biological, environmental and instrumentation origins.
We provide a representation of the Type II Leakage or Connectivity Leakage through the CPSF and CGSF
at the Second Level of Inference. This is a more consistent approach in comparison to previous works, that
have considered the Connectivity Leakage (Type II Leakage) and its Correctors in the limited context of
the linear mixing or crosstalk between Sources Activity estimates (Type I Leakage) (Brookes et al., 2012;
McCoy and Troop, 2013; Colclough et al., 2015; Wens et al., 2015; Colclough et al., 2016; Silva Pereira et
al., 2017). Even though, these approaches constituted a first approximation to the Leakage effect, they
lacked objectivity when ignored more realistic Non-linear Methods in the State of the Art of Sources
Activity and Connectivity, see Table 1. Distinctively, the BC-VARETA framework allowed us to formulate
Non-linear mutually interacting and explicit estimators of the PSF (GSF) and CPSF (CGSF), see Subsection
2.2.1 and Subsection 2.4.1, that evince a bidirectional relationship between the Type I and II Leakage.
Modelling the Source Activity and Connectivity as a whole, as was done with BC-VARETA, provided a
consistent way to implement the correction of the Connectivity Leakage, by incorporating the Corrector
Operator at the Model Priors. In this sense, a more consistent approach was presented by (Pascual-Marqui
et al., 2017; Bosch-Bayard and Biscay, 2018), accounting for the linear de-mixing of LSSM Kalman Filter
Non-linear estimators. Distinctively, in our work we revendicate the concept of Sparsity in the Connectivity
Level as part of the Leakage correction strategy, that might coexist and never collide with different
approaches in the context of Non-linear Methods (like linear de-mixing LSSM Kalman Filter estimators).
A measure of the Type I Leakage in the PSF was built on the Geodesic Distance Spatial Dispersion (SD-PSF)
in the Cortical Manifold, that has been universally adopted for single point spreading like scenarios. We
provide an extension of this concept to multiple points by the GSF and its EMD (EMD-GSF). Remarkably,
we present a generalization to the Cartesian geometry (Product of Cortical Manifolds Spaces) to represent
the measures Type II Leakage given by CPSF and CGSF. To this end we use the Spatial Dispersion of the
Cartesian Geodesic Distance (SD-CPSF), for a single connection, and the Cartesian Earth Mover's distance
(EMD-CGSF), for the extension to multiple connections.
4.3 Analysis of the Results in Simulations
Simulation Aims
Simulation 1 was set up to study the Type I Leakage in an ideal scenario that reflects solely the Volume
Conduction spilling effect on a single point. Simulation 2 was aimed to study the Type II Leakage and its
mutual interaction with the Type I Leakage, in a scenario that reflects the spilling effect of Volume
Conduction on two points Connectivity. Simulation 3 pursues the study the Type II Leakage on four points
where only three of them were connected, reflecting not only the Volume Conduction spilling effect in
Connectivity but also the crosstalk towards not connected points.
Scalp Analysis
For a typical trial of Simulation 1 the projected Activity at the Scalp Sensors of a typical trial, see Figure 4
a), showed the single point large spatial spillover and mismatch of its maximum given to the Volume
Conduction effect, confirming an essential shortcoming of the direct analysis of Sensors data. This has
been pointed out in previous works (Brunner et al., 2016; Van de Steen et al., 2016). Consistently to
Simulation 1 results, the Scalp projections of typical trials in Simulation 2, see Figure 4 c), showed an even
larger spatial spillover and mismatch their maximum given to the Volume Conduction effect. The situation
of Simulation 3, see Figure 6 a) c) e), was more anfractuous given the mixture of the projected Sources in
the conditions of short and middle range distance, thus making impossible deduce the existence of the
third or fourth source in the simulation, just by analyzing its Scalp projection. This happened in lower
degree in the long range condition.
Consequently, this effect is reflected in the Precision Matrix (Connectivity) associated to the Data
Empirical Covariance Matrix at the Cartesian Scalp Sensors space product, see Figure 5 a) and Figure 7 a)
c) e). The Scalp Connectivity analysis based on Precision Matrices (Kaminski and Blinowska, 2014; Kaminski
and Blinowska, 2017; Blinowska, 2011), seems unfavorable as revealed by the visual inspection of these
bidimensional maps: First: A large spillover of the Non-diagonal Block in the two points Connectivity of
Simulation 2, see Figure 5 a). Second: Mixture of the Diagonal Blocks corresponding to the individual
Sources Precisions in the three conditions (short, middle and long range distance) of Simulation 3, see
Figure 7 a) c) e). Third: Crosstalk towards not connected points given the large spillover of the Non-
diagonal Blocks in the three conditions (short, middle and long range distance) of Simulation 3, see Figure
7 a) c) e). Even when we are in presence of highly sparse simulations the scenario for Scalp Connectivity
doesn't show any goodness according to these results.
Sources Analysis
The Sources Localization and Connectivity were distorted qualitatively by the Volume Conduction effect,
across all simulations. For the typical trial of Simulation 1 the reconstruction with eLORETA does not
improves the mentioned situation for the Scalp Data, considering that the Sources were extended along
a larger Cortical area than the simulated Scalp projection, see Figure 4 b). This overestimation of Cortical
activity is a peculiar of Linear Methods such as MNE, see Table 1, but in this case the eLORETA showed
qualitatively similar performance despite its Hybrid estimation formulas. We found that the Cortical
extension of the LCMV reconstruction was much shorter. This was possible due to the Linear Constraints
of the LCMV Hybrid formulas at the Second Level of Inference, that pursue sparsity of the Spatial Filtering
Variances. The BC-VARETA reconstruction was the sparsest confirming our hypothesis, about the effect of
using a Sparse Precision Matrix Model, that underlies this Non-linear Sources estimation Method.
The Sources Activity reconstruction in typical trials of Simulation 2 see Figure 4 d) and Simulation 3, see
Figure 5 b) d) f), for eLORETA was extended across a large Cortical area as expected, and consistently with
the results in Simulation 1 it spills beyond the Scalp projection. The Connectivity, see Figure 5 b) and Figure
7 b) d) f), presented also high spatial spillover, mixing and crosstalk between Sources. This is the cause of
repetitive allusions within the State of the Art to the Type I and II Leakage (Brookes et al., 2012; McCoy
and Troop, 2013; Colclough et al., 2015; Wens et al., 2015; Colclough et al., 2016; Silva Pereira et al., 2017).
The LCMV method achieves better performance in Source Activity and Connectivity reconstruction. The
mixing of Sources and Crosstalk is qualitatively diminished, but still too spilled as compared to the highly
sparse simulated Activity. The Source Activity and Connectivity reconstruction with the BC-VARETA
Method was the sparsest and thus the best according to the properties of the simulations, also evidencing
that the estimation with sparse Precision Matrix model worked as expected. The spatial spillover, mixing
and crosstalk of reconstructed Sources Activity and Connectivity, of the presented typical trials, seems
minimized across all simulations by this Non-linear Sparse Method. This outcome was effective for the
three conditions (short, middle and long range distance) of Simulation 3, whereas expected the results
with all Methods deteriorated with the range shrinking.
The results across the 500 trials of the measures SD-PSF, EMD-GSF, SD-CPSF and EMD-CGSF, for
Simulation 1 and Simulation 2 and Simulation 3 were consistent with the qualitative analysis given for the
typical ones, see Table 5 and Table 6. The measures values were always minimum for the BC-VARETA
Method. The measures of Sources Localization SD-PSF and EMD-GSF were consistent to the SD-CPSF and
EMD-CGSF, in the sense of that when the Connectivity reconstruction performance was higher it was
always higher the performance in Source Localization. The values of the EMD-GSF and EMD-CGSF, along
the three conditions (short, middle and long range distance) of Simulation 3 revealed that reconstruction
performance deteriorated for all methods as the Range decreased, see Figure 6. Even though this
behavior was expected for every Source Localization Method, the BC-VARETA exhibited the most robust
performance.
5 Conclusions
The proposed methodology BC-VARETA allowed us to caulk the "Leakage Effect" in simulation scenario of
MEEG activity that was challenging, according to the high degree of sparsity (super resolution) variability
(different configurations with random positions of sources) and realism (presence of noise in generators
and sensors, and inverse crime evaluation). The BC-VARETA performance was better than well stablished
methods, which operate under different assumptions, i.e. eLORETA and LCMV. These results were
supported by sensitive quality measures (Spatial Dispersion and Earth Movers Distance), that are also
acknowledged to be the most interpretable into the state of the art of MEEG source connectivity analysis.
Remarkably, our Bayesian model and inference (BC-VARETA) constitutes a unification of the state of the
art in the theory of MEEG source activity and connectivity estimation methods and the theory of Gaussian
Graphical Models. We presented fully detailed technical derivations of BC-VARETA, along with its
interpretability and theoretical comparison with those methodologies previously developed. Another
issue addressed into this paper was the rigorous mathematical representation of the Leakage in both
source activity and connectivity. It involved the introduction of new quantities as such the Generalized
Spread Function (explicit activity estimator given multiple sources) and Cartesian Generalized Spread
Function (explicit connectivity estimator given multiple connected sources), and the generalization of
Spatial Dispersion and Earth Movers distance to the connectivity space, i.e. Cartesian product of Cortical
Manifolds.
Acknowledgements
The Grant No. 61673090 from the National Nature Science Foundation of China funded this study.
References
[1] Andrews, D.F. and Mallows, C.L., 1974. Scale mixtures of normal distributions. Journal of the Royal
Statistical Society. Series B (Methodological), pp.99-102.
https://www.jstor.org/stable/2984774
[2] Asadi, N. B., Rish, I., Scheinberg, K., Kanevsky, D., and Ramabhadran, B., 2009. Map approach to
learning sparse Gaussian Markov networks. In Acoustics, Speech and Signal Processing, 2009. ICASSP
2009. IEEE International Conference on, pp.1721-1724.
https://doi.org/10.1109/ICASSP.2009.4959935
[3] Attias, H., 2000. A variational bayesian framework for graphical models. In Advances in neural
information processing systems (pp. 209-215).
https://doi.org/10.1109/ICASSP.2009.4959935 ononon
[4] Babiloni, F., Cincotti, F., Babiloni, C., Carducci, F., Mattia, D., Astolfi, L., Basilisco, A., Rossini, P.M., Ding,
L., Ni, Y. and Cheng, J., 2005. Estimation of the cortical functional connectivity with the multimodal
integration of high-resolution EEG and fMRI data by directed transfer function. Neuroimage, 24(1),
pp.118-131.
https://doi.org/10.1016/j.neuroimage.2004.09.036
[5] Babacan, S.D., Luessi, M., Molina, R. and Katsaggelos, A.K., 2012. Sparse Bayesian methods for low-
rank matrix estimation. IEEE Transactions on Signal Processing, 60(8), pp.3964-3977.
https://doi.org/10.1109/TSP.2012.2197748
[6] Baccalá, L.A. and Sameshima, K., 2001. Partial directed coherence: a new concept in neural structure
determination. Biological cybernetics, 84(6), pp.463-474.
https://doi.org/10.1007/PL000079
[7] Balkan, O., Kreutz-Delgado, K. and Makeig, S., 2014. Localization of more sources than sensors via
Letters, 21(2), pp.131-134.
Signal Processing
jointly-sparse Bayesian
IEEE
https://doi.org/10.1109/LSP.2013.2294862
learning.
[8] Belardinelli, P., Ortiz, E., Barnes, G., Noppeney, U. and Preissl, H., 2012. Source reconstruction
inversion approaches. PloS one, 7(12), p.e51985.
accuracy of MEG and EEG Bayesian
https://doi.org/10.1371/journal.pone.0051985
[9] Bigdely-Shamlo, N., Mullen, T., Kothe, C., Su, K.M. and Robbins, K.A., 2015. The PREP pipeline:
standardized preprocessing for large-scale EEG analysis. Frontiers in neuroinformatics, 9, p.16.
https://doi.org/10.3389/fninf.2015.00016
[10] Blinowska, K.J., 2011. Review of the methods of determination of directed connectivity from
multichannel data. Medical & biological engineering & computing, 49(5), pp.521-529.
https://doi.org/10.1007/s11517-011-0739-x
[11] Bosch-Bayard, J., Valdes-Sosa, P., Virues-Alba, T., Aubert-Vazquez, E., John, E.R., Harmony, T., Riera-
Diaz, J. and Trujillo-Barreto, N., 2001. 3D statistical parametric mapping of EEG source spectra by
means of variable resolution electromagnetic tomography (VARETA). Clinical Electroencephalography,
32(2), pp.47-61.
https://doi.org/10.1177/155005940103200203
[12] Bosch-Bayard, J. and Viscay, R.J., 2018. Several drawbacks in brain connectivity assessment based on
source estimates -elucidations, practical guidelines, and incentives for further research. Submitted to
Brain Topography.
[13] Brookes, M.J., Woolrich, M.W. and Barnes, G.R., 2012. Measuring functional connectivity in MEG: a
multivariate approach insensitive to linear source leakage. Neuroimage, 63(2), pp.910-920.
https://doi.org/10.1016/j.neuroimage.2012.03.048
[14] Brunner, C., Billinger, M., Seeber, M., Mullen, T.R. and Makeig, S., 2016. Volume conduction influences
scalp-based connectivity estimates. Frontiers in computational neuroscience, 10, p.121.
https://doi.org/10.3389/fncom.2016.00121
[15] Colclough, G.L., Brookes, M.J., Smith, S.M. and Woolrich, M.W., 2015. A symmetric multivariate
leakage correction for MEG connectomes. NeuroImage, 117, pp.439-448.
https://doi.org/10.1016/j.neuroimage.2015.03.071
[16] Colclough, G.L., Woolrich, M.W., Tewarie, P.K., Brookes, M.J., Quinn, A.J. and Smith, S.M., 2016. How
reliable are MEG resting-state connectivity metrics?. Neuroimage, 138, pp.284-293.
https://doi.org/10.1016/j.neuroimage.2016.05.070
[17] Danaher, P., Wang, P. and Witten, D.M., 2014. The joint graphical lasso for inverse covariance
estimation across multiple classes. Journal of the Royal Statistical Society: Series B (Statistical
Methodology), 76(2), pp.373-397.
https://doi.org/10.1111/rssb.12033
[18] Daunizeau, J. and Friston, K.J., 2007. A mesostate-space model for EEG and MEG. NeuroImage, 38(1),
pp.67-81.
https://doi.org/10.1016/j.neuroimage.2007.06.034
[19] Da Silva, F.L., Vos, J.E., Mooibroek, J. and Van Rotterdam, A., 1980. Relative contributions of
intracortical and thalamo-cortical processes in the generation of alpha rhythms, revealed by partial
coherence analysis. Electroencephalography and clinical neurophysiology, 50(5-6), pp.449-456.
https://doi.org/10.1016/0013-4694(80)90011-5
[20] Dempster, A.P., Laird, N.M. and Rubin, D.B., 1977. Maximum likelihood from incomplete data via the
EM algorithm. Journal of the royal statistical society. Series B (methodological), pp.1-38.
https://www.jstor.org/stable/2984875
[21] De Munck, J.C., Huizenga, H.M., Waldorp, L.J. and Heethaar, R.A., 2002. Estimating stationary dipoles
from MEG/EEG data contaminated with spatially and temporally correlated background noise. IEEE
Transactions on Signal Processing, 50(7), pp.1565-1572.
https://doi.org/10.1109/TSP.2002.1011197
[22] Drton, M., Massam, H. and Olkin, I., 2008. Moments of minors of Wishart matrices. The Annals of
Statistics, 36(5), pp.2261-2283.
https://doi.org/10.1214/07-AOS522
[23] Elanbari, M., Rawi, R., Ceccarelli, M., Bouhali, O. and Bensmail, H., 2015. Advanced Computation of a
Sparse Precision Matrix HADAP: A Hadamard-Dantzig Estimation of a Sparse Precision Matrix. In
Proceedings of the Sixth International Conference on Computational Logics, Algebras, Programming,
Tools, and Benchmarking.
[24] Faes, L., Erla, S. and Nollo, G., 2012. Measuring connectivity in linear multivariate processes:
definitions, interpretation, and practical analysis. Computational and mathematical methods in
medicine, 2012.
http://dx.doi.org/10.1155/2012/140513
[25] Fan J and Li R.: Variable Selection via Nonconcave Penalized Likelihood and Its Oracle Properties.
Journal of the American Statistical Association. 2001; 96: pp1348-1360.
https://doi.org/10.1198/016214501753382273
[26] Farahibozorg, S.R., Henson, R.N., and Hauk, O., 2018. Adaptive cortical parcellations for source
reconstructed EEG/MEG connectomes. NeuroImage, 169, pp.23-45.
https://doi.org/10.1016/j.neuroimage.2017.09.009
[27] Freeman, W.J., 1980. Use of spatial deconvolution to compensate for distortion of EEG by volume
conduction. IEEE Transactions on Biomedical Engineering, (8), pp.421-429.
https://doi.org/10.1109/TBME.1980.326750
[28] Friedman, J., Hastie, T. and Tibshirani, R., 2008. Sparse inverse covariance estimation with the
graphical lasso. Biostatistics, 9(3), pp.432-441.
[29] Faul, A.C. and Tipping, M.E., 2002. Analysis of sparse Bayesian learning. In Advances in neural
information processing systems (pp. 383-389).
[30] Friston, K., Harrison, L., Daunizeau, J., Kiebel, S., Phillips, C., Trujillo-Barreto, N., Henson, R., Flandin,
G. and Mattout, J., 2008. Multiple sparse priors for the M/EEG inverse problem. NeuroImage, 39(3),
pp.1104-1120.
https://doi.org/10.1016/j.neuroimage.2007.09.048
[31] Fuchs, M., Kastner, J., Wagner, M., Hawes, S. and Ebersole, J.S., 2002. A standardized boundary
element method volume conductor model. Clinical Neurophysiology, 113(5), pp.702-712.
https://doi.org/10.1016/S1388-2457(02)00030-5
[32] Galka, A., Yamashita, O., Ozaki, T., Biscay, R. and Valdés-Sosa, P., 2004. A solution to the dynamical
inverse problem of EEG generation using spatiotemporal Kalman filtering. NeuroImage, 23(2), pp.435-
453.
https://doi.org/10.1016/j.neuroimage.2004.02.022
[33] Hämäläinen, M.S. and Ilmoniemi, R.J., 1994. Interpreting magnetic fields of the brain: minimum norm
estimates. Medical & biological engineering & computing, 32(1), pp.35-42.
https://doi.org/10.1007/BF025124
[34] Hedrich, T., Pellegrino, G., Kobayashi, E., Lina, J. M., and Grova, C., 2017. Comparison of the spatial
resolution of source imaging techniques in high-density EEG and MEG. NeuroImage, 157, pp.531-544.
https://doi.org/10.1016/j.neuroimage.2017.06.022
[35] Hincapié, A.S., Kujala, J., Mattout, J., Pascarella, A., Daligault, S., Delpuech, C., Mery, D., Cosmelli, D.
and Jerbi, K., 2017. The impact of MEG source reconstruction method on source-space connectivity
estimation: a comparison between minimum-norm solution and beamforming. NeuroImage, 156,
pp.29-42.
https://doi.org/10.1016/j.neuroimage.2017.04.038
[36] Honorio, J. and Jaakkola, T.S., 2013. Inverse covariance estimation for high-dimensional data in linear
time and space: Spectral methods for riccati and sparse models. arXiv preprint arXiv:1309.6838.
https://arxiv.org/abs/1309.6838
[37] Huizenga, H.M., De Munck, J.C., Waldorp, L.J. and Grasman, R.P., 2002. Spatiotemporal EEG/MEG
source analysis based on a parametric noise covariance model. IEEE Transactions on Biomedical
Engineering, 49(6), pp.533-539.
https://doi.org/10.1109/TBME.2002.1001967
[38] Jankova, J. and Van De Geer, S., 2015. Confidence intervals for high-dimensional inverse covariance
estimation. Electronic Journal of Statistics, 9(1), pp.1205-1229.
https://doi.org/10.1214/15-EJS1031
[39] Janková, J. and van de Geer, S., 2017. Honest confidence regions and optimality in high-dimensional
precision matrix estimation. Test, 26(1), pp.143-162.
https://doi.org/10.1007/s11749-016-0503-5
[40] Jankova, J. and van de Geer, S., 2018. Inference in high-dimensional graphical models. arXiv preprint
arXiv:1801.08512.
https://arxiv.org/abs/1801.08512
[41] Jordan, M.I. ed., 1998. Learning in graphical models (Vol. 89). Springer Science & Business Media.
https://doi.org/10.1007/978-94-011-5014-9
[42] Kaminski, M. and Blinowska, K.J., 2014. Directed transfer function is not influenced by volume
conduction -- inexpedient pre-processing should be avoided. Frontiers in computational neuroscience,
8, p.61.
https://doi.org/10.3389/fncom.2014.00061
[43] Kaminski, M. and Blinowska, K.J., 2014. Directed transfer function is not influenced by volume
conduction -- inexpedient pre-processing should be avoided. Frontiers in computational neuroscience,
8, p.61.
https://doi.org/10.3389/fncom.2014.00061
[44] Krishnaswamy, P., Obregon-Henao, G., Ahveninen, J., Khan, S., Babadi, B., Iglesias, J.E., Hämäläinen,
M.S. and Purdon, P.L., 2017. Sparsity enables estimation of both subcortical and cortical activity from
MEG and EEG. Proceedings of the National Academy of Sciences, p.201705414.
https://doi.org/10.1073/pnas.1705414114
[45] Kyung, M., Gill, J., Ghosh, M. and Casella, G., 2010. Penalized regression, standard errors, and Bayesian
lassos. Bayesian Analysis, 5(2), pp.369-411.
https://doi.org/10.1214/10-BA607
[46] Lemm, S., Curio, G., Hlushchuk, Y. and Muller, K.R., 2006. Enhancing the signal-to-noise ratio of ICA-
based extracted ERPs. IEEE Transactions on Biomedical Engineering, 53(4), pp.601-607.
https://doi.org/10.1109/TBME.2006.870258
[47] Li, Q. and Lin, N., 2010. The Bayesian elastic net. Bayesian Analysis, 5(1), pp.151-170.
https://doi.org/10.1214/10-BA506
[48] Lim, Y., 2006. The matrix golden mean and its applications to Riccati matrix equations. SIAM Journal
on Matrix Analysis and Applications, 29(1), pp.54-66.
https://doi.org/10.1137/050645026
[49] Liu, C. and Rubin, D.B., 1994. The ECME algorithm: a simple extension of EM and ECM with faster
monotone convergence. Biometrika, 81(4), pp.633-648.
https://doi.org/10.1093/biomet/81.4.633
[50] MacKay D J C. Information Theory, Inference, and Learning Algorithms. Cambridge University Press
2003.
https://doi.org/10.2277/0521642981
[51] Mazumder, R. and Hastie, T., 2012. The graphical lasso: New insights and alternatives. Electronic
journal of statistics, 6, p.2125.
https://doi.org/10.1214/12-EJS740
[52] McLachlan, G. and Krishnan, T., 2007. The EM algorithm and extensions (Vol. 382). John Wiley & Sons.
[53] McCoy, M.B. and Tropp, J.A., 2013. The achievable performance of convex demixing. arXiv preprint
arXiv:1309.7478.
https://arxiv.org/abs/1309.7478v1
[54] Neal, R.M., 1998. Assessing relevance determination methods using DELVE. Nato Asi Series F
Computer And Systems Sciences, 168, pp.97-132.
[55] Palva, J.M., Wang, S.H., Palva, S., Zhigalov, A., Monto, S., Brookes, M.J., Schoffelen, J.M. and Jerbi, K.,
2018. Ghost interactions in MEG/EEG source space: A note of caution on inter-areal coupling
measures. Neuroimage, 173, pp.632-643.
https://doi.org/10.1016/j.neuroimage.2018.02.032
[56] Pascual-Marqui, R.D., Michel, C.M. and Lehmann, D., 1994. Low resolution electromagnetic
tomography: a new method for localizing electrical activity in the brain. International Journal of
psychophysiology, 18(1), pp.49-65.
https://doi.org/10.1016/0167-8760(84)90014-X
[57] Pascual-Marqui, R.D., Esslen, M., Kochi, K. and Lehmann, D., 2002. Functional imaging with low-
resolution brain electromagnetic tomography (LORETA): a review. Methods and findings in
experimental and clinical pharmacology, 24(Suppl C), pp.91-95.
[58] Pascual-Marqui, R.D., 2002. Standardized
low-resolution brain electromagnetic tomography
(sLORETA): technical details. Methods Find Exp Clin Pharmacol, 24(Suppl D), pp.5-12.
[59] Pascual-Marqui, R.D., Pascual-Montano, A.D., Lehmann, D., Kochi, K., Esslen, M., Jancke, L., Anderer,
P., Saletu, B., Tanaka, H., Hirata, K. and John, E.R., 2006. Exact low resolution brain electromagnetic
tomography (eLORETA). Neuroimage, 31(Suppl 1).
[60] Pascual-Marqui, R.D., Biscay, R.J., Bosch-Bayard, J., Faber, P., Kinoshita, T., Kochi, K., Milz, P., Nishida,
K. and Yoshimura, M., 2017. Innovations orthogonalization: a solution to the major pitfalls of
EEG/MEG" leakage correction". arXiv preprint arXiv:1708.05931.
https://doi.org/10.1101/178657
[61] Paz-Linares, D., Gonzalez-Moreira, E., Valdes-Sosa, P.A., 2018. A Technical Note on the Estimation of
Embedded Hermitian Gaussian Graphical Models for MEEG Source Activity and Connectivity Analysis.
arXiv preprint arXiv:submit/2413461.
https://arxiv.org/submit/2413461
[62] Phillips, C., Rugg, M.D. and Friston, K.J., 2002. Systematic regularization of linear inverse solutions of
the EEG source localization problem. NeuroImage, 17(1), pp.287-301.
https://doi.org/10.1006/nimg.2002.1175
[63] Phillips, C., Mattout, J., Rugg, M.D., Maquet, P. and Friston, K.J., 2005. An empirical Bayesian solution
to the source reconstruction problem in EEG. NeuroImage, 24(4), pp.997-1011.
https://doi.org/10.1016/j.neuroimage.2004.10.030
[64] Riera J J and Fuentes M E. Electric lead field for a piecewise homogeneous volume conductor model
of the head. Biomedical Engineering, IEEE Transactions on 1998; 45(6): 746-753.
https://doi.org/10.1109/10.678609
[65] Rish, I., and Grabarnik, G., 2014. Sparse modeling: theory, algorithms, and applications. CRC press.
[66] Sánchez-Bornot J M, Martínez-Montes E, Lage-Castellanos A, Vega-Hernández M and Valdés-Sosa P A.
Uncovering sparse brain effective connectivity: A voxel-based approach using penalized regression.
Statistica Sinica 2008; 18: 1501-1518.
https://www.jstor.org/stable/24308566
[67] Sato, M.A., Yoshioka, T., Kajihara, S., Toyama, K., Goda, N., Doya, K. and Kawato, M., 2004. Hierarchical
Bayesian estimation for MEG inverse problem. NeuroImage, 23(3), pp.806-826.
https://doi.org/10.1016/j.neuroimage.2004.06.037
[68] Silva Pereira, S., Hindriks, R., Mühlberg, S., Maris, E., van Ede, F., Griffa, A., Hagmann, P. and Deco, G.,
2017. Effect of Field Spread on Resting-State Magneto Encephalography Functional Network Analysis:
A Computational Modeling Study. Brain connectivity, 7(9), pp.541-557.
https://doi.org/10.1089/brain.2017.0525
[69] Schmidt, M., 2010. Graphical model structure learning with l1-regularization. University of British
Columbia.
[70] Schoffelen, J.M. and Gross, J., 2009. Source connectivity analysis with MEG and EEG. Human brain
mapping, 30(6), pp.1857-1865.
https://doi.org/10.1002/hbm.20745
[71] Schmolck, A. and Everson, R., 2007. Smooth relevance vector machine: a smoothness prior extension
of the RVM. Machine Learning, 68(2), pp.107-135.
https://doi.org/10.1007/s10994-007-5012-z
[72] Srivastava, M.S., 1965. On the complex Wishart distribution. The Annals of mathematical statistics,
36(1), pp.313-315.
https://www.jstor.org/stable/2238098
[73] Tipping, M.E., 2001. Sparse Bayesian learning and the relevance vector machine. Journal of machine
learning research, 1(Jun), pp.211-244.
https://doi.org/10.1162/15324430152748236
[74] Valdés-Hernández, P.A., von Ellenrieder, N., Ojeda-Gonzalez, A., Kochen, S., Alemán-Gómez, Y.,
Muravchik, C. and Valdés-Sosa, P.A., 2009. Approximate average head models for EEG source imaging.
Journal of neuroscience methods, 185(1), pp.125-132. Valdes-Sosa, P.A., 2004. Spatio-temporal
autoregressive models defined over brain manifolds. Neuroinformatics, 2(2), pp.239-250.
https://doi.org/10.1016/j.jneumeth.2009.09.005
[75] Valdes-Sosa, P., Marti, F., Garcia, F. and Casanova, R., 2000. Variable resolution electric-magnetic
tomography. In Biomag 96 (pp. 373-376). Springer, New York, NY.
https://doi.org/10.1007/978-1-4612-1260-7_91
[76] Valdés-Sosa, P.A., Bornot-Sánchez, J.M., Vega-Hernández, M., Melie-García, L., Lage-Castellanos, A.
and Canales-Rodríguez, E., 2006. 18 granger causality on spatial manifolds: applications to
neuroimaging. Handbook of time series analysis: recent theoretical developments and applications,
pp.461-491.
https://doi.org/10.1002/9783527609970.ch18
[77] Van de Steen, F., Faes, L., Karahan, E., Songsiri, J., Valdes-Sosa, P.A. and Marinazzo, D., 2016. Critical
comments on EEG sensor space dynamical connectivity analysis. Brain topography, pp.1-12.
https://doi.org/10.1007/s10548-016-0538-7
[78] Van Hoey, G., Vanrumste, B., D'have, M., Van de Walle, R., Lemahieu, I. and Boon, P., 2000. Influence
of measurement noise and electrode mislocalisation on EEG dipole-source localisation. Medical and
Biological Engineering and Computing, 38(3), pp.287-296.
https://doi.org/10.1007/BF02347049
[79] Waldorp, L.J., Huizenga, H.M., Dolan, C.V. and Molenaar, P.C., 2001. Estimated generalized least
squares electromagnetic source analysis based on a parametric noise covariance model [EEG/MEG].
IEEE Transactions on Biomedical Engineering, 48(6), pp.737-741.
https://doi.org/10.1109/10.923793
[80] Wan, J., Zhang, Z., Rao, B.D., Fang, S., Yan, J., Saykin, A.J. and Shen, L., 2014. Identifying the
neuroanatomical basis of cognitive impairment in Alzheimer's disease by correlation-and nonlinearity-
aware sparse Bayesian learning. IEEE transactions on medical imaging, 33(7), pp.1475-1487.
https://doi.org/10.1109/TMI.2014.2314712
[81] Wang, H., 2012. Bayesian graphical lasso models and efficient posterior computation. Bayesian
Analysis, 7(4), pp.867-886.
https://doi.org/10.1214/12-BA729
[82] Wang, H., 2014. Coordinate descent algorithm for covariance graphical lasso. Statistics and Computing,
24(4), pp.521-529.
https://doi.org/10.1007/s11222-013-9385-5
[83] Wens, V., Marty, B., Mary, A., Bourguignon, M., Op De Beeck, M., Goldman, S., Van Bogaert, P.,
Peigneux, P. and De Tiège, X., 2015. A geometric correction scheme for spatial leakage effects in
MEG/EEG seed‐based functional connectivity mapping. Human brain mapping, 36(11), pp.4604-4621.
https://doi.org/10.1002/hbm.22943
[84] Wills, A., Ninness, B. and Gibson, S., 2009. Maximum likelihood estimation of state space models from
frequency domain data. IEEE Transactions on Automatic Control, 54(1), pp.19-33.
https://doi.org/10.1109/TAC.2008.2009485
[85] Wipf, D.P., Ramırez, R.R., Palmer, J.A., Makeig, S. and Rao, B.D., 2006. Automatic Relevance
Determination for Source Localization with MEG and EEG Data. Technical Report, University of
California, San Diego.
[86] Wipf, D.P. and Rao, B.D., 2007. An empirical Bayesian strategy for solving the simultaneous sparse
approximation problem. IEEE Transactions on Signal Processing, 55(7), pp.3704-3716.
https://doi.org/10.1109/TSP.2007.894265
[87] Wipf, D. and Nagarajan, S., 2009. A unified Bayesian framework for MEG/EEG source imaging.
NeuroImage, 44(3), pp.947-966.
https://doi.org/10.1016/j.neuroimage.2008.02.059
[88] Wipf, D.P., Owen, J.P., Attias, H.T., Sekihara, K. and Nagarajan, S.S., 2010. Robust Bayesian estimation
of the location, orientation, and time course of multiple correlated neural sources using MEG.
NeuroImage, 49(1), pp.641-655.
https://doi.org/10.1016/j.neuroimage.2009.06.083
[89] Wu, W., Nagarajan, S. and Chen, Z., 2016. Bayesian Machine Learning: EEG\/MEG signal processing
measurements. IEEE Signal Processing Magazine, 33(1), pp.14-36.
https://doi.org/10.1109/MSP.2015.2481559
[90] Yuan, G., Tan, H. and Zheng, W.S., 2017. A Coordinate-wise Optimization Algorithm for Sparse Inverse
Covariance Selection. arXiv preprint arXiv:1711.07038.
https://arxiv.org/abs/1711.07038
[91] Zhang, T. and Zou, H., 2014. Sparse precision matrix estimation via lasso penalized D-trace loss.
Biometrika, 101(1), pp.103-120.
https://doi.org/10.1093/biomet/ast059
[92] Zhang, Z. and Rao, B.D., 2011. Sparse signal recovery with temporally correlated source vectors using
sparse Bayesian learning. IEEE Journal of Selected Topics in Signal Processing, 5(5), pp.912-926.
https://doi.org/10.1109/JSTSP.2011.2159773
[93] Zhang, T. and Zou, H., 2014. Sparse precision matrix estimation via lasso penalized D-trace loss.
Biometrika, 101(1), pp.103-120.
https://doi.org/10.1093/biomet/ast059
[94] Zhang, Y., Zhou, G., Jin, J., Zhao, Q., Wang, X. and Cichocki, A., 2016. Sparse Bayesian classification of
EEG for brain -- computer interface. IEEE transactions on neural networks and learning systems, 27(11),
pp.2256-2267.
https://doi.org/10.1109/TNNLS.2015.2476656
[95] Zhang, Y., Wang, Y., Jin, J. and Wang, X., 2017. Sparse Bayesian learning for obtaining sparsity of EEG
frequency bands based feature vectors in motor imagery classification. International journal of neural
systems, 27(02), p.1650032.
https://doi.org/10.1142/S0129065716500325
Appendix
Mathematical notation
[A.1]
[A.2]
𝒙, 𝐗, 𝕏
𝒙𝓂
[A.3]
X𝑖𝑗,(X)𝑖𝑗, 𝑥𝑖, (𝑥)𝑖
[A.4]
[A.5]
[A.6]
[A.7]
𝑁p(𝒙𝒚, 𝐙)
ℂ(𝒙𝒚, 𝐙)
𝑁p
𝑒𝑥𝑝(𝑥𝑦)
𝐺𝑎𝑚𝑚𝑎(𝑥𝑦, 𝑧)
[A.8]
𝑇𝐺𝑎𝑚𝑚𝑎(𝑥𝑦, 𝑧, (𝑎, 𝑏))
[A.9]
[A.10]
[A.11]
[A.12]
[A.13]
[A.14]
[A.15]
[A.16]
[A.17]
𝐗
𝑡𝑟(𝐗)
𝐗−1
𝐗 𝒯
𝐗†
𝐗, 𝒙
𝐗
𝐗(𝑘), 𝐗(𝑘)
m
∑
𝓂=1
the scalar 𝑥 with
The following symbols denote respectively a vector
(bold italic lowercase) a matrix (bold capital) a set
(double struck capital).
Subscript indicating with lowercase script the 𝓂-th
vector sample.
Subscript indicating with lowercase the 𝑖𝑗 (𝑖) element
of a matrix 𝐗 (vector 𝒙). Light captions denote matrix
(vectors) elements.
Normal distribution of a (p) size vector 𝒙 with mean
𝒚 and Covariance Matrix 𝐙.
Circularly Symmetric Complex Normal distribution
of a (p) size complex vector 𝒙 with complex mean 𝒚
and Complex Covariance Matrix 𝐙.
Exponential distribution of
parameter of shape 𝑦.
Gamma distribution of the scalar 𝑥 with parameters of
shape 𝑦 and rate 𝑧.
Truncated Gamma distribution of the scalar 𝑥 with
parameters of shape 𝑦, rate 𝑧 and truncation interval
(𝑎, 𝑏).
Determinant of a matrix 𝐗.
Trace of a matrix 𝐗.
Inverse of a matrix 𝐗.
Transpose of a matrix 𝐗.
Conjugate transpose of a matrix 𝐗.
Estimator Parameters or Hyperparameters random
matrix (𝐗) or vector (𝒙).
Estimator of auxiliary magnitudes random matrix (𝐗),
dependent on Parameters or Hyperparameters
estimators.
Updates at the 𝑘-th iteration of estimators.
Sum operator along index 𝓂.
[A.18]
[A.19]
[A.20]
∏
m
𝓂=1
𝑝(𝐗)
𝑝(𝐗𝐘)
[A.21]
𝑝(𝐗, 𝐘𝐙)
[A.22]
𝐗𝐘 ∽ 𝑝(𝐗𝐘)
[A.23]
‖𝐗‖𝒾,𝐀𝒾, 𝒾 = 1,2
[A.24]
[A.28]
[A.29]
[A.30]
𝐈p, 𝟏p, 𝟎p
⊙, ⊘
𝑎𝑟𝑔𝑚𝑖𝑛𝐗{𝑓(𝐗)}
or
𝑎𝑟𝑔𝑚𝑎𝑥𝐗{𝑓(𝐗)}
𝑧𝑒𝑟𝑜𝑠𝐗{𝑓(𝐗)}
Definition of variables
Product operator along index 𝓂.
Probability density function of a random variable 𝐗.
Conditional probability density function of a random
variable 𝐗 regarding the state of the variable 𝐘.
Conditional joint probability density function of
random variables 𝐗 and 𝐘 regarding the state of the
variable 𝐙.
Indicates that the variable 𝐗 probability density
function is conditioned to 𝐘.
L1 or L2 norm of the matrix 𝐗 with weights or
elementwise precisions defined by the mask matrix
𝐀𝒾.
Denotes respectively Identity, Ones and Ceros
matrices of size p.
Elementwise matrix product a division operators
(Hadamard).
function 𝑓 ,
Extreme values of
correspondingly minimum or maximum, in the
argument 𝐗.
Zeros of the scalar function 𝑓 in the argument 𝐗.
scalar
the
[B.1]
[B.2]
[B.3]
[B.4]
[B.5]
[B.6]
[B.7]
[B.8]
[B.9]
[B.10]
[B.11]
[B.12]
𝔼
𝕄
𝔾
p
m
q
𝒗𝓂
𝜾𝓂
𝐋
𝝃𝓂
𝚺𝜾𝜾
𝚯𝜾𝜾
Scalp Sensors (Electrodes) Space.
Random Samples space.
Discretized Gray Matter (Generators) Space.
Number of MEEG sensors at the scalp.
Number of data samples obtained from MEEG single
frequency bin Fourier coefficients from a number (m)
of segments.
Number of MEEG generators at the Cortex surface.
Complex size MEEG data Fourier coefficients sample
vector for a single frequency component (observed
variables or Data).
Complex size MEEG source's Fourier coefficients
sample vector for a single frequency component
(unobserved variables or parameters).
Lead Field matrix of n × q size.
Complex Fourier coefficients vector for a single
frequency component from MEEG forward model
residuals (sensors' noise).
Complex size Hermitian and positive semidefinite
matrix of EEG/MEG sources' Fourier coefficients
(unobserved variables or Parameters) Covariance
matrix.
Complex size Hermitian and positive semidefinite
matrix of EEG/MEG source's Fourier coefficients
[B.13]
[B.14]
[B.15]
[B.16]
𝚺𝝃𝝃
𝐑
2
𝜎𝝃
𝚵
[B.17]
𝑄(𝚵, 𝚵)
[B.18]
[B.19]
[B.20]
[B.21]
[B.22]
[B.23]
[B.24]
Π(𝚯𝜾𝜾, 𝐀)
𝜆
𝐓 (𝑘)
𝚺
(𝑘)
𝜾𝜾
𝐒
(𝑘)
𝜾𝜾
(𝑘)
𝚿 𝜾𝜾
𝐒𝒗𝒗
Inverse
2.
(unobserved variables or Parameters)
Covariance matrix.
Complex Hermitian and positive semidefinite matrix
of EEG/MEG forward model residuals' Fourier
coefficients (sensors' noise) Covariance matrix.
Known Complex Hermitian and positive semidefinite
matrix of EEG/MEG sensors correlation structure.
Positive nuisance level hyperparameter 𝜎𝑒
General variable defining the set of hyperparameters.
Data expected log likelihood, obtained after the
expectation operation of the data and parameters log
joint conditional probability density function over the
parameters accounting for the parameters posterior
density function with estimated values of
the
hyperparameters.
Scalar general penalty function or exponent of the
prior distribution Precision matrix 𝚯𝜾𝜾 parametrized in
the regularization parameters or mask matrix 𝐀.
Regularization parameters or tuning hyperparameters
vector of the general penalty function.
MEEG Data
Operator.
Complex Hermitian and positive semidefinite matrix
of MEEG source Fourier coefficients (unobserved
variables or parameters) posterior Covariance matrix.
Complex Hermitian and positive semidefinite matrix
of MEEG sources' Fourier coefficients (unobserved
variables or parameters) empirical Covariance matrix.
Effective Sources Empirical Covariance (ESEC). It
carries the information about sources correlations that
will effectively influence the sources Covariance
matrix estimator in the maximization step (sources
Graphical Model solution), thus, it becomes the
sources Covariance matrix estimator in the especial
case of prior free model.
Complex Hermitian matrix MEEG data Fourier
coefficients Covariance matrix.
to Source activity Transference
|
1708.00138 | 2 | 1708 | 2017-09-26T01:09:13 | The differential geometry of perceptual similarity | [
"q-bio.NC"
] | Human similarity judgments are inconsistent with Euclidean, Hamming, Mahalanobis, and the majority of measures used in the extensive literatures on similarity and dissimilarity. From intrinsic properties of brain circuitry, we derive principles of perceptual metrics, showing their conformance to Riemannian geometry. As a demonstration of their utility, the perceptual metrics are shown to outperform JPEG compression. Unlike machine-learning approaches, the outperformance uses no statistics, and no learning. Beyond the incidental application to compression, the metrics offer broad explanatory accounts of empirical perceptual findings such as Tverskys triangle inequality violations, contradictory human judgments of identical stimuli such as speech sounds, and a broad range of other phenomena on percepts and concepts that may initially appear unrelated. The findings constitute a set of fundamental principles underlying perceptual similarity. | q-bio.NC | q-bio | Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
Abstract
Antonio M Rodriguez1 and Richard Granger1*
Rodriguez & Granger
[arXiv:1708.00138v1]
The differential geometry of perceptual similarity
1Brain Engineering Lab, Dartmouth College, Hanover NH 03755, USA
*[email protected]
Rodriguez A, Granger R (2017) The differential geometry of perceptual similarity. Brain Engineering
Laboratory, Technical Report 2017.2. www.dartmouth.edu/~rhg/pubs/GrangerRGPEGa1.pdf
Human similarity judgments are inconsistent with Euclidean, Hamming, Mahalanobis,
and the majority of measures used in the extensive literatures on similarity and
dissimilarity. From intrinsic properties of brain circuitry, we derive principles of
perceptual metrics, showing their conformance to Riemannian geometry. As a
demonstration of their utility, the perceptual metrics are shown to outperform JPEG
compression. Unlike machine-learning approaches, the outperformance uses no
statistics, and no learning. Beyond the incidental application to compression, the metrics
offer broad explanatory accounts of empirical perceptual findings such as Tversky's
triangle inequality violations (1, 2), contradictory human judgments of identical stimuli
such as speech sounds, and a broad range of other phenomena on percepts and concepts
that may initially appear unrelated. The findings constitute a set of fundamental
principles underlying perceptual similarity.
When do images look alike? All standard Euclidean (and Hamming, and Mahalanobis, and
almost all other) standard measures of similarity turn out to be at odds with human
similarity judgments. We spell out why this is the case, give explanatory principles, and
provide an illustrative application to the widely-used JPEG compression algorithm: JPEG has
been outperformed via extensive learning by neural network and ML methods, whereas we
outperform it with no statistics, and no training. We show that the JPEG findings fall out as
a special case of the underlying broad principles introduced here, which are applicable to a
wide range of unsupervised methods that entail similarity measures.
Euclidean vectors' components are orthogonal, and thus
b = (00010) are
!a = (10000) and
b!c both have Hamming distances of 2,
!a!c and
!c = (00001): distances
equidistant from
and Euclidean distances of 2. However, considered as physical images, the right-hand
b and !c render them more visually similar to each
positioning of the "1" values in vectors
other than either is to !a. When such "neighbor" relations within a vector are considered,
then vector axes are not orthogonal, and non-Euclidean metrics can readily yield smaller
distances between
b and !c than between !a and !c.
Human judgments of similarity imply a particular geometric system, and as in the above
simple example, it is easy to show that human similarity judgments do not conform to
Euclidean, Hamming, Mahalanobis, or the other most commonly used similarity metrics;
rather, we will show that they conform to Riemannian geometry.
Introduction: The fundamentals of perceptual similarity
!
!
!
!
Rodriguez & Granger
Differential geometry of perceptual similarity
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
Moreover, human similarity judgments are not solely a function of the stimuli themselves;
they depend also on the operations internally carried out by the perceiver. Given physical
stimuli (e.g., the English speech sounds /ra/ and /la/) are distinguishable to some
perceivers (native English speakers) but difficult to discriminate by other perceivers (e.g.,
native Japanese speakers), as a result of prior experience of the perceiver (3-7). Thus "re-
coding" the stimuli does not contribute to a solution (8-10). Rather, we wish to identify the
perceptual metrics that a given perceiver uses when judging similarity and difference.
As in the vector "neighbor" example, perceptual brain system anatomy reflects non-
Euclidean metrics: the signal transmitted from one group of neurons to another directly
corresponds to mappings among non-Euclidean spaces. We derive a formalism from
synaptic connectivity patterns and show that the system matches and explains individual
human empirical judgments.
We apply the derived system to the well-studied JPEG compression task (solely to
demonstrate the real-world efficacy of the presented principles).
Conveniently, several statistical machine-learning approaches have recently been shown to
outperform the sturdy hand-constructed JPEG method, via extensive training on image data
(11, 12). Those results may be viewed as calling for potential explanatory principles that
may underlie their successes. What structure in the data is being statistically identified by
these learning approaches?
The method derived in the present manuscript is easily shown capable of outperforming
JPEG as well, with no increase in computational cost over JPEG, and using no statistics and
no training. The results arise from newly posited principles that underlie not just JPEG, but
perceptual similarity in general; JPEG is shown to fall out as a special case of the method.
Beyond the illustrative JPEG example, the metrics are shown to proffer explanatory
accounts of a range of empirical perceptual findings, notably Tversky's triangle inequality
violations (1, 2), contradictory human judgments of stimuli in other modalities such as
speech sounds, and beyond simple percepts to abstract concepts and categories that may
initially appear unrelated.
What JPEG does, and a principled error
Any lossy compression algorithm trades off image quality for image entropy: how the image
appears vs. how much space it can be stored in. The JPEG standard (13, 14) transforms an
image into a frequency basis, and encodes each of the frequency components with a
different amount of precision (tending to encode low-frequency components with more
precision than high-frequency components), thus selectively introducing modest errors
preferentially into high frequency components, yielding a new image with more error but
less entropy than the original; i.e., an image judged (via human viewers) to be of lesser
quality, but capable of fitting into a smaller file size.
An axiomatic error incorporated into JPEG (and indeed into most compression methods,
and most measures of similarity and dissimilarity), is the assumption that the frequency
basis vectors are orthogonal, and thus that changes to any one of them do not impact the
Page 2 of 23
Results
⎡
⎢
⎣
(2x +1)uπ
2N
⎤
⎥
⎦
⎡
⎢
⎣
2N
′Fx,y cos
if x = 0
else
⎧
⎪
⎨
⎩⎪
1
2
1
α(x) =
⎥
⎤
⎦
4α(u)α(v)
N −1
∑
x=0
N −1
∑
y=0
Tu,v = 1
cos (2y +1)vπ
Rodriguez & Granger
Differential geometry of perceptual similarity
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
others. In Euclidean space, these bases are indeed orthogonal; in Riemannian space they
are not. We show that human perceptual similarity judgments are consistent with non-
orthogonal bases, properly treated as a Riemannian space, not Euclidean or affine.
For JPEG, let
!p be a 64-dimensional vector comprised of intensity information from an 8x8
block of pixels (we initially focus on monochrome images for expository simplicity). We
define a matrix Fx,y whose elements are the intensities of the pixels at locations (x, y).
JPEG encoding proceeds via the following four steps:
1) Center image intensities around 0:
′Fx,y = Fx,y −128.
2) Represent image as linear combination of frequency components
The discrete cosine transform (DCT) operator T is given by:
(2.1)
and where N=8 for JPEG.
where
′F )
′F ; i.e., T(
(The DCT is a linear operator acting on the (centered) image
3) Quantization and rounding of frequency components
Each of the 64 dimensions of T are independently scaled by a predetermined
"quantization" matrix QC (with the intended effect of discarding less-relevant
information in the data), with different matrices defined for different "calibers" C of the
error/entropy tradeoff. The operator R simply divides the elements of a transformed
input (T) by the designated quantization matrix QC:
Ru,v = Tu,v QCu,v or, in matrix
notation, R = QC
−1 iT
Full quantization of input
′F, then, is accomplished in JPEG by
(2.2)
′F )
The JPEG standard establishes preëstablished quantization matrices (QC ) for any
given desired compression factor C, i.e., for a given reduction in quality and
commensurate decrease in entropy. These quantization matrices were constructed by
hand (13, 14), with no non-manual method for arriving at its values (15).
4) Entropy encoding
The numerical elements of JPEG's quantized and rounded image Z are encoded via
Huffman coding, such that the most frequently used numerical values are assigned the
shortest bit representation, thus taking advantage of the reduced entropy of the
quantized input, to enable the compressed image to be stored in a file smaller than the
original. Although Huffman coding is used by JPEG, any number of entropy encoding
methods (such as arithmetic encoding) would suffice.
We focus on a specific erroneous assumption underlying JPEG (and other perceptual
compression methods): the use of Euclidean measures of similarity. In fact, any given pixel
Page 3 of 23
Z = round(R i
′F ) = round(QC
′F ) = T i
−1 iT i
Brain connectomes are Riemannian
The Riemannian geometric principles of perception
Three geometric spaces
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
p is not perceptually independent of nor orthogonal to neighboring pixels. As in the
examples in the introduction section, neighboring pixels are perceived by a viewer as being
"closer" to each other than are more distal pixels, and this alters perceived similarity.
Correspondingly, the 64 frequency bases in the DCT are not perceptually orthogonal
(though they are orthogonal in Euclidean space): some are perceptual judged more similar
to each other than others by human viewers. Moreover, these judgments are dependent on
their coefficients, and thus have different similarities in different parts of the DCT basis
space. The curvature of the space thus is not affine, but rather Riemannian (Figure 1).
We introduce an appropriate Riemannian treatment of perceptual similarity. We show that
the resulting method can readily outperform JPEG, but more importantly, it has explanatory
power: JPEG emerges as a special case of the general method, and the underlying geometric
principles of human perception become more closely explained.
Assume an image of 64 pixels (grayscale, for temporary pedagogical simplicity), arranged as
an 8x8 array. In JPEG and all other standard compression mechanisms, the image is treated
as an arbitrarily, but consistently, ordered 64-dimensional vector, such that each vector
entry corresponds to the intensity at one of the 64 pixels in the 8x8 array. This renders the
data into simple vector format, enabling the applicability of vector and matrix operations.
However, it does so at the cost of eliminating the neighbor relations among pixels in the
physical space. (Typically, the 64 pixels are ordered (arbitrarily) with entries 1-8 from the
top row of the array, 9-16 from the next row, and so on.) The elimination of neighbor
relations would be irrelevant if human perception of a pixel were modulated equally, or not
at all, by the characteristics of neighboring and distant pixels alike; this turns out not to be
the case. We forward the alternative in which the image is described in terms of a physical
space Φ with 3 dimensions (for x and y locations, and intensity), for each of the 64 pixels.
This corresponds to a transformation of the "feature" space F into physical space Φ. The
explicit representation of physical pixel positions enables perceptual encodings that use
pixel position as a parameter.
In addition to feature space F and physical space Φ, we introduce a third space, Ψ,
which we term "perceptual space," that includes representation of perceptual geometric
relations among the image elements (Figure 2). Transforms into this perceptual space,
accomplished by differential geometry, will be shown to directly correspond to human
perceptual similarity judgments.
Figure 2 shows sample anatomical connectivity among brain regions, and its formal
properties. Figure 2a shows an instance of typical mammalian thalamocortical and cortico-
cortical synaptic projections (16-20). The projection pattern from one cellular assembly to
another is not perfectly "point to point" (i.e., each cell projecting to exactly one
topographically corresponding target cell) nor completely diffuse (with no topography);
rather, the projection is "radially extended," such that each element contacts a range of
targets roughly within a spatial neighborhood or radius around a target. Figure 2b shows a
simple vector encoding of these projection patterns with corresponding synaptic weights
Page 4 of 23
′n ,
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
′′n , etc. Figure 2c shows examples of typical physiological neural responses in early
sensory cortical areas that can arise from these connectivity patterns. Figure 2d contains
the general form of a Jacobian matrix denoting the overall effect of activity in the neurons of
an input area x on the neurons in target area f; each entry in the Jacobian designates the
change in an element of f as a consequence of a given change in an element of x. Figure
2e is an example instance of such a Jacobian, corresponding to the synaptic connectivity
pattern in Figure 2b.
Intuitively, a Jacobian encodes the interactions among stimulus components. If a vector
contained purely independent entries (as in an imagined perfectly point-to-point
topography with no lateral fan-in or fan-out projections), the Jacobian would be the identity
matrix: ones along the diagonal and all other entries zeros. Each vector dimension then has
no effect on other dimensions: a given input unit affects only a single target unit, and no
others. In actual connectivity, which does contain some radially extended projections, there
are off-diagonal non-zero values in the Jacobian corresponding to the slightly non-
topographic synaptic contacts (Figure 2b).
When the input and output patterns are treated as vectors, any off-diagonal Jacobian
elements reflect influences of one dimension on others: the dimensions are not orthogonal,
and the vectors are Riemannian, not Euclidean (21).
All perceptual systems can be seen to intrinsically express a "stance" on the geometric
relations that occur among the components of the stimuli processed by the system. In the
degenerate case of no off-diagonal elements, the system would act as though it assumes
independence of components (Euclidean vectors). In all pathways characteristic of most
thalamo-cortical and cortico-cortical projections, however, the processing inherently
assumes non-Euclidean neighbor relations among the stimuli.
It is notable that any bank of neuronal elements with receptive fields consisting either of
Gaussians or of first or second derivatives of Gaussians, will have precisely the effect of
computing the derivatives of the inputs in just the form that arises in a Jacobian (see
equation (2.3) below) (22-26). Physiological neuron response patterns thus appear
thoroughly suited to producing transforms into spaces with Riemannian curvatures.
(Notably, this implies that a synaptic change (e.g., LTP) causes specific re-shaping of
neurons' receptive fields, modifying the curvature of the space of the target cells in a given
projection pathway.)
The matrix J in Figure 2d describes the particular transform from an input space to an
output space. This is an instance of specifying the differences between a perceptual input,
versus a percept that is received via this projection pathway. A perceiver will process an
input as though it contains the neighbor relations specified by the Jacobian.
The map from physical to perceptual space
Neighboring entries in a vector, like adjacent notes on a piano keyboard, are closer to each
other than entries from other, non-neighboring dimensions. The features thus do not
constitute independent dimensions (or, put differently, the dimensions are not orthogonal).
In these (extremely common) cases, target perceptual distances are correctly rendered by
Riemannian rather than by Euclidean measures. It is notable that this not an exception but
Page 5 of 23
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
JF→Φ =
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
dΦ0
df2
dΦ1
df2
dΦ2
df2
!
!
!
dΦ0
df1
dΦ1
df1
dΦ2
df1
dΦ0
df0
dΦ1
df0
dΦ2
df0
dΦ0
df63
dΦ1
df63
dΦ2
df63
µF→Φ the Jacobian
Rodriguez & Granger
[arXiv:1708.00138v1]
the normal case for perception. Euclidean vectors do not treat constituents as having
neighbors, but perceivers do.
We wish to determine, then, how changes to an image will be perceived. A change in
physical space (i.e., the image) can be directly measured. The corresponding predicted
change in perceptual space can then be computed via a metric tensor which measures
distance in the perceptual space with respect to positions in physical space. This metric
tensor is computed via a Jacobian that maps from distances in physical space to distances in
JF→Φ will be a 3x64
perceptual space (Figure 3). For the map
matrix:
(2.3)
(Supplemental sections §2.5-§2.11 give sample values used to generate specific Jacobians
for image processing, as in the examples shown in Figure 4).
This Jacobian enables identification of a distance metric for the feature space F with
respect to its embedding in physical space Φ in terms of the metric tensor g (see
Supplemental section §2.5 for examples of values used):
( !x) i JF→Φ( !x)
(2.4)
i.e., the metric tensor g is using measures in space Φ applied to objects in space F, or, put
differently, the tensor measures distances in F with respect to measures in space Φ.
Then, mapping physical space Φ to perceptual space Ψ (via Jacobian operator JΦ→Ψ)
defines how features in the physical space are perceived by a viewer, enabling a formal
description of how changes in the physical image are registered as perceptual changes.
Specific construction of the Jacobian mapping JΦ→Ψ
Information from the physical stimulus or from the perceiver (or both) enables
construction of a Jacobian to map from physical vectors to the perceptual space a perceiver
may use. Such a Jacobian, JΦ→Ψ, can be obtained directly from either synaptic
connectivity patterns or from psychophysics – either by a priori assumptions or from
empirical measurements.
(i) Synaptic Jacobian:
a) Empirical
Measure anatomical connections and synaptic strengths, if known; the Jacobian is
directly obtained from those data as in Figure 2. These measures typically are
unavailable, but as will be seen, approximations may be drawn from a set of simple
connectivity assumptions.
b) Estimated
Page 6 of 23
gΦ:F ( !x) = JF→Φ
T
Differential geometry of perceptual similarity
Rodriguez & Granger
Brain Engineering Lab Tech Report 2017.2
[arXiv:1708.00138v1]
Assume radius of projection fan-out from a cortical region to a target region
(Figure 2a), based on measures of typical such projections in the literature (16-20),
and estimate a factor by which distances among stimulus input features (e.g.,
pixels) influence relatedness of the features, and resulting curvature of the
Riemannian space in which they are thus assumed to be perceptually embedded.
(ii) Psychophysical Jacobian:
a) Empirical
Measure constituent physical features of the stimuli and calculate distances among
stimulus features, such as pixel size, pixel disparity, viewing distance, and obtain
empirical measures of human-reported distances; the Jacobian is the set of
relations among psychological and physical distances as in Equation (2.3).
b) Estimated
Assume Gaussian fall-off of relatedness of neighboring pixels in a stimulus;
measure constituent features as in (ii a) and estimate a factor by which distances
between stimulus input features (e.g., pixel distances in x and y directions)
influence the relatedness of the features (and the resulting curvature of the
Riemannian space in which they are assumed to be perceptually embedded).
In the present work we proceed with method (ii b), i.e., measuring (Euclidean) physical
distances among pairs of inputs and positing a range of candidate factors by which the
physical disparity among features may give rise to perceived feature interactions. We show
a series of resulting findings corresponding to this range of different hypothesized factors
(Supplemental section §2.5).
Having obtained a Jacobian by any of the above means, we compute metric tensor g as in
equation (2.4). (The tensor alternately may be obtained in condition (ii) using the
covariance matrix Σ from psychophysical experimental data: gΨ:Φ = ΣΨ:Φ
−1 .) (See
supplemental section §1.4).
We may move the obtained metric gΨ:Φ from physical space to feature space, obtaining a
gΨ:F that measures distances in the feature space with respect to the
new metric
perceptual space:
(2.5)
This new metric in the feature space now computes the Riemannian distances among
dimensions that hold in the feature space.
The metric can be used to compute the matrix of all distances among all pairs of features
xi,x j in a column vector !x of dimensionality k:
(2.6)
)
(where
!a is the determinant of !a). (Sample distance matrices for selected specific
)−1 2
measured visual parameters are shown in Supplemental section §2.7; tables §2-§4).
Page 7 of 23
i gΨ:Φ i JF→Φ
dist(xi,x j) = (2π)k gΨ:F
T
gΨ:F = JF→Φ
(
exp − 1
2 (xi − x j)T gΨ:F (xi − x j)
(
Methods
Derivation of Riemannian geometric JPEG (RGPEG)
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
The dimensions of a feature vector in (Euclidean) space F are orthogonal, but the
dimensions of the corresponding vector in (Riemannian) perceptual space Ψ are not
orthogonal; rather, the pairwise distances among the dimensions are described by Eq. (2.6).
JPEG modifies the image feature vector, introducing error (the distance between the
original and modified vector), such that the modified vector has lower entropy, and thus can
be stored with a smaller description.
Correspondingly, we too will modify the feature vector, introducing error in order to lower
entropy, but in this case using graph-based operations on non-Euclidean dimension
distances (Eq. (2.6)). We introduce the Riemannian geometric perceptual encoding graph
(RGPEG) method.
JPEG uses a (hand-constructed) quantization ("Q") matrix that specifies the amount by
which each of the 64 DCT dimensions will be perturbed, such that when they are subjected
to integer rounding, they will exhibit lower entropy.
We replace the JPEG quantization operations with a principled formula that computes
perturbations of basis dimensions to achieve a desired entropy reduction and
commensurate error – but in perceptual space rather than in feature space. Specifically, the
surrogate quantization step moves the image in perceptual space along the gradient of the
eigenvectors of the Hamiltonian of the basis space. We show that the resulting computation
can outperform JPEG operations (or any operations that take place in feature space rather
than in perceptual space).
We define a graph whose nodes are the 64 basis dimensions of the feature space. (For JPEG
this basis is the set of 64 2-d discrete cosine transforms; for RGPEG we derive the
generalization of this basis for perceptual space, showing the DCT to be a special case).
Activation patterns in the graph can be thought of as the state of the space, and operations
on the graph are state transitions. We define Ω(x,s) as the state describing the intensity of
each of the pixels in the (8x8) image, such that Ω(x,0)is the original image, and any Ω(x,s)
for non-zero s values is an altered image, including the possible compressed versions of the
image. We define the s values to be in units of bits x length; corresponding to the number of
bits required to store a given image, and thus commensurate with entropy (see
Supplemental section §2.3). We wish to know how to change the image such that the
entropy will be reduced. Changes to the image with respect to entropy are expressed as
We treat the problem of such image alterations in terms of the heat equation (see, e.g., (27,
28), and see Supplemental section §2.3). We equate the second derivative of the image
state with respect to distance, with the derivative of the image with respect to entropy:
Page 8 of 23
Derivation of entropy constraint equation
∂Ω(x,s)
∂s
=ω(x)
dφ(s)
ds
ω(x)
=
∂2Ω(x,s)
∂x 2
d 2ω(x)
=
∂Ω(x,s)
∂s
∂Ω(x,s)
∂s
∂2Ω(x,s)
∂x 2
dx 2 φ(s)
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
(2.7)
We term eq. (2.7) the entropy constraint equation; we want to identify Ψ(x,s) that satisfies
this equation, such that we can generate modifications of an image to achieve a new image
exhibiting a reduction entropy (and correspondingly increase in error).
Via separation of variables we assume a solution of the form
(2.8)
Ω(x,s) =ω(x)φ(s)
where the function ω is only in terms of position information x and the function φ only in
terms of entropy s. Thus the former connotes the "position" portion of the solution, i.e.,
values of image pixels regardless of entropy values, whereas φ is the entropy portion of the
solution.
We can formulate two ordinary differential equations corresponding to the two sides of the
partial differential equation in equation (2.7):
and
which both equal the same value and can thus be equated:
dx 2 φ(s)
(2.9)
which can be simplified
Since the two functions are equal, they are equal to some quantity (which cannot be a
function of x or s, since the equality would then not consistently hold). We call that quantity
λ. There can be a distinct λ value for each candidate solution i. For any such given
solution, the entropy term is:
ds =φi(s)λi
whose solution is
(2.10)
φi(s) = eλis
As mentioned, there will be i solutions for each value of λ. (See supplemental section §2.3).
For the position term:
(2.11)
=ωi(x)λi
the solution is in the form of the Fourier decomposition
Page 9 of 23
d 2ω(x)
dx 2
1
φ(s)
dφ(s)
ds =
dφ(s)
ds =
1
ω(x)
d 2ω(x)
dx 2
dφ(s)
d 2ω(x)
!γi(x)
ci
n
j=1
Aij
!Lg =
ωi(x) =
∑
i
Application of entropy constraint equation to image feature space
!A of that graph. We compute the degree matrix
!D via
2.
!A, and the normalized graph Laplacian is then L = D 1
Differential geometry of perceptual similarity
Rodriguez & Granger
Brain Engineering Lab Tech Report 2017.2
[arXiv:1708.00138v1]
(2.12)
!γi terms are the eigenvectors of the Laplacian of the position term, eq. (2.11),
where the
and where the ci terms correspond to the coefficients of the eigenvector basis of the initial
condition of the state Ω(x,s) corresponding to the initial image itself, Ω(x,0). (Precise
formulation of the ci is shown in the next section). The 64 solutions of the Fourier
decomposition form the basis space into which the image will be projected. (For JPEG, this
is the discrete cosine transform or DCT set, as mentioned; we will see that this corresponds
to one special case of the solution, for a specific set of values of the entropy constraint
equation.)
Consider the graph (Figure 4a) whose nodes are dimensions of feature space F and whose
edges are the pairwise Riemannian distances between those dimensions as defined by the
distance matrix of equation (2.6) in section IIIc. The distance matrix can be treated as the
adjacency matrix
∑ for
!A with row indices i = 1,…,m and column indices j = 1,…,n. The graph Laplacian is
The total energy of the system can be expressed in terms of the Hamiltonian H, taking the
form H = L + P where L is the Laplacian and P (corresponding to potential energy) can be
neglected as a constant for the present case; the hamiltonian is thus equivalent for this
purpose to the laplacian:
(2.13)
Intuitively, the Hamiltonian expresses the tradeoffs among different possible states of the
system (Figure 4); applied to images, the Hamiltonian can be measured for its errors
(distance from the original) on one hand, and its entropy or compactness on the other: a
more compact state (lower entropy) will be less exact (higher error), and vice versa.
The aim is to identify an operator that begins with a point in feature space (an image) and
moves it to another point such that the changes in error and entropy can be directly
measured not in feature space but in perceptual space (Fig 3). Thus the desired operator
will move the image from its initial state (with zero "error," since it is the original image,
and an initial entropy value corresponding to the information in the image state) to a new
state with a new tradeoff between the now-increased error and corresponding entropy
decrease.
The Hamiltonian enables formulation of such an operator. The eigenvectors of the
Hamiltonian (Figure 4c) constitute a candidate basis set for the image vector (Figure 4d),
and since HΩ = λΩ, the eigenvalues λ of the Hamiltonian can provide an operator U(s)
corresponding to any given desired entropy s (see Supplemental section §2.3). As we will
Page 10 of 23
Dii =
∂2Ω(x,s)
∂x 2
H =
!D −
2Lg D 1
∑
i
!γi(x)
ci
ωi(x) =
Ω(x,s) =ωi(x)φi(s) =
Ω(x,s) =ω(x)φ(s)
φi(s) = eλis
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
see, the φi(s) function is the desired update operator, moving the point corresponding to
the state (the image) to a higher-error location in perceptual space to achieve the decreased
entropy level corresponding to s.
The separated components of the state equation (2.8) form the position solution and
entropy solution to the equation, respectively.
The position portion, ω(x), was shown in equation (2.12) to be
and the entropy portion, φ(s), was shown in equation (2.10) to be
The former expresses the set of positional configurations for each given solution and the
latter provides the foundation for the update operator for states i, to achieve entropy level s,
where the λ values correspond to the eigenvalues of the Hamiltonian.
Combining the terms, we obtain
(2.14)
Γ composed of the column
To put these operations in matrix form, we define the matrix
vectors
!γi(x), i.e., the eigenvectors of equation (2.12). We define the final form of the
update operator, U(s), to be the matrix composed of column vectors
φi(s)
φi(s). (Each
has only a single non-zero entry, in the vector location indexed by i, and thus U(s) is a
diagonal matrix).
The transformation steps for altering an image to a degraded image with lowered entropy
and increased error, then, begins with the image vector (
f ), and projects that vector into
the perceptual space defined by the eigenvector basis from equation (2.12), such that
(2.15)
f
The vector
forms the initial conditions of the original image, transformed into
Γ matrix, composed of the
!γi eigenvectors from equation (2.12)
perceptual space (by the
Γ). The values fi of
constitute the values of the ci coefficients that
as the columns of
will be used in equation (2.14).
Having transformed the vector into perceptual space, the update operator is then applied
(2.16)
f
The initial image now has been moved into perceptual space (
′f ), and moved within
), with a corresponding
that space to a point corresponding to entropy level s (
increase in error (which will be measured).
Page 11 of 23
!γi(x)!
φi(s)
ci
!
!
f →
!
′′f
!
f = U(s)⋅
′′
!
′f = U(s)⋅
!
Γ ⋅
!
!
f = Γ ⋅
′
!
!
!
!
′f →
!
f
′
!
!
!
′f
∑
i
!
!
!
′′f
!
!
Γ ⋅
!
f = round(
′′′
!
f
) = round( U(s)⋅
′′
Update operator moves image to lower entropy state and minimizes error increase
Rodriguez & Granger
Differential geometry of perceptual similarity
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
The lower-entropy image,
, thus has been scaled such that it now can be encoded via
rounding into a more compact version:
f )
(2.17)
Any subsequent encoding step may then be applied, such as Huffman or arithmetic coding,
operating on the rounded result. These are equivalently applicable to any other method
(JPEG, RGPEG, or other) of arriving at a transformed image, and are thus irrelevant to the
present formulation. We instead focus on the direct measures of error and of entropy. We
proceed to compare these measures directly for JPEG and for the newly introduced RGPEG.
The image
f now has been moved from feature space to the perceptual space defined by
the eigenvector basis of equation (2.12), as in Figure 4c, selecting a quality level (see
Supplemental section §2.3), applying the appropriate update operator, and rounding,
resulting in equation (2.17).
As described in section IIId, these computations depended on construction of a Jacobian
either via knowledge of (or estimated approximation of) the anatomical paths from input to
percept (synaptic Jacobian), or via empirical psychophysical measures (psychophysical
Jacobian).
We carried out several instances of computed compression via an estimated synaptic
Jacobian, composed by measuring distances between pixels on a screen image, measuring
viewing distance from the screen, converting these to viewing angle, and measuring all
pixels in terms of viewing angles and the distances among them (Supplemental section §2.5,
and supplemental table §1). Examples of computed Hamiltonians and eigenvector bases are
shown in Figure 4e and 4g for a particular empirical pixel size and viewing distance
(Supplemental section §2.5); the formulae show how any empirically measured features
give rise to a corresponding Hamiltonian. A set of several additional sample Hamiltonians
and eigenvector bases are shown in Supplemental figures §9-§13.
In sum, JPEG assumes its basis vectors (discrete cosine transforms) to be orthogonal, which
they are in feature (Euclidean) space, but not in perceptual (Riemannian) space. As shown,
the perceptual non-zero distances among basis dimensions can be either empirically
ascertained via psychophysical similarity experiments, as in the psychophysical-jacobian
method, or assumed on the basis of presumptive measures of anatomical distances (or
approximations thereof) as in the synaptic-jacobian method, or calculated on the basis of
physically measured distances in the physical space, as in the physical-distance-jacobian
method (see Supplemental section §2.5). In the present paper we have predominantly
tested the estimated psychophysical jacobian method (method ii b above), which (perhaps
surprisingly) is shown, by itself, to outperform JPEG. From these methods, we derived
Hamiltonians from the image space, and eigenvector bases from the Hamiltonians, and
showed that the JPEG DCT basis was a special case with particular settings shown in
Supplemental figure §12.
Page 12 of 23
Side by side comparison of JPEG / RGPEG
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
Performing compression with multiple sets of parameters (see Supplemental Figures §15-
§24) yielded empirical results enabling comparisons of the error and entropy measures for
the JPEG method and the method (RGPEG) derived from the Riemannian geometric
principles described herein. We have shown that for specific assumptions of geometric
distance and of perceived intensity difference, the JPEG method occurs as a special case of
the general RGPEG principles (Supplemental section §2.11.1). It is intriguing to note that,
using simple estimations of geometric distance and log scale intensity differences, the
generalized RGPEG method typically outperforms the JPEG special case, as expected; Figure
5 shows one such detailed side by side comparison; many more are shown in Supplemental
figures §15-§24). It also is notable that the computational space and time costs for the
RGPEG method are identical to those for JPEG (Supplemental section §2.12).
Figure 5 shows a range of compressed versions of a sample image (from the Caltech256
dataset), along with the measures of error (er) and entropy (en) for each image. The
method can most clearly be seen to produce fewer artifacts when compared at relatively
high compression levels (high entropy and high error); these are clear to qualitative visual
inspection; the figure also shows quantitative plots of the tradeoffs of values among error
and entropy for a set of selected quality levels. Across a range of quality settings, the error
and entropy values for RGPEG outperform those for JPEG.
Of primary interest is not the fact that JPEG compression can readily be outperformed by
the generalized RGPEG method; rather, the reason for the outperformance is that RGPEG
embodies a novel set of principles of perceptual similarity, and that these principles have
explanatory power for the set of perceptual phenomena described (of which JPEG
compression is one instance). We briefly discuss these explanatory principles.
Standard distance measures (Euclidean, Mahalanobis, etc.) (29) do not match human
similarity and dissimilarity judgments (e.g., Section IIIc above). To address this, some
standard approaches "re-code" the stimuli to more accurately reflect typical subjects'
reported perceived similarity or dissimilarity among stimuli (30, 31). Yet different
individual perceivers can differently register dissimilarity among identical physical stimuli,
such as the incompatible similarity judgments of speech sounds by native speakers of
different languages (4, 6). The solution is not to re-code the stimuli, but rather to separately
represent physical stimuli (e.g., speech sounds) on one hand, and the particular perceptual
mappings of those stimuli on the other, via a metric operation that transforms distances
from the reference frame of the physical stimulus space into distances in any given
perceiver's perceptual reference frame (Section IIId).
Euclidean vector distances assume orthogonality of constituent vector dimensions. This
could in theory hold but it is in general not the case for perceptual stimuli. The constituent
dimensions of a vector do not distinguish between "nearby" or "distant" dimensions, but
human perceptual judgments typically do. Riemannian space can intuitively be thought of
as having "curved" axes (relative to a tangent space) such that some regions of a given axis
are "closer" to some axes and farther from others, quite distinct from Euclidean space. The
tools from differential geometry presented here enable stimuli in Euclidean feature space to
be mapped to physical and perceptual spaces; we forward the principle that these mappings
Page 13 of 23
Physical stimulus similarity is distinct from perceptual stimulus similarity.
Discussion: derivation of principles
Perceptual distances are intrinsically Riemannian.
Further principles arise from study of perceptual transforms.
Application to a well-studied perceptual anomaly.
Synaptic plasticity changes the curvature of perceptual space.
Perceptual mappings arise directly from anatomical structure and physiological operation.
Rodriguez & Granger
Differential geometry of perceptual similarity
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
underlie judgments of perceptual similarity. This paper focuses on examples in visual
domains; additional extensions to auditory stimuli, and to abstract concept categorization
are separate findings being pursued.
a) A perceptual system cannot "neutrally" process stimuli; any system contains intrinsic
assumptions about the relations that occur among the components of any stimuli. A
perceptual system connectome encodes a Jacobian either with or without off-diagonal
entries, causing it to treat stimulus components (e.g., neighboring pixels in an image) as
dependent or independent, respectively, and the nature of any off-diagonal entries
determines the exact dependency relations among the components, corresponding to the
specific curvature of the metric perceptual space).
b) Cortical neuron receptive fields are often characterized in terms of Gaussians (22-25, 32,
33). Such components produce outputs that compute the partial derivatives of their inputs
in just the form needed for the Jacobian and tensor computations posited here; i.e., typical
neural assemblies appear tailored to computing transforms into Riemannian target spaces.
Re-shaping neurons' receptive fields via synaptic modification directly changes the Jacobian
mapping and the curvature of the target space. Every synaptic "learning rule" corresponds
to a mechanism by which existing metric transforms (arising from the connectome) are
modified in response to stimuli. All learning rules can be cast in terms of changing
curvature of the projection from input to perceptual space.
Connectomes are almost entirely unmapped in sufficient detail to construct a Jacobian, and
in any event perceptual spaces are formed by a combination of successive feedforward
stages as well as feedback top-down influences. A given perceiver's perceptual space may
nonetheless be elicited empirically by psychophysical measures (section IIId).
Unsupervised learning rules can readily educe statistical distribution characteristics of data,
and typically are judged by measures such as within-category vs. between-category
distances (34-36). But the discovery of unsupervised structure is not neutral with respect
to metric spaces: in response to a given set of data, different rules cause different changes to
the Jacobian, discovering different structure in the data (illustrated in the special case of
JPEG encoding, but broadly applicable to learning structure in data). Recent neural net
approaches have identified learning methods that can outperform JPEG; the present work,
by contrast, outperforms JPEG with no training and no statistics, by identifying previously
unnoted fundamentals of perceptual encoding that underlie similarity judgments.
In the psychophysical Jacobian method (Section IIId), for instance, perceptual distances
arise from the minimum distance within the target Riemannian space, i.e., the geodesic. It
could have been the case that other distances might instead have been involved. We
forward the principle that perceived distances are predicted by measures of Riemannian
minimum distance. Other underlying principles may similarly emerge from further study.
Page 14 of 23
Transforms can be computed from observed behavior.
Machine learning is based on the same geometric principles.
Differential geometry of perceptual similarity
Rodriguez & Granger
[arXiv:1708.00138v1]
Brain Engineering Lab Tech Report 2017.2
Tversky and colleagues (1, 2) showed that perceived similarity judgments of some classes of
stimuli violated the triangle inequality: even though stimuli A and B may physically share
more features than A and C, the latter may be judged more similar than the former. The
present studies suggest that subjects in these experiments are perceiving the stimuli in a
Riemannian space (Figure 6), in which a seemingly-direct path from one point to another
may entail proceeding via curved Riemannian coördinates, making that (perceived) path
longer than alternative paths.
In sum, the new formalism presented here is proposed as a general method for describing
and predicting perceptual and cognitive similarity judgments, as a complement to standard
vector distance metrics (Euclidean, Mahalanobis, etc.), which are applicable only to
measures in non-curved spaces. The results are equally applicable to visual, auditory, and
other modalities, as well as to abstract concept data.
At the core of the work are the twin principles that i) sensory stimuli (and arbitrary data)
may have internal Riemannian structure, i.e., dependence relations among their
(dimensional) component features; and ii) any system, natural or artificial, that processes
such data contains intrinsic assumptions or biases about the nature of those dependence
relations. Such a system may assume that input data are Euclidean and that their
components are thus independent, or the system may assume the presence of any of a very
wide variety of inter-component dependencies (such as neighbor or topography relations).
We formalize such premises, laying groundwork for extended study of natural perceptual
systems and of artificial algorithms for processing, representing, and identifying structure
in arbitrary data. Ongoing work is focused on extending the findings to domains beyond
vision, with the aim of identifying additional useful applications as well as identifying
further fundamental principles of representation.
The authors gratefully acknowledge helpful discussions with Eli Bowen. This research was
supported in part by grant N00014-15-1-2132 from the Office of Naval Research and grant
N000140-15-1-2823 from the Defense Advanced Research Projects Agency.
Acknowledgements
Page 15 of 23
Rodriguez & Granger
[arXiv:1708.00138v1]
!a
!b
!c
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
d(!a, !b):dE =1.4;dH = 2;dR = 4.1
d(!a, !c):dE =1.4;dH = 2;dR = 4.6
d(!b, !c):dE =1.4;dH = 2;dR =1.3
!
!
!
!
!a,
Figure 1. Illustration of the Riemannian nature of perceptual similarity. (Top) The
b, !c) are
transposes of three vectors (1 0 0 0 0 0), (0 0 0 0 1 0), and (0 0 0 0 0 1) (
rendered as images with empty space for zeros and dark spots for ones. The Euclidean
pairwise distances between any two of !a,
b, and !c are equal (distances of 2). Their
Hamming distances also are equal (distances of 2). If we measure the distances between
the dark spots, the answers (in mm) come out to be similar from !a to
b and from !a to !c,
but quite different (much smaller) from
b to !c. This "ruler distance" matches the evoked
perceptual similarity judgments empirically elicited from human viewers: all judge
b and
!c to be more similar than either is to !a. (Bottom left) The 64 vectors of the two
dimensional discrete cosine transform form an orthogonal basis in Euclidean space; they
are equidistant from each other. Perceptual similarity judgments between them, however,
exhibit wide variations; some are judged far more similar to each other than others by
human perceivers. (Bottom right) Taking just the first and 64th DCT entries (upper left and
lower right corners of the DCT, respectively) as an example, when viewed with unit
coefficients (as on the left), they are judged quite distinct; however, when viewed with
intermediate coefficients they are judged to be somewhat similar (right side). Thus the
perceptual metric being used by human viewers apparently is not uniform across this basis
space. Thus not only is the space non-Euclidean, it also is non-affine. Throughout this
paper, we assume full Riemannian curvature in this basis space.
!
Page 16 of 23
Rodriguez & Granger
[arXiv:1708.00138v1]
a)
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
c)
b)
n′′
n′
n
n′
n′′
n′′
n′
n
n′
n′′
d)
J(x) ≡
!x
∂y1
∂x1
∂y2
∂x1
"
∂yq
∂x1
input
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
!y
!
output
∂y1
∂y1
∂x p
∂x2
∂y2
∂y2
∂x p
∂x2
" ! "
∂yq
∂yq
∂x2
∂x p
!
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
e)
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
n
′n
′′n
0
0
0
0
0
′n
n
′n
′′n
0
0
0
0
′′n
′n
n
′n
′′n
0
0
0
receptive fields
(Gaussian)
0
0
⎤
⎥
0
0
⎥
0
0
⎥
0
⎥
′′n
⎥
0
′n
⎥
n
′′n
⎥
′n
′n
⎥
⎥
′′n
n
⎦
0
0
′′n
′n
n
′n
′′n
0
0
0
0
0
′′n
′n
n
′n
0
′′n
′n
n
′n
′′n
0
0
!
Figure 2. Brain connectomes are Riemannian. a) Simple example of anatomical
projections between two regions. b) Simple vector encoding of an anatomical projection
with synaptic weights. c) Examples of physiological neural responses in early visual areas
(gaussians). d) A Jacobian matrix denoting the overall effect of activity in the neurons of an
input area (x) on the neurons in a target area (f); each entry denotes the change in an
element of f as consequence of a given change in an element of x. e) Example instance of
such a Jacobian, corresponding to the synaptic connection pattern in part (b).
Page 17 of 23
Rodriguez & Granger
[arXiv:1708.00138v1]
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
µΦ→Ψ
µF→Φ
feature
space
(Z64)
F
physical
space Φ
(R3)
Ψ
perceptual
space
(Z3)
Figure 3. The map from physical to perceptual space. The three relevant projection
spaces for image compression. For all the examples in this paper, we adopt the JPEG
assumption of an 8x8 pixel image. The image consists of a set of intensity settings for each
pixel at a given x and y coordinate; this corresponds to Euclidean "physical space" Φ.
Images are mapped into feature space, listing the 8x8 pixels as a 64-dimensional vector
with integer intensity values from -255 to +255. Human judgments of the similarity of two
images (such as an original and a compressed image) correspond to a distinct (Riemannian)
space accounting for geometric neighbor relations among the pixels (absent from feature
space representation), along with just-noticeable differences (JND) of intensity values at
any given pixel. The mapping functions (µ) map from feature to physical space ( F → Φ)
and from physical to perceptual space ( Φ → Ψ) as shown.
Page 18 of 23
Rodriguez & Granger
[arXiv:1708.00138v1]
a)
...
...
...
13 16
61
11 10 16
16
55
12 12 14 19
56
24
14
62
14 17
22 29
77
18
22 37 56
24 35 55 64
92
49 64 78 87 103 121 120 101
99
72 92 95 98 112 100 103
51
40
60
58
69
57
87
80
109 103
104 113
24
26
40
51
68
81
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
d)
e)
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
f)
g)
35
44
90
40
45
97
42
46
97
42
46
99
40
45
93
35
44
93
15
43
43
78
90
99
101 101 102 102 103 107 107 110
111 111 114 114 115 115 116 116
117 117 117 117 119 120 120 121
121 122 122 125 125 125 126 126
128 128 129 129 130 132 132 133
⎡
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎢
⎣
⎤
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎥
⎦
Figure 4. Treatment of image as graph, and derivation of Hamiltonian. (a) Basis
vectors in feature space F treated as a graph with whose nodes are the dimensions of the
basis and whose edges are the pairwise distances between dimensions (see Eq (2.6)). From
that graph, the adjacency and degree matrices, and thus the graph Laplacian, can be directly
computed. (b) Q matrix for JPEG (quality level 50%). (c) Computed Q matrix for RGPEG.
(d) Hamiltonian for JPEG. (e) Hamiltonian for RGPEG (see Supplemental section §2.7, table
§5. (f) Eigenvectors of Hamiltonian for JPEG. (g) Eigenvectors of Hamiltonian for RGPEG.
(See Supplemental sections §2.7-2.11).
b)
G
E
P
J
Q
c)
G
E
QRGPEG =
P
G
R
Q
Page 19 of 23
Rodriguez & Granger
[arXiv:1708.00138v1]
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
Figure 5. Side by side comparison of JPEG (J) and RGPEG (R) compression on a
sample image. (For more instances see Supplemental figures §15-§24.) (Top) Examples
of images (alongside corresponding computed Jacobians) for given values of desired quality
(and corresponding Q matrices), at quality levels 30, 50, 60, and 80, for JPEG (J) and RGPEG
(R). For each image, the computed error (er) and entropy (en) are given below the image.
For comparable error measures, the entropy for RGPEG is consistently lower than for JPEG.
(Bottom left) Receiver operating characteristic for entropy-error tradeoff for JPEG (boxes)
and RGPEG (circles). At comparable entropy values, RGPEG error values are consistently
equivalent or smaller. (Bottom right) Sample measures of entropy (blue) and error
(purple) for JPEG (dotted) and RGPEG (solid) at distinct quality settings. (All images from
Caltech-256 (37).
Page 20 of 23
Rodriguez & Granger
[arXiv:1708.00138v1]
t i m u l u
s
e
c
a
p
s
t
u
p
i n
s
A
d3
C
d1
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
d3 < d1 + d2
d2
d3′
B
perceptual
mapping
d3′ > d1′ + d2′
d2′
B
c e
s p a
c e p t u a l
p e r
d1′
A
C
Figure 6. Interpretation of the triangle inequality violation (initially described by Tversky
and Gati 1982). In a physical stimulus, the distance from A to B is less than the combined
distances from A to C to B, i.e., d3 ≤ d1 + d2, obeying the triangle inequality in the stimulus
input space. A perceiver, however, measures those distances not in the input space but in
her own perceptual reference frame, which is a Riemannian space (see text). The curvature
of that space may render different geodesic distances; specifically, the geodesic from A to B
may be longer than the geodesic from A to C to B; thus ′d3 > ′d1 + ′d2, violating the triangle
inequality in perceptual space.
Page 21 of 23
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
References
Rodriguez & Granger
[arXiv:1708.00138v1]
1.
Tversky A & Gati I (1982) Similarity, separability, and the triangle inequality.
Psychological Review 89:123-154.
2.
Tversky A (1977) Features of similarity. Psychological Review 84:327-352.
3.
Logan J, Lively S, & Pisoni D (1991) Training Japanese listeners to identify English
/r/ and /l/: a first report. J Acoust Soc Am. 89:874-886.
4.
Kuhl P (2004) Early language acquisition: cracking the speech code. Nat Rev
Neurosci 5:831-843.
5.
Guion S, Flege J, Akahane-Yamada R, & Pruitt J (2000) An investigation of current
models of second language speech perception: the case of Japanese adults'
perception of English consonants. . J Acoust Soc Am. 107:2711-2724.
6.
Iverson P, et al. (2003) A perceptual inference account of acquisition difficulties for
non-native phonemes. . Cognition 87:B47-B57.
7.
McCandliss B, Fiez J, Protopapas A, Conway M, & McClelland J (2002) Success and
failure in teaching the /r/ - /l/ contrast to Japanese adults: tests of a Hebbian model
of plasticity and stabilization in spoken language perception. Cogn Affect Behav
Neurosci 2:89-108.
8.
Nosofsky R (1992) Similarity scaling and cognitive process models. Ann Rev Psychol
43:25-53.
9.
Nosofsky R & Johansen M (2000) Exemplar-based accounts of "multiple system"
phenomena in perceptual categorization. Psychonomic Bulletin & Review 7:375-402.
10.
Ashby F & Perrin N (1988) Toward a unified theory of similarity and recognition.
Psychological Review 95:124-150.
11.
Toderici G, et al. (2016) Full resolution image compression with recurrent neural
networks. arXiv.
12.
Svoboda P, Hradis M, Barina D, & Zemcik P (2016) Compression artifacts removal
using convolutional neural networks. arXiv.
13. Wallace G (1992) The JPEG still picture compression standard. IEEE Trans.
Consumer Electronics 38(1).
14.
Pennebaker W & Mitchell J (1993) JPEG: Still image data compression standard (Van
Nostrand Reinhold).
Klein S, Silverstein D, & Carney T (1992) Relevance of human vision to jpeg-dct
15.
compression. SPIE/IS&T Symposium on Electronic Imaging 10.1117/12.135968:200-
215.
16.
kishan A, Lee C, & Winer J (2008) Branched projections in the auditory
thalamocortical and corticocortical systems. Neuroscience 154:283-293.
17.
Lee C, Schreiner C, Imaizumi K, & Winer J (2004) Tonotopic and heterotopic
projection systems in physiologically defined auditory cortex. Neuroscience
128:871-887.
Goldman-Rakic P (1988) Topography of cognition: parallel distributed networks in
18.
primate association cortex. Ann Rev Neurosci 11:137-156.
19. Winkowski D & Kanold P (2013) Laminar transformation of frequency organization
in auditory cortex. J Neurosci 33:1498-1508.
20. Watkins P, Kao J, & Kanold P (2014) Spatial pattern of intra-laminar connectivity in
supragranular mouse auditory cortex. Frontiers in Neural Circuits 8:1-18.
Eisenhart L (1997) Riemannian geometry (Princeton University Press).
21.
Olshausen B & Field D (1996) Emergence of simple-cell receptive field properties by
22.
learning a sparse code for natural images. Nature 381:607-609.
Page 22 of 23
Differential geometry of perceptual similarity
Brain Engineering Lab Tech Report 2017.2
Soodak R (1986) Two-dimensional modeling of visual receptive fields using
Gaussian subunits. Proceedings of the National Academy of Science 83:9259-9263.
Field D & Tolhurst D (1986) The structure and symmetry of simple-cell receptive-
field profiles in the cat's visual cortex. Proc Royal Acad Sci B 228:379-400.
Rao R & Ballard D (1999) Predictive coding in the visual cortex: a functional
interpretation of some extra-classical receptive-field effects. Nature Neurosci 2(79-
87).
Lindeberg A (2013) A computational theory of visual receptive fields. Biological
Cybernetics 107:589-635.
Perona P & Malik J (1990) Scale-space and edge detecion using anisotropic diffusion.
IEEE TPAMI 12:629-639.
Mumford D & Shah J (1989) Optimal approximations by piecewise smooth functions
and associated variational problems. Communications on Pure and Applied
Mathematics 42:577-685.
Duda R, Hart P, & Stork D (2012) Pattern classification (Wiley).
Shepard R (1962) The analysis of proximities: Multidimensional scaling with an
unknown distance function (II). Psychometrika 27:219-246.
Attneave F (1950) Dimensions of similarity. American Journal of Psychology 63:516-
556.
Brown E, Nguyen D, Frank L, Wilson M, & Solo V (2001) An analysis of neural
receptive field plasticity by point process adaptive filtering. Proceedings of the
National Academy of Science 98:12261-12266.
Fritz J, Shamma S, Elhilali M, & Klein D (2003) Rapid task-related plasticity of
spectrotemporal receptive fields in primary auditory cortex. Nature Neurosci
6:1216-1223.
Smith E, Patalano A, & Jonides J (1998) Alternative strategies of categorization.
Cognition 65:167-196.
Everitt B (1993) Cluster Analysis (E Arnold).
Jain A (2010) Data clustering: 50 years beyond k-means. Pattern Recog Lett 31:651-
666.
Griffin G, Holub A, & Perona P (The Caltech 256) . (Caltech Tech Report).
Rodriguez & Granger
[arXiv:1708.00138v1]
23.
24.
25.
26.
27.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
Page 23 of 23
|
1803.10753 | 1 | 1803 | 2018-03-28T17:44:44 | The comparison of Higuchi fractal dimension and Sample Entropy analysis of sEMG: effects of muscle contraction intensity and TMS | [
"q-bio.NC"
] | The aim of the study was to examine how the complexity of surface electromyogram (sEMG) signal, estimated by Higuchi fractal dimension (HFD) and Sample Entropy (SampEn), change depending on muscle contraction intensity and external perturbation of the corticospinal activity during muscle contraction induced by single-pulse Transcranial Magnetic Stimulation (spTMS). HFD and SampEn were computed from sEMG signal recorded at three various levels of voluntary contraction before and after spTMS. After spTMS, both HFD and SampEn decreased at medium compared to the mild contraction. SampEn increased, while HFD did not change significantly at strong compared to medium contraction. spTMS significantly decreased both parameters at all contraction levels. When same parameters were computed from the mathematically generated sine-wave calibration curves, the results show that SampEn has better accuracy at lower (0-40 Hz) and HFD at higher (60-120 Hz) frequencies. Changes in the sEMG complexity associated with increased muscle contraction intensity cannot be accurately depicted by a single complexity measure. Examination of sEMG should entail both SampEn and HFD as they provide complementary information about different frequency components of sEMG. Further studies are needed to explain the implication of changes in nonlinear parameters and their relation to underlying sEMG physiological processes. | q-bio.NC | q-bio | The comparison of Higuchi's fractal dimension and Sample Entropy
analysis of sEMG: effects of muscle contraction intensity and TMS
Milena B. Čukić1,*, PhD, Mirjana M. Platiša2, PhD, Aleksandar Kalauzi3, PhD, Joji Oommen4,
MS, Miloš R. Ljubisavljević4, MD, PhD
1Department for Physiology with Biophysics, School of Biology, University of Belgrade,
Belgrade, Serbia
2Institute of Biophysics, School of Medicine, University of Belgrade, Belgrade, Serbia
3Department for Life Sciences, Institute for Multidisciplinary Research, University of Belgrade,
Belgrade, Serbia
4Department for Physiology, College of Medicine and Health Sciences, UAE University, Al Ain,
UAE
Keywords: Electromyogram (EMG); Sample Entropy; Higuchi's fractal dimension; Transcranial
Magnetic Stimulation; Non-linear analysis
Abstract
The aim of the study was to examine how the complexity of surface electromyogram
(sEMG) signal, estimated by Higuchi's fractal dimension (HFD) and Sample Entropy (SampEn),
change depending on muscle contraction intensity and external perturbation of the corticospinal
activity during muscle contraction induced by single-pulse Transcranial Magnetic Stimulation
(spTMS). HFD and SampEn were computed from sEMG signal recorded at three various levels of
voluntary contraction before and after spTMS. After spTMS, both HFD and SampEn decreased at
medium compared to the mild contraction. SampEn increased, while HFD did not change
significantly at strong compared to medium contraction. spTMS significantly decreased both
parameters at all contraction levels. When same parameters were computed from the
mathematically generated sine-wave calibration curves, the results show that SampEn has better
accuracy at lower (0-40 Hz) and HFD at higher (60-120 Hz) frequencies. Changes in the sEMG
complexity associated with increased muscle contraction intensity cannot be accurately depicted
by a single complexity measure. Examination of sEMG should entail both SampEn and HFD as
they provide complementary information about different frequency components of sEMG. Further
studies are needed to explain the implication of changes in nonlinear parameters and their relation
to underlying sEMG physiological processes.
Introduction
Surface EMG (sEMG) is a record of electrical activity of underlying muscle fibers. It is a
complex, nonlinear and non-stationary signal, influenced by factors like neuron discharge rates,
recruitment patterns of motor units, muscle architecture, as well as various other factors
(Nieminen and Takala, 1996). The analysis of sEMG has extended beyond the traditional
diagnostic applications to also include applications in diverse areas such as biomedical,
prosthesis or rehabilitation devices, human machine interfaces, and more. Traditionally the
sEMG analysis, particularly in the medical field, is dominated by spectral or amplitude measures.
Nevertheless, it was repeatedly shown that nonlinear analyzes of sEMG can provide additional
information on the underlying motor strategies (Del Santo et al., 2007; Farina et al., 2002),
hidden rhythms (Filligoi and Felici, 1999), fatigue (Ikegawa et al., 2000) as well as detection of
pathological changes in the system (Meigal et al., 2012, 2009). Furthermore, several studies
showed that nonlinear methods are potentially more sensitive than classical sEMG analysis
methods, being able to capture very subtle changes in the signal under study. For example,
fractal analysis (Mesin et al., 2009; Ravier et al., 2005) was used to show the relationship
between muscle force changes during contraction and the complexity of the sEMG signal (Gitter
and Czerniecki, 1995; Gupta et al., 1997) and to detect the lowest level of the voluntary
activation of the muscle under study (Arjunan and Kumar, 2010). Nonlinear methods have also
proven useful in distinguishing between sEMG of patients with Parkinson's disease and healthy
controls (Meigal et al., 2009), potentially providing useful biomarkers of system dysfunction in
aging and disease.
Transcranial Magnetic Stimulation (TMS) is a noninvasive method used to stimulate brain
cortex. When applied over motor cortex it induces synchronous activation of corticospinal
neurons, which in sEMG evokes characteristic activation, termed Motor Evoked Potential (MEP).
After MEP a temporary silence of sEMG activity called silent period (SP) occur. It was accepted
that the effects of this single-pulse TMS (spTMS) of the motor cortex do not induce lasting changes
in cortical activity beyond those immediate ones. However, our recent study showed that the
complexity of the sEMG signal after single-pulse TMS decreases suggesting that the overall
corticospinal activity after spTMS becomes less complex (Cukic et al., 2013). The study used
Higuchi's fractal dimension (HFD) (Higuchi, 1988) which serves as a measure of signal
complexity. Fractal dimension (FD) refers to a non-integer or fractional dimension of a geometric
object; it could be used for phase space approach to estimate the FD of an attractor in the state-
space domain. Applications of HFD within this framework pertain primarily to time domain since
the signal itself is considered as a geometric figure (Kalauzi et al., 2012). However, it has been
suggested that nonlinear measures can overestimate or underestimate subtle changes in the signal
complexity with different algorithms yielding different results for a single nonlinear measure
(Ferenets et al., 2006; Goldberger et al., 2002; Mesin et al., 2009).
Sample entropy (SampEn), another nonlinear method, was first introduced as a 'regularity'
statistics (Richman and Moorman, 2000) rather than a direct index of physiological complexity.
SampEn quantifies the probability that sequences of patterns in a dataset that are initially closely
related remain close on the next incremental comparison, within a specified tolerance. Thus,
SampEn appears to be a potentially useful method to unravel the complexity of the sEMG signal
and its relation to underlying physiological processes (Cashaback et al., 2013; Lake et al., 2002;
Molina-Picó et al., 2011; Ruonala et al., 2014). Very few studies explored the use of SampEn to
analyze EMG signal. For example, the complexity of biceps sEMG was shown to exhibit a
complex relation to contraction intensity (Cashaback et al., 2013). Muscle fatigue decreased the
complexity of sEMG towards the end of fatiguing-exhausting contraction (Cashaback et al., 2013),
as well as the complexity of submaximal and maximal voluntary contraction (Pethick et al., 2015).
Furthermore, various nonlinear complexity parameters were shown to be significantly less variable
in differentiating non-fatiguing and fatiguing muscle contraction (Karthick et al., 2014), making
them potentially useful for automated analysis of neuromuscular activity in normal and
pathological conditions. Nevertheless, the mechanisms influencing sEMG complexity are still
poorly understood. Therefore, to further explore the use of non-linear methods in sEMG analysis,
potentially broadening their clinical applications, in this study, we compared SampEn with HFD
using the same interference spTMS paradigm. We examined the changes in sEMG complexity by
SampEn during voluntary contraction of different intensities before and after application of
spTMS. Finally, to elucidate the difference in obtained results we compared the results of SampEn
and HFD analysis of theoretical mathematically constructed calibration curves (Kalauzi et al.,
2012).
Materials and Methods
Participants
Surface electromyogram (EMG) was recorded from the first dorsal interosseous muscle
(FDI) of each participant. The sample comprised of ten participants, five women, five men (age
22-48 +/- 6.7 years). All participants were healthy volunteers, without a prior history of
neuromuscular disorders.They were all righthanded, according to Edinburgh Inventory (Oldfield,
1971). Participants gave their written informed consent before the experimental procedure. The
study was approved by the Al Ain Medical District Human Research Ethics Committee (Protocol
No. 12/44) and performed in accordance with the ethical standards laid down in the Declaration of
Helsinki.
Experimental protocol and conditions
Participants were comfortably seated in an armchair, resting the right hand on a handhold.
They were asked to exert voluntary FDI contraction by abducting the index finger against elastic
resistant adjusted for each subject and contraction level. The intensity of contraction was
provisionally expressed as a percentage of maximal voluntary contraction (MVC) and scaled as
follows: mild (10-20%), medium (20-40%) and strong (40-70%). The contraction intensity was
randomly varied. Subjects received up to 20 TMS pulses at each intensity, out of which 15 were
used for the analysis. Sufficient time was given after each muscle contraction-spTMS trial, with
longer time availed after strong contractions to avoid development of muscle fatigue.
Subsequently, individual trials were grouped based on contraction intensity for further analysis.
Transcranial Magnetic Stimulation and sEMG
Transcranial magnetic stimulation (MagPro, R100, MagVenture, Denmark) was performed
using with a figure-of-eight coil optimally positioned over the left hemisphere to evoke MEPs in
the right FDI muscle (45 degrees to the central line). The optimal stimulation spot was marked
with a semi-permanent marker to allow maintenance of stable stimulation coil position during the
experiment. The resting motor threshold (RMT) was determined to the nearest 1% of the stimulator
output and was the minimum intensity required to evoke MEPs of 50 μV in five out of ten
consecutive trials (Rossini et al., 1994). The mean RMT for all subjects was 46.3 ± 8.6% of the
maximal stimulator output. Subsequently, spTMS stimulus intensity was set at 1.3 above resting
motor threshold.
Ag-AgCl electrodes were used to record the surface EMG from the right FDI muscle
(electrode diameter 9 mm). The raw EMG signal was amplified and filtered with the band-pass
filter in the range of 20 Hz – 1 KHz (CED 1902 isolated pre-amplifier, Cambridge Electronic
Design, UK). Each recording was 8-10 s long with spTMS delivered approximately 4 s after the
contraction onset. sEMG signals were digitized with the sampling rate of 1 KHz (CED 1401,
Cambridge Electronic Design, Cambridge, UK) and stored for further off-line analysis. During the
experiment, a Root Mean Square of the sEMG signal was computed and together with an audio
signal shown to subjects.
Data analysis
From the recorded sEMG two epochs were selected for analysis: PRE TMS activity starting
from the onset of the stable contraction up to approximately 10ms before TMS artifact (see Figure
1) and POST TMS activity, starting from the onset of uninterrupted sEMG (ignoring occasional
mid-SP EMG bursts) after the SP. The beginning of post-epoch was estimated visually by the same
examiner, a method shown to yield consistent detection not different from automated routines
(Julkunen et al., 2013). Irrespective of individual variations in SP duration and time needed to
develop stable contraction before spTMS, the length of both epochs used for analysis was set to
3.5 s. The epoch length did not differ more than 5 ms, which could not influence the analysis. In
total, approximately 900 individual epochs (15 PRE and 15 POST-TMS, from 10 subjects, and
three levels of contraction) were analyzed, and the output was used for the statistical analysis.
Higuchi Fractal Dimension (HFD)
The fractal dimension of sEMG was calculated by using Higuchi's algorithm (Higuchi,
1988), as a measure of signal complexity in the time domain. Higuchi proposed an algorithm for
the estimation of fractal dimension directly in the time domain without reconstructing the strange
attractor. This method gives a reasonable estimate of the fractal dimension even in the case of short
signal segments and is computationally fast. EMG was analyzed in time, as a sequence of samples
x(1), x(2),..., x(N), and k new self-similar time series
were constructed as:
(1)
for m = 1, 2, ..., k where m is the initial time; k = 2, ...., kmax, where k is the time interval.
According to previous studies which dealt with the application of Higuchi's algorithm with varying
kmax (Spasic et al., 2005), for this type of signals, the best option is kmax= 8. Int[r] is the integer
part of a real number r. The length of every Lm(k) was calculated for each time series or curves
as:
(2)
has to be averaged for all m, therefore forming an average value of a curve length
L(k) for each k=2,...,
(3)
Fractal dimension was evaluated as the slope of the best-fit form of ln(L(k)) vs. ln(1/k):
FD=ln(L(k))/ln(1/k). (4)
Fractal dimensions were calculated separately for each epoch (PRE and POST TMS) using
Matlab 7.0, using the computation reported earlier by Kalauzi et al. (Spasic et al., 2005) (The Math
Works, Natick, Massachusetts, USA).
mkX)]/)int[(()....,2(),(),(:kkmNmxkmxkmxmxXmkmkXkkmNNkimxikmxkkLkmNim]int[1))1(()(1)(]int[1)(kLmmaxkkkLkLkmm1)()( Sample Entropy (SampEn)
Sample Entropy (SampEn) was computed according to the procedure published by
Richman and Moorman
(Richman and Moorman, 2000). Given a
finite sequence
; we constructed vectors of length m,
to
, defined as
,
(5)
Compute distance between
and
, denoted by d(yi,yj), a Chebyshev distance which
have to be r , as
d(yi, yj) = max
(6)
For
calculating the probability that any vector
that is similar to within
r as
(7)
Where ni (m,r) is the number of vectors
that are similar to
subject to the criterion of
similarity d (yi, yj) ≤ r . Calculate
(8)
SampEn
(9)
SampEn quantifies the irregularity of a time series and estimates the conditional probability
that two sequences of m consecutive data points, which are similar to each other (within given
tolerance r), will remain similar when one consecutive point is included. The SampEn algorithm
),...,,(21NNxxxx1ymNy],...,,[11miiiixxxymNi1iyjy,kjkixx10mkijmNi,...,1jyiy1,,mNrmnrmPiijyiymNiirmPmNrmA1,1),(rmArmArmxN,,1ln,, considers two parameters: tolerance level r and pattern length m. According to previous studies,
we chose a tolerance level of r = 0.15 times the standard deviation of the time series and m = 2
(Matlab 7.0). Also, the data were analyzed by another SampEn algorithm, in-house written in Java
programming language, confirming the initial results.
Calibration curves
To elucidate the difference in results between the two nonlinear parameters we used series
of surrogate mathematically generated sinusoids with frequencies ranging from 1 to 116 Hz (step
5Hz), while their amplitudes were kept constant. The calibration curves were then analyzed using
both algorithms (HFD and SampEn). The breaking point used to construct the calibration curves
was defined as fb =0.117 x fs (Kalauzi et al., 2012). Since the exact position of the calibration curve
depends on sinusoid's sampling frequency (Kalauzi et al., 2012), the sampling frequency of the
surrogates was set to 1 KHz, to match the actual sEMG sampling rate.
Statistical Analysis
The normality of the distribution of all data-sets was examined using Shapiro-Wilk test
(Pre and Post-TMS at three levels of contraction). None of the datasets had a normal distribution.
Wilcoxon non-parametric rank-sum test was used to compare HFD and SampEn (SPSS Statistical
Package for the Social Sciences, Chicago IL release 17.0). P<0.05 value was considered
statistically significant.
Results
Figure 1 shows the raw signal recorded at three different contraction levels form FDI
muscle.
Figure 1. Raw sEMG signal at three different contraction levels. Top panel (A) shows mild, middle panel
(B) shows medium, and the lower panel (C) shows strong contraction. Arrows indicate the beginning and
the end of the segment used for analysis PRE (left side of the recording) and POST TMS (right side of the
recording). The HFD of the same segments were: 1.0921/1.0914 (mild), 1.0495/1.047 (medium) and
1.0167/1.015 (strong). The SampEn values of the same segments were: 0.036877/0.035281 (mild),
0.071382/0.056896 (medium) and 0.121227/0.104262 (strong).
Figure 2 and 3 are showing changes in mean SampEn PRE and POST TMS at three
different levels of voluntary contraction.
Figure 2. (A) Changes in SampEn before and after spTMS. SampEn of sEMG time series before and after spTMS
at three levels of contraction. * p< 0.001 PRE vs. POST TMS. # p< 0.001 medium vs. mild contraction,
0.001 strong vs. medium contraction. (B) Changes in HFD before and after spTMS. HFD of sEMG time series
before and after spTMS at three levels of contraction. * p< 0.001 PRE vs. POST TMS, # p< 0.001 medium and
strong vs. mild contraction.
The range of calculated values of SampEn was 0.34 – 0.47 while HFD range was 1.072 –
1.091. SampEn significantly decreased between mild and medium contraction (both PRE and
POST comparison) (p<0.001), but then significantly increased between medium and strong
contraction level (p<0.001). There was no significant difference in SampEn between the mild and
strong level of contraction (p>0.05). SpTMS induced a significant decrease in SampEn (PRE vs.
POST comparison, Wilcoxon test) at all levels of muscle contraction (p<0.001). The results of
Higuchi's Fractal analysis of the same sEMG epochs previously analyzed by SampEn are shown
in Figure 2. HFD significantly decreased between mild and medium and mild and strong
contraction (both PRE and POST comparison) (p<0.001), but not between medium and strong
contraction (p>0.05). Similarly, to SampEn spTMS induced a significant decrease in HFD POST
compared with PRE at all contraction levels (p<0.001).Thus, both nonlinear methods show that
spTMS induced reduction of the complexity of the signal, irrespective of the contraction intensity
while they depicted different changes in complexity between various contraction levels.
To elucidate the difference between HFD and SampEn results of sEMG complexity, and
test whether it can be related to their differential sensitivity to the frequency content of the signal,
we analyzed mathematically constructed calibration curves (see method), which are shown in
Figure 3.
Figure 3. Comparison of FD and SampEn application on calibration curves. Calibration
curves, showing how Higuchi FD (A) and SampEn (B) depend on frequencies of surrogate sinusoids.
Based on earlier results (Kalauzi et al., 2012) the breaking point for construction of
theoretical calibration curves was set to fb=0.117 x fs , corresponding to the sampling frequency of
1 KHz. SampEn (fi)=φ2(Si(fi)) mapped more linearly in the range 0 < fi< 60 Hz (Fig. 4b), whereas
HFD(fi)=φ1(Si(fi)) mapped the values to more limited range (Fig. 4a). On the other hand, for
frequencies 60 < fi< 120 Hz SampEn values were mapped to a relatively narrow region
(0.8<SampEn<1), while HFD mapped input values to a region occupying approximately 4/5 of the
whole HFD range (1.2<HFD<2). This indicates that SampEn and HFD are sensitive to the
frequency content of the signal, showing different potential to detect changes in complexity, in a
different frequency range (SampEn for lower, HFD for higher frequencies), of a signal under study.
Discussion
The results of the study confirmed our earlier results that the complexity (HFD) of sEMG
decreases with increasing intensity of muscle contraction and is further decreased by spTMS.
However, SampEn showed different changes in sEMG complexity, as compared to HFD, while
similarly to HFD the complexity was further reduced by spTMS. To elucidate the difference in
results we compared the outcome of the analysis of two methods applied on simulated calibration
curves and showed that this difference appears to be related to their different sensitivity to the
frequency content of the sEMG signal within examined frequency range. The results should be
taken into consideration when these nonlinear methods are applied for sEMG analysis.
Our previous study that used HFD showed that the complexity of an sEMG signal
significantly decreases with the increase of muscle contraction intensity (Cukic et al., 2013). The
results of this study confirm these earlier findings. The reduction of sEMG complexity with
increased contraction intensity may appear counterintuitive when compared to changes in sEMG
Root Mean Square (RMS) values, characterized by linear relationship between the contraction
force and the RMS value. Accordingly, it may be assumed that the increase in force, associated
with the increase in muscle unit discharge rate and recruitment, would yield more complex
signal. Furthermore, the results of the current study appear at variance to some of the earlier
studies, which showed that fractal dimension of the EMG signal rises with the increase of muscle
force (Gupta et al., 1997), so that FD could even provide a reasonably good quantification of
contraction intensity (Anmuth et al., 1994; Arle and Simon, 1990; Glenny et al., 1991). As
argued earlier (Cukic et al., 2013) some of the differences between current and earlier studies
may be related to recruitment strategies deployed by CNS when varying muscle force production
in different muscles examined in these studies. Namely, it is well-established that muscle force
production is largely regulated by motor unit recruitment, (Milner-Brown et al., 1973) with
stronger intensities achieved by increased discharge rates (Kukulka and Clamann, 1981; Milner-
Brown et al., 1973). In different muscles, the majority of muscle units are recruited at various
levels of force production. In biceps brachii 95% of units are recruited at 70% of maximal force
production, whereas in FDI muscle, almost all units are recruited at a much lower intensity of
around ~ 30% of MVC, with further force increment being generated by frequency modulation
(Carpentier et al., 2001; Riley et al., 2008; Staudenmann et al., 2014). Thus, an increase in FD
with the rise of voluntary contraction could reach saturation at 70% of MVC (Carpentier et al.,
2001; Gitter and Czerniecki, 1995).
Unlike HFD, which decreased between 20% and 40% of MVC and did not further change
at 70% MVC, SampEn initially decreased, but then increased between 40% and 70% of the
maximal voluntary force. The relationship between muscle force and complexity measured by
SampEn was rarely addressed in the past (Cashaback et al., 2013; Zhang et al., 2016). Cashaback
et al., (2013) showed that short-term biceps brachii sEMG complexity was moderately
influenced by contraction intensity, while the long- term sEMG complexity did not reach
statistical significance. On the other hand Zhang et al. (2016) showed strong correlation between
SampEn values and the amplitude measurements of the surface EMG signal. At present, it is not
clear what may be the reason for this discrepancy although differences between muscles (biceps
brachii in Cashaback's study) and subjects (amputees in Zhnag's study) cannot be excluded. It
should also be noted that muscle fatigue also decreases the complexity of sEMG (Karthick et al.,
2014). Nevertheless, it does not seem likely that it played a role in this study as the contractions
were randomized and extra time was availed after each strong contraction to prevent the
development of fatigue.
To further examine the difference between HFD and SampEn in relation to muscle
contraction intensity we computed the complexity of the signal containing defined frequency
spectrum. The analysis of a series of surrogate mathematical sinusoids, Si(fi), with monotonously
increasing frequencies fi = 1, 2, ..., 120 Hz (Kalauzi et al., 2012) suggested that HFD and SampEn
are influenced by the frequency of the underlying signal. SampEn (fi)=φ2(Si(fi)) is more linear in
the range 0 < fi < 60 Hz (Fig. 4b), where HFD(fi)=φ1(Si(fi)) tends to map the values to a very
limited HFD range (Fig. 4a). On the other hand, for frequencies 60 < fi < 120 Hz SampEn values
are being mapped into a relatively narrow region (0.8<SampEn<1), while sensitivity of HFD is
increased, mapping its input values to a region occupying approximately 4/5 of the whole HFD
range (1.2<HFD<2). It is well established that the sEMG frequency spectrum changes with the
contraction although the major part of the spectrum are lower than 100Hz, (De Luca, 1984;
Knaflitz et al., 1990) other components (harmonics) representing higher frequencies have been
described during strong levels of contraction (Christensen et al., 1984; Timmer et al., 1998b).
Thus, the difference in detected changes between different intensities of contraction may be related
to the differential sensitivity of these two methods to prevailing frequency content of sEMG
associated with different contraction intensity. It should be stressed that the connection between
fractal dimension and spectral content of the signal, concerning the previous practice of application
of solely spectral measures in electrophysiology, was extensively investigated in the past, but none
provided the exact mathematical relationship (Weiss et al., 2011). A recent publication by Kalauzi,
(Kalauzi et al., 2012) showed the exponential dependence of fractal exponents on the frequency
and characterized the relation mathematically. This finding is important as it provides the
theoretical framework for the analysis applied in this study. It should also be noted that this allows
direct estimation of signal fractal dimension from its Fourier components and establishes that FD
does not depend on sinusoid's amplitude and initial phase, but on the frequency of the waveform
and sampling frequency.
Finally, the present results confirmed our earlier findings that the complexity of the sEMG
signal measured by HFD decreases after cortical spTMS irrespective of the intensity of muscle
contraction (Cukic et al., 2013). The present results extend them by demonstrating that an sEMG
complexity decreased after spTMS also when estimated by SampEn. As argued earlier the
reduction of sEMG complexity after spTMS suggest changes in corticospinal activity, most likely
due to transient TMS-induced synchronization of descending excitatory signal (Harris et al., 2008;
Marsden et al., 2000; Rosler et al., 2002; Timmer et al., 1998a). Namely, it seems as if spTMS has
interupted CNS regulatory mechanisms, making the system less variable and adaptable. It is
possible that CNS under these circumstances cannot precisely gauge the state of the corticospinal
networks causing the voluntary drive to overshoot, thus causing synchronization, after a sudden
externally caused brake of voluntary activity.
Until recently, HFD was thought to be the most sensitive measure for analyzing the
complexity of sEMG as suggested by results of analysis that compared the sensitivity of both
spectral and nonlinear measures applied on artificially generated EMG (Mesin et al., 2009).
However, it has repeatedly been demonstrated that the analysis of complex physiological signals
is best performed if different measures/algorithms are used (Ferenets et al., 2006; Kronholm et al.,
2007; Stam, 2005) since they may be sensitive to various features of the signal. Thus, the results
further reiterate the notion that multiple methods and algorithms should be used to survey the
complexity of the signal (Arle and Simon, 1990; Eke et al., 2002; Ravier et al., 2005).
Conclusions
The results of this study further support increasing body of evidence showing that
multiscale approach can quantify subtle information content in physiological time series. They
also confirm earlier results that spTMS decreases the complexity of sEMG beyond its immediate
electrophysiological effects. Importantly, the data show that SampEn and HFD have different
sensitivity in different frequency ranges, making them methodologically complementary for the
analysis of sEMG. Finally, based on current results, it could be argued that both methods should
be used to elucidate comprehensively changes in complexity of the sEMG signal and thus
corticospinal activity. Further studies are needed to explore the duration of spTMS influence on
changes in sEMG complexity during voluntary muscle contraction and at different levels of muscle
contraction and TMS intensity in healthy and diseased nervous systems. This may provide greater
insight into control processes of voluntary control of force in health and disease further expanding
the understanding of CNS pathologies.
References
Anmuth, C.J., Goldberg, G., Mayer, N.H., 1994. Fractal dimension of electromyographic signals
recorded with surface electrodes during isometric contractions is linearly correlated with
muscle activation. Muscle Nerve 17, 953–954.
Arjunan, S.P., Kumar, D.K., 2010. Decoding subtle forearm flexions using fractal features of
surface electromyogram from single and multiple sensors. J. Neuroeng. Rehabil. 7, 53.
doi:10.1186/1743-0003-7-53.
Arle, J.E., Simon, R.H., 1990. An application of fractal dimension to the detection of transients
in the electroencephalogram. Electroencephalogr. Clin. Neurophysiol. 75, 296–305.
Carpentier, A., Duchateau, J., Hainaut, K., 2001. Motor unit behaviour and contractile changes
during fatigue in the human first dorsal interosseus. J. Physiol. 534, 903–12.
Cashaback, J.G. a, Cluff, T., Potvin, J.R., 2013. Muscle fatigue and contraction intensity
modulates the complexity of surface electromyography. J. Electromyogr. Kinesiol. 23, 78–
83. doi:10.1016/j.jelekin.2012.08.004
Christensen, H., Lo, M.M., Dahl, K., Fuglsang-Frederiksen, A., 1984. Processing of electrical
activity in human muscle during a gradual increase in force.
Electroencephalogr.Clin.Neurophysiol. 58, 230–239.
Cukic, M., Oommen, J., Mutavdzic, D., Jorgovanovic, N., Ljubisavljevic, M., 2013. The effect
of single-pulse transcranial magnetic stimulation and peripheral nerve stimulation on
complexity of EMG signal: fractal analysis. Exp. brain Res. 228, 97–104.
doi:10.1007/s00221-013-3541-1
De Luca, C.J., 1984. Myoelectrical manifestations of localized muscular fatigue in humans. Crit
Rev.Biomed.Eng 11, 251–279.
Del Santo, F., Gelli, F., Mazzocchio, R., Rossi, A., 2007. Recurrence quantification analysis of
surface EMG detects changes in motor unit synchronization induced by recurrent inhibition.
Exp. brain Res. 178, 308–315. doi:10.1007/s00221-006-0734-x
Eke, A., Herman, P., Kocsis, L., Kozak, L.R., 2002. Fractal characterization of complexity in
temporal physiological signals. Physiol Meas. 23, R1-38.
Farina, D., Fattorini, L., Felici, F., Filligoi, G., 2002. Nonlinear surface EMG analysis to detect
changes of motor unit conduction velocity and synchronization. J. Appl. Physiol. 93, 1753–
63. doi:10.1152/japplphysiol.00314.2002
Ferenets, R., Lipping, T., Anier, A., Jäntti, V., Melto, S., Hovilehto, S., 2006. Comparison of
entropy and complexity measures for the assessment of depth of sedation. IEEE Trans.
Biomed. Eng. 53, 1067–77. doi:10.1109/TBME.2006.873543
Filligoi, G., Felici, F., 1999. Detection of hidden rhythms in surface EMG signals with a non-
linear time-series tool. Med. Eng. Phys. 21, 439–48.
Gitter, J.A., Czerniecki, M.J., 1995. Fractal analysis of the electromyographic interference
pattern. J. Neurosci. Methods 58, 103–108.
Glenny, R.W., Robertson, H.T., Yamashiro, S., Bassingthwaighte, J.B., 1991. Applications of
fractal analysis to physiology. J. Appl. Physiol. 70, 2351–67.
Goldberger, A.L., Amaral, L.A.N., Hausdorff, J.M., Ivanov, P.C., Peng, C.-K., Stanley, H.E.,
2002. Fractal dynamics in physiology: Alterations with disease and aging. Proc. Natl. Acad.
Sci. 99, 2466–2472. doi:10.1073/pnas.012579499
Gupta, V., Suryanarayanan, S., Reddy, N.P., 1997. Fractal analysis of surface EMG signals from
the biceps. Int.J.Med.Inform. 45, 185–192.
Harris, J. a, Clifford, C.W.G., Miniussi, C., 2008. The functional effect of transcranial magnetic
stimulation: signal suppression or neural noise generation? J. Cogn. Neurosci. 20, 734–40.
doi:10.1162/jocn.2008.20048
Higuchi, T., 1988. Approach to an irregular time series on the basis of the fractal theory. Phys. D
Nonlinear Phenom. 31, 277–283.
Ikegawa, S., Shinohara, M., Fukunaga, T., Zbilut, J.P., Webber, C.L., 2000. Nonlinear time-
course of lumbar muscle fatigue using recurrence quantifications. Biol. Cybern. 82, 373–82.
Julkunen, P., Kallioniemi, E., Könönen, M., Säisänen, L., 2013. Feasibility of automated analysis
and inter-examiner variability of cortical silent period induced by transcranial magnetic
stimulation. J. Neurosci. Methods 217, 75–81. doi:10.1016/j.jneumeth.2013.04.019
Kalauzi, A., Bojić, T., Vuckovic, A., 2012. Modeling the relationship between Higuchi's fractal
dimension and Fourier spectra of physiological signals. Med. Biol. Eng. Comput. 50, 689–
99. doi:10.1007/s11517-012-0913-9
Karthick, P.A., Makaram, N., Ramakrishnan, S., 2014. Analysis of progression of fatigue
conditions in biceps brachii muscles using surface electromyography signals and
complexity based features. Conf. Proc. ... Annu. Int. Conf. IEEE Eng. Med. Biol. Soc.
IEEE Eng. Med. Biol. Soc. Annu. Conf. 2014, 3276–9. doi:10.1109/EMBC.2014.6944322
Knaflitz, M., Merletti, R., De Luca, C.J., 1990. Inference of motor unit recruitment order in
voluntary and electrically elicited contractions. J. Appl. Physiol. 68, 1657–1667.
Kronholm, E., Virkkala, J., KÄRKI, T., KARJALAINEN, P., LANG, H., HÄMÄLÄINEN, H.,
Karki, T., KARJALAINEN, P., LANG, H., Hamalainen, H., 2007. Spectral power and
fractal dimension: Methodological comparison in a sample of normal sleepers and chronic
insomniacs. Sleep Biol. Rhythms 5, 239–250. doi:10.1111/j.1479-8425.2007.00317.x
Kukulka, C.G., Clamann, H.P., 1981. Comparison of the recruitment and discharge properties of
motor units in human brachial biceps and adductor pollicis during isometric contractions.
Brain Res. 219, 45–55.
Lake, D.E., Richman, J.S., Griffin, M.P., Moorman, J.R., 2002. Sample entropy analysis of
neonatal heart rate variability. Am. J. Physiol. Regul. Integr. Comp. Physiol. 283, R789-
797. doi:10.1152/ajpregu.00069.2002
Marsden, J.F., Ashby, P., Rothwell, J.C., Brown, P., 2000. Phase relationships between cortical
and muscle oscillations in cortical myoclonus: electrocorticographic assessment in a single
case. Clin. Neurophysiol. 111, 2170–2174.
Meigal, A.Y., Rissanen, S.M., Tarvainen, M.P., Georgiadis, S.D., Karjalainen, P. a, Airaksinen,
O., Kankaanpää, M., 2012. Linear and nonlinear tremor acceleration characteristics in
patients with Parkinson's disease. Physiol. Meas. 33, 395–412. doi:10.1088/0967-
3334/33/3/395.
Meigal, a I., Rissanen, S., Tarvainen, M.P., Karjalainen, P. a, Iudina-Vassel, I. a, Airaksinen, O.,
Kankaanpaa, M., Kankaanpää, M., Kankaanpaa, M., 2009. Novel parameters of surface
EMG in patients with Parkinson's disease and healthy young and old controls. J.
Electromyogr. Kinesiol. 19, 206–213. doi:http://dx.doi.org/10.1016/j.jelekin.2008.02.008
Mesin, L., Cescon, C., Gazzoni, M., Merletti, R., Rainoldi, A., 2009. A bi-dimensional index for
the selective assessment of myoelectric manifestations of peripheral and central muscle
fatigue. J. Electromyogr. Kinesiol. 19, 851–863. doi:10.1016/j.jelekin.2008.08.003
Milner-Brown, H.S., Stein, R.B., Yemm, R., 1973. Changes in firing rate of human motor units
during linearly changing voluntary contractions. J. Physiol. 230, 371–90.
Molina-Picó, A., Cuesta-Frau, D., Aboy, M., Crespo, C., Miró-Martínez, P., Oltra-Crespo, S.,
2011. Comparative study of approximate entropy and sample entropy robustness to spikes.
Artif. Intell. Med. 53, 97–106. doi:10.1016/j.artmed.2011.06.007
Nieminen, H., Takala, E.P., 1996. Evidence of deterministic chaos in the myoelectric signal.
Electromyogr. Clin. Neurophysiol. 36, 49–58.
Oldfield, R.C., 1971. The assessment and analysis of handedness: the Edinburgh inventory.
Neuropsychologia 9, 97–113.
Pethick, J., Winter, S.L., Burnley, M., 2015. Fatigue reduces the complexity of knee extensor
torque fluctuations during maximal and submaximal intermittent isometric contractions in
man. J. Physiol. 593, 2085–96. doi:10.1113/jphysiol.2015.284380
Ravier, P., Buttelli, O., Jennane, R., Couratier, P., 2005. An EMG fractal indicator having
different sensitivities to changes in force and muscle fatigue during voluntary static muscle
contractions. J. Electromyogr. Kinesiol. 15, 210–221. doi:10.1016/j.jelekin.2004.08.008
Richman, J.S., Moorman, J.R., 2000. Physiological time-series analysis using approximate
entropy and sample entropy. Am J Physiol Hear. Circ Physiol 278, H2039-2049.
Riley, Z.A., Maerz, A.H., Litsey, J.C., Enoka, R.M., 2008. Motor unit recruitment in human
biceps brachii during sustained voluntary contractions. J. Physiol. 586, 2183–2193.
doi:10.1113/jphysiol.2008.150698
Rosler, K.M., Petrow, E., Mathis, J., Aranyi, Z., Hess, C.W., Magistris, M.R., 2002. Effect of
discharge desynchronization on the size of motor evoked potentials: an analysis.
Clin.Neurophysiol. 113, 1680–1687.
Rossini, P.M., Barker, A.T., Berardelli, A., Caramia, M.D., Caruso, G., Cracco, R.Q.,
Dimitrijević, M.R., Hallett, M., Katayama, Y., Lücking, C.H., 1994. Non-invasive electrical
and magnetic stimulation of the brain, spinal cord and roots: basic principles and procedures
for routine clinical application. Report of an IFCN committee. Electroencephalogr. Clin.
Neurophysiol. 91, 79–92.
Ruonala, V., Meigal, A., Rissanen, S.M., Airaksinen, O., Kankaanpää, M., Karjalainen, P.A.,
2014. EMG signal morphology and kinematic parameters in essential tremor and
Parkinson's disease patients. J. Electromyogr. Kinesiol. 24, 300–6.
doi:10.1016/j.jelekin.2013.12.007
Spasic, S., Kalauzi, A., Grbic, G., Martac, L., Culic, M., 2005. Fractal analysis of rat brain
activity after injury. Med. Biol. Eng. Comput. 43, 345–8.
Stam, C.J., 2005. Nonlinear dynamical analysis of EEG and MEG: review of an emerging field.
Clin.Neurophysiol. 116, 2266–2301. doi:http://dx.doi.org/10.1016/j.clinph.2005.06.011
Staudenmann, D., van Dieën, J.H., Stegeman, D.F., Enoka, R.M., 2014. Increase in heterogeneity
of biceps brachii activation during isometric submaximal fatiguing contractions: a
multichannel surface EMG study. J. Neurophysiol. 111, 984–90. doi:10.1152/jn.00354.2013
Timmer, J., Lauk, M., Pfleger, W., Deuschl, G., 1998b. Cross-spectral analysis of physiological
tremor and muscle activity. II. Application to synchronized electromyogram. Biol. Cybern.
78, 359–368.
Timmer, J., Lauk, M., Pfleger, W., Deuschl, G., 1998a. Cross-spectral analysis of physiological
tremor and muscle activity. I. Theory and application to unsynchronized electromyogram.
Biol. Cybern. 78, 349–57.
Weiss, B., Clemens, Z., Bódizs, R., Halász, P., 2011. Comparison of fractal and power spectral
EEG features: effects of topography and sleep stages. Brain Res. Bull. 84, 359–75.
doi:10.1016/j.brainresbull.2010.12.005
Zhang, X., Ren, X., Gao, X., Chen, X., Zhou, P., 2016. Complexity analysis of surface EMG for
overcoming ECG interference toward proportional myoelectric control. Entropy 18, 1–12.
doi:10.3390/e18040106
|
1703.00223 | 1 | 1703 | 2017-03-01T10:48:45 | New Empirical Evidence on Disjunction Effect and Cultural Dependence | [
"q-bio.NC"
] | We perform new experiment using almost the same sample size considered by Tversky and Shafir to test the validity of classical probability theory in decision making. The results clearly indicate that the disjunction effect depends also on culture and more specifically on gender (females rather than males). We did more statistical analysis rather that putting the actual values done by previous authors. We propose different kind of disjunction effect i.e. strong and weak based on our statistical analysis. | q-bio.NC | q-bio |
New Empirical Evidence on Disjunction Effect and
Cultural Dependence
Indranil Mukhopadhyay1
Human Genomics Unit, Indian Statistical Institute, Calcutta-700035.
Nithin Nagaraj2
Consciousness Studies Programme, National Institute of Advanced Studies, IISc. Campus,
Bengaluru 560012.
Sisir Roy3
Conciousness Studies Programme, National Institute of Advanced Studies,IISC Campus,
Bengaluru 560012.
Abstract
We perform new experiment using almost the same sample size considered by
Tversky and Shafir to test the validity of classical probability theory in decision
making. The results clearly indicate that the disjunction effect depends also
on culture and more specifically on gender(females rather than males). We did
more statistical analysis rather that putting the actual values done by previous
authors. We propose different kind of disjunction effect i.e. strong and weak
based on our statistical analysis.
Keywords: Disjunction effect, decision making, quantum probability,
cognition
1. Introduction
Tversky and Shafir (1992) [1] discovered a phenomenon called the disjunction
effect, while following the process of testing a rational axiom of decision theory.
1Email: [email protected].
2Email: [email protected].
3Corresponding author, email: [email protected].
Preprint submitted to arXiv
July 22, 2018
It is also called as the sure thing principle (Savage, 1954) [2]. First consider the
states A and B that belong to the state of the world X. This principle states
that, if action A over B is preferred, and under the complementary state of the
world, again, action A over B is preferred, it is expected that one should prefer
action A over B even when the state of the world is not known. Symbolically,
this can be expressed as
If
and
Then
i.e.
P (A ∩ X) > P (B ∩ X)
P (A ∩ X C) > P (B ∩ X C)
P (A) = P (A ∩ (X ∪ X C) > P (B ∩ (X ∪ X C) = P (B)
occurs always.
P (A) > P (B)
With the aim of testing this principle, in their experiment, Tversky and
Shafir (1992) [1] performed the test considering a two stage gamble by presenting
98 students. They adopted a two stage gamble, i.e., it is possible to play the
gamble twice. The gamble is done under the following two conditions:
• The students are informed that they lost the first gamble.
• The students remained unaware of the outcome of the first gamble.
The gamble to be played, had an equal stake, i.e., of wining 200 or loosing
100 for each stage of taking decision, i.e., whether to play or not to play the
gamble.
Interestingly, the results of these experiments can be described into the fol-
lowing manner:
• The students who won the first gamble 69% choose to play at the second
stage;
2
• The students who lost then 59 % choose to play again;
• The students who are unaware whether they won or lost 36% of them (i.e.,
less of the majority of the students) choose to play again.
Explaining the findings in terms of choice based on reasons, Tversky and
Shafer (1992) [1] did raise some questions for these surprising results: whenever,
the persons, related to the play, knew that if they win, then as they would have
extra house money, they can play again, because if they lose, they can play again
to recover the loss. Now the students who did not know the outcome, then the
main issue is why sizable fraction of the students want to play again the game
since either they win or lose and cannot be anything else? Thus they arrived at
the key result, but faced the problem of explaining the outcome just as either
a win or loss. Busemeyer, Wang and Townsand (2006) [3] originally suggested
that this finding could be an example of an interference effect, similar to that
found in the double slit experiments conducted in modern particle physics.
Let us consider the following analogy between the disjunction experiment
and the classic double slit type of experiment in physics: Both the cases involve
two possible paths: here, in the disjunction experiment, the two paths are infer-
ring the outcome of either a win or a loss with the first gamble; for the double
split experiment, the two paths are splitting the photon into the upper or lower
channel applying a beam splitter. The path taken can be known (observed) or
unknown (unobserved), in both the experiments. Finally in both the cases, the
fact is that when the case of gambling for disjunction experiment and hence,
the detection at a location for the two slit experiments are considered for the
chosen unknown, i.e., unobserved conditions, the resultant probability, meant
for observing interference phenomena, are found to be much less than each kind
of the probabilities which is observed for the known (observed) cases. Under
these circumstances, we can speculate that during the disjunction experiment,
under the unknown condition, instead of being definite, so far as the win or loss
state is concerned, the student enters a superposition state. In fact, this state
prevents finding a reason for choosing the gamble. In double-slit experiments,
3
the law of additivity of probabilities of occurring two mutually exclusive events
(particle aspect or wave aspect) is violated i.e. total probability
PAB = PA + PB
for two mutually exclusive events A and B. This is due to the existence of
interference effects, known as Formula of Total Probability (FTP). It has al-
ready been established fact that the two slit (interference) experiment is the
basic experiment which violates FTP. Feynman (1951) [4], in many of his works
in physics, presented his points with detailed arguments about this experiment.
There, the results, i.e., the appearance of interference fringes appeared to him,
not at all surprising phenomena. He explained it as follows: In principle, inter-
action with slits placed on the screen may produce any possible kind of distri-
bution of points on the registration screen. Now, let us try to explain following
quantum probabilistic features which appear only when one considers following
three kind of different experiments [5]:
• When only the first slit open i.e., the case B = +1, in an equivalent
manner.
• When only the second slit is open i.e., B = −1, in this case.
• In the particular case, both slits being open, it is the random variable B
determining the slit to pass through.
At this stage, let us now choose any point at the registration screen. Then
resultant scenario will be as follows:
• the random variable A if A = +1.
• the opposite case happens if a particle hits the screen at this point, i.e.,
A = −1.
Now for classical particles, FTP should predict the probability for the ex-
periment (both slits are open), supposed to be provided by the (1) and (2)
4
experiments. But, it has already been mentioned that, in case of quantum par-
ticles, FTP is violated:
for the additional cosine-type term appearing in the
right-hand side of FTP, it is the interference effect in probabilities which is re-
sponsible. Feynman characterized this particular characteristic feature of the
two slit experiment as the most profound violation of laws of classical probability
theory. He explained it the following way:
In an ideal experiment, where there is no presence of any other external un-
certain disturbances, the probability of an event, called probability amplitude
is the absolute square of the complex quantity. But, when there is possibility of
having the event in many possible ways, the probability amplitude is the sum of
the probability amplitude considered separately. Following their experimental
results, Tversky and Shafir (1992) [1] demand that this violation of classical
probability is also possible to be present, happening in their experiments with
cognitive systems. Though, due to the possible restrictions present in quantum
mechanics, we could not start from Hilbert formalism at the start, in the labo-
ratory, this formalism was justified by experiments. Let us now try to interpret
the meaning of interference effect within the context of the experiments on gam-
bling, described above. We will follow, here, Busemayer's formulation (2011) [6]
which is as follows:
Two different judgment tasks A and B are considered in this case. The
task A is considered having j (taking j = 2, binary choice) different levels of
response variable and B, with k (say, k = 7 points of confidence rating) levels
of a response measure. Two groups, being randomly chosen, out of the total
participants we have:
• Group A gets task A only.
• Group BA gets task B followed by task A.
The response probabilities can be estimated as follows:
• From the group A, let p(A = j) be estimated; This denotes the probability
of choosing level j out of the response to task A .
5
• And then from the group B, let p(B = k) be estimated; the corresponding
probability denoted by choosing level k from the task B.
• Now, it is possible to estimate the conditional probability p(A = jB = k);
which can be stated as the probability of responding with level j from the
task A, given the person responded with level k on earlier task B.
So, we can write the estimated interference for level j to task A (produced,
when responding to task B) as,
A(j) = p(A = j)pT (A = j)
, where, pT (A = j) denotes the total probability for the response to task A. In
defense of using Quantum formalism in case of human judgments, Busemeyer
and Truebold did put four following reasons, beautifully, in their famous paper
(Busemeyer et al., 2011 [6]):
(a) judgment is not a simple readout from a preexisting or recorded state,
but instead it is constructed from the current context and question; from this
first point it then follows that (b) making a judgment changes the context which
disturbs the cognitive system; and the second point implies that (c) changes in
context produced by the first judgment affects the next judgment producing order
effects, so that (d) human judgments do not obey the commutative rule of classic
probability theory(Busemeyer et al; 2011 [6]).
In fact, the existence of interference term for microscopic entities or quan-
tum entities clearly indicates the existence of three valued or non-Boolean logic.
This is popularly known as Quantum logic. It is mathematically shown that
a set of propositions which satisfies the different axiomatic structures for the
non-Boolean logic generates Hilbert space structures. The quantum probability
associated with this type of Quantum logic can be applied to decision making
problems in cognitive domain. It is to be noted that, up till now, no quantum
mechanical framework is taken as valid description of the anatomical structures
and function of the brain. This framework of quantum probability is very ab-
stract and devoid of any material content. So it can be applied to any branch
6
of knowledge like Biology, Social science etc. Of course, it is necessary to un-
derstand the issue of contextualization, for example, here, in case of decision
making in brain. It is worth mentioning that the decision making may depend
also on culture. For example, the students participated in the above mentioned
gamble are mainly taken from the west. So far as the authors' knowledge, con-
cern of this kind of experiment has not been performed taking the subjects from
the east. We presume that this is an important factor one should consider in this
kind of experiment involving gambling since the very concept of gambling may
depend on culture. While experimenting to find the effect of culture, we have
found an additional interesting fact. It is not only culture but also gender that
might play an important role on disjunction effect. It is natural to expect and
consider that the gender dependency is closely dependent on culture. Recently,
we performed the above mentioned two stage gambling considering almost the
same sample size in India. The results clearly show that the violation of classi-
cal probability rule depends very much on the variation of gender. In the next
section we will describe the experiment and the results.
2. Materials and Methods
In order to get an insight of the above discussion we have performed a gam-
bling experiment in the line of Tvaersky and Shafir (1952). Our main objective
is to see whether there is any disjunction effect, especially in Indian context.
2.1. Participants
As we believe that there might be a cultural effect, we carefully select the
participants. The participants should consist of both males and females as cul-
tural effect with respect to gender is sometimes significantly observable. We
have selected the population of experimental objects as a homogeneous groups
with respect to age so that there would not be any effect that can mask our
objectives. From Raja Rajeswari Engineering College, Bengaluru, India, we se-
lect 101 college students randomly. It is seen that there are 50 female and 51
7
male students participating the experiment. The students belong 1st, 2nd, or
3rd year of their engineering curriculum and naturally belong to a very homo-
geneous age group. Prior permission is taken from the principal of the college
and the studetns gave consents to this experiment.
2.2. Design and procedure
In the experiment we toss a coin and the experiment depends on the outcome.
We first divide students into two groups: one consisting 70 students and the
other 31 students. We perform the experiment similar to that described in
Busemeyer.
If a students wins the first toss, he/she will receive Rs. 100; if
loses, he/she will lose Rs. 50. After the first toss, we ask a student whether
he/she wants to play again.
If he/she wants to play again, he/she will win
Rs. 200 if his/her guess about he outcome matches with the outcome of the
second toss; otherwise will lose Rs. 100. We play this game under two different
schemes.
In scheme 1, each student belonging to the first group (70 students), will
declare his/her guess about the result of a coin tossing experiment and will be
informed the result after the toss. We then ask that student whether he/she
wants to play again and record his/her response. In scheme 2, we perform this
experiment for the second group of students (31 students) slightly differently
than the first scheme. Here, each student will declare his/her guess about the
result if the first toss, but he/she will not be informed the result. However,
we then ask him/her whether he/she wants to play again and record his/her
response. We have used only one coin throughout experiment. Before the
beginning of this experiment, we checked whether the coin is unbiased using
a binomial experiment and confirmed that it is unbiased. We also make sure
that the student who has perfumed the experiment has no way to disclose the
result or his/her guess and attitude towards this experimental result. Moreover,
there is no exchange of information between the two groups of students for two
different schemes.
8
• We have selected randomly 101 college students from Raja Rajeswari En-
gineering College, Bengaluru, India. The students belong 1st, 2nd, or
3rd year of their engineering curriculum and naturally belong to a very
homogeneous age group.
• We used an unbiased coin for the experiment. Before the experiment we
have checked that the coin used is unbiased.
• We then divide them into two groups: one consisting 70 students and the
other 31 students.
• We performed the experiment similar to that described in Busemeyer. If
a students wins the first toss, he/she will receive Rs. 100; if loses, he/she
will lose Rs. 50.
• Each student belonging to the first group (70 students), will declare his/her
guess about the result of a coin tossing experiment and will be informed
the result after the toss. We record if he/she wants to play again.
• Each student belonging to the second group (31 students) will declare
his/her guess about the result if the first toss, but he/she will not be
informed the result. We then record if he/she wants to play again.
Probability theoretic explanation:
We have already discussed that P (A) > P (B) is true always under the
conditions that P (A ∩ X) > P (B ∩ X) and P (A ∩ X C ) > P (B ∩ X C).
Thus, if you prefer A when the event X is known, and if you prefer B when
the complementary event X C is known, then it will imply that you will always
prefer A over B irrespective of the events X or X C. Any violation of this is
termed as disjunction effect.
We have performed experiment similar to that given in Busemeyer and we
observed marked violation of this probabilistic claim and its explanation de-
pends on several factors, not reported or discussed in literature. Not only the
psychological factors, but also other factors like sex, culture, geographical region
etc, might have important role to play.
9
3. Results
We have observed a few interesting facts from the experiment. Instead on
looking at the numerical figures of the outcomes of the experiment only as
in Tversky and Shafir (1952), we have done a detailed statistical analysis in
order to strengthen the interpretation of our observations. As described in
the previous section, our experiment consists of two different schemes. Under
the first scheme, the students were informed the outcome of the first toss. If
they know that they won the first gamble, 76.47% want to play again, i.e.
majority want to play again. However we would like to be sure that the result
is statistically significant. For this, we perform a test of null hypothesis H0 :
p = 0.5 against an alternative hypothesis H1 : p > 0.5. The p-value associated
with this test is 0.0004 indicating that the majority wants too play again. On
the other hand, if they did not know, whether they won or lost, 58.33% want
to play again. Although it seems that majority wants to play again, but the
result is not statistically significant (p-value= 0.1215). Under scheme 2, when
the students did not know the outcome of the first toss, i.e. when they did
not know whether they won or lost, 54.84% want to play again. So, in this
case, although it seems that majority wants to play again, but the result is not
statistically significant since the associated p-value is 0.3601.
In each case majority wants to play again. But in Busemeyer the corre-
sponding figures are 69%, 59%, and 36% respectively, while in our experiment
the figures are 76.47%, 58.33%, and 54.84% respectively. The last figure differs
widely while the second figure matches surprisingly. We think that probably
p-value corresponding to the second figure (i.e. 59%) is not significant. So
Busemeyer inference needs to be revisited. However, the idea of paying again
when they know that whether they won or lost still remains valid if we go only
by the actual figures compared to 50%, which indicates the state of indifference.
So we did an exact test of hypothesis that H0 : p = 0.5 against H1 : p > 0.5 i.e.
to see absence of disjunction effect. The p-value is 0.3601 indicating that they
are indifferent to the decision. So either disjunction effect is not observed here,
10
or very weak, but they do not prefer to play when they do not know the state.
It seems clear that the observations by Busemeyer is not matching with ours.
We conjecture that this discrepancy may be due to the effect of different cultural
settings in which the experiments are conducted. However, we proceed further
to examine whether there exists any other factor that might play an implicit role
in the experimental results. Indian culture and gender are intermingled always.
Hence it would be a wise idea to revisit the experimental results incorporating
gender factor. We observed that among females, 75% (p-value= 0.0106) females
want to play again if they know that they won, 52.94% (p-value= 0.3145) females
want to play again if they know that they lost. This result is not at all significant
indicating that once lost, females do not want to take one more risk. Moreover,
41.18% (p-value= 0.6855) females want to play again if they do not the first
result.
This picture is markedly different among males. 77.78% (p-value= 0.0038)
males want to play again if they know that they won, 63.16% (p-value= 0.0835)
males want to play again if they know that they lost. However, this result
is marginally significant, although not strong. Moreover, 71.43% (p-value=
0.0288) males want to play again when they do not know the outcome of the
first experiment. Since majority wants to play again whether they won or lost
or uninformed, there is no disjunction effect observed among males.
4. Discussions
The above results of our experiment raise the following important issues:
• If we consider only absolute values, males do not show any disjunction
effect, but females show strong disjunction effect. Now we have to explain
this in Indian context, if possible.
• However, if we go by the p-values of the corresponding tests of significance,
interesting observations can be made using the combination of tests re-
sults. For all individuals, males and females taken together, let us first
11
combine the results of the conditions that first result is known to be win
and that to be loss. The combined p-value for this is 0.00053 indicating
that majority of the individuals want to play again if they know the result.
But the p-value for the result when they do not know the first result is
0.2366 indicating that there is disjunction effect.
• For males, disjunction effect is not observed if irrespective of whether we
go by the actual values or by the p-values; combined p-value is 0.0029
whereas p-value when the first result is not known is 0.0288. However,
the general effect is moving towards the disjunction effect although not
established statistically or by observations.
• For females, although actual observations suggest only weak disjunction
effect, but comparing p-values for the combined p-value (0.0223) to the
p-value (0.6855) when the first result is unknown shows clear disjunction
effect.
• Disjunction effect is although a clear concept and is realized in a number
of experiments, cultural effects are strong especially among males and
females. In Indian context, probably disjunction effect for females is so
strong that they overcome the absence of disjunction effect among males.
• Busemeyer's experimental results are based only on the actual values and
not tested statistically. This is highly influenced by specific cases and
specific scenarios considered in the experiment. It is not wise to declare
that the effect is present or absent whenever the number of observations
is less or greater than 50%; statistical test must be employed to validate
and confirm the findings.
• The missing part of all other previous experiments is the appropriate sta-
tistical analysis of the results. Based on our analysis, we propose different
kinds of disjunction effect, e.g. strong and weak.
This can be stated in the following manner.
12
Table 1: Categorisation of disjunction effect based on actual observation and statistical tests.
Actual observation
significant
Not significant
Statistical test
Significant
Strong effect
--
(p-value is small)
Not significant Weak effect
No effect
(p-value is large)
The above categorization triggers to rethink the classification of disjunction
effect with respect toothed factor like gender etc. Here we propose a general
categorization based on actual observations and results of statistical significance
tests in presence of another factor 'gender'. This is given in Table 2.
Table 2: A general categorization of disjunction effect based on actual observation and statis-
tical tests.
Actual observation
Significant in both
Significant in males Not significant
males and females
but not in females
in both
Statistical test
Significant in both
or vice versa
males and females
Strong effect
--
(p-value is small)
Significant in males but
not in females or vice
Moderate effect
Moderate effect
--
--
versa (large p-value)
Not significant
in both
Weak effect
Weak effect
No effect
(p-value is large)
In this manner we may come up with more logical categorization.
13
Acknowledgements
The authors acknowledge the Rector, RR Group of institutions, the principal
R.R.Engineering College and the participating students from R.R.Engineering
college for their help and cooperation. We would also like to thank Mr. Nepal
Banerjee for his help in conducting the experiments. This work is under the
project SB/S4/MS:844/2013 approved by SERB, DST, Government of India.
References
References
[1] A. Tversky, E. Shafir, Choice under conflict: The dynamics of deferred de-
cision, Psychological science 3 (6) (1992) 358 -- 361.
[2] J. Savage Leonard, The foundations of statistics, NY, John Wiley (1954)
188 -- 190.
[3] J. R. Busemeyer, Z. Wang, J. T. Townsend, Quantum dynamics of human
decision-making, Journal of Mathematical Psychology 50 (3) (2006) 220 -- 241.
[4] R. P. Feynman, R. B. Leighton, M. Sands, R. B. Lindsay, The feynman
lectures on physics, vol. 3: Quantum mechanics (1966).
[5] E. Conte, A. Y. Khrennikov, O. Todarello, A. Federici, L. Mendolicchio,
J. P. Zbilut, Mental states follow quantum mechanics during perception
and cognition of ambiguous figures, Open Systems & Information Dynamics
16 (01) (2009) 85 -- 100.
[6] J. R. Busemeyer, E. M. Pothos, R. Franco, J. S. Trueblood, A quantum the-
oretical explanation for probability judgment errors., Psychological review
118 (2) (2011) 193.
14
|
1606.09277 | 1 | 1606 | 2016-06-29T20:42:29 | Response and noise correlations to complex natural sounds in the auditory midbrain | [
"q-bio.NC"
] | How natural communication sounds are spatially represented across the inferior colliculus, the main center of convergence for auditory information in the midbrain, is not known. The neural representation of the acoustic stimuli results from the interplay of locally differing input and the organization of spectral and temporal neural preferences that change gradually across the nucleus. This raises the question how similar the neural representation of the communication sounds is across these gradients of neural preferences, and whether it also changes gradually. Multi-unit cluster spike trains were recorded from guinea pigs presented with a spectrotemporally rich set of eleven species-specific communication sounds. Using cross-correlation, we analyzed the response similarity of spiking activity across a broad frequency range for similarly and differently frequency-tuned neurons. Furthermore, we separated the contribution of the stimulus to the correlations to investigate whether similarity is only attributable to the stimulus, or, whether interactions exist between the multi-unit clusters that lead to correlations and whether these follow the same representation as the response similarity. We found that similarity of responses is dependent on the neurons' spatial distance for similarly and differently frequency-tuned neurons, and that similarity decreases gradually with spatial distance. Significant neural correlations exist, and contribute to the response similarity. Our findings suggest that for multi-unit clusters in the mammalian inferior colliculus, the gradual response similarity with spatial distance to natural complex sounds is shaped by neural interactions and the gradual organization of neural preferences. | q-bio.NC | q-bio |
Response and noise correlations to complex natural sounds
in the auditory midbrain
Dominika Lyzwa ∗1,2,3 and Florentin Worgotter3,4
1Max Planck Institute for Dynamics and Self-Organization, Gottingen, Germany
2Institute for Nonlinear Dynamics, Physics Dep., Georg-August-University, Gottingen
3Bernstein Focus Neurotechnology, Gottingen
4Institute for Physics-Biophysics, Georg-August University, Gottingen
November 10, 2018
Abstract
How natural communication sounds are spatially represented across the inferior colliculus, the main
center of convergence for auditory information in the midbrain, is not known. The neural representation of
the acoustic stimuli results from the interplay of locally differing input and the organization of spectral and
temporal neural preferences that change gradually across the nucleus. This raises the question how similar
the neural representation of the communication sounds is across these gradients of neural preferences,
and whether it also changes gradually. Multi-unit cluster spike trains were recorded from guinea pigs
presented with a spectrotemporally rich set of eleven species-specific communication sounds. Using
cross-correlation, we analyzed the response similarity of spiking activity across a broad frequency range
for similarly and differently frequency-tuned neurons. Furthermore, we separated the contribution of the
stimulus to the correlations to investigate whether similarity is only attributable to the stimulus, or, whether
interactions exist between the multi-unit clusters that lead to correlations and whether these follow the
same representation as the response similarity. We found that similarity of responses is dependent on
the neurons' spatial distance for similarly and differently frequency-tuned neurons, and that similarity
decreases gradually with spatial distance. Significant neural correlations exist, and contribute to the
response similarity. Our findings suggest that for multi-unit clusters in the mammalian inferior colliculus,
the gradual response similarity with spatial distance to natural complex sounds is shaped by neural
interactions and the gradual organization of neural preferences.
1 Introduction
A neuron's response is shaped by all the inputs it receives, as well as by the integration and processing
of this input, hence by the neuron's stimulus preferences. The inferior colliculus is the main center of
convergence in the auditory midbrain (Irvine, 1992). It receives and integrates diverse preprocessed inputs
from essentially all ascending auditory brainstem nuclei (Aitkin and Phillips, 1984b; Malmierca et al.,
2002) that terminate on different locations within the central inferior colliculus (ICC) (Oliver, 2005). Dif-
ferences exist e.g. for high and low frequency regions as well as for caudal or rostral regions. Information
about interaural time differences from the medial superior olive for example is mainly projected to low
and middle frequency regions (Oliver, 2005). Neural preferences to stimulus frequency and modulation
are mainly organized gradually within the ICC (Merzenich and Reid, 1974; Schreiner and Langner, 1988;
∗Corresponding author: [email protected]
1
Langner et al., 2002). In the tonotopic gradient (Fig.1) low frequencies are represented dorsol laterally and
high frequencies ventral medially (Rose et al., 1963; Merzenich and Reid, 1974). Along this tonotopic
gradient, the stimulus frequency which elicits the highest spiking response gradually increases. For a given
intensity this is called the best frequency (BF) and for the overall lowest spike-eliciting intensity this is the
characteristic frequency (CF). Oriented approximately orthogonal to this frequency gradient are laminae
that contain neurons with very similar best frequencies within a range of 1/3 octave, the isofrequency
laminae (Schreiner and Langner, 1997). Strong indications for a concentric gradient within laminae of
preferred amplitude modulation frequencies for the sound envelope have been provided (Schreiner and
Langner, 1988; Langner et al., 2002; Baumann et al., 2011). The ICC has also been shown to be essential
for extracting time-varying spectrotemporal information (Escab´ı and Schreiner, 2002) and therefore might
be important for processing of complex sounds.
Figure 1: Schematic of the central inferior colliculus displaying the tonotopic gradient and isofrequency laminae.
The inputs to the inferior colliculus are very diverse (Oliver, 2005), e.g.
the Medial Superior Olive (MSO) projects
mainly to low and middle frequency regions.
The neural response, the representation of acoustic stimuli in the ICC results from the interplay of the locally
differing and heterogeneous input and the spatially gradual change of spectrotemporal neural preferences.
Along these gradients the sound is filtered either for the same spectral content or for same amplitude mod-
ulations. Thus, the question arises how for complex, spectrotemporally rich sounds, such as speech or
vocalizations, the neurons' output is organized across this main convergence center. Neural interactions
of connected neurons could further contribute to the specific organization of the neural representation in
this nucleus, hence the output of the neurons to vocalization stimuli. Comparing response correlations and
neural correlations which result from connected neurons allows the evaluation of whether similarity of re-
sponses is shaped by the underlying neural structure rather than by the stimulus input to the neurons.
In this work we investigate the dependence of neural response similarity on location, thus on the spatial
distance between the neurons, for the representation of vocalizations in the central inferior colliculus. The
hypothesis is tested that despite the locally differing various input, the spectrotemporal gradients induce a
gradually changing neural representation of these natural complex sounds. To this end, neural responses
are compared by cross-correlation for differently and similarly frequency-tuned neurons, with respect to the
spatial distance between the neurons. Simultaneous and non-simultaneous recordings were compared to
obtain indications whether similarity of responses is mainly induced by the stimulus or by neuronal interac-
tions that can induce neural correlations; the latter might suggest cooperative processing of the multi-units
clusters. Whether neural correlations for vocalizations exist in the mammalian inferior colliculus has not
been investigated before. Neural correlations can be beneficial, detrimental or not to affect encoding of
sensory stimuli, and might depend on the specific neuronal structure (Averbeck et al., 2006).
We analyzed simultaneous recordings from 32 sites in the ICC of guinea pigs in response to monaurally
presented conspecific vocalizations. The set of eleven behaviorally relevant sounds (Berryman, 1976) dis-
plays a wide spectrum of acoustic properties, such as amplitude and frequency modulations, harmonics and
temporal correlations.
2
It was suggested that neurons are adapted to process natural sounds (Rieke et al., 1995), therefore, these
might trigger responses which are not elicited by artificial or simple acoustic stimuli. Response similarity
between multi-unit clusters is obtained by pairwise cross-correlation analysis of the spiking activity i.e. the
processed output of the ICC neurons. The correlation of spiking responses is additionally compared to the
correlation of long range activity, the local field potential (LFP). Multi-unit cluster activity is the combined
activity mainly from neighboring single neurons that span one order of magnitude. This integrated activity
could allow one to investigate local population processing in the ICC. It has also been shown that multi-unit
clusters respond stronger to natural sound than single neurons (Grace et al., 2003) and that the natural stim-
uli can be more accurately discriminated based on these responses than based on single neuron responses
(Engineer et al., 2008).
Whereas in a previous study in the mammalian ICC correlation of single neuron spike trains was investi-
gated in dependence of the neuron's spectrotemporal properties with the finding that the best frequency is
the most correlated parameter (Chen et al., 2012) and a microcircuitry was suggested, in the present study
similarity is investigated at the level of multi-unit clusters which likely display a different correlation struc-
ture. Here, natural communication sounds are used instead of artificial sounds and a wider spatial range of
up to 1600 µm is probed. Dependencies on spatial distance have not been found for the grass frog midbrain
(Epping and Eggermont, 1987), but have been shown in the primary auditory cortex (Eggermont, 2006).
Since the grass frog midbrain only displays a weak tonotopic gradient, these findings do not transfer to
the mammalian inferior colliculus with a clear tonotopic organization and substantial differences in neural
structure.
In summary, we find that neural correlations exist in the mammalian inferior colliculus and that the neu-
ral and response correlations for spiking and long range activity gradually decrease with spatial distance
for similarly and differently frequency-tuned neurons. This suggests the gradual neural representation of
vocalizations is shaped by interactions between the neurons and their spectral and temporal preferences.
2 Materials & Methods
2.1 Electrophysiology
Neural activity was collected from the central nucleus of the inferior colliculus (ICC) of 11 adult male and
female Dunkin Hartley guinea pigs. Recordings were acquired from 11 guinea pigs in 3 to 4 electrode
insertion positions (taken altogether 36 positions), with activity recorded simultaneously from 32 recording
sites. The electrophysiological recordings and experimental setup are described in detail elsewhere (Rode
et al., 2013; Lyzwa et al., 2016).
For the recording, either a linear double-shank array (shank distance was 500 µm with 16 contacts linearly
spaced at 100 µm, on each shank) or a 4-double-tetrode array (shank distance of 500 µm, contact distance
within a tetrode of 25-82 µm) were used to measure activity simultaneously from 32 sites (impedances were
0.5-1 MΩ at 1 kHz; NeuroNexus Technologies, Ann Arbor, MI). Whereas the linear-double shank array
mainly records responses along the best frequency gradient, and covers a broad range of best frequencies,
the 4-double-tetrodes record from a few neighboring isofrequency laminae, and several multi-unit clusters
have similar frequency tuning (Fig.2). They have similar best frequencies, but might have different prefer-
ences for amplitude modulations (AM) depending on their spatial distance within the ICC (Schreiner and
Langner, 1988; Baumann et al., 2011). The electrode array was introduced under an angle of 45◦ dorso-
laterally along the tonotopic gradient of the ICC, into 3 to 4 different insertion positions for each animal.
The animals were anesthetized with Ketamine and stereotactically fixed with ear tubes through which the
sound was presented directly to the left eardrum. While acoustically presenting vocalization stimuli to the
left ear, neural activity was recorded from the contralateral ICC at a sampling rate of 24.414 kHz using a
TDT Tucker Davis System. For each vocalization 20 trials were recorded with intensities of 30-70 dB SPL
in steps of 10 dB SPL. For the analysis of this study recordings are those at 70 dB SPL stimulus intensity,
as these show the strongest response.
3
Frequency response maps (FRM) were obtained from spiking responses to pure tone stimuli. A total of
40 stimulus frequencies, ranging between 0.5 to 45 kHz, with a ramp rise and fall time of 5 ms each and
a duration of 50 ms were presented. From the FRMs, the best and characteristic frequencies were ob-
tained. These tone-evoked preferred frequencies, the frequency eliciting the highest spike-rate at the lowest
spike-eliciting intensity (the characteristic frequency, CF) and at a given presented stimulus intensity (the
best frequency, BF) ranged from 0.5 to 45 kHz. The frequency response maps for a linear double-shank
recording along the tonotopic gradient and for a 4-double-tetrode recording are given in Fig. 2.
Figure 2: Frequency tuning along the tonotopic gradient and within a few isofrequency lamina. a) Frequency
response maps (FRMs) recorded from 32 sites along the tonotopic gradient with a linear double-shank electrode. The
characteristic frequency (CF) covers a range from 0.5-29 kHz. The CF increases gradually from sites higher up (dor-
solateral) to lower ones (ventromedial), also the shape of the FRM changes from a broader symmetric shape to an
elongated and skewed shape for higher frequencies. The topmost sites do not show strong responses and might be lying
outside the central IC. b) FRMs of multi-units recorded with a double-tetrode electrode from two isofrequency laminae
with CFs of about 7 kHz (left column) and about 4 kHz (right). The CF does not change visibly within one isofrequency
lamina, but the frequency tuning varies in sensitivity, e.g. spike rates from site 23 and site 5 vary by over 150 Hz.
2.2 Vocalization Stimuli
The 11 vocalization stimuli used in this study are a representative set of guinea pig communication calls and
give information about the animal's behavioral state (Berryman, 1976). Figure 3 shows the spectrograms
and waveforms for four examples of this set: the 'tooth chatter', 'drr', 'scream' and 'squeal'. These natural
complex sounds display a variety of frequency modulations, frequency ranges and envelope types. Some
vocalizations such as the 'drr' (b) contain mainly frequencies below 3 kHz and have periodicities in the
waveform. The 'tooth chatter' also has a periodic waveform but a frequency content of up to 30 kHz.
Others such as the 'scream' (c) and 'squeal' (d) have complex waveforms, cover a broad spectral range and
display harmonics. Vocalizations with a complex waveform and almost all energy at low frequencies are
also present.
4
Figure 3: Vocalizations Spectrograms and waveforms of four representative examples of the entire set of eleven
guinea pig vocalizations. The 'tooth chatter' (a) and 'drr' (b) have periodicities in the waveform, with (a) containing
frequencies up to 30 kHz, and (b) containing only low frequencies below 3 kHz. The 'scream' (c) and 'squeal' (d) have
complex waveforms, cover a broad spectral range and show harmonics.
The vocalizations were played 20 ms after recording onset and vary in duration between 300 ms and
1,300 ms. They were recorded from male and female Dunkin Hartley guinea pigs at a sampling rate of
97.656 kHz. Details on the stimuli recording including sound calibration and the frequency content for the
full vocalization set can be found in (Rode et al., 2013; Lyzwa et al., 2016).
2.3 Preprocessing of Neurophysiological Data
In order to investigate responses to vocalizations in the ICC, the spiking activity is employed for the analysis,
because this is the processed output of the neurons (Pettersen et al., 2012). Additionally, the analysis is
extended to the local field potential (LFP), which is long-range activity and contains i.a. the synaptic input
to the neurons (Pettersen et al., 2012). The local field potentials were obtained by Butterworth filtering the
voltage traces in a range between 0.5-500 Hz (Pettersen et al., 2012).
To obtain spiking multi-unit activity the recorded voltage traces are Butterworth-filtered for 300-3,000 Hz
and thresholded at z = 3 standard deviations exceeding the ongoing activity (Θ = µ + zσ), with the mean
µ and the standard deviation σ of the ongoing activity. This spontaneous or ongoing activity was acquired
from the first 20 ms of each recording, during which no stimulus was presented, in order to account for
adaptation effects over time and different spontaneous rates of the neurons. Due to the low impedance of
0.5− 1MΩ, detected spikes likely originate from different single neurons, and therefore no refraction time
between spikes was assumed. The recorded activity is multi-unit cluster activity, which is the compound
spiking response mainly of several neighboring single neurons recorded from one site. We used the offline
spike-sorting program WaveClus (Quian Quiroga et al., 2004) to sort and separate spikes according to the
spike waveform on a subset of the passband-filtered multi-unit cluster responses, with details in (Lyzwa et
al., 2016). Separation into single units was not possible because within sorted clusters a significant fraction
of the spikes displayed inter-spike-intervals of less than 3 ms; and in some cases spikes could not be sorted
into clusters. The responses investigated here are from neural groups comprising at least 3-5 single neurons
and smaller contributions from neurons that are farther away from the recording electrode and are not
distinguishable. Note that it is possible that different sub-groups of multi-unit clusters respond to different
vocalizations. Multi-unit spike trains were binned at 1 ms and convolved with an exponential filter function,
f (t)=t· exp(α · t), with time t, to mimic the time course of excitatory postsynaptic potentials (EPSP) (van
Rossum, 2001), as used by (Machens et al., 2003). The full width at half maximum α, of the EPSP-like
function was chosen to be 3 ms (Lyzwa et al., 2016). Averaging the binned spike trains across all ntrial =20
trials yielded the post-stimulus time histogram (PSTH).
5
2.4 Cross-Correlation Analysis
In order to test similarity of responses from different multi-unit clusters to the same vocalization, we em-
ployed cross-correlation which yields a compact description for the large set of neurons analyzed in this
work. Similarity of responses to each vocalization was tested for pairs of multi-unit clusters from one
recording which allows including the spatial distance between the neurons into the analysis. A correlation-
based similarity measure of spike trains (Schreiber et al., 2003) has been employed earlier for neural dis-
crimination of single neurons and groups of neurons (Wang et al., 2007). It has been shown that for the
comprehension of speech (Shannon et al., 1995) which is spectrotemporally varying complex sound as are
vocalizations, temporal information is crucial. As an index of the responses' temporal information, here,
we computed the degree of correlation between the spike trains from different multi-unit clusters. Pairwise,
EPSP-spike trains or LFP from two multi-unit clusters x(t),y(t) of length n, were cross-correlated and the
highest value within a maximum possible delay of τ between the responses was selected:
(cid:32)
(cid:112)∑n
Corr(τ) = max
∑n−τ
t=1 (x(t + τ)−(cid:104)x(cid:105))· (y(t)−(cid:104)y(cid:105))
t=1(x(t)−(cid:104)x(cid:105))2 ∑n
t=1(y(t)−(cid:104)y(cid:105))2
(cid:33)
(1)
with a lag of τ = [−10 ms,10 ms]. This delay is within the range of maximum response latencies in the
ICC (Langner et al., 1987). The correlation values were computed with a lag, because response latencies
do vary across multi-unit clusters with different spectral preferences (Langner et al., 1987). The correlation
value of ntrial = 20 trials for this multi-unit pair was then averaged.
2.5 Neural Correlations
Simultaneously recorded spiking responses from interacting neurons can show correlated trial-to trial vari-
ability (Averbeck et al., 2006). If no stimulus is present the correlated activity is termed neural or noise
correlation (N). For stimulus-driven simultaneous responses, the measured response correlation contains
Corr, in addition to the neural correlation, a contribution which is due to the neurons responding to and
following the same stimulus (stimulus correlations, D). These stimulus correlations are present for simul-
taneous and non-simultaneous recordings, as long as the same stimulus is used. In order to separate the
stimulus from the response correlations, simultaneously recorded trials of the multi-unit clusters were ran-
domly shuffled (Abeles, 1982) before correlating them, and thus only stimulus correlations remain. To
investigate whether neural correlations exist and to obtain them, the averaged stimulus correlation is sub-
tracted from the response correlation for each trial and the average is taken N = ∑n=20
This approach (Abeles, 1982) has been developed for single neurons, and attempts to infer functional con-
nectivity from the computed neural correlations. Here, we use multi-unit clusters, and we do not attempt to
make inferences about functional neural connections but to test whether neural correlations exist. If corre-
lated activity is present then there is a significant difference between response correlations and the stimulus
correlations which is likely due to neural interactions. Significance was assessed using the Student's t-test
for normal distributions, and the Wilcoxon-Mann-Whitney test for comparison of non-normal distributions.
In order to visualize the effect of noise correlations on the encoded stimulus information, scatter plots are
often used (Averbeck et al., 2006). They display the distribution of averaged spike rates of two neurons
or neural groups to different stimuli. The more the response distributions to the different stimuli overlap,
the less information is carried by them. Comparing the scatter plot of simultaneous trials to the one with
shuffled trials yields information whether the neural correlation affect encoding. If, for example, separabil-
ity of responses to different stimuli increases when removing neural correlations, these are detrimental for
encoding.
Corr(i)− S(i).
i
6
2.6 Frequency Tuning and Spatial Distance of Neuronal Pairs
Correlation is investigated for neuronal pairs of two multi-unit clusters. The two multi-unit clusters are
characterized by their spectral and relative spatial distance. The spectral distance is the absolute differ-
ence between the characteristic frequencies of the two multi-unit clusters. Pairs differing by more than
1 /3 octave in their characteristic are most probably from different isofrequency lamina along the tonotopic
gradient (Schreiner and Langner, 1988). These pairs were assigned to the group of differently frequency-
tuned neurons. Those pairs that have the same preferred frequency within an interval of 1 /3 octave are
likely from the same isofrequency lamina (Schreiner and Langner, 1988) and were assigned to the group of
similarly frequency-tuned neurons.
The spatial distances between the multi-unit clusters of all pairs were mapped according to the channels
on the electrode. Distances between all 32 channels were either obtained directly from the NeuroNexus
manual (NeuroNexus Catalog, Ann Arbor, MI 48108, 2014) or calculated using the Pythagorean theorem,
yielding a 32 × 32 matrix, respectively for the linear double-shank and the 4-double-tetrode electrode ar-
ray. Minimum and maximum distances for the double-shank and double-tetrode array were respectively
Dmin =100 µm, Dmax =1581 µm and Tmin =25 µm, Tmax =1372 µm.
The correlation analysis was performed for pairs
yielding ide-
ally (32× 31)/2 =496 pairs, which do not include autocorrelations or correlations counted twice. How-
ever, only pairs with a minimum distance of 200 µm were considered for the analysis of similarly tuned
neurons (respectively 100 µm for differently tuned neurons) (Buzs´aki, 2004; Malmierca et al., 1995), in
order to assure that no multi-unit clusters are taken from adjacent recording sites, which possibly share the
same neurons and therefore yield excessive high correlation values. The distribution of spatial distances
was not uniform and also varied for double-shank and tetrode recordings. Therefore, the number of multi-
unit clusters, for which correlation values were averaged for one spatial distance, varied across recordings.
Differences in response similarity based on different amplitude modulation preferences for multi-unit clus-
ters with the same frequency tuning might be averaged out when taking the mean across multi-unit pairs.
Note that the use of multi-unit clusters for the study could limit the ability to assess the degree to which
correlated firing may encode temporal features in the vocalizations.
Correlation is computed between multi-unit responses to the same vocalization, from one recording set
which consists of responses from 32 multi-unit clusters. The analysis is repeated for all 36 recordings and
analyzed for each recording. It was verified that the observed trend is consistent across recording sets. In
the following, results for an individual example recording set are shown and the average across all recording
sets (1,152 multi-unit clusters) is displayed. When averaging values across multi-unit pairs, the displayed
error bars were chosen to represent one standard deviation to indicate correlation variability across multi-
unit pairs. The error of the correlation values were computed via error propagation and are minor.
from the same recording,
3 Results
We analyzed response and neural correlation from 1,152 multi-unit clusters across a wide frequency range
of the central inferior colliculus of 11 guinea pigs for a spectrotemporally rich set of 11 species-specific
vocalizations. Using cross-correlation, for spiking and LFP activity, we tested variation of similarity for
individual vocalizations across the ICC and investigated whether correlation values depend on the spatial
distance between multi-unit cluster. We compare response correlations and the contributions respectively
due to the stimulus and due to neural interactions. At first, we display time-averaged neural responses to
vocalizations, and then show correlations for similarly and differently tuned neurons.
7
Figure 4: Vocalization Post-stimulus time histograms. PSTHs in response to the vocalizations a) 'tooth chat-
ter', b) 'purr', and c) 'long scream'; for multi-unit clusters from of a linear double-shank recording, spanning a best
frequency range of 0.6-20 kHz. On top, the waveforms and spectrograms are displayed.
3.1 Neural Responses
The post-stimulus time histogram (PSTH) represents the trial-averaged (ntrial =20) temporal neural re-
sponse. Responses to vocalizations vary for differently frequency-tuned multi-unit clusters and follow the
spectrally matching components in the stimulus. Figure 4 displays the PSTHs of multi-unit clusters along
the best frequency (BF) gradient in response to three vocalizations. Responses to the 'tooth chatter' phase-
lock to the stimulus envelope for multi-unit clusters throughout the whole best frequency range (Fig. 4a),
because the stimulus has spectral energy in this range (Fig. 3a). However, the response in general becomes
broader for high frequency neurons. Responses to the 'purr', on the other hand, phase-lock accurately only
for low frequency neurons, but then become broad and unspecific (Fig. 4b). Spectral energy is present for
frequencies up to 3 kHz in the 'purr' vocalization. At the start and the end of the stimulus presentation,
onset and rebound responses are more pronounced for middle and high best frequency neurons. The re-
sponses' dependence on the match of the best frequency and the spectral content of the stimulus (Suta et
al., 2003) is clearly illustrated by responses to the 'scream' (Fig. 4c). In the beginning of the stimulus,
only low frequencies are present, and only low-BF multi-unit clusters respond. Subsequently, the stimulus
contains frequencies up to 25 kHz and middle-BF multi-unit clusters respond. High-BF multi-unit clusters
respond to a high frequency peak at 600 ms. In some cases, these multi-unit responses can be approximated
by the bandpass filtered waveform of the vocalization, filtered around the best frequency of the multi-unit
cluster (Lyzwa, 2015), hence it would be likely that clusters with the same best frequency also have higher
correlations of their responses. Similarity of responses to the same stimulus by different multi-unit clusters
can be directly obtained by comparing the PSTHs. In order to quantify similarities across the large data set
of multi-unit clusters used in this work, a more compact measure is employed. To this end, responses are
cross-correlated and the correlation value indicates the degree of response similarity.
3.2 Dependence on Similarity of Frequency Tuning
Responses correlations to vocalizations are compared for multi-unit cluster with similar and different fre-
quency tuning.
8
Time [ms] Time [ms] Time [ms] Best Frequency [kHz] a Figure 5: Correlations for similarly and differently frequency-tuned neurons. Averaged correlation values for
multi-unit pairs a) from one recording (n = 155), and b) averaged across all recordings. For comparison, the number of
pairs was kept constant for the two cases, n = 4223. Top: Spiking activity, bottom: local field potentials. Responses for
similarly frequency-tuned neurons are mostly significantly more correlated than for differently frequency-tuned neurons
(two-sided Wilcoxon-Mann-Whitney test p = 0.05).
In general, neurons within the same isofrequency lamina have similar preferred frequency (within 1/3 octave
(Schreiner and Langner, 1988)) but possibly different preferences for amplitude modulations (Schreiner and
Langner, 1988; Langner et al., 2002; Baumann et al., 2011). Neurons with different frequency tuning prop-
erties are in different laminae and along the tonotopic gradient. As an example for multi-unit pairs from
only one recording, Figure 5a displays the correlation values for each vocalization, for spiking activity and
local field potentials. Correlation values show large variability within one recording set, as depicted by the
error bars which correspond to one standard deviation. However, correlation values are significantly higher
for similarly frequency-tuned pairs than for neurons with different best frequencies as assessed by the two-
sided Wilcoxon-Mann-Whitney test, p < 0.05, for all vocalizations (except for the 'tooth chatter' in LFPs).
These correlation values vary across the different recording sets, because of the different distributions of
frequency tuning similarity and spatial distances between the neuron pairs. For the average correlation
values across all recordings (n =4223 multi-unit pairs), differences in the correlation are significant for all
vocalizations for LFP activity, Fig. 5b. Differences in correlation for similar and different frequency tuning
are smaller for spiking activity than for LFP and not significant for the 'squeal' and 'tooth chatter', Fig. 5b.
The 'tooth chatter' shows strong phase-locking across the tonotopic gradient, the 'squeal' displays partly
spectral energy across frequency bands and therefore also a neural response across frequency bands. The
variability of correlation across multi-unit pairs, which is large, is depicted by the error bars. The errors
computed via error propagation are very small. Significance is evaluated by the two-sided Wilcoxon-Mann-
Whitney test, p < 0.05.
Correlation values are significantly higher for LFP than for spiking activity for each vocalization, for sim-
ilarly and differently frequency-tuned neurons (p < 0.05). The LFP responses are long-range activity, and
spread throughout wider spatial and frequency regions than spiking responses which are confined to the
range of one multi-unit cluster. Across vocalizations, response similarity varies significantly, except for
(2,4;9,10), respectively for spiking (LFP) responses for similar and different neural frequency preference.
The responses to the 'tooth chatter' are significantly larger than to all other vocalizations, which is due to
the stimulus' spectral content across a wide frequency range and the responses' phase-locking throughout
this range (see Fig. 4a).
Higher correlation within similar frequency regions than across the frequency gradient might point to vo-
calizations being processed by isofrequency laminae as functional units (Schreiner and Langner, 1997),
however, even though the differences are significant, they are minor and variability exists across multi-unit
clusters. In the following, correlation dependence on spatial distance will be investigated separately for
these two groups.
9
0.10.40.7spiking activityCorrelation 0.10.40.71Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrLFPCorrelation ns0.10.4spiking activityCorrelation diff. iso. laminaesame iso.lamina0.10.40.71Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrLFPCorrelation nsnsabDifferent Freq.SimilarFreq.Figure5:Correlationsforsimilarlyanddifferentlyfrequencytunedneurons.Averagedcorrelationvaluesformulti-unitpairsa)fromonerecording(n=155),andb)averagedacrossallrecordings.Forcomparison,thenumberofpairswaskeptconstantforthetwocases,n=4223.Top:Spikingactivity,bottom:localfieldpotentials.Responsesforsimilarlyfrequencytunedneuronsaremostlysignificantlymorecorrelatedthanfordifferentlyfrequencytunedneurons(two-sidedWilcoxon-Mann-Whitneytestp=0.05).Ingeneral,neuronswithinthesameisofrequencylaminahavesimilarpreferredfrequency(within1/3octave(SchreinerandLangner,1988))butpossiblydifferentpreferencesforamplitudemodulations(SchreinerandLangner,1988;Langneretal.,2002;Baumannetal.,2011).Neuronswithdifferentfrequencytuningprop-ertiesareindifferentlaminaeandalongthetonotopicgradient.Asanexampleformulti-unitpairsfromonlyonerecording,Figure5adisplaysthecorrelationvaluesforeachvocalization,forspikingactivityandlocalfieldpotentials.Correlationvaluesshowlargevariabilitywithinonerecordingset,asdepictedbytheerrorbarswhichcorrespondtoonestandarddeviation.However,correlationvaluesaresignificantlyhigherforsimilarlyfrequencytunedpairsthanforneuronswithdifferentbestfrequenciesasassessedbythetwo-sidedWilcoxon-Mann-Whitneytest,p<0.05,forallvocalizations(exceptforthe'toothchatter'inLFPs).Thesecorrelationvaluesvaryacrossthedifferentrecordingsets,becauseofthedifferentdistributionsoffrequencytuningsimilarityandspatialdistancesbetweentheneuronpairs.Fortheaveragecorrelationvaluesacrossallrecordings(n=4223multi-unitpairs),differencesinthecorrelationaresignificantforallvocalizationsforLFPactivity,Fig.5b.DifferencesincorrelationforsimilaranddifferentfrequencytuningaresmallerforspikingactivitythanforLFPandnotsignificantforthe'squeal'and'toothchatter',Fig.5b.The'toothchatter'showsstrongphase-lockingacrossthetonotopicgradient,the'squeal'displayspartlyspectralenergyacrossfrequencybandsandthereforealsoaneuralresponseacrossfrequencybands.Thevariabilityofcorrelationacrossmulti-unitpairs,whichislarge,isdepictedbytheerrorbars.Theerrorscomputedviaerrorpropagationareverysmall.Significanceisevaluatedbythetwo-sidedWilcoxon-Mann-Whitneytest,p<0.05.CorrelationvaluesaresignificantlyhigherforLFPthanforspikingactivityforeachvocalization,forsim-ilarlyanddifferentlyfrequencytunedneurons(p<0.05).TheLFPresponsesarelong-rangeactivity,andspreadthroughoutwiderspatialandfrequencyregionsthanspikingresponseswhichareconfinedtotherangeofonemulti-unitcluster.Acrossvocalizations,responsesimilarityvariessignificantly,exceptfor(2,4;9,10),respectivelyforspiking(LFP)responsesforsimilaranddifferentneuralfrequencypreference.Theresponsestothe'toothchatter'aresignificantlylargerthantoallothervocalizations,whichisduetothestimulus'spectralcontentacrossawidefrequencyrangeandtheresponses'phase-lockingthroughoutthisrange(seeFig.4a).Highercorrelationwithinsimilarfrequencyregionsthanacrossthefrequencygradientmightpointtovo-calizationsbeingprocessedbyisofrequencylaminaeasfunctionalunits(SchreinerandLangner,1997),however,eventhoughthedifferencesaresignificant,theyareminorandvariabilityexistsacrossmulti-unitclusters.Inthefollowing,correlationdependenceonspatialdistancewillbeinvestigatedseparatelyforthesetwogroups.9 Figure 6: Correlation dependence on spatial distance for differently frequency-tuned neurons. Averaged corre-
lations values for each distance from one recording set are displayed for all vocalizations, for a) spiking activity, and
b) Local field potentials. Correlations decrease almost linearly with spatial distance, with a much bigger decrease for
spiking activity than for LFP. Exponential regression of this decrease ( -- ), with overlaps of 61-100 %, see Tab. 1. On top,
the number of multi-unit pairs for each spatial distance over which the average correlation was computed is displayed.
The y-axis for each vocalization shows a correlation range of 0-1.
3.3 Dependence on Spatial Distance
In order to display the relation between correlation values and the spatial distance between neurons, the
values were averaged for multi-unit clusters for each spatial distance of one recording set. The relations
are shown for spiking activity and LFPs, an example for one recording set of pairs with different (Fig. 6)
and similar frequency tuning (Fig. 7) is given. Correlations decrease with spatial distance and are almost
zero for distances of 400 µm for the spiking responses. LFP correlations are overall higher than the ones
for spiking activity and display a less rapid decrease with distance, correlation values of about 0.5 still exist
for the maximum measured distance of 1600 µm (Fig. 6). LFP is long range activity and correlations are
present over long distances. For multi-unit clusters from one recording set, that have similar frequency
tuning this decrease is also observed (Fig. 7). These findings are consistent across all 36 analyzed recording
sets. Neural and stimulus correlations (3.4) follow the same decrease as the response correlations.
The decrease can be approximated with an exponential function f (x) = a· e−bx, with x the spatial distance.
In Tab. 1 values for a, b and the match of data and fit χ are given.
Activity
spike
LFP
spike
LFP
Frequency Tuning
a
different
different
similar
similar
0.01-0.59
0.76-1
0.01-0.39
0.73-1
b[1/µm]
0.57 -0.0069
0.012-0.003
0.48-0.0041
0.042-0.002
χ [%]
84-100
61-94
95-100
42-93
(cid:104)χ(cid:105) [%]
91
80
98
67
Table 1: Parameter for exponential regression. f (x) = a· e−bx, with spatial distance x, for Fig. 6a,b, Fig. 7a,b.
Dependencies on the spatial distance differ across vocalizations. Figure 8 shows the correlations displayed
in Fig. 6, 7 in a smaller window for distances up to 650 µm for all eleven vocalizations for spiking and
LFP activity. The 'tooth chatter' shows overall highest response correlations. A clear correlation decrease
is observed from 200-500 µm for differently (Figure 8a, c) and similarly ( 8b, d) frequency-tuned neurons
for all vocalizations.
10
Lyzwaetal.Correlationstonaturalsoundsinmidbrain0.10.40.7spiking activityCorrelation 0.10.40.71Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrLFPCorrelation ns0.10.4spiking activityCorrelation diff. iso. laminaesame iso.lamina0.10.40.71Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrLFPCorrelation nsnsabDifferent Freq.SimilarFreq.Figure5.Correlationsforsimilarlyanddifferentlyfrequencytunedneurons.Averagedcorrelationvaluesformulti-unitpairsa)fromonerecording(n=155),andb)averagedacrossallrecordings.Forcomparison,thenumberofpairswaskeptconstantforthetwocases,n=4223.Top:Spikingactivity,bottom:localfieldpotentials.Responsesforsimilarlyfrequencytunedneuronsaremostlysignificantlymorecorrelatedthanfordifferentlyfrequencytunedneurons(two-sidedWilcoxon-Mann-Whitneytestp=0.05).13260 300 600 900 12001500Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrSpatial distance [um]0 300 600 900 12001500Spatial distance [um]LFP# Pairsspiking activityabSpaƟaldistance[μm]SpaƟaldistance[μm]1CorrelaƟon11111111111Figure6.Correlationdependenceonspatialdistancefordifferentlyfrequencytunedneurons.Averagedcorrelationsvaluesforeachdistancefromonerecordingsetaredisplayedforallvocalizations,fora)spikingactivity,andb)Localfieldpotentials.Correlationsdecreasealmostlinearlywithspatialdistance,withamuchbiggerdecreaseforspikingactivitythanforLFP.Exponentialregressionofthisdecrease( -- ),withoverlapsof61-100%,seeTab.1.Ontop,thenumberofmulti-unitpairsforeachspatialdistanceoverwhichtheaveragecorrelationwascomputedisdisplayed.(They-axisforeachvocalizationshowsacorrelationrangeof0-1.)distance.InTab.1,valuesfora,bandamatchofdataandfitχaregivenfortheexamplesdisplayedin350Fig.6,7.351Dependenciesonthespatialdistancedifferacrossvocalizations.Figure8showsthecorrelationsdisplayed352inFig.6,7inasmallerwindowfordistancesupto650µmforallelevenvocalizations.Aclearcorrelation353decreaseisobservedfrom200-500µmfordifferently(Figure8a,c)andsimilarly(8b,d)frequencytuned354neurons.355Thisisaprovisionalfile,notthefinaltypesetarticle10 Figure 7: Correlation dependence on spatial distance for similarly frequency-tuned neurons. Averaged correla-
tions values for each distance from one recording set are displayed for all vocalizations, for a) spiking activity, and
b) LFPs. Exponential regression of this decrease ( -- ), with overlaps of 41-100 % see Tab. 1. On top, the number of
multi-unit pairs for each spatial distance over which the average correlation was computed is displayed. Distances 200
µm-1000 µm for similarly tuned neurons are analyzed, see 2.2.
Figure 8: Comparison of correlations for all vocalizations. Correlations are shown for differently (a,b) and similarly
(c,d) frequency-tuned neurons for spiking activity (a,c) and LFPs (b,d), displayed in Fig. 6, 7. Differences in total
value of response correlations exist across vocalizations. For each vocalization all correlations display an almost linear
decrease with spatial distance between neurons.
11
Lyzwaetal.Correlationstonaturalsoundsinmidbrain8 160 300 600 900 Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurr0 300 600 900 spiking activityLFPb# PairsaSpaƟaldistance[μm]SpaƟaldistance[μm]SpaƟaldistance[μm]SpaƟaldistance[μm]111111111111CorrelaƟonSpaƟaldistance[μm]SpaƟaldistance[μm]11111111# Pairs# PairsFigure7.Correlationdependenceonspatialdistanceforsimilarlyfrequency-tunedneurons.Averagedcorrelationsvaluesforeachdistancefromonerecordingsetaredisplayedforallvocalizations,fora)spikingactivity,andb)LFPs.Exponentialregressionofthisdecrease( -- ),withoverlapsof41-100%seeTab.4.2.Ontop,thenumberofmulti-unitpairsforeachspatialdistanceoverwhichtheaveragecorrelationwascomputedisdisplayed.(Onlydistances≥200µ−1000µforsimilarlytunedneuronsareanalyzed,see2.2.)ActivityFreq.tuningab/µmχ[%]hχi[%]spikedifferent0.01-0.590.57-0.006984-10091LFPdifferent0.76-10.012-0.00361-9480spikesimilar0.01-0.390.48-0.004195-10098LFPsimilar0.73-10.042-0.00242-9367Table1Parameterforexponentialregression.Forf(x)=a·e−bx,withxthespatialdistance,respectivelyforFig.6a,bandFig.7a,b.Thisisaprovisionalfile,notthefinaltypesetarticle16Figure7:Correlationdependenceonspatialdistanceforsimilarlyfrequency-tunedneurons.Averagedcorrela-tionsvaluesforeachdistancefromonerecordingsetaredisplayedforallvocalizations,fora)spikingactivity,andb)LFPs.Exponentialregressionofthisdecrease( -- ),withoverlapsof41-100%seeTab.1.Ontop,thenumberofmulti-unitpairsforeachspatialdistanceoverwhichtheaveragecorrelationwascomputedisdisplayed.Onlydistances200µm-1000µmforsimilarlytunedneuronsareanalyzed,see2.2.10020030040050000.10.20.3Correlation1002003004005000.40.50.60.70.80.9 Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrLFPspiking activity10020030040050000.10.20.30.40.5Correlation1002003004005000.50.60.70.80.91 Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrspiking activityLFPbacdSpaƟaldistance[μm]SpaƟaldistance[μm]Figure8:Comparisonofallvocalizationsforsmallspatialdistances.Correlationsareshownfordifferently(a,b)andsimilarly(c,d)frequencytunedneuronsforspikingactivity(a,c)andLFPs(b,d),displayedinFig.6,7.Differencesintotalvalueofresponsecorrelationsexistacrossvocalizations.Butforeachvocalizationallcorrelationsdisplayanalmostlineardecreasewithspatialdistancebetweenneurons.11 Similarity of responses decreases almost linearly with spatial distance, and is almost zero for distances
above 400 µm for spiking activity, whereas LFP shows stronger and longer range correlations even for
the maximum measured distance.The parameter for the decrease are within the same range for similarly
and differently frequency-tuned neurons, but the decrease is much smaller for LFP than for spiking activity
(Tab. 1). This gradual decrease within similar and different frequency regions suggests that the neural
response is strongly influenced by the gradually organization of spectral and temporal preferences in the
ICC.
3.4 Neural Correlations
Comparison of correlations from simultaneous responses to those of non-simultaneous responses gives an
indication of the amount of correlations due to neural interactions between the multi-unit clusters. It has
been shown previously that the sum of the stimulus-driven and neural correlations does not necessarily yield
the response correlation (Melssen and Epping, 1987).
Figure 9 displays all three types: the averaged response correlations, the correlations due to the stimulus
and the neural correlations, for multi-unit pairs with either similar or different frequency tuning, both, for
spiking (a) and LFP (b) activity. The displayed error was obtained via error propagation. Correlation values
vary across multi-unit clusters as displayed in Fig. 5.
Response correlations of the spiking activity are significantly larger than non-simultaneous correlations,
and these are larger than the neural correlations. The difference between response, stimulus-driven and
neural correlations is significant for all vocalizations for spiking and LFP activity (Fig. 9). Only the 'whis-
tle', which has the overall lowest response correlations, does not show significant difference between the
stimulus and neural correlation of spiking activity for differently frequency-tuned multi-unit pairs, Fig. 9a,
top. Differences between response and stimulus correlations are not significant for all vocalizations within
each recording set, but are significant when averaged across all multi-unit cluster.
Figure 9: Response, stimulus-driven and neural correlations, averaged across all recordings. Correlations for spiking
activity, for (a) differently (n=10,775 pairs); and (c) similarly frequency-tuned neurons (n=4230 pairs); and correlations
for LFP activity for (b) differently, and (d) similarly frequency-tuned neurons. Non-significant differences are denoted
by ns, all other differences are significant (p = 0.05).
Differences between response, stimulus-driven and neural correlations are larger for LFP than for spiking
activity. These are significant for correlations of all vocalizations in single recording sets and for averaged
values across all multi-unit clusters. For the LFP activity, neural correlations are even significantly higher
than stimulus correlations for all vocalizations but three, for both, similarly and differently frequency tuned
multi-unit pairs (Fig. 9b).
12
Similarityofresponsesdecreasesalmostlinearlywithspatialdistance,andisalmostzerofordistancesabove400µforspikingactivity,whereasLFPshowsstrongerandlongerrangecorrelationsevenforthemaximummeasureddistance.Theparameterforthedecreasearewithinthesamerangeforsimilarlyanddifferentlyfrequencytunedneurons,butthedecreaseismuchsmallerforLFPthanforspikingactivity(Tab.1).ThisgradualdecreasewithinsimilaranddifferentfrequencyregionssuggeststhattheneuralresponseisstronglyinfluencedbythegraduallyorganizationofspectralandtemporalpreferencesintheICC.3.4NeuralCorrelationsComparisonofcorrelationsfromsimultaneousresponsestothoseofnon-simultaneousresponsesgivesanindicationoftheamountofcorrelationsduetoneuralinteractionsbetweenthemulti-unitclusters.Ithasbeenshownpreviouslythatthesumofthestimulus-drivenandneuralcorrelationsdoesnotnecessarilyyieldtheresponsecorrelation(MelssenandEpping,1987).Figure9displaysallthreetypes:theaveragedresponsecorrelations,thecorrelationsduetothestimulusandtheneuralcorrelations,formulti-unitpairswitheithersimilarordifferentfrequencytuning,both,forspiking(a)andLFP(b)activity.Thedisplayederrorwasobtainedviaerrorpropagation.Correlationvaluesvaryacrossmulti-unitclustersasdisplayedinFig.5.Responsecorrelationsofthespikingactivityaresignificantlylargerthannon-simultaneouscorrelations,andthesearelargerthantheneuralcorrelations.Thedifferencebetweenresponse,stimulus-drivenandneuralcorrelationsissignificantforallvocalizationsforspikingandLFPactivity(Fig.9).Onlythe'whistle',whichhastheoveralllowestresponsecorrelations,doesnotshowsignificantdifferencebetweenthestimulusandneuralcorrelationofspikingactivityfordifferentlyfrequencytunedmulti-unitpairs,Fig.9a,top.Differencesbetweenresponseandstimuluscorrelationsarenotsignificantforallvocalizationswithineachrecordingset,butaresignificantwhenaveragedacrossallmulti-unitcluster.0.10.3Correlation 0.10.3Scream longScream shortSquealTooth ChatterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrCorrelation nsspiking activityac0.10.50.9LFP 0.10.50.9Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurr bd ResponseStimulusNeuralFigure9:Response,stimulus-drivenandneuralcorrelations,averagedacrossallrecordings.a)correlationsforspikingactivity,respectivelyacrossdifferentandwithinthesameisofrequencylamina(n=10,775pairs);andb)correlationsforLFPactivity(n=10,800pairs).Non-significantdifferencesaredenotedbyns,allotherdifferencesaresignificant(p=0.05).Differencesbetweenresponse,stimulus-drivenandneuralcorrelationsarelargerforLFPthanforspikingactivity.Thesearesignificantforcorrelationsofallvocalizationsinsinglerecordingsetsandforaveragedvaluesacrossallmulti-unitclusters.FortheLFPactivity,neuralcorrelationsareevensignificantlyhigherthanstimuluscorrelationsforallvocalizationsbutthree,forboth,similarlyanddifferentlyfrequencytunedmulti-unitpairs(Fig.9b).12 The three exceptions, the 'tooth chatter', 'short chutter', and 'squeal', show energy across the whole fre-
quency range at short time intervals in the vocalization spectrograms, Fig. 3 and (Lyzwa et al., 2016). For
the spiking activity, these three vocalizations also show the overall smallest amount of neural correlations
relative to the response correlations.
To summarize, neural correlations which are most likely due to interactions between multi-unit clusters
exist in the central inferior colliculus for spiking activity and for local field potentials. These correlations
exist between similarly and differently frequency-tuned multi-unit activity. They are significant but minor
for spiking activity, and much larger for LFPs which is long range activity (Pettersen et al., 2012). Thus,
the responses are also shaped by the interactions between neurons.
Figure 10: Separation of spike rates with correlated trial-variability. The scatter plots of the average spike rates
in response to the 11 vocalizations for simultaneous (UnShuffled) and non-simultaneous (Shuffled) trials are compared
for 4 multi-unit pairs, for units with different (a, c) or similar frequency tuning (b, d); and which are spatially relatively
close (a, b) or far away (c, d). a) BF1,2 = 5 kHz, 8 kHz; 360 µm;b) BF1,2 = 2.25 kHz, 2.83 kHz; 400 µm;
c) BF1,2 = 8 kHz, 2.25 kHz; 1270 µm;d) BF1,2 = 2 kHz, 2.25 kHz; 800 µm. The colors for the different vocaliza-
tions correspond to those in Fig. 8.
Neural correlations exist in the inferior colliculus. Interactions between neurons can lead to a co-variation
of their trial-to-trial variability of the spiking responses, which might also manifest in a correlated trial-
variability of their spike-rates. The averaged stimulus-elicited spike rate for each trial (n=20), for multi-unit
pairs are compared for simultaneous and non-simultaneous responses to investigate whether a decorrela-
tion (shuffling) induces better separability of the spike rates to different vocalizations. Thus, we explored
whether the contribution of neural correlations to separate spike rates of different vocalizations differed
among multi-unit pairs with a small or large difference either in frequency tuning, or in spatial distance.
Figure 10 displays examples of scatter plots for multi-unit pairs with similar and different frequency tuning,
and for distant and relatively close-by multi-unit clusters. Although shuffling changes the distribution of
responses, separability of the 11 vocalizations does not change substantially in any of the cases. This is true
across all recording sets and was shown here for 4 examples.
Correlated trial-variability due to interactions between neurons in the ICC does exist but does not alter
separability of the multi-units' spike rates to vocalizations, although at a single neuron level this might be
different.
13
Lyzwaetal.CorrelationstonaturalsoundsinmidbrainmammalianICC.393Dependenciesonspatialdistancehavenotbeenfoundforthegrassfrogmidbrain(EppingandEggermont,3941987),buthavebeenshowninthemammalianICC(Chenetal.,2012)andprimaryauditorycortex395(Eggermont,2006).Responsecorrelationsandneuralcorrelationshavebeenanalyzedforsingleneuronsin396themammalianICC,onlyforrelativelyshortdistancesbelow370um,(Chenetal.,2012),showingthat397nearbyneuronshaveahigherprobabilityofdisplayingsimilarneuralpreferencesandresponses,suggesting398amicrocircuitry.399Contrarytothesestudies,intheherepresentedwork,alargerspatialdistanceiscovered,andresponses400ofmulti-unitstonaturalsoundstimuliareinvestigated.401EppingandEggermont(EppingandEggermont,1987)analyzedneuralactivityof150sortedunitsinthe402auditorymidbrainofthegrassfrog.Usingcross-correlation,theyfoundresponsecorrelationsfor60%403oftheneuronsand15%oftheneuronsdisplayedcorrelationsduetoneuralinteractionsandconnections.404Theseneuralcorrelationswererestrictedtopairswithdistanceslessthan300µm,whereastheresponse405correlationswereindependentofthespatialdistance.Theauthorssuggestedthatthehighamountof406responsecorrelationsrelativetothesmallamountofneuralcorrelationsindicatesthatinordertocreate407theneuralresponse,thestimulusinputplaysapredominantroleoverneuralmechanisms(Eppingand408Eggermont,1987).Theyattributedthistoaspreadprojectionofthestimulusinput,ratherthantorestricted409areas,whichwouldbeinlinewithaweaktonotopicorganizationinthegrassfrogmidbrainandtothe410findingthatneighboringneuronsoftendisplaydifferentspectralpreferences.Incontrasttothegrasshopper411midbrain,themammalianICChasacleartonotopicstructurewithneighboringneuronsdisplayingsimilar412spectralpreferences.(Thus,thecorrelationstructureasdescribedbyEppingandEggermontislikelyvery413differentforthemammalianICC.)414-Responsecorrelationsdiffer,asmanydifferentinoutsandthespectrotemporalproertoeschangeclearly,415suchasthetontopicgradient,hencetheoverallresponse.Correlationsexistforall.-similarforsame416frequencyanddifferentfrequency.417100200300400500100200300400500UnShuffled same iso. lamina100200300400500Shuffled Scream longScream shortSquealTooth ChutterWhistleChutter longChutter shortLow WhistleLow ChutterDrrPurrb100200300400100200300400Unit 1 Spike Rate [Hz]100200300400Unit 1 Spike Rate [Hz] d100200300400100200300400Unit 1 Spike Rate [Hz]Unit 2 Spike Rate [Hz]100200300400Unit 1 Spike Rate [Hz] c100200300400500100200300400500Unit 2 Spike Rate [Hz]UnShuffled different iso. laminae100200300400500Shuffled aFigure11.Separationofspikerateswithcorrelatedtrial-variability.Thescatterplotsoftheaveragespikeratesinresponsetothe11voca-lizationsforsimultaneous(UnShuffled)andnon-simultaneous(Shuffled)trialsarecomparedfor4multi-unitpairs,forunitswithdifferent(a,c)orsimilarfrequencytuning(b,d);andwhicharespatiallyrelativelyclose(a,b)orfaraway(c,d).a)BF1,2=5kHz,8kHz;360µm;b)BF1,2=2.25kHz,2.83kHz;400µm;c)BF1,2=8kHz,2.25kHz;1270µm;d)BF1,2=2kHz,2.25kHz;800µm.ThecolorsforthedifferentvocalizationscorrespondtothoseinFig.??.Thisisaprovisionalfile,notthefinaltypesetarticle14 4 Discussion
In this work, we analyzed similarity of multi-unit responses to a set of 11 vocalizations across the mam-
malian central inferior colliculus. Our findings are based on a large set of multi-unit clusters (N =1,152), of
which the best frequencies span a range between 0.5-45 kHz. The studied vocalizations are a representative
set of behaviorally relevant stimuli (Berryman, 1976).
Neurons with similar frequency tuning (1/3 octave) have significantly higher correlation values than differ-
ently tuned neurons (Fig.5) though differences were minor and correlation varied across multi-unit cluster.
Regions of similar frequency tuning, the isofrequency laminae have been suggested earlier to be functional
processing units (Schreiner and Langner, 1997). We investigated response similarity separately for these
two groups.
4.1 Spatial Dependance
In contrast to previous studies (Chen et al., 2012; Epping and Eggermont, 1987; Eggermont, 2006), this
work is a first systematic investigation of response similarity to vocalizations in dependance of spatial dis-
tance in the mammalian ICC.
Response correlations and neural correlations to artificial sound have been analyzed for single neurons in
the cat ICC, for relatively short distances up to 370 µm (Chen et al., 2012), showing that nearby neurons
have a higher probability of displaying similar neural preferences and responses, and suggesting a microcir-
cuitry. Epping and Eggermont (Epping and Eggermont, 1987) analyzed neural activity of 150 sorted units
in the auditory midbrain of the grass frog. Using cross-correlation, they found response correlations for 60
% of the neurons and 15 % of the neurons displayed correlations due to neural interactions and connections.
These neural correlations were restricted to pairs with distances less than 300 µm, whereas the response
correlations were independent of the spatial distance. The authors suggested that the high amount of re-
sponse correlations relative to the small amount of neural correlations indicates that in order to create the
neural response, the stimulus input plays a predominant role over neural mechanisms (Epping and Egger-
mont, 1987). They attributed this to a spread projection of the stimulus input, rather than to restricted areas,
which would be in line with a weak tonotopic organization in the grass frog midbrain and to the finding that
neighboring neurons often display different spectral preferences.
In contrast to the grasshopper midbrain, the mammalian ICC has a clear tonotopic structure with neighbor-
ing neurons displaying similar spectral preferences. The correlation structure as described by Epping and
Eggermont is very different for the mammalian ICC. We showed that response similarity depends on the
spatial distance between two multi-unit clusters, and decreases exponentially with increasing distance. In
general, for distances above 400 µm (Fig. 6a, 7a, 7a,c), very little (≤ 0.11) correlation of spiking responses
is present. Correlations vary across vocalizations but all display the linear dependence on spatial distance
between neurons (Fig. 8). The 'tooth chatter' which shows phase-locking to the stimulus envelope through-
out a very large frequency range (Fig. 4), displays the highest correlations. Responses from local field
potentials have much higher correlations, and a more flat, also linear, decrease and display high correlations
( 0.5) for distances as high as 1600 µm (Fig. 6b, 7b, 7b,d). These large correlations lengths are due to the
local field potential being long range activity.
The decrease could be exponentially fitted with amplitude and decay parameters ranging between a =
0.01 − 0.59 and b = 0.041 − 0.57/µm for spiking activity, and the ranges for similarly and differently
frequency-tuned neurons overlapped in parts (Tab. 1). In a higher auditory processing station, in the cor-
tex, a correlation dependance on the spatial distance has also been demonstrated. Eggermont (Eggermont,
2006) analyzed neural groups, that reflect patched activity and were termed 'clusters', in the cat primary
auditory cortex with the use of cross-correlation matrices of spontaneous activity. The author found that the
correlation followed an exponential decrease with distance in mm, a=0.05, b=0.24/mm.
14
Our obtained values for the maximum amplitude fall within the same order, but the decay parameter b
obtained in our study is at least a 17 times larger than the one found in (Eggermont, 2006). Eggermont
found this dependence for spontaneous activity of neural groups in the cat primary auditory cortex which
is also tonotopically organized. The steeper decrease in our work might be explained by the smaller size
and mapping space available for the ICC compared to the primary auditory cortex, and by the different
and smaller animal studied. It has been suggested that the functional organization is dynamic, and that the
functional connections depend on the particular stimulus applied (Epping and Eggermont, 1987); thus, the
use of responses to natural communications sounds instead of responses to spontaneous activity could also
account for this difference in spatial decrease.
In this analysis, correlations for two multi-unit clusters with similar frequency tuning within 1/3 octave that
are separated by at least 200 µm were analyzed (respectively 100 µm for differently tuned neurons), in order
to ensure comparison of different signals. However, neural responses from a radius of more than 200 µm
might be picked up by the recording electrode.
In order to exclude this possibility, and to investigate
correlation dependencies for smaller distances (<100 µm), single neuron recordings could be used. In
the present analysis, we compare within 1/3 octave similarly and differently frequency-tuned multi-unit
cluster. This allows one to make inferences about the organization of response similarity within the ICC
insofar as neurons within an isofrequency laminae have the same best frequency within this interval of 1/3
octave, hence if the neurons differ in best frequency by more than this interval, they are likely in different
isofrequency laminae. Histological stains would allow to obtain the exact location of the the multi-unit pairs
within the ICC. Furthermore, knowing the positions of the neurons would allow mapping out and testing the
dependence of the response similarity on the specific location within the ICC. Additionally, this study could
be complemented by single neuron responses which would further reduce ambiguity of the exact position
of the neural response. Another interesting and challenging investigation would be to label synaptic input in
the ICC from the different ascending brainstem nuclei and measure the position of the neuron pairs relative
to these main inputs in the ICC.
We analyzed response similarity for neurons tuned either similarly or differently to the stimulus frequency
and showed a minor but significantly higher correlation for similarly frequency-tuned neurons (Fig 5) with
variability.
It would be informative to study similarity dependence on spatial distance with respect to
neural modulation preferences, the best amplitude modulation frequency of the stimulus envelope. These
neural preferences has been suggested to be arranged in a concentric gradient within isofrequency laminae
(Schreiner and Langner, 1988; Langner et al., 2002; Baumann et al., 2011). This might help to further
discriminate in detail response similarity within and across isofrequency laminae. An ideal experiment to
measure this would be to record simultaneously from several neurons within an isofrequency lamina using
a tetrode array recordings with relatively high impedance to capture single neuron activity. Recordings
would be made in response to vocalizations and in response to dynamic moving ripple sound (Escab´ı and
Schreiner, 2002) which allows to compute the receptive fields and modulation preferences for the single
neurons. Histological stains would inform about the actual positions of the recording sites.
In summary, despite the vocalizations displaying very diverse and inhomogeneous spectral contents, and
despite the locally diverse inputs to the ICC, the neural representations of vocalizations exhibit gradual
organization of the similarity of the neurons' responses. Multi-unit clusters with similar spiking responses
are spatially spread for distances <400 µm, however, differences might exist for single neurons and cannot
be captured at this resolution level. These findings give indications for applications in auditory midbrain
prosthesis, e.g. for the sufficient spatial separation between stimulating electrodes and contacts.
15
4.2 Neural Correlations
For our large set of 1,152 multi-unit cluster, we find that simultaneous response correlations are significantly
higher than the respective non-simultaneous response correlations. Correlations due to interactions between
the multi-unit cluster exist in the mammalian ICC. This finding is very different from a study of Epping and
Eggermont (Epping and Eggermont, 1987) in which they showed that for the grass frog midbrain only 15
% of the units displayed neural correlations. In contrast to the grasshopper midbrain, it has been shown that
the mammalian ICC has a clear tonotopic structure with neighboring neurons displaying similar spectral
preferences, and a gradient of amplitude modulation frequency. We showed significant neural correlations
across the ICC in response. They also show a linear decrease with spatial dependence. Thus, in the mam-
malian ICC neural interactions contribute significantly to shape the neural organization of vocalizations and
the output of this nucleus.
Epping and Eggermont (Epping and Eggermont, 1987) also found stimulus dependencies for half of the
neural correlations, indicating that the functional organization is dynamic, thus the functional connections
depend on the particular stimulus applied. In this work, we analyzed correlations separately for each vo-
calization, in order to account for such possible stimulus dependencies. The amount of neural correlations
varies across vocalization stimuli (Fig. 9), which is also true for the response correlations (Fig. 5).
The neural correlations though much smaller than the response correlations follow the same spatial de-
pendence, as do the stimulus correlations. This could indicate that both effects contribute to the gradual
decrease of response similarity, the gradual changing of spectrotemporal properties with spatial distance
as well as neural interactions which decrease with spatial distance, since the interactions are most likely
between nearby and connected neurons. The neural interactions do not contribute to a better separability
between the time-averaged spike rates across similarly and differently frequency-tuned, distant or close-by
neural pairs (Fig.10), but might further shape the organization of the neural representation within the ICC.
Single neuron responses could be additionally recorded for this study of large spatial distances. Further-
more, single neuron recordings would allow making inferences of connectivities and quantify the amount
of correlations due to neural interactions. However, limitations of the shift predictor exist even for single
neurons such as the obscuring of neural interaction due to deterministic responses or temporal overlap of
stimulus and neural response (Epping and Eggermont, 1987). Probabilistic models based on pairwise inter-
action, describing the weights of interactions in a network could be established (Schneidman et al., 2006).
Anesthesia has been shown to affect neural responses (Astl et al., 1996), and by fluctuating or slowly chang-
ing anesthesia levels also changes the brain state which might affect the cross-correlation strength. Thus,
response similarity and particularly neural correlations might differ in awake animals, however, recordings
in awake animals also bear difficulties and biases. Since anesthesia has non-negligible effects on the neural
activity (Astl et al., 1996), correlation values are likely to improve in awake animals.
In summary, it was found that multi-unit clusters in the ICC display significant neural correlations that are
due to interactions of neurons. These exist for similarly and differently frequency-tuned neurons, decrease
with spatial distance and differ across vocalizations. For LFPs, the neural correlations are even larger than
stimulus correlations for most of the vocalizations. These findings suggests that the neural interactions
shape the spiking output of this nucleus, and the neural representation of vocalizations with gradually de-
creasing similarity in the central IC.
In conclusion, we showed that despite the diverse inputs to the ICC from all ascending projections, termi-
nating in different spatial locations in the ICC (Oliver, 2005), and the rich spectrotemporal properties of
vocalizations, the neural representation shows an organization of gradual decrease in response similarity
with spatial distance, for spiking activity and for local field potentials. Thus, when comparing responses
of neuronal groups to complex sound, such as vocalizations across the ICC, it is important to take their
spatial separation into account and not only their frequency tuning. Our findings suggest that for multi-unit
cluster in the mammalian inferior colliculus, the gradual response similarity with spatial distance to natural
complex sounds is shaped by neural interactions and the gradients of neural preferences.
16
Acknowledgments
We thank Thilo Rode, Tanja Hartmann, Thomas Lenarz and Hubert H. Lim for providing the neural data and
vocalizations. This work was supported by Grant # 01GQ0810 and # 01GQ0811 of the Federal Ministry of
Education and Research within the Bernstein Focus of Neural Technology Gottingen.
References
Abeles M (1982) Local cortical circuits: An electrophysiological study Springer, Berlin.
Aitkin L, Phillips S (1984b) Is the inferior colliculus an obligatory relay in the cat auditory system? Neu-
roscience Letters 44:259 -- 264.
Astl J, Popel´ar J, Kvasnak J, Syka J (1996) Comparison of response properties of neurons in the inferior
colliculus of guinea pigs under different anesthetics. Audiology 35(6):335 -- 345.
Averbeck B, Latham P, Pouget A (2006) Neural correlations, population coding and computation. Nature
reviews. Neuroscience 7:358 -- 366.
Baumann S, Griffiths T, Sun L, Petkov C, Thiele A, Rees A (2011) Orthogonal representation of sound
dimensions in the primate midbrain. Europe PMC Funders Group 14:423 -- 425.
Berryman J (1976) Guinea pig vocalizations: Their structure, causation and function. Z. Tierpsycholo-
gie 41:80 -- 106.
Buzs´aki G (2004) Large-scale recording of neuronal ensembles. Nature Neuroscience 7:446 -- 451.
Chen C, Rodriguez F, Read H, Escab´ı M (2012) Spectrotemporal sound preferences of neighboring inferior
colliculus neurons: implications for local circuitry and processing. Frontiers Neural Circuits. 6:1 -- 16.
Eggermont J (2006) Properties of correlated neural activity clusters in cat auditory cortex resemble those
of neural assemblies. Journal of Neurophysiology 96:746 -- 64.
Engineer CT, Perez C, Chen YH, Carraway R, Reed A, Shetake J, Jakkamsetti V, Chang K, Kilgard M
(2008) Cortical activity patterns predict speech discrimination ability. Nature Neuroscience 11:603 -- 608.
Epping W, Eggermont J (1987) Coherent neural activity in the auditory midbrain of the grassfrog. Journal
of Neurophysiology 57:1464 -- 83.
Escab´ı M, Schreiner C (2002) Nonlinear Spectrotemporal Sound Analysis by Neurons in the Auditory
Midbrain. J Neurosci 22:4114 -- 4131.
Grace J, Amin N, Singh N, Theunissen F (2003) Selectivity for conspecific song in the zebra finch auditory
forebrain. Journal of Neurophysiology 89:472 -- 87.
Irvine D (1992) Physiology of the auditory brainstem. In: Handbook of Auditory Research, The Mammalian
Auditory Pathway: Neurophysiology Springer.
Langner G, Albert M, Briede T (2002) Temporal and spatial coding of periodicity information in the inferior
colliculus of awake chinchilla (Chinchilla laniger). Hearing Research 168:110 -- 30.
Langner G, Schreiner C, Merzenich M (1987) Covariation of latency and temporal resolution in the inferior
colliculus of the cat. Hearing Research 31:197 -- 202.
Lyzwa D (2015) Neural responses to natural sounds in the auditory midbrain: A model comparison. Inter-
national Conference on Neural Networks (IJCNN) pp. 1 -- 7.
17
Lyzwa D, Herrmann J, Worgotter F (2016) Natural vocalizations in the mammalian inferior colliculus are
broadly encoded by a small number of independent multi-units. Front. Neural Circuits 9:1 -- 19.
Machens C, Schutze H, Franz A, Kolesnikova O, Stemmler M, Ronacher B, Herz A (2003) Single auditory
neurons rapidly discriminate conspecific communication signals. Nature Neuroscience 6:341 -- 2.
Malmierca M, Merchan M, Henkel C, Oliver D (2002) Direct projections from cochlear nucleus to auditory
thalamus in the rat. J Neurosci 22:10891 -- 10897.
Malmierca M, Rees A, Le Beau F, Bjallie F (1995) Laminar Organization of Frequency-Defined Local
Axons Within and Between the Inferior Colliculi of the Guinea Pig. The Journal of Comparative Neu-
rology 357:124 -- 144.
Melssen W, Epping W (1987) Detection and estimation of neural connectivity based on crosscorrelation
analysis. Biological Cybernetics 57:403 -- 414.
Merzenich M, Reid M (1974) Representation of the cochlea in the inferior colliculus of the cat. Brain
Res. 77:397 -- 415.
Oliver D (2005) Neuronal organization in the inferior colliculus In Winer J, Schreiner C, editors, The
Inferior Colliculus. Springer, New York.
Pettersen K, Linden H, Dale A, Einevoll G (2012) Extracellular spikes and CSD. In: Handbook of Neural
Activity Measurement, Vol. 1 Cambridge University Press.
Quian Quiroga R, Nadasdy Z, Ben-Shaul Y (2004) Unsupervised spike detection and sorting with wavelets
and superparamagnetic clustering. Neural Computation 16:1661 -- 1687.
Rieke F, Bodnar D, Bialek W (1995) Naturalistic stimuli increase the rate and efficiency of information
transmission by primary auditory afferents. Proceedings of the Royal Society of London. Series B: Bio-
logical Sciences 262:259 -- 265.
Rode T, Hartmann T, Hubka P, Scheper V, Lenarz M, Lenarz T, Kral A, Lim H (2013) Neural representation
in the auditory midbrain of the envelope of vocalizations based on a peripheral ear model. Frontiers in
Neural Circuits 7.
Rose J, Greeenwood D, Goldberg J (1963) Some discharge characteristics of single neurons in the inferior
colliculus of the cat. I. tonotopical organization, relation of spike-counts to intensity, and firing patterns
of single neurons. Journal of Neurophysiology 26:294 -- 320.
Schneidman E, Berry M, Segev R, Bialek W (2006) Weak pairwise correlations imply strongly correlated
network states in a neural population. Nature 440:1007 -- 12.
Schreiber S, Fellous J, Whitmer D, Tiesinga P, Sejnowski T (2003) A new correlation-based measure of
spike timing reliability. Neurocomputing 52 -- 54:925 -- 931.
Schreiner C, Langner G (1988) Periodicity Coding in the Inferior Colliculus of the Cat. II. Topographical
organization. J Neurophysiol 60:1823 -- 1840.
Schreiner C, Langner G (1997) Laminar fine structure of frequency organization in auditory midbrain.
Nature pp. 383 -- 386.
Shannon R, Zeng F, Kamath VW J, Ekelid M (1995) Speech recognition with primarily temporal cues.
Science 270:303 -- 304.
Suta D, Kvasnak E, Popel´ar J, Syka J (2003) Representation of species-specific vocalizations in the inferior
colliculus of the guinea pig. Journal of Neurophysiology 90:3794 -- 3808.
18
van Rossum M (2001) A novel spike distance. Neural Computation 13:751 -- 763.
Wang L, Narayan R, Grana G, Shamir M, Sen K (2007) Cortical discrimination of complex natural stimuli:
can single neurons match behavior? The Journal of Neuroscience 27:582 -- 9.
19
|
1602.00258 | 1 | 1602 | 2016-01-31T15:12:57 | Placebo Response is Driven by UCS Revaluation: Evidence, Neurophysiological Consequences and a Quantitative Model | [
"q-bio.NC"
] | Despite growing scientific interest in the placebo effect and increasing understanding of neurobiological mechanisms, theoretical modeling of the placebo response remains poorly developed. The most extensively accepted theories are expectation and conditioning, involving both conscious and unconscious information processing. However, it is not completely understood how these mechanisms can shape the placebo response. We focus here on neural processes which can account for key properties of the response to substance intake. It is shown that placebo response can be conceptualized as a reaction of a distributed neural system within the central nervous system. Such a reaction represents an integrated component of the response to open substance administration (or to substance intake) and is updated through "unconditioned stimulus (UCS) revaluation" learning. The analysis leads to a theorem, which proves the existence of two distinct quantities coded within the brain, these are the expected or prediction outcome and the reactive response. We show that the reactive response is updated automatically by implicit revaluation lerning, while the expected outcome can also be modulated through conscious information processing. Conceptualizing the response to substance intake in terms of UCS revaluation learning leads to the theoretical formulation of a potential neuropharmacological treatment for increasing unlimitedly the effectiveness of a given drug. | q-bio.NC | q-bio | Placebo Response is Driven by UCS Revaluation: Evidence, Neurophysiological
Consequences and a Quantitative Model
aDepartment of Engineering "Enzo Ferrari" , University of Modena and Reggio Emilia, Via Vivarelli 10, int.1-41125 Modena , Italy
bMD at Local Health Unit of Modena, Via S.Giovanni del Cantone 23, 41121 Modena , Italy
Luca Puviania, Sidita Ramab
Abstract
Despite growing scientific interest in the placebo effect and increasing understanding of neurobiological mechanisms, theo-
retical modeling of the placebo response remains poorly developed. The most extensively accepted theories are expectation
and conditioning, involving both conscious and unconscious information processing. However, it is not completely under-
stood how these mechanisms can shape the placebo response. We focus here on neural processes which can account for key
properties of the response to substance intake. It is shown that placebo response can be conceptualized as a reaction of a dis-
tributed neural system within the central nervous system. Such a reaction represents an integrated component of the response
to open substance administration (or to substance intake) and is updated through "unconditioned stimulus (UCS) revaluation
learning". The analysis leads to a theorem, which proves the existence of two distinct quantities coded within the brain, these
are the expected or prediction outcome and the reactive response. We show that the reactive response is updated automatically
by implicit revaluation lerning, while the expected outcome can also be modulated through conscious information processing.
Conceptualizing the response to substance intake in terms of UCS revaluation learning leads to the theoretical formulation of
a potential neuropharmacological treatment for increasing unlimitedly the effectiveness of a given drug.
6
1
0
2
n
a
J
1
3
]
.
C
N
o
i
b
-
q
[
1
v
8
5
2
0
0
.
2
0
6
1
:
v
i
X
r
a
Email address: [email protected] (Luca Puviani)
Preprint submitted to xxxx
Introduction
What is the neural substrate of placebo response? Are there neurons endowed with a special location or specific neurobi-
ological pathways that are necessary or sufficient for placebo response? Recent approaches have attempted to narrow the
focus, and a major insight from the recent publications is that there seems not be a single neurobiological or psychobiologi-
cal mechanism which is able to explain the placebo effect in general. Instead, different mechanisms exist by which placebo
(and nocebo) responses originate and exert their actions. Here, we pursue a different approach. Instead of arguing whether a
particular neurobiological pathway, mechanisms, or group of neurons contributes to placebo response or not, our strategy is
to characterize the kinds of neural processes (at the system-level) that might account for key properties of the more general
response to substance intake. With this aim we start analyzing the response to (open or overt) pharmacological administration.
Drug administration is said to be open (or overt, opposite to hidden or covert) whenever the administration itself is correctly
perceived by a given subject; so that the individual is conscious of the administration act, and, generally speaking, he/she
might also be aware of the effects that such a substance will determine on the organism (e.g. see [1, 2]). We emphasize three
properties: pharmacological induced response is integrated (the response is given by the integral contribution of different
components, which are the active pharmacological effect, the reactive contribution, due to self-induced response, which, in
turn is governed by UCS revaluation, conditioning and higher cognitive processing, such as beliefs), at the same time it is
highly differentiated (one can experience any of a huge number of different pharmacologically-induced response). Further-
more, a third property is represented by the fact that the reactive response mimics the active pharmacological effect (we denote
this property as reactive mimicking). We first consider neurobiological data indicating that neural processes associated with
drug administration response are integrated and differentiated, and, finally, that a distributed system (denoted reactive system)
mimics the active pharmacological response. We then provide mathematical tools describing the above mentioned properties
of the pharmacological response; this leads us to formulate theoretical and operational criteria for determining whether the
activity of a group of neurons contributes to placebo response and with which intensity. Moreover, the derived model sug-
gests a pharmacological strategy for increase unlimitedly the response to a drug treatment (and hence the drug effectiveness),
exploiting the reactive system dynamics which governs the implicit placebo response.
Methods
General Properties of (Overt) Drug Administration-Induced Response
Generally speaking, the effect of a drug administration is due to the integration of different processes and components which,
together, exert their effect on a given organism.
Integration - Integration is a property shared by every response induced by substance intake, irrespective of its specific
pharmacological target: Each response comprises a single overall effect that cannot be decomposed into independent compo-
nents. More specifically, drug administration, and generally speaking substance consumption, exerts an overall effect wich
is perceived as unique, conveying pharmacological effects, rewarding, expectations, beliefs and prior experiences (e.g., see
[3, 1, 4]). A striking demonstration is given by the decreased effectiveness of hidden treatments. In particular, hidden drug
administrations eliminate the psychosocial (placebo) component, so that only the pharmacodynamic effect of the treatment
(free of any psychological contamination) can be accomplished. More specifically, in the open versus hidden treatments
paradigm [3, 1, 5, 2] drugs are administered through hidden infusions by machines; such infusions can be administered using
computer-controlled infusion pumps that are preprogrammed to deliver drugs at a desidred time. The crucial factor is that the
patients do not know that the drug is being injected. The difference between the responses following the administration of the
overt and covert therapy is the placebo (psychological) component, even though no placebo has been given [1]. Furthermore,
in [3] are reported experimental results which show how the overall response to an analgesic pharmacological treatment is due
to the contribution of 1) prior conditioning, 2) expectations and 3) active pharmacological effect; moreover, the three above
mentioned contributions are indirectly evaluated through the involvement of different treatment groups in double-blind design,
exploiting the open versus hidden paradigm.
Differentiation - While the drug administration response is given by an integral contribution of different components, there
are experimental evidences which show how the mammalian brain can finely discriminate between administration (and/or the
intake) of different substances. For instance, neurons in rats can discriminate between cocaine and liquid rewards [6, 7], or
between cocaine and heroin [8], possibly even better than between natural rewards [9, 10].
At this point it is useful to introduce some definitions. Provided that a given drug administration represents an uncondi-
tioned stimulus (UCS), an active response is defined as the response determined by the active pharmacological effect of the
2
UCS within the central nervous system (CNS). Moreover, a reactive response is defined as any response which is determined
by a self-induced reaction within the CNS, without any active pharmacological contribution. For instance, a conditioned
response (CR) induced after some pairings of an active pharmacological UCS and a neutral conditioned stimulus (CS) cor-
responds to a reactive response, since the elicited response is not substained by any pharmacological active effect. For this
reason, a reactive response could represent a specific form of placebo. Furthermore, the UCS implicit learning is defined as
the implicit (or automatic) UCS response (UCR) evaluation and re-valuation (i.e., UCR inflation and devaluation or deflation;
e.g., see [11, 12, 13]).
Difference between (pharmacological) conditioning and implicit UCS revaluation. When a neutral CS is paired with an
active drug, the associative learning process is called classical conditioning [14], (for instance when a tone is paired with
the self administration of cocaine in rats; [15]); furthermore, when a drug response (UCR) is evaluated and revaluated over
trials the learning process is represented by UCS revaluation [12]. In other words, from one hand, classical conditioning
learning occurs whenever the CS and the UCS corresponds to different discriminable elements, such that the organism can
learn the simultaneously (or causal) co-occurrence of both elements (in other words the "association" of two distinct elements
is learned) [14].
Instead, on the other hand, in successive trials involving drug administration there is a unique element
(i.e. the drug administration itself, which represents the UCS), which determines an UCR which, in turn, could vary during
successive administration trials; more specifically, if the dose of a drug (i.e. the UCS active stimulation intensity) changes
over successive open administration trials, the organism updates the evaluation of the experienced outcome through UCS
revaluation (for instance, in [11, 13, 16] experiments related to UCR revaluation, independently from classical conditioning,
are reported; furthermore, a quantitative analysis of UCS revaluation and its relation with classical conditioning are derived
in [12]). It is worth noting that, generally speaking, both classical conditioning and UCS revaluation could occur during drug
administration, since the context and other cues could be associated as CSs to the drug administration (UCS).
Finally, a generic induced response within the CNS can be represented by the superposition of the activity of different
neural populations; more specifically, assuming that the CNS consists of N different neuronal populations, the response,
denoted with the vector y, can be expressed as:
y =
yivi,
(1)
N
∑
i=1
where yi represents the i-th neuronal population activity1 and {vi;i = 1,2, ...,N} represents a set of versors, being asso-
ciated with different neuronal populations, which form a complete basis B for the CNS space. It is worth noting that the
different neuronal populations could be interdependent (i.e., B does not represent an orthonormal basis), and, the CNS is a
nonlinear system, so that even small variations of the activity (or the elicitation) of a single response component could lead to
a very different outcomes (for instance, a small variation of the mean firing rate of a population, or of the quantity of a given
neurotransmitter, can even lead to an opposite behavioral outcome; [17]).
Reactive mimicking.
In a growing body of literature it has been reported that pharmacological conditioning (and im-
plicit UCS revaluation) determines reactive responses which mimic the original active pharmacological effects. For instance,
experimental results reported in [15] show that an increase in dopamine release in the ventral striatum, measured through mi-
crodialysis, are observed not only when rats self administer cocaine (UCS), but also when they are solely presented with a tone
(CS) that has been previously paired with cocaine administration. Furthermore, in [3] it is shown that pharmacological condi-
tioning, like conditioning with opioids, produces placebo analgesia mediated via opioid receptors, and that, if conditioning is
performed with nonopioid drugs, other nonopioid mechanisms are involved, so that conditioning activates the same neuronal
populations as the active pharmacological treatment. Similar results and conclusions have been obtained in mice experiments
[18]. Moreover, brain imaging data, evidence that placebos can mimic the effect of active drugs and activate the same brain
areas; this occurs for placebo-dopamine in Parkinson's disease [19, 20], for placebo-analgesics [3, 21, 18, 22, 23, 24] or
antidepressants, and for placebo-caffeine in healthy subjects [25]. It is worth noting that since the active pharmacological
response is highly differentiate, the reactive mimicking property leads to the highly differentiation of the reactive system.
Such a system can be thought as a distributed network within the CNS, which generates reactions (or self-made responses) to
relevant stimuli and substances.
The above mentioned properties of pharmacological responding, that are, integration, differentiation and reactive mimick-
ing, can be described from a quantitative perspective; this should lead to the derivation of the laws governing the origin and
1In particular, yi is a real quantity representing the product between the mean number of elicited neurons and their mean firing rates for the i-th neuronal
population (with i = 1,2, ...,N); consequently, yi takes on a positive (negative) value if the response produces an increase of (a decrease or inhibition of) the
activity for the i-th population, and is equal to zero whenever the response does not involve any adjustment for the baseline activity of the population.
3
the dynamics of the placebo response. To this aim it is also necessary describing the dynamics of the UCS revaluation, that is
the implicit UCR evaluation over successive administration trials.
Error-Driven Learning
From a growing body of literature emerges that learning occurs through the computation of specific error-signals (or prediction
errors) (see [26, 27] for a review). Generally speaking, the prediction error is defined as the difference between the response
expected from a given stimulation and the response actually perceived by the elicited organism. This definition relies on
experimental observations acquired in functional imaging studies [28, 29, 26], or directly measured in dopaminergic circuits
(e.g., in the ventral tegmental area, VTA) or in other fear-related circuits [30, 31, 32, 33, 10, 27, 34, 35, 36].
Different mathematical models describing classical conditioning learning (e.g., Rescorla-Wagner model [37, 38] or tempo-
ral difference (TD) models [29, 39, 40, 41]), or describing learning in general, such as the probabilistic (Bayesian) "perception"
and "action" learning models (i.e., the predictive coding (PC) [42, 43] and the active inference [44, 45]), assume that coding
behavioral responses involves the computation of an error-signal. More specifically, the brain makes predictions in relation to
a given stimulus and, on the basis of the experienced outcome, the prediction is updated through the prediction error. If the
experienced outcome is greater (lower) than the prediction, the computed error signal is positive (negative) and corrects the
new prediction; furthermore, if the experienced response coincides with the expected outcome, the error signal is zero and no
prediction updatings take place.
Theoretical Concepts and Quantities
Differentiation. Eq. (1) describes the elicited response within the CNS accounting for all the involved neuronal populations
(or response components); however, we can consider one single component to ease the reading. This choice, however, does
not entail any loss of generality, since our model can be applied to any component.
Integration. As previously shown, the response to a drug administration can be expressed by three main contributions
(each of which might be further decomposed in more detail), these are: 1) the active pharmacological response, denoted x;
2) the reactive (self-induced) response due to implicit UCS revaluation learning and classical conditioning, denoted iR; 3) the
reactive response due to expectations and, generally speaking, to higher cognitive information processing (such as beliefs,
social observations and verbal suggestions), denoted γ. Hence, the integration property leads to the following expression for
the general elicited neuronal population:
(2)
It is important to note that the above described "response decomposition" is only a theoretical abstraction, since, because
y = x + iR + γ.
the integration property, the response is unique and the brain cannot discriminate between the different contributions.
Reactive mimicking. Since the reactive response (iR) associated with a given substance determined by implicit UCS
revaluation or by pharmacological conditioning mimics the active pharmacological effect (x), the reactive response for a given
neuronal population can be expressed as:
(3)
where the term α takes on real values and represents the efficiency (or the strength) of the reactive system for the given
iR = α · x,
neuronal component.
Implicit UCS revaluation. If multiple successive open drug administration trials are considered, provided that no conscious
information are available (such as verbal suggestions or beliefs) about the drug effects, it can be assumed that in every trial
the expected (or predicted) response coincides with the last experienced outcome (wich, in turn, coincides with the response
experienced in the previous trial), or, alternatively, that the expected response converges over successive trials to the actual
experienced outcome. Without any loss of the generality, we can assume that the predicted outcome is equal to the last
experienced outcome. Provided that the active pharmacological effect (x) is constant over successive drug administration
trials, and that the higher cognitive information processing are avoided (in other words, γ = 0, see Eq.(2)), the reactive
response due to implicit learning (i.e., due to the UCR, or y evaluation and revaluation over trials), iR, is updated through
the error-signal computation. In the first trial the reactive response is equal to zero, since no previous learning occured for
the given drug administration (UCS); hence, y1 = x. Since the expected outcome was equal to zero for that UCS, the error
signal after the first administration trial is equal to x and, such a prediction error updates the reactive response, according
4
to Eq. (3). In the second open drug administration trial the response is given by Eq. (2) with γ = 0, that is y2 = x + αx;
moreover, a new error signal is computed as y2 − y1 = αx, and the reactive response updated at the end of the second trial is
given by iR2 = αx + α2x. It easy to demonstrate (see Supplemental Information) that the response elicited in the n-th drug
administration trials can be expressed as:
αk(cid:17)
(cid:16)
yn = x + x
n−1
∑
k=1
.
(4)
Eq. (4) shows that the response to a given administration drug diverges (to infinity) or converges to an asymptotic value,
as the number of trials increases, depending on the magnitude of α. The experimental observations (and the ordinary life
experiences) show that the pharmacological responding tends to reach an asymptotic value over successive trials and does not
diverge; hence the absolute value of α has to be comprises between zero and the unity (i.e., 0 < α < 1). We denote the above
mathematical condition as stability of the reactive system. Under such a condition it is easy to demonstrate that the asymptotic
response can be expressed as follows:
y∞ = x + x· α
1− α ,
(5)
where the active pharmacological effect (x) and the reactive component learned during successive trials (i.e., the quantity given
by iR∞ = x· α
1−α ), are highlighted. At this stage, if in a successive trial an inerte substance is administered (in other words, if
a placebo is given, so that the active pharmacological effect is brought to zero, x = 0), the placebo response coincides with
the reactive response iR∞, implicitly learned over trials. We argue that such a reactive response represents the unconscious or
implicit placebo response. If the placebo is administered during successive trials, it is easy to demonstrate that the response
tends asymptotically to zero, driven by the error signal (see Supplemental Information). It is important pointing out that
the open administration (UCS) allows the individual to perceive the UCS (i.e., the drug administration itself), and hence to
trigger the learned reactive response associated with that UCS; furthermore, an error signal is computed whenever the UCS
outcome is revaluated. Conversely, if the administration was hidden, no UCS can be perceived and no reactive response can be
triggered. Moreover, it is arguably that different features of the drug administration (or substance intake), such as the flavor of
a given pill (or even conditioned stimulus), contribute to perception and differentiation of the given substance intake (UCS).
Finally, the integration, the reactive mimicking and the stability of the reactive system properties lead to the following
theorem:
Theorem 1. It is necessary that two distinct quantities are encoded within the CNS for the stability of the reactive system,
these are the reactive response and the predicted (expected) outcome.
The demonstration is given on the Supplemental Information.
In practice, the theorem proves that both the reactive
response (iR) and the expected (or predicted) outcome have to be computed and independently updated in order to assure the
stability of the reactive system. In other words, the theorem proves that the reactive response cannot be represented by the
predicted (i.e., expected) response.
Outcome Simulation and Conscious Placebo/Nocebo Response
Since the above mentioned theorem proves that both the reactive response and the expected outcome have to be coded
within the CNS, the response to drug administration (more specifically the reactive component of it) should be a function
of these quantities. In the previous Section it is shown how, in the absence of cognitive information processing, the reactive
response by UCS revaluation learning through the computation of the error signal. In this Section it is proposed how it is
possible modulating the reactive response through cognitive and conscious information processing, such as verbal suggestions,
beliefs, expectations, and even by social observational learning [46]. More specifically, we hypothesize that cognitively
processed information, which permits to infer (i.e., to simulate) the outcome of a given substance intake, lead to a process that
we denoted outcome simulation. Such a process consists of high level cognitive processing of different pieces of information
related to the outcome, and lead to the computation of the most probable outcome. Certainly such a computation depends
also on the previous experiences which are evoked by the provided information. If the simulated outcome is different from
the expected outcome (i.e., from the previously experienced outcome) an error signal is computed and the reactive response
(iR) is updated accordingly. From a mathematical and modeling perspective, the outcome simulation process corresponds
to a virtual substance administration trial, in which the experienced outcome is the simulated (predicted or inferred through
5
cognitive processing) outcome, and the error signal is computed as the difference between the simulated and the expected. It
easy to prove that the reactive response is updated by the error signal through the following recursive expression [12]:
(6)
where en represents the error signal computed in the n-th trial (real or virtual), provided that α does not change over time;
otherwise, the more general expression for iR in the n-th trial is as follows:
iR,n = iR,n−1 + α · en,
iR,n = αn
n−1
∑
k=1
ek.
(7)
For these reasons, verbal suggestions or beliefs which let a given individual inferring a positive (negative) outcome, lead
to the computation of a positive (negative) error signal which increase (decrease) the reactive response associated with a given
drug or treatment. We argue that the simulation process takes place not only in drug administration but even in more general
situations, such as in stimulus-outcome learning (or UCS evaluation); for instance animals could learn that a given stimulus is
dangerous observing others facing with such a stimulus [47]. Moreover, cognitive information processing could influence the
reactive response associated with different stimuli and substances; for instance, in [48] is shown, by functional MRI studies,
how prices of wine bottles (which represent here a piece of information related to the outcome) can affect the experienced
pleasantness of the wine intaking (UCS). Indeed, the experienced pleasantness (which represents here the UCR) is due to
the integration of the active component, x, and the reactive (self-induced) response iR, which is due to prior learning and/or
cognitive evaluation. Concluding, cognitive information processing could lead to an outcome simulation process which, in
turn, modulates the reactive response through the computation of an error signal, uderstood as the difference between the
simulated (inferred) outcome and the expected (previously experienced) outcome.
Operational Measures and Prediction of Implicit Placebo Response
Given a specific target neuronal population, the active response to a pharmacological administration (x) could vary across
individuals, and, more important, the reactive response component (i.e., the implicit placebo or reactive response, iR) could be
even more variable. Indeed, the parameter α, which drives the reactive dynamics, should be a function of different parameters,
such as the specific neural population involved (such as dopaminergic, opioidergic, serotonergic, and so on) and the number,
the types and the sensitivity of the receptors within the given population. In turn, these variables depend on different parame-
ters, for instance these are functions of the individual's genetic makeup. Indeed, there is growing evidence that the individual's
genetic makeup (a stable trait) influences clinical outcomes and potentially may allow for identification of placebo responders
(see [49] for a review). More specifically, treatment outcomes in the placebo arms of trials that have assessed genetic variation
in the dopaminergic, opioid, cannabinoid, and serotonergic pathways suggest that genetic variation in the synthesis, signaling,
and metabolism of these neurotransmitters contributes to variation in the placebo response [49]. For these reasons, it could be
useful to estimate (i.e., measure) the quantities x and α for a given pharmacological target and for a specific genetic makeup.
With such measures it should be possible to optimize the combination drug-genetic makeup and considering the implicit
placebo response as a stable and reliable contribution of the pharmacological responding and, hence, as an indistinguishable
part of the "true" pharmacological treatment effect. Operationally, the measure of the active pharmacological response can
be assessed by hidden or covert drug administration. The response should be assessed quantitatively adopting, for instance,
functional brain imaging or neurons activity recording, estimating the activity (understood as the product of the mean number
of firing neuron and the mean firing rates) increase of the target neuronal population. Furthermore, the parameter α can be
assessed by open (overt) drug administration, provided that the subject does not know which is the pharmacological target
[3] (in other words the reactive response due to implicit UCS evaluation is considered, and the expectations, beliefs of clin-
ical benefit or higher cognitive information processes are avoided; i.e. γ = 0, see Eq. (2)). More specifically, assessing the
increase of the given neuronal population activity in successive drug administration trials permits to estimate the increase of
the overall response (which comprises the active and the reactive response) which will converge to an asymptote, according to
Eq. (5). Moreover, the asymptotic convergence to zero of the activity of the given neuronal population can be assessed over
successive placebo administration trials. Hence, the parameter α could be reliably estimated from these data through simple
mathematical computations.
Moreover, since an important prediction of the model is the dynamics of the error signals and of the reactive responses
over successive open (active and placebo) administration trials, these quantities can be experimentally measured or estimated.
In particular, experimental measures obtained recording dopamine neurons (e.g., see [39, 50, 10, 27, 34, 36]) reveal that
6
dopamine neurons in specific brain regions (such as the ventral tegmental area, VTA, or the substantia nigra) code prediction
and prediction error signals related to rewards (in particular related to rewarding substances, drugs or food delivery). Despite
these experiments were designed to study dopamine neurons activity in classical conditioning and in the unexpected (and
predicted) delivery of rewarding susbstances, these suggest the possibility to monitor the activity of the VTA dopamine
neurons (for instance in rats) during the administration of drugs which determine a rewarding response, such as cocaine
or morphine. In particular, it is expected that in hidden drug administration trials no phasic activity of VTA dopamine neurons
has to be observed; moreover, the same result should be observed during the first open drug administration trial. Furthemore,
starting from the second administration trial, the coding of the reactive response associated with the drug administration
(UCS) has to be observed, and such a response should increase asymptotically over successive trials. Finally, the VTA
dopamine neurons activity should tend asymptotically to zero over successive placebo administration trials. From the above
mentioned experimental measures it is possible to estimate the α parameter for the dopaminergic neuronal population related
to a specific rewarding drug (indeed, different rewarding signals and neuronal populations exist, related to liquid or solid
rewarding substances, drugs, and other rewarding stimuli; [10]). It is worth pointing out that the VTA neurons activity have
to be monitored for all the duration of the drug effect within each trial, since it is arguably that the reactive response mimicks
also the temporal trend of the active drug effect. In fact, in the above mentioned measures performed on the VTA neurons,
time represents a fundamental variable which is coded by neurons (since, for instance, the time of the reward occurence plays
and important role). In other words, it should be possible that, if the active drug effect starts after a time delay from the
administration, and it vanishes with a specific temporal trend, the reactive response could mimick the same temporal trend.
Obtaining Increasing and Resistant-to-Extinction Reactive Responses
Since the mammalian brain is able to induce reactive responses similar to those obtained by pharmacological active effects,
exploiting the implicit reactive system dynamics for obtaining a desired outcome should be feasible, in principle. More
specifically, our system-level model suggests that it could be possible obtaining resistant-to-extinction reactive responses ma-
nipulating the error signal computation. Indeed, even avoiding and neglecting cognitive processing (e.g., verbal suggestions)
during drug administration or substance intake, as previously shown, the error signal is implicitly computed, and it drives (or
updates) the reactive response iR associated with the representation of the given drug (i.e. the UCS; see also [12] for a more
comprehensive analysis). Provided that, for a given genetic makeup and for a specific target CNS component, the reactive
system efficiency α is greater than zero, after some active drug administration trials the reactive response reaches the asymp-
tote iR∞ = x· α
1−α (see Eq. (5)) through implicit UCS revaluation learning, and, hence, through the successive updatings of the
reactive response by error signal computations (see Eqs. (6-7)). Furthermore, if a placebo administration trial follows, then
the implicit placebo response, which is equal to the asymptotic value iR∞, is experienced, moreover, the experienced outcome
is updated and an error signal is computed, leading to a decrease of the reactive (placebo) response, which vanishes over
successive placebo trials. However, if the error signal computation is blocked or "degraded" during placebo (or lower active
dose) administration, the experienced outcome is updated; conversely the reactive response (and hence the implicit placebo
response) does not decrease (or decrease by little), since no error signal (or a degraded error signal) updated it. Furthermore,
a resistant-to-extinction reactive response can be obtained, since no error signals can be computed over successive trials, be-
cause the expected outcome concides with the reactive response, which, in turn, coincides with the experienced outcome; in
other words, a reactive response which does not tend asymptotically to zero over successive placebo trials is obtained. We
denote such a response iR0, since it will be elicitated by the reactive system at every substance intake trial (as it were origi-
nally generated by a fictitious trial zero), regardless of whether an active pharmacological effect exists. Indeed, it is easy to
demonstrate that, if a resistant-to-extinction reactive response iR0 has been associated to a given substance, the expression for
the response to substance intake driven through implicit learning in the generic n-th trial is2:
yn = iR0 + x + α (yn−1 − iR0) .
(8)
The resistant-to-extinction property of a reactive response is the key element for increasing unlimitedly an implicit re-
sponse; indeed, repeating the cycle of trials (i.e., the above described protocol) adopted for obtaining iR0, will lead to the
increase of the reactive response, through an accumulation principle (see Supplemental Information for the computational
details). Hence, through multiple cycles of active pharmacological drug administration followed by inerte (or a lower active
dose) substance administration, together with a pharmacological treatment able to degrade the error signal computation (more
2In [12] it is shown that different resistant-to-extinction emotional reactive responses are natively coded in the mammalian brain from birth, and these are
associated with phylogenetic (i.e., prepared biological and evolutionary fear relevant) stimuli representations [51, 52].
7
specifically able to lower the precisions associated with error signals, see [44, 43]), it is possible to "force" the CNS to react
to such a drug with an increasing self-induced response which mimics the active pharmacological treatment. After some
cycles it is expected that the reactive response iR will be greater than the original active pharmacological response, and the
pharmacological effect can be increased indefinitely.
In practice, it is expected that the error signal cannot be completely eliminated in the placebo administration trials, so that
multiple placebo administration trials with an "error signal degradation treatment" will be required, and the resulting iR0 will
be smaller than iR∞. Nevertheless, the proposed implicit learning model describes the reactive system dynamics at a system
(macro) level and, even if it predicts the possibility of obtaining a resistant-to-extinction (i.e., stable) and increasing reactive
response, it does not give any indications of how obtaining it, nor which pharmacological targets have to be considered for
the degradation of the error signal. For these reasons, a specific model, which can describe the error signal computation at a
deeper level (i.e., at the neurons level) is required. We argue such a model could be represented by the so called predictive
coding model (and its version applied to action learning; active inference) [42, 43, 44, 45].
Predictive Coding, Active Inference and Error Signal Computation
In predictive coding [42, 43] is formalized the notion of the Bayesian brain, in which neural representations in the higher levels
of cortical hierarchies generate predictions of representations in lower levels. These top-down predictions are compared with
representations at the lower level to compute a prediction error. The resulting error-signal is passed back up the hierarchy
to update higher representations; this recursive exchange of signals lead to the minimization (ideally the suppression) of
the prediction error at each and every level to provide a hierarchical explanation for sensory inputs that enter at the lowest
(sensory) level. In the Bayesian jargon neuronal activity encodes beliefs or probability distributions over states in the world
that causes sensations [43]. In predictive coding the notion of precision (or confidence, which represents the inverse of the
variance) of the error signals is also formalised, and the mechanism through which the brain has to estimate and encode the
precision associated with the prediction errors is explained. The prediction errors are then weighted with their precision before
being assimilated at a high hierarchical level. More specifically, active inference accounts posit that the brain deals with noisy
prediction errors by decreasing the gain on cortical pyramidal neurons that function like precision units to regulate outputs
from signal error computations, thereby reducing the influence of these outputs [53, 42, 54]. It is important pointing out that
the resulting error signal depends on the relative precision of prediction errors at each level of the neural hierarchy. Hence,
a neuropharmacological alteration of the precision on different levels in the cortical hierarchy could lead to the modulation
of the error signal computation. Furthermore, it is supposed [55, 56] that such an alteration may naturally occur in some
pathologies, such as in Parkinson's desease, where the depletion of dopamine (which is supposed to encode the precision of
prediction errors by altering their synaptic gain, as also other neurotransmitters do [56]) at different levels, would alter the
balance of precision at higher (sensorimotor) relative to lower (primary sensory) levels in the cortical hierarchy; moreover, the
symptoms vary according to the site (hierarchical level) of changes in precision [55].
Neuronal Populations within the Reactive System
Which are the neuronal populations belonging to the reactive system? In other words, for which CNS response components
does the reactive mimicking property hold? If a given neuronal population belongs to the reactive system, then it is possible to
obtain a reactive response (i.e., an implicit placebo response) for this neuronal population. Experimental results from the liter-
ature evidence that at least the following CNS components could generate a reactive response: 1) emotional system responses
[12] (which include, for instance, the dopaminergic mesolimbic and mesocortical system, [57, 58], the endocannabinoid and
opioid system in placebo analgesia, [21, 23, 59, 24, 60, 61, 62], the serotoninergic system, the target neuronal systems of de-
pression, anxiety and addiction; see [1]); 2) the dopaminergic motor system [19, 20]; 3) the humoral immune response system
(in particular the components of the CNS such as the hypothalamic-pituitary-adrenal axis, HPA, or the sympathetic nervous
system, SNS; [1, 63, 64, 65]); 4) the endocrine system; (see [1, 66] for a review). We speculate that the main components of
the reactive system are the emotional/motivational and the humoral immune systems, since, from an evolutionary perspective,
the organism has to be able to learn and to automatically react whenever "important" stimuli are perceived, and, implicitly
recognize "relevant" substances, such as food, in order to approach or avoid them or to quickly and automatically react (for
instance through an automatic trigger of a first immune response). More specifically, a stimulus is "important or relevant" for
an organism if it previously elicited a CNS response belonging to the reactive system, such as a painful stimuli (which are
related to emotional components), a rewarding substance (such as food), or primary immune responses.
8
Discussion
We propose a model describing response to substance intake based on implicit UCS revauation learning (which describes the
automatic evaluation and revaluation of an unconditioned stimulus, that represents the given substance intake). Such a model
formalizes the key properties of drug administration response based on well established empirical evidence. The theoretical
conceptualization leads to a theorem which proves the necessity of the encoding of two fundamental quantities: the reactive
response and the expected (or predicted) outcome. We have derived a model that meets this requirement and is able to predict
the response to drug administration over successive trials. The placebo response is mathematically formalized as a reactive
response, which represents an implicit and inevitable component of the overall response to substance intake; moreover its
effect becomes naturally evident when an organism responds to a substance administration whose active (pharmacological)
effect has been eliminated.
Modeling the dynamics of the reactive system, from which the placebo response originates, permits also the definition
of quantities (e.g., the reactive efficiency, α) and of operational measures which could lead to the maximization and the
stabilization of the reactive response (iR) for a given genetic makeup, even when the pharmacological active effect (x) is
kept constant. Furthermore, the model shows that it should be possible to create a resistant-to-extinction reactive response
blocking (or degrading) the error signal computation in specific trials. More specifically, after some trials in which the active
drug is administered in order to increase the implicit reactive response, successive trials have to follow, in which placebo
administration is given together with a pharmacological treatment able to degrade the precisions of the error signal (and hence
the overall error signal computation). The targets of such a pharmacological treatment have to be specific neurotransmitters on
precise level of the neural hierarchy, depending on the target component of the original active drug administration; moreover,
predictive coding and active inference model provides the theoretical and computational basis for the assessment of such
targets.
In this manuscript, the theoretical feasibility study of a strategy for the unlimited increase of a given drug effectiveness
has been developed. Further computational and experimental studies are needed in order to explore the mentioned strategy.
We also argue that the generality of our model permits the modeling of a broad range of phenomena and pathologies which
involve the interaction between an organism and a substance intake, such as, drug addictions and eating disorders, and for the
development of novel strategies for decreasing pathological resistant-to-extinction reactive responses, such as in phobias or
post-traumatic stress disorders.
Supplemental Information
In the following sections the equations describing the reactive system dynamics (and, hence, the implicit placebo response
dynamics) are derived. It is assumed that no cognitive information processes occur (in other words, verbal suggestions, beliefs
and expectations on clinical benefit are avoided); however, drug administration, or substance intake, has to be "overt", so
that the considered individual is aware that the same substance is administered over successive trials (but he/she does not
know which is the pharmacological target of such a substance). Furthermore, to ease the computations, it is assumed that the
expected (predicted) response, in a given k-th trial, coincides with the last experienced outcome (i.e., yexpected,k = yk−1).
Implicit Response Acquisition to a Substance Intake
In this section the responses elicited during succesive trials in which an active drug is administered are derived.
In Table 1 the variables of interest are mathematically described through succesive trials. A single CNS component (i.e.,
a specific target neural population) is considered for the computations, provided that such a component belongs also to the
reactive system.
9
Table 1. In this Table the computations leading to the automatic central nervous system (CNS) response to a drug intake over successive trials are
derived. It is assumed that the active pharmacological effect (x) is constant over successive trials.
The response to drug intake in the n-th trial can be written as:
n−1
∑
yn = x + x
k=1
(cid:16)
αk(cid:17)
(9)
Considering the reactive system stability property, it follows that the term α has to be less than one in magnitude (i.e.
0 < α < 1). Hence the summation term in eq. (9) converges to the value α/(1−α), for the property of the geometric series,
when the number of trials tends to infinity. Thus, the asymptotic response due to the active drug administration can be written
as:
(10)
(11)
The eq. (9) can also be expressed as:
y∞ =
x
1− α = x +
α · x
1− α .
which shows the recursive nature of the response acquisition process over successive trials.
yn = x + α · yn−1
Reactive or Implicit Placebo Response Dynamics
In this section the response elicited during succesive trials where a placebo is administered (i.e. the active pharmacological
component is absent) is derived.
In Table 2 the variables of interest are mathematically described over succesive trials. A single generic component is
considered in the computations, moreover, it is assumed that the asymptotic response has been reached over the previous
active drug administration trials (see Eq.(5)).
Table 2. - In this Table the computations leading to the reactive (implicit placebo) response extinction over successive trials are derived. It is assumed
that, previously to placebo admnistration trials, the pharmacological effect was equal to x.
10
Trial number Active pharmacological response Reactive response CNS response Error signal kx eαii 1k1kR,kR,kRkkixy,1kkkyyexxxixxixiiinnRRRRR12,23,2,1,0......00xxxxxxxxxn...03210xxxxyxxxyxxyxyynn1223210......0xexexexeenn123210...0n...3210Trial number Active pharmacological response Reactive response CNS response Error signal kx eαii 1k1kR,kR,kRkkixy,1kkkyyexxixxixxixixinnnRRRRR11,223,2,1,0)1/(...)1/()1/()1/()1/(0...0003210nxxxxxx)1/(...)1/()1/()1/()1/(332210nnxyxyxyxyxyxexexexeenn123210... 0n...3210 The response to placebo administration in the n-th trial can be written as:
(cid:40)
yn = α · yn−1
y0 = Y0
,
(12)
where, provided that the asymptote had been reached during the previous active pharmacological trials, Y0 = x·α/(1−α).
Theorem: It is necessary that two distinct quantities are encoded within the CNS for the stability of the reactive system,
these are the reactive response and the predicted (expected) outcome.
Proof. Proof by contradiction (reductio ad absurdum)
Hypothesis 1: The reactive response associated with a given substance intake coincides with the expected (predicted)
outcome, and the expected or predicted outcome converges to the experienced outcome.
Hypotheses 2: The properties integration, reactive mimiking and reactive system stability hold.
Hypotheses 3: Cognitive information processing are avoided (in other words, no beliefs, expectations for clinical benefits,
verbal suggestions and so on, are involved).
Hypothesis 1 asserts that a unique reactive signal predicting the drug administration outcome exists, and that this signal
coincides with the reactive response elicited when the drug is administered; furthermore, the predicting signal converges (by
learning) to the actual experienced elicitation. The last assumption has been formulated to include a more general scenario than
that considered in our initial assumptions, in which the expected outcome coincides with the last experienced outcome. From
a mathematical viewpoint, the expected outcome can be computed using any supervised learning method (or, alternatively,
TD methods [40]) in which the predicted outcome is evaluated on the basis of the past m predictions (i.e., of the predicted
outcomes in the last m trials) and of the actual outcome, minimizing the error between the prediction and the experienced
outcome. Otherwise it can be assumed that the predicted outcome coincides with the last experienced outcome.
1. Let the UCS be the drug administration.
2. Drug administration to a given subject takes place on successive trials, so that it exerts an active elicitation (i.e., it elicits
the active response x). During the first trial the response is exclusively due to the active pharmacological component,
that is y1 = x. After the first trial (for instance, during the drug administration in the second trial), the predicted (reactive)
response, called ypredicted,1, is computed.
3. In the second trial, after the drug administration, the predicted outcome (ypredicted,1) adds up to the successive UCS
active elicitation, so that the outcome can be expressed as y2 = yexpected,1 + x. Furthermore, since ypredicted,1 does not
coincide with the actual experienced outcome, the new prediction ypredicted,2 is computed after the second trial; it can
be easily proved that ypredicted,2 > ypredicted,1 (since the experienced outcome has been strengthened and the error signal
has to be minimized).
4. In the third trial the experienced outcome can be written as y3 = ypredicted,2 + x; since y3 > y2 ≥ ypredicted,2 a new value
5. In the n-th trial the outcome can be expressed as yn = ypredicted,n−1 + x; it is easy to prove that yn > yn−1 ≥ ypredicted,n−1.
for the predicted response is computed, called ypredicted,3, such that ypredicted,3 > ypredicted,2.
Moreover, if the number of trials tends to infinity, the outcome grows indefinitely (i.e., lim
n→∞
yn = ∞).
6. The last statement is absurd, as it contradicts Hypothesis 2, in particular it contradicts the stability system principle.
Increasing a Reactive Response through Resistant-To-Extinction Increments
Provided that a resistant-to-extinction reactive response, denoted iR0, has been previously obtained (see Section "Obtaining
Increasing and Resistant-to-Extinction Reactive Response"), it is easy to show that the drug intake response is expressed as:
(13)
where the term iR0 represents the inextinguishable reactive response obtained during the first protocol cycle. The generic
yn = iR0 + x + α (yn−1 − iR0) ,
response at the n-th trial can also be expressed as:
11
Eqs. (13-14) can be obtained from the computations shown in Tab. 3.
yn = iR0 + x + x
n−1
∑
k=1
αk.
(14)
Table 3. - In this Table the computations leading to the automatic central nervous system (CNS) response to a drug intake over successive trials are
derived. It is assumed that the active pharmacological effect is constant and equal to x, furthermore, a resistant-to-extinction reactive response (iR0)
associated to the given substance intake has been previously obtained, through the strategy described in the previous sections.
12
Trial number Active pharmacological response Reactive response CNS response Error signal kx eαii 1k1kR,kR,kRkkixy,1kkkyyexxxiixxiixiiiiiinRnRRRRRRRRR120,203,02,01,00,......xxxxxxxxxn...03210012023020100......RnnRRRRixxxxyixxxyixxyixyiyxexexexeenn123210... 0n...3210 References
References
[1] Benedetti F. Mechanisms of placebo and placebo-related effects across diseases and treatments. Annual Review of
Pharmacology and Toxicology. 2008;48:33 -- 60.
[2] Colloca L, Lopiano L, Lanotte M, Benedetti F. Overt versus covert treatment for pain, anxiety, and Parkinson's disease.
Lancet Neurology. 2004;3(11):679 -- 684.
[3] Amanzio M, Benedetti F. Neuropharmacological Dissection of Placebo Analgesia: Expectation-Activated Opioid Sys-
tems versus Conditioning-Activated Specific Subsystems. Journal of Neuroscience. 1999;19(1):484 -- 94.
[4] Bingel U, Colloca L, Vase L. Mechanisms and clinical implications of the placebo effect: is there a potential for the
elderly? A mini-review. Gerontology. 2011;57(4):354 -- 363.
[5] Benedetti F, Carlino E, Pollo A. Hidden administration of drugs. Clinical Pharmacology and Therapeutics.
2011;90(5):651 -- 661.
[6] Bowman EM, Aigner TG, Richmond BJ. Neural signals in the monkey ventral striatum related to motivation for juice
and cocaine rewards. Journal of Neurophysiology. 1996;75(3):1061 -- 1073.
[7] Carelli RM, Deadwyler SA. A comparison of nucleus accumbens neuronal firing patterns during cocaine self-
administration and water reinforcement in rats. Journal of Neuroscience. 1994;14(12):7735 -- 7746.
[8] Chang JY, Janak PH, Woodward DJ. Comparison of mesocorticolimbic neuronal responses during cocaine and heroin
self-administration in freely moving rats. Journal of Neuroscience. 1998;18(8):3098 -- 3115.
[9] Carelli RM, Ijames SG, Crumling AJ. Evidence that separate neural circuits in the nucleus accumbens encode cocaine
versus "natural" (water and food) reward. Journal of Neuroscience. 2000;20(11):4255 -- 4266.
[10] Schultz W. Multiple reward signals in the brain. Nature Reviews Neuroscience. 2000;1(3):199 -- 207.
[11] Hosoba T, Iwanaga M, Seiwa H. The effect of UCS inflation and deflation procedures on 'fear' conditioning. Behaviour
Research and Therapy. 2001;39(4):465 -- 475.
[12] Puviani L, Rama S, Vitetta G. Prediction Errors Drive UCS Revaluation and not Classical Conditioning: Evidence and
Neurophysiological Consequences. arXiv:160107766. 2016;.
[13] Rescorla RA. Effect of inflation of the unconditioned stimulus value following conditioning. Journal of Comparative
and Physiological Psychology. 1974;86(1):101 -- 106.
[14] Pavlov IP. Conditioned Reflexes: An Investigation of the Physiological Activity of the Cerebral Cortex. Oxford Univer-
sity Press. 1927;.
[15] Ito R, Dalley JW, Howes SR, Robbins TW, Everitt BJ. Dissociation in conditioned dopamine release in the nucleus ac-
cumbens core and shell in response to cocaine cues and during cocaine-seeking behavior in rats. Journal of Neuroscience.
2000;20(19):7489 -- 7495.
[16] Schultz DH, Balderston NL, Geiger JA, Helmstetter FJ. Dissociation between implicit and explicit responses in post-
conditioning UCS revaluation after fear conditioning in humans. Behavioral Neuroscience. 2013;127(3):357 -- 368.
[17] Brunton LL, Lazo JS, Parker KL. Goodman & Gilman's The Pharmacological Basis of Therapeutics. 11th ed. McGraw-
Hill; 2006.
[18] Guo JY, Wang JY, Luo F. Dissection of placebo analgesia in mice: the conditions for activation of opioid and non-opioid
systems. J Psychopharmacol (Oxford). 2010;24(10):1561 -- 1567.
[19] De la Fuente-Fernandez R, Ruth TJ, Sossi V, Schulzer M, Calne DB, Stoessl AJ. Expectation and dopamine release:
mechanism of the placebo effect in Parkinson's disease. Science. 2001;293(5532):1164 -- 1166.
13
[20] De la Fuente-Fernandez R, Stoessl AJ. The placebo effect in Parkinson's disease. Trends in Neurosciences.
2002;25(6):302 -- 6.
[21] Eippert F, Bingel U, Schoell ED, Yacubian J, Klinger R, Lorenz J, et al. Activation of the opioidergic descending pain
control system underlies placebo analgesia. Neuron. 2009;63(4):533 -- 543.
[22] Lui F, Colloca L, Duzzi D, Anchisi D, Benedetti F, Porro CA. Neural bases of conditioned placebo analgesia. Pain.
2010;151(3):816 -- 824.
[23] Nolan TA, Price DD, Caudle RM, Murphy NP, Neubert JK. Placebo-induced analgesia in an operant pain model in rats.
Pain. 2012;153(10):2009 -- 2016.
[24] Petrovic P, Kalso E, Petersson KM, Ingvar M. Placebo and opioid analgesia -- imaging a shared neuronal network.
Science. 2002;295(5560):1737 -- 1740.
[25] Haour F. Mechanisms of the placebo effect and of conditioning. Neuroimmunomodulation. 2005;12(4):195 -- 200.
[26] Garrison J, Erdeniz B, Done J. Prediction error in reinforcement learning: a meta-analysis of neuroimaging studies.
Neuroscience and Biobehavioral Reviews. 2013;37(7):1297 -- 1310.
[27] Schultz W, Dickinson A. Neuronal coding of prediction errors. Annual Review of Neuroscience. 2000;23(1):473 -- 500.
[28] Berns GS, McClure SM, Pagnoni G, Montague PR. Predictability modulates human brain response to reward. Journal
of Neuroscience. 2001;21(8):2793 -- 2798.
[29] O'Doherty JP, Dayan P, Friston K, Critchley H, Dolan RJ. Temporal difference models and reward-related learning in
the human brain. Neuron. 2003;38(2):329 -- 337.
[30] Bray S, O'Doherty J. Neural coding of reward-prediction error signals during classical conditioning with attractive faces.
Journal of Neurophysiology. 2007;97(4):3036 -- 3045.
[31] Delgado MR, Li J, Schiller D, Phelps EA. The role of the striatum in aversive learning and aversive prediction errors.
Philosophical Transactions of the Royal Society of London Series B, Biological Sciences. 2008;363(1511):3787 -- 3800.
[32] Li SS, McNally GP. The conditions that promote fear learning: prediction error and Pavlovian fear conditioning. Neu-
robiology of Learning and Memory. 2014;108:14 -- 21.
[33] McNally GP, Johansen JP, Blair HT. Placing prediction into the fear circuit. Trends in Neurosciences. 2011;34(6):283 --
292.
[34] Schultz W. Behavioral theories and the neurophysiology of reward. Annual Review of Psychology. 2006;57:87 -- 115.
[35] Steinberg EE, Keiflin R, Boivin JR, Witten IB, Deisseroth K, Janak PH. A causal link between prediction errors,
dopamine neurons and learning. Nature Neuroscience. 2013;16(7):966 -- 973.
[36] Waelti P, Dickinson A, Schultz W. Dopamine responses comply with basic assumptions of formal learning theory.
Nature. 2001;412(6842):43 -- 48.
[37] Miller RR, Barnet RC, Grahame NJ.
1995;117(3):363 -- 386.
Assessment of the Rescorla-Wagner model.
Psychological Bulletin.
[38] Rescorla RA, Wagener AR. 3. In: Black AH, Prokasy WF, editors. A theory of Pavlovian conditioning: Variations in
the effectiveness of reinforcement and nonreinforcement. Appleton-Century-Crofts, New York; 1972. p. 64 -- 99.
[39] Schultz W, Dayan P, Montague PR. A neural substrate of prediction and reward. Science. 1997;275(5306):1593 -- 1599.
[40] Sutton RS. Learning to predict by the methods of temporal differences. Machine Learning. 1988;p. 3, 9 -- 44.
[41] Sutton RS, Barto AG. Time-derivative models of pavlovian reinforcement. In: Gabriel M, J Moore E, editors. Learning
and Computational Neuroscience: Foundations of Adaptive Networks. MIT Press.; 1990. p. 59, 229 -- 243. 497 -- 537.
[42] Friston K. Learning and inference in the brain. Neural Networks. 2003;16(9):1325 -- 1352.
14
[43] Friston K. Hierarchical models in the brain. PLoS Computational Biology. 2008;4(11):e1000211.
[44] Friston KJ, Daunizeau J, Kiebel SJ. Reinforcement learning or active inference? PLoS ONE. 2009;4(7):e6421.
[45] Friston KJ, Daunizeau J, Kilner J, Kiebel SJ. Action and behavior: a free-energy formulation. Biological Cybernetics.
2010;102(3):227 -- 260.
[46] Colloca L, Benedetti F. Placebo analgesia induced by social observational learning. Pain. 2009;144(1-2):28 -- 34.
[47] Olsson A, Nearing KI, Phelps EA. Learning fears by observing others: the neural systems of social fear transmission.
Social Cognitive and Affective Neuroscience. 2007;2(1):3 -- 11.
[48] Plassmann H, O'Doherty J, Shiv B, Rangel A. Marketing actions can modulate neural representations of experienced
pleasantness. Proceedings of the National Academy of Sciences of the United States of America. 2008;105(3):1050 --
1054.
[49] Hall KT, Loscalzo J, Kaptchuk TJ. Genetics and the placebo effect: the placebome. Trends in Molecular Medicine.
2015;21(5):285 -- 294.
[50] Schultz W. Predictive reward signal of dopamine neurons. Journal of Neurophysiology. 1998;p. 80: 1 -- 27.
[51] Esteves F, Parra C, Dimberg U, Ohman A. Nonconscious associative learning: Pavlovian conditioning of skin conduc-
tance responses to masked fear-relevant facial stimuli. Psychophysiology. 1994;p. 31:375 -- 385.
[52] Ohman A, Soares JJ. On the automatic nature of phobic fear: conditioned electrodermal responses to masked fear-
relevant stimuli. Journal of Abnormal Psychology. 1993;102(1):121 -- 132.
[53] Barrett LF, Simmons WK. Interoceptive predictions in the brain. Nature Reviews Neuroscience. 2015;16(7):419 -- 429.
[54] Pezzulo G, Rigoli F, Friston K. Active Inference, homeostatic regulation and adaptive behavioural control. Progress in
Neurobiology. 2015;134:17 -- 35.
[55] Adams RA, Shipp S, Friston KJ. Predictions not commands: active inference in the motor system. Brain Structure and
Function. 2013;218(3):611 -- 643.
[56] Feldman H, Friston KJ. Attention, uncertainty, and free-energy. Frontiers in Human Neuroscience. 2010;4:215.
[57] Scott DJ, Stohler CS, Egnatuk CM, Wang H, Koeppe RA, Zubieta JK.
Individual differences in reward responding
explain placebo-induced expectations and effects. Neuron. 2007;55(2):325 -- 336.
[58] Colloca L. Emotional modulation of placebo analgesia. Pain. 2014;155(4):651.
[59] De Pascalis V, Chiaradia C, Carotenuto E. The contribution of suggestibility and expectation to placebo analgesia
phenomenon in an experimental setting. Pain. 2002;96(3):393 -- 402.
[60] Wager TD, Scott DJ, Zubieta JK. Placebo effects on human mu-opioid activity during pain. Proceedings of the National
Academy of Sciences of the United States of America. 2007;104(26):11056 -- 11061.
[61] Watson A, El-Deredy W, Iannetti GD, Lloyd D, Tracey I, Vogt BA, et al. Placebo conditioning and placebo analgesia
modulate a common brain network during pain anticipation and perception. Pain. 2009;145(1-2):24 -- 30.
[62] Zubieta JK, Bueller JA, Jackson LR, Scott DJ, Xu Y, Koeppe RA, et al. Placebo effects mediated by endogenous opioid
activity on mu-opioid receptors. Journal of Neuroscience. 2005;25(34):7754 -- 7762.
[63] Cacioppo JT, Tassinary LG, Berntson G. Handbook of psychophysiology. Cambridge University Press; 2007.
[64] Goebel MU, Trebst AE, Steiner J, Xie YF, Exton MS, Frede S, et al. Behavioral conditioning of immunosuppression is
possible in humans. FASEB Journal. 2002;16(14):1869 -- 1873.
[65] Vits S, Cesko E, Enck P, Hillen U, Schadendorf D, Schedlowski M. Behavioural conditioning as the mediator of placebo
responses in the immune system. Philosophical Transactions of the Royal Society of London Series B, Biological
Sciences. 2011;366(1572):1799 -- 1807.
[66] Enck P, Benedetti F, Schedlowski M. New insights into the placebo and nocebo responses. Neuron. 2008;59(2):195 -- 206.
15
|
1501.00481 | 3 | 1501 | 2019-03-24T02:15:47 | Modeling the ballistic-to-diffusive transition in nematode motility reveals variation in exploratory behavior across species | [
"q-bio.NC",
"q-bio.PE"
] | A quantitative understanding of organism-level behavior requires predictive models that can capture the richness of behavioral phenotypes, yet are simple enough to connect with underlying mechanistic processes. Here we investigate the motile behavior of nematodes at the level of their translational motion on surfaces driven by undulatory propulsion. We broadly sample the nematode behavioral repertoire by measuring motile trajectories of the canonical lab strain $C. elegans$ N2 as well as wild strains and distant species. We focus on trajectory dynamics over timescales spanning the transition from ballistic (straight) to diffusive (random) movement and find that salient features of the motility statistics are captured by a random walk model with independent dynamics in the speed, bearing and reversal events. We show that the model parameters vary among species in a correlated, low-dimensional manner suggestive of a common mode of behavioral control and a trade-off between exploration and exploitation. The distribution of phenotypes along this primary mode of variation reveals that not only the mean but also the variance varies considerably across strains, suggesting that these nematode lineages employ contrasting ``bet-hedging'' strategies for foraging. | q-bio.NC | q-bio |
Modeling the ballistic-to-diffusive transition in
nematode motility reveals variation in
exploratory behavior across species
Stephen J. Helms∗1, W. Mathijs Rozemuller∗1, Antonio Carlos
Costa∗2, Leon Avery3, Greg J. Stephens2,4, and Thomas S. Shimizu
†1
1AMOLF Institute, Amsterdam, The Netherlands
2Dept. of Physics & Astronomy, Vrije Universiteit, Amsterdam,
The Netherlands
3Dept. of Physiology and Biophysics, Virginia Commonwealth
Univ., Richmond, VA, USA
4Okinawa Institute of Science and Technology, Onna-son,
Okinawa, Japan
Abstract
A quantitative understanding of organism-level behavior requires pre-
dictive models that can capture the richness of behavioral phenotypes, yet
are simple enough to connect with underlying mechanistic processes. Here
we investigate the motile behavior of nematodes at the level of their trans-
lational motion on surfaces driven by undulatory propulsion. We broadly
sample the nematode behavioral repertoire by measuring motile trajecto-
ries of the canonical lab strain C. elegans N2 as well as wild strains and
distant species. We focus on trajectory dynamics over timescales span-
ning the transition from ballistic (straight) to diffusive (random) move-
ment and find that salient features of the motility statistics are captured
by a random walk model with independent dynamics in the speed, bear-
ing and reversal events. We show that the model parameters vary among
species in a correlated, low-dimensional manner suggestive of a common
mode of behavioral control and a trade-off between exploration and ex-
ploitation. The distribution of phenotypes along this primary mode of
variation reveals that not only the mean but also the variance varies con-
siderably across strains, suggesting that these nematode lineages employ
contrasting "bet-hedging" strategies for foraging.
∗These authors contributed equally to this work.
†Electronic address: [email protected]; Corresponding author
1
Keywords:
species comparisons -- phenotyping
behavior -- dimensionality reduction -- random walk -- cross-
Introduction
A ubiquitous feature of biological motility is the combination of stereotyped
movements in seemingly random sequences. Capturing the essential character-
istics of motion thus requires a statistical description, in close analogy to the
random-walk formulation of Brownian motion in physics. A canonical exam-
ple is the "run-and-tumble" behavior of E. coli bacteria, in which relatively
straight paths (runs) are interspersed by rapid and random reorientation events
(tumbles) [1]. The random walk of E. coli can thus be characterized by two
random variables (run length and tumble angle) and two constant parameters
(swimming speed and rotational diffusion coefficient), and detailed studies over
decades have yielded mechanistic models that link these key behavioral param-
eters to the underlying anatomy and physiology [2 -- 5]. Random-walk theory
has been fruitfully applied also to studies of eukaryotic cell migration in both
two [6 -- 8] and three [9] dimensions.
Can a similar top-down approach be fruitfully applied to more complex
organisms -- for example, an animal controlled by a neural network? Animal be-
havior is both astonishing in its diversity and daunting in its complexity, given
the inherently high-dimensional space of possible anatomical, physiological, and
environmental configurations.
It is therefore essential to identify appropriate
models and parameterizations to succinctly represent the complex space of be-
haviors -- a non-trivial task that has traditionally relied on the insights of
expert biologists. In this study, we ask if one can achieve a similar synthesis
by an alternative, physically-motivated approach [10]. We seek a quantitative
model with predictive power over behavioral statistics, and yet a parameteriza-
tion that is simple enough to permit meaningful interpretations of phenotypes
in a reduced space of variables. As an example, we focus on the motile behavior
of nematodes, which explore space using a combination of random and directed
motility driven by undulatory propulsion.
The nematode C. elegans has long been a model organism for the genetics
of neural systems [11, 12], and recent advances in imaging have made it feasible
to record a large fraction of the worm's nervous system activity at single-cell
resolution [13 -- 15]. These developments raise the compelling possibility of elu-
cidating the neural basis of behavioral control at the organism scale, but such
endeavors will require unambiguous definitions of neural circuit outputs and
functional performance. The worm's behavioral repertoire [16, 17] is commonly
characterized in terms of forward motion occasionally interrupted by brief re-
versals [18 -- 20], during which the undulatory body wave that drives its move-
ment [21] switches direction. In addition, worms reorient with a combination of
gradual curves in the trajectory ("weathervaning") [22,23] and sharp changes in
body orientation (omega-turns [19] and delta-turns [24]). These elementary be-
haviors are combined in exploring an environment [22, 25]. Environmental cues
2
such as chemical, mechanical, or thermal stimuli [26] lead to a biasing of these
behaviors, guiding the worm in favorable directions [22,25,27]. Finally, in prac-
tical terms, the worm's small size (∼1 mm in length), moderate propulsive speed
(∼100 µm s−1) and short generation time (∼2 days) allow a considerable fraction
of its behavioral repertoire to be efficiently sampled in the laboratory [18, 28].
An influential example of such an analysis is the "pirouette" model proposed
by Pierce-Shimomura and Lockery [25] which describes the worm's exploratory
behavior as long runs interrupted occasionally by bursts of reversals and omega
turns that reorient the worm, in close analogy to the run-and-tumble model of
bacterial random walks [1]. Later work by Iino et al. identified that worms also
navigate by smoother modulations of their direction during long runs ("weath-
ervaning") [22], and Calhoun et al. have suggested that C. elegans may track
the information content of environmental statistics in searching for food [29], a
motile strategy that has been termed 'infotaxis' [30]. A recent study by Roberts
et al. [20] analyzed high (submicron) resolution kinematics of C. elegans loco-
motion and developed a stochastic model of forward-reverse switching dynamics
that include the short-lived (∼0.1 s) pause states that were identified between
forward and reverse runs.
Importantly, while these previous studies have illuminated different modes
of behavioral control, they were not designed to obtain a predictive model of the
trajectory statistics and thus a succinct parameterization of C. elegans motility
remains an important open problem. A quantitative parameterization capturing
the repertoire of C. elegans' behavioral phenotypes would facilitate data-driven
investigations of behavioral strategies: for example, whether worms demonstrate
distinct modes of motility (characterized by correlated changes in parameters)
over time, or in response to changes in environmental conditions [28, 31 -- 33].
Variation in the obtained parameters among individuals can inform on the dis-
tribution of behavioral phenotypes within a population, and reveal evolutionary
constraints and trade-offs between strategies represented by distinct parameter
sets [34].
C. elegans is a member of the Nematoda phylum, one of the largest and most
diverse phylogenetic groups of species [35, 36]. Despite the diversity of ecologi-
cal niches these animals inhabit [35], comparisons of nematode body plans have
revealed a remarkable degree of conservation, even down to the level of individ-
ual neurons [37]. This combination of highly conserved anatomy and ecological
diversity makes nematode motility a compelling case for studies of behavioral
phenotypes. Anatomical conservation suggests it might be possible to describe
the behavior of diverse nematodes by a common model, and identifying the
manner in which existing natural variation is distributed across the parame-
ter space of the model could reveal distinct motility strategies resulting from
optimization under different environmental conditions.
In this study, we develop a simple random walk model describing the trans-
lational movements of a diverse collection of nematode species, freely-moving
on a two-dimensional agar surface. In addition to providing a quantitative and
predictive measure of trajectory dynamics, the parameters of our model define
a space of possible behaviors. Variation within such a space can occur due to
3
changes in individual behavior over time (reflecting temporal variation in the
underlying sensorimotor physiology, or "mood"), differences in behavior among
individuals (reflecting stable differences in physiology, or "personality") and
differences between strains and species (reflecting cumulative effects of natural
selection). By quantitative analyses of such patterns of variation, we seek to
identify simple, organizing principles underlying behavior.
Results
Nematodes Perform Random Walks Off-Food with a Broad
Range of Diffusivities Across Strains
In order to identify conserved and divergent aspects of motility strategies, we
sampled motile behavior over a broad evolutionary range. We selected a phylo-
genetically diverse collection of nematodes with an increased sampling density
closer to the laboratory strain C. elegans (Figure 1A and Supplementary In-
formation, SI). To sample individual variation, we recorded the motility of up
to 20 well-fed individuals per strain and each individual for 30 minutes on a
food-free agar plate at 11.5 Hz with a resolution of 12.5 µm/px (see SI).
We measured the centroid position ((cid:126)x(t)) and calculated the centroid veloc-
ity ((cid:126)v(t)), using image analysis techniques (Figure S11 and SI). We chose the
centroid as the measure of the worm's position because it effectively filters out
most of the dynamics of the propulsive body wave. There was considerable
variation in the spatial extent and degree of turning visible in the trajectories
both within and across strains (Figure 1A, S2).
As previously seen in C. elegans [1], the measured mean-squared displace-
ment,
(cid:104)[∆x(τ )]2(cid:105) ≡ (cid:104)(cid:126)x(t + τ ) − (cid:126)x(t)2(cid:105),
(1)
revealed a transition from ballistic to diffusive motion within a 100 s timescale
(Figure 1B, S3). Over short times, the worm's path was relatively straight, with
the mean-squared displacement scaling quadratically with the time lag τ and
speed s as (cid:104)s2(cid:105)τ 2 (i.e. a log-log slope of 2). Over longer times, the slope de-
creased with τ reflecting the randomization of orientation characteristic of diffu-
sion, and an effective diffusivity Deff was extracted by fits to (cid:104)[∆x(τ )]2(cid:105) = 4Deffτ
(see SI). For times (cid:38) 100 s, the slope of the mean-squared displacement de-
creased yet further due to encounters of the worm with walls of the observation
arena. We confirmed that this confinement did not affect detection of the bal-
listic to diffusive transition (Figure S1). This analysis revealed that the visible
differences in the spatial extent of these 30-minute trajectories stem from vari-
ation by nearly an order of magnitude in speed and two orders of magnitude in
diffusivity (Figure 1C, Tables S1 & S2).
4
The Random Walk of Nematodes Can Be Decomposed into
Speed, Turning and Reversal Dynamics
The broad range of observed speeds and diffusivities suggest that these diverse
nematodes have evolved a variety of strategies for spatial exploration. To gain
further insights into the manner in which such contrasting behaviors are imple-
mented by each strain, we sought to extract a minimal model of the nematodes'
random walk by further decomposing the trajectory statistics of all nine mea-
sured strains. In this and the following three sections, we illustrate our analy-
sis and model development with data from three contrasting strains: CB4856
and PS312, which demonstrated two of the most extreme phenotypes, and the
canonical laboratory strain N2 (see SI for equivalent data for all strains).
The translational motion of the worm can be described by the time-varying
centroid velocity (cid:126)v(t) which can in turn be decomposed into speed s(t) and
direction of motion (hereafter referred to as its "bearing") φ(t):
(cid:126)v(t) =
d(cid:126)x(t)
dt
= s(t) [cos φ(t), sin φ(t)]
(2)
To account for head-tail asymmetry in the worm's anatomy, we additionally
define the body orientation (ψ(t); hereafter referred to simply as "orientation")
by the angle of the vector connecting the worm's centroid to the head (Figure
2A). The centroid bearing is related to this orientation of the worm by
φ(t) = ψ(t) + ∆ψ(t)
(3)
where the difference ∆ψ(t) is a measure of the alignment of the direction of
movement with the worm's body orientation (hereafter referred to simply as
"alignment"). We found for all strains that the distribution of ∆ψ(t) was bi-
modal with peaks at 0◦and 180◦(Figure 2C, S7A). These match the forward and
reverse states of motion described in C. elegans [18, 19].
Each of the three components of the worm's motility (speed, orientation,
and alignment) varied considerably over time and in qualitatively different ways
between strains (Figure 2B). For example, the three strains shown in Figure 2B
differed not only in their average speed, but also in the amplitude and timescale
of fluctuations about the average speed. Similarly, the statistics of orientation
fluctuations about the drifting mean also differed visibly between strains. Fi-
nally, transitions between forward and reverse runs were far more frequent in
PS312 as compared to N2 and CB4856. Given the apparently random manner
in which these motility components varied over time, we proceeded to analyze
the dynamics of each of these three components as a stochastic process.
Speed Dynamics
Speed control has not been extensively studied in C. elegans, but it is known that
worms move with a characteristic speed that is influenced by stimuli [26]. When
intervals corresponding to transitions between forward and reverse runs were
5
excluded from the time series, we found that the autocorrelation in speed fluc-
tuations decayed exponentially over a few seconds (Figure 3A, S5A), a timescale
similar to the period of the propulsive body wave. These dynamics are natu-
rally captured by an Ornstein-Uhlenbeck process [39], which describes random
fluctuations arising from white noise (increments of a diffusive Wiener process,
2Ds that relax with timescale τs back to an average
dWt [39]) with magnitude
value, µs = (cid:104)s(cid:105):
√
(cid:112)
−1
ds(t) = τ
s
[µs − s(t)] dt +
2DsdWt
(4)
Numerical integration of this equation closely reproduced the observed speed
distributions during runs (Figure S5B).
Diffusive Turning with Drift
The orientation ψ(t) captures turning dynamics that are independent of abrupt
changes in bearing φ(t) due to reversals. To change orientation, C. elegans
executes a combination of large, ventrally-biased [40] sharp turns [18, 24] and
gradual "weathervaning" [22], both of which contribute to randomization of ori-
entation over time. This random walk in orientation was not purely diffusive:
the orientation correlation Cψ(τ ) = (cid:104)cos [ψ(t + τ ) − ψ(t)](cid:105) does not decay expo-
nentially (Figures 3B Inset, S6B), and the mean-squared angular displacement,
MSAD(τ ) = (cid:104)[ψ(t + τ ) − ψ(t)]2(cid:105), increases nonlinearly with time (Figures 3B,
S6A).
We found that this nonlinear MSAD of ψ(t) could be well fit by a quadratic
function of the time delay τ : MSAD(τ ) = k2
ψrmsτ 2 + 2Dψτ , corresponding to a
diffusion-and-drift model with root-mean-square (rms) drift magnitude kψrms
and angular diffusion coefficient Dψ (see Supporting Information for derivation).
A non-zero drift magnitude kψrms (cid:54)= 0 indicates that in addition to purely ran-
dom (diffusive) changes in orientation, there is an underlying bias (i.e. direc-
tional persistence) in the worms' turning over 100 s windows, consistent with
previous studies in larger arenas [23].
These observations lead to a simple model for the orientation dynamics that
combines drift (approximated as a deterministic linear process over a 100 s win-
dow) with stochastic diffusion:
(cid:112)
dψ(t) = kψdt +
2DψdWt,
(5)
where we set the drift magnitude kψ = kψrms and dWt represents increments
of a Wiener process [39].
We note that while this model described well the orientation dynamics within
100 s windows, over longer timescales additional dynamics may be relevant. The
magnitude of kψ in our data (∼1 ◦ s−1) was similar to that of weathervaning
excursions reported for C. elegans navigating in salt gradients [22].
Forward and Reverse Runs
The observation that motion during runs switched abruptly between forward
and reverse states (with ∆ψ ≈ {0◦, 180◦}, respectively; Figures 2B,C,S7A)
6
suggested that reversals could be described as a discrete stochastic process.
The manner in which reversals contribute to randomization of bearing over a
time lag τ is captured by the autocorrelation function of ∆ψ(t), C∆ψ(τ ) ≡
(cid:104)cos(∆ψ(t + τ )− ∆ψ(t))(cid:105). We found that C∆ψ(τ ) decayed nearly exponentially
to a non-zero baseline (Figure 3C, Figure S7C). This is the predicted behav-
ior for the autocorrelation function of the simplest of two-state processes (a
"random telegraph process"):
P (Tfwd > t) = exp(−t/τfwd)
P (Trev > t) = exp(−t/τrev),
(6)
(7)
in which the distribution of forward and reverse run intervals (Tfwd and Trev) are
completely determined by a single time constant (τfwd and τrev, respectively).
The random telegraph process yields an autocorrelation function that decays ex-
ponentially as C∆ψ(τ ) = C∆ψ(∞) + (1 − C∆ψ(∞)) e−τ /τRT to a minimum value
C∆ψ(∞) ≡ ((τfwd − τrev)/(τrev + τfwd))2 with a timescale τRT ≡(cid:0)τ
(cid:1)−1
−1
fwd + τ−1
rev
[41]. Results obtained from fitting the autocorrelation function are consistent
with those obtained from the distribution of time intervals between detected
switching events (figure S7, SI). In principle, the forward and reverse states
could be characterized by differences in motility parameters of our model other
than these transition times, as forward and reverse motion are driven by distinct
command interneurons in C. elegans [42,43]. However, we found that run speeds
were nearly identical between forward and reverse runs (Figure S8). While we
expect that this symmetry will be broken under some specific conditions, such
as the escape response [44], the strong speed correlation between the two states
motivates the assumption, adopted in our model, that reversals change only the
bearing (by 180◦) and the propensity to reverse direction, represented in our
model by the time constants τfwd and τrev.
A Model with Independent Speed, Turning and Reversals
Captures the Ballistic-to-Diffusive Transition in Nematode
Motility
Given that the dynamics of the worm's speed, turning and reversals could be de-
scribed as simple stochastic processes, we asked whether combining them as in-
dependent components in a model of the worms' random walk could sufficiently
describe the observed motility statistics (Figure 4A). We simulated trajectories
of worms by numerically integrating equations (4)-(7) for the speed, orientation,
and reversal dynamics, respectively, which yields the worm's velocity dynamics
through equations (2) and (3), with ∆ψ(t) equal to 0◦ during forward runs and
180◦ during reverse runs. Simulations of this model using parameters fit to indi-
vidual worms produced trajectories that qualitatively resembled real trajectories
and varied considerably in their spatial extent (Figure 4B).
Next, we quantitatively assessed the performance of the model in reproduc-
ing the statistics of the observed trajectories over the time scale of 100 s, within
7
mean-squared displacement through (d/dt)(cid:104)[∆x(τ )]2(cid:105) = 2(cid:82) τ
which all strains completed the transition from ballistic to diffusive motion (Fig-
ure 4C). We found that the model based on independent speed, turning and
reversal dynamics closely reproduced not only the diffusivity of each strain but
also the time evolution of the mean-squared displacement ((cid:104)[∆x(τ )]2(cid:105)) across
the ballistic-to-diffusive transition (Figure 4C, top). A closer inspection of the
dynamics across this transition is possible by examining the velocity autocorre-
lation function (C(cid:126)v(τ )), the time integral of which determines the slope of the
0 dτ(cid:48)C(cid:126)v(τ(cid:48)), a vari-
ant of the Green-Kubo relation [45,46]. The transition from ballistic to diffusive
motion is characterized by the manner in which the normalized velocity auto-
correlation C(cid:126)v(τ )/C(cid:126)v(0) decays over the time lag τ from unity (at τ = 0) to zero
(as τ → ∞). We found that C(cid:126)v(τ ) varied considerably across strains, not only
in the overall ballistic-to-diffusive transition time, but also in the more detailed
dynamics of the autocorrelation decay over time (Figure 4C, middle). Salient
features, such as the transition time, of the measured velocity autocorrelation
functions C(cid:126)v,obs were reproduced closely by the simulated velocity autocorre-
lation functions C(cid:126)v,model, but there were also subtle deviations in the detailed
dynamics for a number of strains.
Given our model's simplifying assumption that dynamics for s(t), ψ(t), and
∆ψ(t) are independent stochastic processes, we asked whether the remaining
discrepancies between the simulated and measured velocity autocorrelation dy-
namics could be explained by violations of this assumption of independence. As
a model-free assessment of the degree of non-independence, we first calculated
the predicted velocity autocorrelation for the case that the dynamics of all three
components are independent, C(cid:126)v,indep(τ ) = Cs(τ )Cψ(τ )C∆ψ(τ ), where Cs(τ ),
Cψ(τ ), and C∆ψ(τ ) are the autocorrelation functions of the measured data for
each of the components (see Supporting Information for derivation). We then
compared the differences C(cid:126)v,obs−C(cid:126)v,indep (blue curve in Figure 4C, bottom) and
C(cid:126)v,obs − C(cid:126)v,model (red curve in Figure 4C, bottom). Indeed, there were subtle
differences both on shorter (∼1 s) and longer timescales (∼10 s). However, these
errors for the simulated model were very similar to, or less than, those for the
model-free prediction from the data under the assumption of independence (i.e.,
C(cid:126)v,obs − C(cid:126)v,model (cid:46) C(cid:126)v,obs − C(cid:126)v,indep). These results demonstrate that modeling
s(t), ψ(t), and ∆ψ(t) as independent stochastic processes provides a very good
approximation to trajectory statistics across the ballistic-to-diffusive transition.
The relatively subtle differences between the data and model arise primarily
in instances where this assumption of independence between the three motil-
ity components breaks down. Consistent with these conclusions, inspection of
cross-correlation functions computed from the data revealed that correlations
between s(t), ψ(t), and ∆ψ(t) are largely absent, with only weak correlations
between speed (s) and reversals (∆ψ) in a subset of strains (Fig. S9).
Variation of exploratory behavior across Species
The results presented in the previous sections demonstrate that a random-walk
model with seven parameters describing independent speed, turning and reversal
8
dynamics, provides a good approximation of the worms' motile behavior over the
∼100 s timescale spanning the ballistic-to-diffusive transition. The model pa-
rameters thus define a seven-dimensional space of motility phenotypes in which
behavioral variation across strains and species can be examined. If components
of behavior were physiologically regulated or evolutionarily selected for in a co-
ordinated manner, we would expect to find correlated patterns in the variation
of these traits.
We fit our model to the trajectory statistics of each individual worm and
built a phenotype matrix of 106 worms x 7 behavioral parameters (summarized
in Tables S2-S4). The correlation matrix for these 7 parameters demonstrates
that the strongest correlation were the forward and reverse state lifetimes (τf wd,
τrev), followed by those describing speed and forward state life times (µs, τf wd).
More broadly, there were extensive correlations among the model parameters,
not only within the parameters of each motility component (speed, orientation,
reversals) but also between those of different components.
We looked for dominant patterns in the correlations using principal compo-
nent analysis [47] (Figure 5B), uncovering a single dominant mode of correlated
variation (Figure 5B, left). This principal mode (mode 1), capturing nearly 40%
of the total variation, described significant correlations among all the parame-
ters except for Ds and Dψ (Figure 5B, right, Table S5). We did not attempt
to interpret higher modes since, individually, they either did not significantly
exceeded the captured variance under a randomization test (mode 3 and higher;
see SI, and Figure 5B, left) or were found upon closer inspection to be dominated
by parameter correlations arising from fitting uncertainties (mode 2).
We used numerical simulations to determine the effects on motile behavior of
varying parameters along the principal mode. The measured trajectory pheno-
types projected onto this mode in the range {−4, 2} centered about the average
phenotype located at the origin, and we performed simulations for parameter
sets evenly sampled along this range. These largely reproduced the observed
variation in the measured diffusivities Deff as a function of the projection along
the first mode. The agreement was particularly good at higher values (> −1) of
the mode projection, but at lower values we noted a tendency for the Deff from
simulations to exceed that of the data. The latter discrepancy can be explained
by elements of behavior not captured by our model (see Discussion). Neverthe-
less, as illustrated by simulated trajectories (Figure 5C, bottom), trajectories
became more expansive as the mode projection increased, as did Deff by nearly
two orders of magnitude over the tested range. This suggested that the princi-
pal mode indicates exploratory propensity (Figure 5C), and we confirmed that
it is indeed more strongly associated with changes in Deff than expected for
randomly generated parameter sets (Figure S10).
Interestingly, this mode of
variation we found across individual phenotypes is reminiscent of "roaming"
and "dwelling" behavioral variability that has been shown within individuals
across time, in C. elegans [28, 32] as well as other organisms [48, 49].
9
Specialized and Diversified Behavioral Strategies Across
Strains
The principal behavioral mode discussed in the preceding section was identified
by analyzing variation across all individual worms measured in this study, com-
ing from diverse strains and species that differ in their average behavior (see
Tables S2 - S4). How does the variability among individuals of a given strain
compare to differences between the average phenotypes of strains/species? On
the one hand, each strain might be highly "specialized", with relatively small
variation within strains as compared to that across strains. On the other hand,
strains might implement "diversified" strategies in which genetically identical
worms vary strongly in their behavior. To address these two possibilities, we
analyzed the distribution of individual phenotypes within each strain, as well
as that of the set of averaged species phenotypes.
For each measured individual, we computed the projection of its motility
parameter set along the principal behavioral mode and estimated strain-specific
distributions of this reduced phenotype (Figure 6, Table S6). In principle, any
detail in the shape of these distributions could be relevant for evolutionary
fitness, but here we focused our analysis on the mean and standard deviation,
given the moderate sampling density (≤ 20 individuals per strain). Further,
we computed the principal-mode projection of the average phenotype of each
species to define an interspecies phenotype distribution (Figure 6).
Strains varied considerably in both the position and breadth of their pheno-
typic distributions along the principal behavioral mode. Remarkably, variation
across individuals within each strain was comparable in magnitude to that for
the set of average phenotypes across species (Figure 6). Some strains were
specialized towards roaming or dwelling behavior, such as CB4856 and PS312,
respectively, with a strong bias in their behavior and comparatively low individ-
ual variability. Others, such as QX1211 and PS1159, appeared more diversified
with an intermediate average phenotype and higher individual variability. These
considerable differences in phenotype distributions across strains reveal the evo-
lutionary flexibility of population-level heterogeneity in nematodes, and suggest
a possible bet-hedging mechanism for achieving optimal fitness in variable en-
vironments [50, 51].
In assessing such variability of phenotypes, it is essential to ask how uncer-
tainty in the determined parameters (obtained from model fits) contribute to
the observed variability in phenotypes. We therefore computed the contribu-
tion of uncertainties in the individual phenotype determination by bootstrap
resampling of the 100 s windows of each individual's recorded trajectory (see
SI). The uncertainties thus computed reflect contributions from both parameter
uncertainties in curve fitting of data, as well as temporal variability in an indi-
vidual's parameters over timescales longer than the window size (100 s). With
the exception of two strains (sjh2 and CB4856), this measure of uncertainty ac-
counted for less than half of the individual variation within each strain (Figure
6B). These findings support the view that the phenotypic variation estimated in
the current analysis largely represented stable differences in individual behavior.
10
Discussion
We have presented a comparative quantitative analysis of motile behavior across
a broad range of strains and species of the nematode phylum, ranging from the
lab strain C. elegans N2 to Plectus sjh2 at the base of the chromadorean ne-
matode lineage. Despite the vast evolutionary distances spanned by strains in
this collection [52], we found that a behavioral model described by only seven
parameters could account for much of the diversity of the worms' translational
movement across the ∼100 s timescale spanning the ballistic-to-diffusive transi-
tion. This simple model provides a basis for future studies aiming to capture
more detailed aspects of nematode behavior, or to connect sensory modulation
of behavior to the underlying physiology. More generally, our results demon-
strate how quantitative comparisons of behavioral dynamics across species can
provide insights regarding the design of behavioral strategies.
The Minimal Model: What Does It Capture, and What
Does It Miss?
We focused on a high-level output of behavior -- translational and orientational
trajectory dynamics -- and sought to build the simplest possible quantitative
model that could capture the observed behavioral statistics. We found that a
model with only three independent components -- (1) speed fluctuations that
relax to a set point on a timescale of a few seconds, (2) orientation fluctuations
with drift, and (3) stochastic switching between forward and reverse states of
motion -- describes well, overall, the trajectory statistics of all tested nematode
species across the ballistic-to-diffusive transition (Figure 4).
Notably, we have not included explicit representations of some reorientation
mechanisms that have been studied in the past, such as the deep turns (omega-
and delta-turns) [18,24], or the combination of such turns with reversals (pirou-
ettes) [25]. In our data, we find that the timing of the initiation and termination
of reversals, which would both count as runs in the pirouette description, fol-
low exponential distributions with similar time constants as previously reported
for the pirouette run distribution. While omega and delta turns must indeed
be mechanistically distinct from gradual turns, we have chosen here not to ex-
plicitly model their occurrence since orientation changes in our trajectory data
were adequately described by a continuous diffusion-drift process (Figures 3B,
S6A). It is possible, however, that explicit representations of pirouettes and/or
omega turns would be important in other experimental scenarios, e.g. those
that include navigation in the presence of gradient stimuli.
In our model, "roaming" and "dwelling" were not assigned discrete behav-
ioral states (as was done e.g. in [28,31,32]), but instead emerged as a continuous
pattern of variation among motility parameters describing the worm's random
walk. However, robust extraction of motility parameters required pre-filtering
of trajectory data that likely biased them towards more "roaming" phenotypes
(see SI), which we believe account for the noted tendency of model simulations
to overestimate Deff that was more pronounced for trajectories at the "dwelling"
11
end of the spectrum (Figure 5C).
In its current form, our simple model does not account for possible corre-
lations between the dynamics of the three motility components (speed, orien-
tation, and reversals). Indeed, at least weak correlations do exist between the
components (Figure S9). Comparisons of simulated versus measured trajecto-
ries demonstrated that the effects of such correlations on the motility statistics
are small but detectable (Figure 4C). The differences were most significant for
the velocity-autocorrelation dynamics on a ∼10 s timescale, and were similar to
those for model-free predictions obtained by combining component-wise correla-
tion functions under the assumption of independence. Discrepancies on this in-
termediate timescale occurred most often in fast-moving strains that frequently
approached the repellent boundary. Therefore, we suspect that the discrep-
ancy arises from a stereotyped sequence, such as the escape response [44], that
introduces temporal correlations between speed changes, turning, and reversals.
While here we have focused on the transition to diffusive motion, some recent
experiments suggest that C. elegans might engage in superdiffusive behavior on
timescales longer than 100 s [23, 33]. Superdiffusive behavior could arise from
nonstationarities in motile behavior, such as the roaming/dwelling transitions
on timescales of several minutes [32]. Another mechanism for superdiffusion is
directed motility [23] in response to external stimuli such as chemical or thermal
gradients.
In such environments, nematodes are known to use at least two
distinct mechanisms for navigation [22,25] and the model here could be extended
by studying the dependence of motility parameters on environmental statistics.
Information about the body shape can be incorporated to build a more
complete behavioural model that also includes dynamics hidden by centroid
behaviour [1, 53]. Indeed, work by Brown et al. showed that a rich repertoire
of dynamics can be identified as temporal "motifs" in the postural time series
of C. elegans and used to classify mutants with high discriminatory power [54].
We have found that all of the species tested here can also be described with a
common set of postural modes (not shown), suggesting future directions on the
evolutionary space of postural dynamics.
The exploratory behavioral mode: Variability and its Phys-
iological Basis
While we found that a single behavioral model could be used to character-
ize nematode motility across the chromadorean lineage, the parameters of the
model varied extensively from strain to strain. Quantitatively, about 37% of
the variation corresponded to a correlated change in the parameters underlying
the timing of forward and reverse runs and the dynamics controlling speed and
turning (Figure 5B). We find that this principal mode of variation is associated
with strong changes in exploratory propensity, as characterized by Deff (Figure
5C). This pattern of parameter variation drove a change from low speed short
runs to high speed long runs, resembling the canonical descriptions of roaming
and dwelling in C. elegans [32].
12
Roaming and dwelling are thought to represent fundamental foraging strate-
gies reflecting the trade-off between global exploration and local exploitation of
environmental resources [55]. Recent work has suggested that such archetypal
strategies can be recovered by quantitatively analyzing the geometry of phe-
notypic distributions in parameter space [31, 34]. The motility phenotypes we
found in the present study were biased along one principal dimension, with
the extremes corresponding to roaming and dwelling behaviors. This obser-
vation compels us to suggest that an exploration-exploitation trade-off is the
primary driver of phenotypic diversification in the motility of chromadorean ne-
matodes in the absence of stimuli. Interestingly, a recent study on the motility
of a very different class of organisms (ciliates) yielded a similar conclusion [49]:
across two species and different environments, the diversity of motility pheno-
types was found to be distributed principally along an axis corresponding to
roaming and dwelling phenotypes. The emergence of roaming/dwelling as the
principal mode of variation in such disparate species underscores the idea that
the exploration-exploitation trade-off is a fundamental constraint on biological
motility strategies.
A surprising finding in our study was that, for a majority of strains, the ex-
tent of behavioral variability across individuals within a strain was comparable
to that for variation of phenotypes across species (Figure 6). In slowly changing
environments, the most evolutionarily successful species are those that consis-
tently perform well in that environment. This can be achieved by evolving a spe-
cialized, high fitness phenotype that varies little among individuals (such as with
PS312 and sjh2). However, increased phenotypic variability among individuals
can improve fitness in more variable environments if some individuals perform
much better in each condition -- a so-called "bet-hedging" strategy [50, 51]. The
large variability we observed among individual phenotypes within each strain
might reflect such a bet-hedging strategy in nematode exploratory behavior.
The observation that the variation among genetically identical individuals
can be comparable to that between disparate species raises the intriguing possi-
bility that there exist conserved molecular and/or physiological pathways driv-
ing diversification of spatial exploration strategies. Analogous variation in ex-
ploratory behavior was also detected in an analysis of nonstationarity in the
behavior of wild-type and mutant C. elegans under various nutritional condi-
tions [31]. Physiologically, protein kinase G (PKG) signaling and DAF-7 (TGF-
β) signaling from the ASI neuron are thought to be major mechanisms control-
ling roaming and dwelling in C. elegans [28, 31]. PKG signaling is also involved
in controlling foraging in Drosophila and other insects as well as many aspects
of mammalian behavior [56, 57]. Flavell et al. also elucidated a neuromod-
ulatory pathway involving serotonin and the neuropeptide pigment dispersing
factor (PDF) controlling the initiation and duration of roaming and dwelling
states [32].
Perturbations to the molecular parameters of such pathways underlying
global behavioral changes might provide a mechanism for the observed cor-
related variations at the individual, intra-, and inter-species levels. The identi-
fication of such conserved pathways affecting many phenotypic parameters is of
13
fundamental interest also from an evolutionary perspective, as they have been
proposed to bias the outcome of random mutations towards favorable evolution-
ary outcomes [58,59]. Our simple model provides a basis for future investigations
to uncover conserved mechanisms that generate behavioral variability, by defin-
ing a succinct parameterization of behavior that can be combined with genetic
and physiological methods.
Acknowledgments
We thank Massimo Vergassola, Vasily Zaburdaev, Alon Zaslaver, and Jeroen
van Zon for helpful suggestions and critical reading of the manuscript, Will
Ryu, Aravi Samuel and Andre Brown for inspiration and encouragement, and
members of the Shimizu lab for discussions. Casper Quist and Hans Helder of
Wageningen University provided wild nematodes isolated from soil and useful
information regarding the ecology of nematodes.
14
Figure 1: Nematodes perform random walks off-food with a mean speed and
effective diffusivity that varies across strains. (A) Phylogenetic tree with the
strains used in this study. The bold numbers are the major clades of Nematoda.
The gray box indicates genetically distinct wild isolates of C. elegans. A repre-
sentative worm image and 30 minute trajectory are shown to the right. Shaded
regions indicate a 95% confidence interval. (B) The average mean-squared dis-
placement, MSD, across N2 individuals is shown in black. For comparison, we
show the MSD expected from ballistic (blue) and diffusive (red) dynamics. The
motility transitions from a ballistic to diffusive regime within a time scale of tens
of seconds. (C) Mean speed (cid:104)s(cid:105) and effective diffusivity Deff (mean and 95%
confidence intervals) for each strain, calculated from fits of the mean-squared
displacement as in B. Across strains, both (cid:104)s(cid:105) and Deff vary by orders of mag-
nitude.
15
Aτ (s) BCMSD (µm2)0.111010010110210310410510610710810100300103104105(µm2/s)DPS312QX1211PS1159DF5020N2JU757CB4856JU775sjh2s(µm/s)DiffusiveDataBallistics2τ24DeffτCaenorhabditis elegansN2Caenorhabditis elegansCB4856Caenorhabditis elegansJU775Caenorhabditis elegansQX1211Caenorhabditis briggsaeJU757Rhabditis myriophilaDF5020Pristionchus pacificusPS312Panagrolaimus sp. PS1159Intraspecies Variation111287654321Outgroups: Arthropoda, Nematomorpha, etc.109Interspecies VariationPlectus sp.sjh2 (wild)VIVIIIIIIChromadoriaEnoplida and TriplonchidaDorylaimida250 µm10 mm Figure 2: The random walk of nematodes is composed of speed, turning, and
reversal dynamics. (A) We describe the motility of the worm by the time-varying
quantities s(t) (speed; black), ψ(t) (orientation; red), and ∆ψ(t) (alignment;
green) which measures the difference between the alignment of the velocity φ(t)
(blue) and ψ(t). (B) One minute examples of speed, orientation, and velocity
alignment time series for individuals from three exemplar strains.
(C) The
probability distribution of ∆ψ(t) reveals bimodality corresponding to forward
and reverse motion. Shaded regions indicate a 95% confidence interval.
16
N2CB4856PS312ABCForwardReverse10sx(t)3002001000ψ (deg)cos ∆ψ∆ψ(t) = φ(t) - ψ(t)v(t) =dxdt= s(t)[cos φ(t), sin φ(t)]180-18090-900s (µm/s)1-10090180∆ψ(deg)021N2CB4856PS312ProbabilityDensitytime Figure 3: Statistical characterization of the motility dynamics. (A) The auto-
correlation of the speed indicated that fluctuations decayed exponentially over a
few seconds. (A, inset) Speed distributions for three exemplar strains. (B) The
mean-squared angular displacement (MSAD) increased quadratically. (B, in-
set) The orientation autocorrelation function did not decay exponentially, with
some worms demonstrating significant undershoots below zero. (C) The veloc-
ity alignment autocorrelation decayed exponentially over tens of seconds to a
positive constant. In each plot, the ensemble average for all individuals from
the strains are shown with solid lines and trends are shown with dashed liens.
Shaded regions indicate a 95% confidence interval.
17
A010150246810Prob. Dens. (×10-3) 01020300100200300s(µm/s)τ(s)505025015105τ(s)50250τ(s)B0.00.51.0MSAD (rad2)C0501000.01.00.5τ(s)Cs (×102µm2/s2)CψC∆ψ Figure 4: A model consisting of independent speed (Ornstein-Uhlenbeck pro-
cess), turning (drift and diffusion), and reversal dynamics (random telegraph
process) quantitatively captures nematode motility. (A) Summary of the model.
(B) Simulated trajectories for the three exemplar strains. (C) Statistical com-
parison of the data (black) and simulations (red), ensemble averaged across in-
dividuals for each strain. (C, top) The mean-squared displacement (MSD) was
closely reproduced in all cases. (C, middle) The normalized velocity autocorre-
lation, C(cid:126)v(τ )/C(cid:126)v(0), (VACF) was less well captured. (C, bottom) The relatively
small errors in the simulated VACF (red) can be traced to the assumption of
independence in the dynamics of the speed, orientation, and velocity alignment
(blue). Shaded regions indicate a 95% confidence interval.
18
N2CB4856N2CB4856JU775QX1211JU757DF5020PS312PS1159sjh21021081061040.00.51.0ModelspeedOrnstein-Uhlenbeck processreversalsRandom telegraph process1mmMSD(µm(cid:31))0.00.51.0v(t)s(t)ψ(t)∆ψ(t)φ(t)(t)x(t)τ(s)0.11101000.11101000.11101000.11101000.11101000.11101000.11101000.11101000.1110100DataIndependentABCvCError in CvPS312 Figure 5: Motility parameters co-vary along an axis controlling exploratory
behavior. (A) Correlation matrix of the behavioral parameters across the whole
dataset. (B, left) Fraction of variance captured by each mode and the amount
expected for an uncorrelated dataset (red line). (B, right) The loadings on the
top eigenvector. (C) The effective diffusivity (top) and a 30 minute trajectory
(bottom, colors match points on graph) from simulations in which the loading
on the top eigenvector was varied; the principal mode can be used as an effective
phenotype from a more dwelling to a more roaming behavior. The projections
and effective diffusivity of the measured trajectories are shown as black points,
and the average of each strain is shown as a square.
19
Deff (µm2/s)1 cm-4-202More dwellingMore roaming% VarianceModeslog10µslog10τslog10Dslog10kψlog10Dψlog10τfwdlog10τrevSpeedDynamicsDynamicsReversal StateDynamicslog10µslog10τslog10Dslog10kψlog10Dψlog10τfwdlog10τrevABCWeightMode 10.00.5-0.5PS312QX1211N2JU757PS1159JU775sjh2DF5020CB4856 Figure 6: Variation of model parameters reveal specialized and diversified be-
havioral strategies across strains. (A) Distribution of the average phenotype for
each species (interspecies variation, blue) or individuals within a strain (red).
Note that the variation of individual phenotypes in some strains (e.g. QX1211,
PS1159) is comparable in magnitude to that of interspecies variation. Obser-
vations are indicated with colored ticks. (B) Comparison of the width of the
phenotype distributions (colored bars), quantified as the bootstrapped standard
deviation of the data points in (A), with the uncertainty in the determination of
the phenotype (black bars), quantified as the standard deviation of individual
phenotype determinations over bootstrapped 100 s time windows. Error bars
correspond to 95% confidence intervals across bootstrap samples.
20
0123σ Behavioral modeSpecies MeansIndividuals−6−3036Mode Projec�onMore dwellingMore roamingN2CB4856JU775QX1211JU757DF5020PS312PS1159sjh2C. elegansIntraspecies Varia�onInterspeciesBAProbability Densityσ Behavioral Mode0123 References
[1] Berg HC, Brown DA. Chemotaxis in Escherichia coli analysed by Three-
dimensional Tracking. Nature. 1972 Oct;239(5374):500 -- 504.
[2] Lovely PS, Dahlquist FW. Statistical measures of bacterial motility and
chemotaxis. J Theor Biol. 1975;50:477 -- 496.
[3] Schnitzer M, Block S, Berg H, Purcell E. Strategies for chemotaxis. Symp
Soc Gen Microbio. 1990;46:15 -- 34.
[4] De Gennes PG. Chemotaxis: The role of internal delays. Eur Biophys J.
2004;33:691 -- 693.
[5] Celani A, Shimizu TS, Vergassola M. Molecular and Functional Aspects of
Bacterial Chemotaxis. J Stat Phys. 2011;144:219 -- 240.
[6] Gail MH, Boone CW. The locomotion of mouse fibroblasts in tissue culture.
Biophys J. 1970 Oct;10(10):980 -- 993.
[7] Tranquillo RT, Lauffenburger DA, Zigmond SH. A stochastic model for
leukocyte random motility and chemotaxis based on receptor binding fluc-
tuations. J Cell Biol. 1988 Feb;106(2):303 -- 309.
[8] Selmeczi D, Mosler S, Hagedorn PH, Larsen NB, Flyvbjerg H. Cell motility
as persistent random motion: theories from experiments. Biophys J. 2005
Aug;89(2):912 -- 931.
[9] Wu PH, Giri A, Sun SX, Wirtz D. Three-dimensional cell migration does
not follow a random walk. PNAS. 2014 Mar;111(11):3949 -- 3954.
[10] Brown AE, De Bivort B. Ethology as a physical science. Nature Physics.
2018;p. 1.
[11] Brenner S.
The Genetics of Caenorhabditis elegans.
Genetics.
1974;77(1):71 -- 94.
[12] Bargmann CI, Marder E. From the connectome to brain function. Nature
methods. 2013;10(6):483 -- 490.
[13] Kato S, Kaplan HS, Schrodel T, Skora S, Lindsay TH, Yemini E, et al.
Global brain dynamics embed the motor command sequence of Caenorhab-
ditis elegans. Cell. 2015;163(3):656 -- 669.
[14] Venkatachalam V, Ji N, Wang X, Clark C, Mitchell JK, Klein M,
et al. Pan-neuronal imaging in roaming Caenorhabditis elegans. PNAS.
2016;113(8):E1082 -- E1088.
[15] Nguyen JP, Shipley FB, Linder AN, Plummer GS, Liu M, Setru SU, et al.
Whole-brain calcium imaging with cellular resolution in freely behaving
Caenorhabditis elegans. PNAS. 2016;113(8):E1074 -- E1081.
21
[16] Gjorgjieva J, Biron D, Haspel G. Neurobiology of Caenorhabditis elegans
Locomotion: Where Do We Stand? Bioscience. 2014 May;64(6):476 -- 486.
[17] Cohen N, Sanders T. Nematode locomotion: dissecting the neuronal --
environmental loop. Curr Opin Neurobiol. 2014 Apr;25:99 -- 106.
[18] Croll NA. Components and patterns in the behaviour of the nematode
Caenorhabditis elegans. J Zool. 1975;176:159 -- 176.
[19] Croll NA. Behavioural analysis of nematode movement. Adv Parasitol.
1975 Jan;13:71 -- 122.
[20] Roberts WM, Augustine SB, Lawton KJ, Lindsay TH, Thiele TR, Izquierdo
EJ, et al. A stochastic neuronal model predicts random search behaviors
at multiple spatial scales in C. elegans. Elife. 2016;5:e12572.
[21] Gray J, Lissmann HW. The locomotion of nematodes. J Exp Biol. 1964
Mar;41:135 -- 54.
[22] Iino Y, Yoshida K. Parallel use of two behavioral mechanisms for chemo-
taxis in Caenorhabditis elegans. J Neurosci. 2009 Apr;29(17):5370 -- 80.
[23] Peliti M, Chuang JS, Shaham S. Directional locomotion of C. elegans in
the absence of external stimuli. PLoS One. 2013 Jan;8(11):e78535.
[24] Broekmans OD, Rodgers JB, Ryu WS, Stephens GJ. Resolving coiled
shapes reveals new reorientation behaviors in C-elegans. eLife. 2016;5.
[25] Pierce-Shimomura JT, Morse TM, Lockery SR. The fundamental role
J Neurosci. 1999
of pirouettes in Caenorhabditis elegans chemotaxis.
Nov;19(21):9557 -- 69.
[26] Faumont S, Lindsay T, Lockery S. Neuronal microcircuits for decision
making in C. elegans. Curr Opin Neurobiol. 2012 Jun;22(4):580 -- 591.
[27] Ryu WS, Samuel ADT. Thermotaxis in Caenorhabditis elegans analyzed
J Neurosci. 2002
by measuring responses to defined thermal stimuli.
Jul;22(13):5727 -- 33.
[28] Fujiwara M, Sengupta P, McIntire SL. Regulation of body size and be-
havioral state of C. elegans by sensory perception and the EGL-4 cGMP-
dependent protein kinase. Neuron. 2002 Dec;36(6):1091 -- 102.
[29] Calhoun AJ, Chalasani SH, Sharpee TO. Maximally informative foraging
by Caenorhabditis elegans. Elife. 2014 Jan;3.
[30] Vergassola M, Villermaux E, Shraiman BI.
'Infotaxis' as a strategy for
searching without gradients. Nature. 2007;445(7126):406 -- 409.
[31] Gallagher T, Bjorness T, Greene R, You YJ, Avery L. The geometry of lo-
comotive behavioral states in C. elegans. PLoS One. 2013 Jan;8(3):e59865.
22
[32] Flavell SW, Pokala N, Macosko EZ, Albrecht DR, Larsch J, Bargmann
CI. Serotonin and the neuropeptide PDF initiate and extend opposing
behavioral states in C. elegans. Cell. 2013 Aug;154(5):1023 -- 1035.
[33] Salvador LCM, Bartumeus F, Levin SA, Ryu WS. Mechanistic analy-
sis of the search behaviour of Caenorhabditis elegans. J R Soc Interface.
2014;11:20131092 -- 20131092.
[34] Shoval O, Sheftel H, Shinar G, Hart Y, Ramote O, Mayo A, et al. Evolu-
tionary trade-offs, Pareto optimality, and the geometry of phenotype space.
Science. 2012 Jun;336(6085):1157 -- 60.
[35] De Ley P. A quick tour of nematode diversity and the backbone of nematode
phylogeny. WormBook. 2006;.
[36] Corsi AK, Wightman B, Chalfie M. A Transparent Window into Biology:
A Primer on Caenorhabditis elegans. Genetics. 2015 Jun;200(2):387 -- 407.
[37] Rabinowitch I, Schafer W. Neuronal remodeling on the evolutionary
timescale. J Biol. 2008 Jan;7(10):37.
[38] Stephens GJ, Johnson-Kerner B, Bialek W, Ryu WS. From modes to
movement in the behavior of Caenorhabditis elegans. PLoS One. 2010
Jan;5(11):e13914.
[39] Kampen NGV. Stochastic Processes in Physics and Chemistry. North
Holland; 2007.
[40] Gray JM, Hill JJ, Bargmann CI. A circuit for navigation in Caenorhabditis
elegans. PNAS. 2005 Mar;102(9):3184 -- 91.
[41] Papoulis A. Probability, Random Variables, and Stochastic Processes. 2nd
ed. New York: McGraw-Hill; 1984.
[42] Chalfie M, Sulston J, White J, Southgate E, Thomson J, Brenner S. The
neural circuit for touch sensitivity in Caenorhabditis elegans. J Neurosci.
1985 Apr;5(4):956 -- 964.
[43] Piggott BJ, Liu J, Feng Z, Wescott SA, Xu XZS. The neural circuits and
synaptic mechanisms underlying motor initiation in C. elegans. Cell. 2011
Nov;147(4):922 -- 933.
[44] Culotti JG, Russell RL. Osmotic avoidance defective mutants of the nema-
tode Caenorhabditis elegans. Genetics. 1978 Oct;90(2):243 -- 56.
[45] Green MS. Markoff Random Processes and the Statistical Mechanics of
Time-Dependent Phenomena. II. Irreversible Processes in Fluids. J Chem
Phys. 1954 Dec;22(3):398.
23
[46] Kubo R. Statistical-Mechanical Theory of Irreversible Processes. I. General
Theory and Simple Applications to Magnetic and Conduction Problems. J
Phys Soc Japan. 1957 Jun;12(6):570 -- 586.
[47] P Murphy K. Machine Learning: A Probabilistic Perspective. Cambridge,
MA: The MIT Press; 2012.
[48] Osborne Ka, Robichon A, Burgess E, Butland S, Shaw RA, Coulthard A,
et al. Natural behavior polymorphism due to a cGMP-dependent protein
kinase of Drosophila. Science. 1997 Aug;277(5327):834 -- 6.
[49] Jordan D, Kuehn S, Katifori E, Leibler S. Behavioral diversity in microbes
and low-dimensional phenotypic spaces. PNAS. 2013 Aug;110(34):14018 --
23.
[50] Slatkin M.
Hedging one's
evolutionary bets.
Nature. 1974
Aug;250(5469):704 -- 705.
[51] Philippi T, Seger J. Hedging one's evolutionary bets, revisited. Trends
Ecol Evol. 1989 Feb;4(2):41 -- 44.
[52] Kiontke K, Fitch DHA. Nematodes. Curr Biol. 2013 Oct;23(19):R862 --
R864.
[53] Stephens GJ, Johnson-Kerner B, Bialek W, Ryu WS. Dimensionality
and dynamics in the behavior of C. elegans. PLoS Comput Biol. 2008
Apr;4(4):e1000028.
[54] Brown AEX, Yemini EI, Grundy LJ, Jucikas T, Schafer WR. A dictio-
nary of behavioral motifs reveals clusters of genes affecting Caenorhabditis
elegans locomotion. PNAS. 2013 Jan;110(2):791 -- 6.
[55] Davies NB, Krebs JR, West SA. An Introduction to Behavioural Ecology.
4th ed. Wiley-Blackwell; 2012.
[56] Reaume CJ, Sokolowski MB. cGMP-dependent protein kinase as a modifier
of behaviour. Handb Exp Pharmacol. 2009 Jan;191:423 -- 43.
[57] Kaun KR, Sokolowski MB. cGMP-dependent protein kinase: linking for-
aging to energy homeostasis. Genome. 2009 Jan;52(1):1 -- 7.
[58] Kirschner M, Gerhart J. Evolvability. Proc Natl Acad Sci USA. 1998
Jul;95(15):8420 -- 7.
[59] Gerhart J, Kirschner M. The theory of facilitated variation. Proc Natl
Acad Sci U S A. 2007 May;(suppl 1):8582 -- 9.
24
Supporting Information
SI Materials and Methods
Selection of Strains
nematode
divided
into
classically
phylum is
A phylogenetic tree with the strains used in this study is shown in Figure 1A.
The
three major
branches -- chromadorea, enoplea, and dorylaimia -- that are broken into a total
of five major B-clades [S2] and twelve minor H-clades [S3]. The chromadorean
lineage is the largest, spanning B-clades III-V and H-clades 3-12 [S2,S3]. C.
elegans is located in clade V9 (the rhabditids), one of the most diverse clades
[S4]. In addition to the lab strain N2, we selected three of the most genetically
distinct wild isolates of C. elegans (CB4856, JU775, and QX1211) to sample
intraspecies variation [S5]. From H-clade 9 in order of increasing evolution-
ary distance, we selected Caenorhabditis briggsae JU757, Rhabditis myriophila
DF5020, and Pristionchus pacificus PS312. The next closest major group, B-
clade IV, contains H-clades 10-12. H-clade 12 contains the plant parasitic ty-
lenchs and was thus not included in this study. H-clades 10 and 11 contain many
bacterial feeders, of which we selected Panagrolaimus sp. PS1159. Finally, from
the basal chromadorea, we obtained Plectus sp. sjh2, a member of H-clade 6.
C. elegans N2, CB4856 and JU775 were provided by the Caenorhabditis
Genetics Center, which is funded by NIH Office of Research Infrastructure Pro-
grams (P40 OD010440). C. elegans QX1211 was kindly provided by Erik An-
dersen (Northwestern Univ.). Plectus sp. sjh2 was isolated from a soil sample
using morphological criteria by Casper Quist and Hans Helder (Wageningen
Univ.). SJH then isolated a single species by starting cultures with a single
worm. The remaining strains were used in previous studies by Avery [S6].
Cultivation of Worms
Worms were grown on NGM-SR plates (3 g NaCl, 24 g agar, 2.5 g peptone,
1 mL 5 mg mL−1 cholesterol in EtOH in 975 mL water, with 1 mL 1 m CaCl2,
1 mL 1 m MgSO4, 25 mL 1 m K2PO4 pH 6, 1 mL 200 mg mL−1 streptomycin in
water, and 0.23 g 5 mL 40 mg mL−1 nystatin in DMSO, added after autoclaving)
seeded with E. coli HB101, as previously described [S7]. E. coli HB101 was first
cultured in M9 minimal media (3 g KH2PO4, 6 g Na2HPO4, 5 g NaCl, 1 mL
1 m MgSO4 in 1 L water) supplemented with 10% Luria broth and 10 mg mL−1
streptomycin [S8]. Plates were incubated with a light circle of HB101 culture
for a day at 37 ◦C and then stored at 4 ◦C. For Plectus sp. sjh2, low salt plates
(2% agar supplemented with 5 mg L−1 of cholesterol from a 5 mg mL−1 EtOH
solution) were used as previously described [S9]. On NGM-SR plates, these
worms became shriveled and died. As the plates did not have nutrients for the
bacteria to grow, HB101 was grown to high density in Luria broth overnight at
37 ◦C, washed 3X in water, resuspended at 10X concentration, and applied to
the plates.
25
Nematodes were cultured by either transferring a few worms by worm pick or
a chunk of agar to a new plate after the worms reached adulthood. The plates
were then incubated at 20 ◦C. The growth rate varied considerably among
strains, with Plectus sp. sjh2 taking nearly two weeks to reach adulthood. We
avoided starving the worms at any point during their cultivation, especially
in the period before behavioral experiments were performed, as this can induce
transgenerational phenotypic changes [S10,S11], and we have observed transient
effects on motility lasting at least a couple of generations (data not shown).
Imaging
The imaging experiments were done on 3.5 cm plates containing the same me-
dia used for cultivation. A 2×2 10 mm repellant grid was made by etching the
plate with a tool dipped in 1% sodium dodecyl sufate, a detergent that C. el-
egans and most other nematodes avoided. (Whereas many C. elegans studies
have used copper rings as a repellant boundary [S12], we found that it did not
sufficiently repel other nematodes; data not shown). Four young adult, well-
fed nematodes were transferred individually by worm pick to a 10 µL drop of
M9 (water for Plectus sp. sjh2) to remove bacteria stuck to the worms. The
worms were then transferred by pipette in a minimal amount of buffer to the
imaging plate, and excess buffer was removed as much as possible. The plate
was imaged 10-20 minutes after picking the worms, minimizing most transient
behaviors. The plate was placed on a custom imaging rig in an inverted, uncov-
ered configuration with illumination by a Schott MEBL-CR50 red LED plate.
The behavior was recorded for 30 minutes using a Point Grey Grasshopper Ex-
press GX-FW-60S6M-C camera equipped with an Edmund Optics NT54-691
lens (set to a magnification of 0.5X) at a resolution of 2736x2192 (12.5 µm/px)
at 11.5 frames/s using a custom National Instruments LabView acquisition pro-
gram. The video was subsequently compressed using the open-source XVid
MPEG-4 compression algorithm using maximal quality settings.
Tracking and Image Analysis
The behavioral videos were analyzed using a custom automated analysis pro-
gram in MathWorks Matlab. The average background was calculated from 50
frames evenly sampled across the entire video. The background was then sub-
tracted from each frame and a global threshold was applied. The thresholded
image was cleaned by applying a series of morphological operations: Incomplete
thresholding of the worm was smoothed by applying morphological closing with
a disk with a similar radius as the worm. Any remaining holes were filled in using
a hole-filling algorithm. Small holes or ones with a low perimeter to area ratio
were excluded as they sometimes fill in worms undergoing an omega turn, as
described in [S13]. Finally, regions in which the worm was just barely touching
itself were split by sequentially applying open, diagonal fill, and majority mor-
phological operations. The worm was then identified as the largest connected
component with an area within 2-fold of the expected value. The centroid was
26
tracked across frames to obtain (cid:126)x(t). In addition, the image skeleton was cal-
culated. Sample images from each of the processing steps are shown in Figure
S11.
The head of the worm was automatically identified using two statistical
properties of the worm's behavior, namely (i) on average, the head of the worm
moves more than the tail, and (ii) on average, worms spend more time moving
forward (in the direction of their head) than they do moving in reverse. The
procedure is based on skeletonization and centroid detection of the worm image,
which can fail in situations where image contrast is low (e.g. due to non-uniform
background), so trajectories were first divided into segments that contain no
more than 3 frames missing the skeleton and centroid information, and the
head orientation was assigned within each segment based on local behavioral
statistics. Finding statistical criteria that allow unambiguous assignment of
head orientation across all strains studied here was challenging because of the
diversity in their behavior, but the following procedure was found to work well
empirically. The identity of the two ends of the skeleton across image frames
were accounted for by a simple tracking algorithm based on minimizing the total
distance between skeleton points. For segments longer than 150 frames (with
no more than ten consecutive missing skeletons), we found that we could apply
property (i) by computing the variance in body angles within 10% of the body
length from the ends, and assigning the head to the end with the greater summed
variance. However, manual inspection revealed that this sporadicly resulted in
misassignment of the head, identifiable as long reversals interrupted by short
forward runs. Therefore, in addition, for segments longer than 200 frames (with
no more than five consecutive missing centroids), we used property (ii), defining
the head as the end of the skeleton that spent the majority of the trajectory at
the leading edge of movement. Segments shorter than 150 frames were discarded
from further analysis.
The velocity (cid:126)v(t) was calculated from the centroid position (cid:126)x(t) using the
derivative of a cubic polynomial fit to a sliding 1 s window. The direct esti-
mation of the velocity using a symmetrized derivative had a large δ-correlated
component that interfered with later analysis. The use of the cubic polyno-
mial did not noticeably distort the correlation functions (Figure S12). When
the worm's speed s(t) = (cid:126)v(t) is very low, its projections on the lab-frame x-
and y-axes vx = (cid:126)v(t) · x and vy = (cid:126)v(t) · y become dominated by discretization
(pixelation) noise, and the bearing φ(t) = tan−1(vy/vx) is poorly defined. This
in turn leads to large fluctuations in ∆ψ(t) = φ(t) − ψ(t), which can introduce
a large number of false reversal events, noticeable as a steep decrease in the au-
tocorrelation C∆ψ(τ ) = (cid:104)cos(∆ψ(t + τ )− ∆ψ(t))(cid:105) at small values of the delay τ .
We therefore exclude segments of the trajectories corresponding to run intervals
shorter than six frames (less than half a second). When these artifacts are fil-
tered out in this manner, the ∆ψ autocorrelation functions were well described
by single exponentials (Figure S7C). We note that the exclusion of short runs
effectively excludes segments of data in which the worm remains stopped (or at
a very low speed) -- a feature that is more pronounced in some strains than
others -- and this leads to a systematic bias for simulated model trajectories
27
to have a higher effective diffusivity Deff than the data for the corresponding
strain (as can be seen in Figure 5C).
Calculation of Behavioral Statistics
The worm's behavior fluctuated or sometimes drifted over long times (Figure
S4), but the average statistics over 100 s windows were approximately stationary.
In order to focus on dynamics within the 100 s timescale, the mean-squared dis-
placement and all auto- and cross-correlation functions were calculated for 100 s
windows and then averaged. This reduced the influence of longer timescale
fluctuations in the speed and reversal rate. For all calculations, observations
near the boundaries and pairs of points between which the worm approached
the boundary were excluded. The uncertainty of each individual's phenotype
projection on the principal behavioral mode was computed by projecting the
motility parameters after bootstrapping over the 100s windows of each individ-
ual's trajectory. The standard deviation of the bootstrapped projections is used
as uncertainty.
Calculation of Effective Diffusivity, Deff
To estimate the effective diffusivity Deff, we fit the mean-squared displacement
(cid:104)[∆x(τ )]2(cid:105) over the diffusive regime. For this purpose, we defined the diffu-
sive regime as the time-lag interval after which the normalized velocity auto-
correlation C(cid:126)v(τ )/C(cid:126)v(0) decayed to below 0.1. We note that in some cases
(especially for fast-moving strains such as CB4856, JU775 and sjh2) the fit to
(cid:104)[∆x(τ )]2(cid:105) = 4Deffτ in this regime was poor due to boundary effects arising
from the finite size of the behavioral arena. For these strains, Deff should be
regarded as a lower bound for the true diffusivity.
Reversal Analysis
The reversal state was assigned as described in the main text by analysis of
∆ψ(t). Assuming a random telegraph process that generates states ∆ψ = 0
(forward) and ∆ψ = π (reverse) with probabilities 1−frev and frev, respectively,
the autocorrelation at long time lags is C∆ψ(τ → ∞) = (1 − 2frev)2. For the
proposed telegraph process, each state has an exponentially distributed lifetime
. The expected correlation timescale
(τfwd, τrev) and therefore frev =
for the mixture of the two states is τRT (τrev, τfwd) =(cid:0)τ
(cid:1)−1
τrev + τfwd
. The ∆ψ
τrev
rev
−1
fwd + τ−1
(cid:21)
autocorrelation function was therefore fit to
C∆ψ(τ ) = [1 − C∆Ψ∞(τrev, τfwd)] exp
τfwd − τrev
τfwd + τrev
where C∆Ψ∞(τrev, τfwd) = (
is: frev = 0.5 −(cid:112)
−
(S8)
)2. The fraction of time spent reversing
C∆Ψ∞(τrev, τfwd)/4, where frev ∈ [0, 0.5]. The transition time
τ
τRT (τrev, τfwd)
+ C∆Ψ∞(τrev, τfwd)
(cid:20)
28
constants are then τrev =
τRT (τrev, τfwd)
1 − frev
and τfwd =
τRT (τrev, τfwd)
.
frev
To validate our approach, we compared the parameters obtained with our
fitting procedure with those obtained from the distribution of time intervals
between detected switching events (Figure S7). For both forward and reverse
states, the distribution of time intervals between detected switching events (Fig-
ure S7B) were well-fit by a biexponential distribution P (Trun > t) = C∆Ψ∞
exp(−t/τshort) + (1 − C∆Ψ∞) exp(−t/τlong) with the time constants τshort and
τlong typically separated by > 10-fold, and the fraction of short intervals C∆Ψ∞
varying broadly over its full range, 0 ≤ C∆Ψ∞ ≤ 1.0 (Figure S7D,E). Values for
τshort were typically below 1 s (Figure S7D). While some fraction of these short
intervals might represent true runs, they could also arise from spurious detec-
tion of switches in velocity bearing due to noise in estimating the centroid (see
legend of Figure S7D) and in any event, contribute little to the overall dynamics
of bearing decorrelation.
Values for τfwd and τrev obtained by fitting equation S1 to the measured au-
tocorrelation functions correlated well with τlong (Figure S7E), thus confirming
that τlong contributes to bearing randomization. We conclude that the for-
ward/reverse switching dynamics are well described by equations (6) and (7),
with parameters τfwd, and τrev.
Speed Analysis
Transitions between forward and reverse runs tended to be excluded from the
analysis because the speed crosses zero, rendering φ a noisy variable generating
many short runs below our exclusion threshold of 6 frames (see above). The
speed set point µs was fit by taking the mean. The remaining parameters
of the speed dynamics (3) were fit by its analytical autocorrelation function:
Cs(τ ) = Dsτs exp (−τ /τs).
Orientation Analysis
Changes in orientation during runs (i.e. intervals between reversal events) were
analyzed with respect to their mean-squared angular displacements (MSAD)
over time, corresponding to a model for angular diffusion with drift.For an object
lying on a two-dimensional plane, rotational diffusion about an axis normal to
the plane leads to fluctuations in the orientation (an angle measured in the lab
frame) ψ(t) over time according to:
dψ(t) =
2DψdWt,
(S9)
where Dψ is the rotational diffusion coefficient, and dWt represents increments
of a Wiener process. Bias in these fluctuations over time can be captured, to
first order, by adding a linear drift term so that
(cid:112)
(cid:112)
dψ(t) = kψdt +
29
2DψdWt,
(S10)
If kψ and Dψ are constant in time, the mean-squared angular displacement
with kψ the drift coefficient.
MSAD(τ ) = (cid:104)[ψ(t + τ ) − ψ(t)]2(cid:105), is a quadratic function of the time delay τ :
(cid:104)[ψ(t + τ ) − ψ(t)]2(cid:105) = (cid:104)[kψτ +
(S11)
where (cid:104)·(cid:105) denotes averaging over all time pairs separated by τ and the last
equality follows from the Wiener process properties (cid:104)Wt+τ − Wt(cid:105) = 0 and
(cid:104)[Wt+τ − Wt]2(cid:105) = τ .
2Dψ(Wt+τ − Wt)]2(cid:105) = k2
ψτ 2 + 2Dψτ,
(cid:112)
More generally, if kψ(t) and Dψ(t) are time-varying quantities, we can still
approximate within a finite time window (centered about time tw) the "local"
values kψ,w ≈ kψ(tw) and Dψ,w ≈ Dψ(tw). In this study, we extract estimates
of these (possibly time varying) parameters from fits to the averaged MSAD
computed over time windows:
w=W(cid:88)
w=1
−1
W
(cid:104)[ψ(t + τ ) − ψ(t)]2(cid:105) = (cid:104)k2
ψ(cid:105)wτ 2 + 2(cid:104)Dψ(cid:105)wτ,
where W is the number of windows and (cid:104)x(cid:105)w = W −1(cid:80)w=W
w=1 xw represents
√
averages over windows. By fitting this averaged MSAD by a quadratic function
b = (cid:104)k2
aτ + bτ 2, we thus obtain the estimates a/2 = (cid:104)Dψ(cid:105)w and
w . Note
that a/2 obtained by this procedure yields an estimate of the mean value for
Dψ, but
b corresponds to an estimate not of the mean value, but the root-
mean-square (rms) value for kψ. Throughout the text, we therefore explicitly
refer to the latter estimate as kψrms (and refer to the former simply as Dψ).
ψ(cid:105)1/2
(S12)
√
Simulations
Reversals, orientation, and speed dynamics were all simulated independently
using the model described. Forward and reverse run durations were chosen
according to equations (5) and (6) by drawing exponential random numbers with
mean value τfwd or τrev. During reverse runs, ∆ψ was set to π. The orientation
(4) and speed (3) dynamics were simulated using the Euler-Maruyama method
[S14] with a time step that matched the frame rate. To prevent negative speeds,
a reflective boundary condition was imposed by taking the absolute value of
the speed at each simulation step. The velocity was then calculated from the
decomposition in (1) and trapezoidally integrated to give the centroid position
(cid:126)x(t).
Behavioral Mode Analysis
The model parameters were fit to each trajectory to give a phenotypic matrix T.
The phenotypic matrix was centered by subtracting the mean phenotype, T =
T − (cid:104)T(cid:105)indiv.. The correlation matrix was then calculated, CT = corr T, and
decomposed into eigenvalues λ and eigenvectors (behavioral modes) b, CTb =
λb. To reduce any bias coming from a single trajectory, this calculation was
30
bootstrapped 1000 times. The significance of the k-th top mode is assessed
by a comparison with the expected variance explained of the k-th top mode
of randomly chosen directions in the behavioral space. We use the explained
variance of the k-th mode of a newly created set of modes where the first k − 1
modes are equal to the top behavioral modes and the remaining modes are
pointing in randomly chosen orthogonal directions. This process is repeated
1000 times.
The projections of each trajectory on these behavioral modes were calculated
by P = Tb. The uncertainty in the locus of each individual phenotype along
the behavioral mode was computed by projecting the motility parameters after
bootstrapping over the 100 second windows and taking the standard deviation.
Statistics
Unless otherwise indicated, errorbars and confidence intervals represent the 2.5%
and 97.5% percentiles (spanning the 95% confidence interval) estimated from
1000 bootstrap samples. All probability distributions were empirically estimated
using kernel density methods in Python's Seaborn package with a bandwidth
automatically selected using Scott's rule of thumb [S15]. Tabulated mean values
of the effective diffusivity model and the motility model (Table S1-S4) represent
geometric rather than arithmetic means was used as the parameters varied log-
normally.
Derivation of the Velocity Autocorrelation Function Under
the Assumption of Independence
The velocity autocorrelation function can be written in terms of the motility
components,
C(cid:126)v(τ ) = (cid:104)(cid:126)v(0) · (cid:126)v(τ )(cid:105)
= (cid:104)s(0) [cos [ψ(0) + ∆ψ(0)] , sin [ψ(0) + ∆ψ(0)]]×
×s(τ ) [cos [ψ(τ ) + ∆ψ(τ )] , sin [ψ(τ ) + ∆ψ(τ )]](cid:105)
(S13)
The expected value of the product of independent random variables is the
product of the expected value of each variable, i.e. (cid:104)xy(cid:105) = (cid:104)x(cid:105)(cid:104)y(cid:105). Therefore we
can factor out Cs = (cid:104)s(0)s(τ )(cid:105), leaving the vector product with ψ and ∆ψ. The
expanded vector product is:
C(cid:126)v(τ ) = Cs(τ ) × (cid:104)cos [ψ(0) + ∆ψ(0)] cos [ψ(τ ) + ∆ψ(τ )]
+ sin [ψ(0) + ∆ψ(0)] sin [ψ(τ ) + ∆ψ(τ )](cid:105)
(S14)
The trigonometric functions on ψ(t) + ∆ψ(t) can be rewritten as products of
trigonometric functions of the terms:
cos [ψ(t) + ∆ψ(t)] = cos ψ(t) cos ∆ψ(t) − sin ψ(t) sin ∆ψ(t)
sin [ψ(t) + ∆ψ(t)] = sin ψ(t) cos ∆ψ(t) + cos ψ(t) sin ∆ψ(t)
31
However, since ∆ψ(t) = {0, π}, sin ∆ψ(t) = 0:
cos [ψ(t) + ∆ψ(t)] = cos ψ(t) cos ∆ψ(t)
sin [ψ(t) + ∆ψ(t)] = sin ψ(t) cos ∆ψ(t)
Substituting into (S14),
C(cid:126)v(τ ) = Cs(τ ) × (cid:104)cos ψ(0) cos ψ(τ ) cos ∆ψ(0) cos ∆ψ(τ )+
sin ψ(0) sin ψ(τ ) cos ∆ψ(0) cos ∆ψ(τ )(cid:105)
(S15)
We can now factor out Cψ(τ ) = (cid:104)cos [ψ(τ ) − ψ(0)](cid:105) = (cid:104)cos ψ(0) cos ψ(τ ) +
sin ψ(0) sin ψ(τ )(cid:105) to get:
C(cid:126)v(τ ) = Cs(τ )Cψ(τ )(cid:104)cos ∆ψ(0) cos ∆ψ(τ )(cid:105)
Finally, we substitute (again dropping sin ∆ψ(t) terms):
C∆ψ(τ ) = (cid:104)cos [∆ψ(0) − ∆ψ(τ )](cid:105) = (cid:104)cos ∆ψ(0) cos ∆ψ(τ )(cid:105)
to get:
C(cid:126)v,indep(τ ) = Cs(τ )Cψ(τ )C∆ψ(τ )
32
Figure S1: Confinement by the boundary affects the mean-squared displace-
ment (MSD) at long times, but does not impair resolution of the ballistic to
diffusive transition. We compare the statistical behavior of C. elegans N2 in
the experiments presented here within small (1-cm) arenas (black) and a pre-
viously reported dataset that used larger (5-cm) arenas [S1] (red). The MSD
(A), defined as (cid:104)[∆x(τ )]2(cid:105) ≡ (cid:104)(cid:126)x(t + τ ) − (cid:126)x(t)2(cid:105), of our small-arena dataset is
similar to that of the large-arena dataset at short times, but does show mild ef-
fects of confinement at long times ((cid:38) 100 s). The ballistic to diffusive transition
can be more closely studied by examining decay of the velocity autocorrelation
function (VACF), defined as Cv(τ ) ≡ (cid:104)(cid:126)v(t) · (cid:126)v(t + τ )(cid:105) (B), which is related to
0 dτ(cid:48)C(cid:126)v(τ(cid:48)) [S16]. The decay
MSD (i.e.
of the VACF to zero, which indicates orientation randomization and hence the
transition from the ballistic to diffusive regime, is not significantly affected by
the presence of the confining boundary.
[∆x(τ )]2) by (d/dτ )(cid:104)[∆x(τ )]2(cid:105) = 2(cid:82) τ
33
MSD (µm2)1.0 (s)0.1110100AB0.1110100102103104105106107108 (s)−0.20.00.20.40.60.8Normalized VACF1cm arena5cm arena(Stephens et al. 2010)MSD (µm2)¿¿ Figure S2: An overview of the dataset. Trajectories of all worm included in the
study. Each box represents a 10 mm by 10 mm chamber. In blue, we highlight
points excluded from the analysis because they were influenced by the boundary.
Figure S3: The ballistic to diffusive transition for all strains. We show the
average mean-squared displacment (MSD), calculated across individual trajec-
tories, for each strain (black). The expected ballistic (blue) and diffusive MSD
curves (red), as in Figure 1B.
34
N2CB4856JU775QX1211JU757DF5020PS312PS1159sjh2N2CB4856JU775QX1211JU757DF5020PS312PS1159sjh20.11101000.11101000.11101000.11101000.11101000.11101000.11101000.11101000.1110100102104106108MSD (¹m2)¿ (s) BallisticDiffusiveData Figure S4: The worms' behavior was approximately stationary. For each strain,
we show the average speed (top) and fraction of time spent reversing (bottom)
calculated over 100 s sliding windows and averaged across individuals.
Figure S5: Characterization of speed statistics across strains. (A) The speed
autocorrelation (black) of each strain decays exponentially (red). (B) The speed
distribution (black) of each strain is closely reproduced by model simulations
(red).
35
0.00.51.0Time (min)010203001020300102030010203001020300102030010203001020300102030Mean Frac(cid:31)onReversingMean Speed(µm/s)N2CB4856JU775QX1211JU757DF5020PS312PS1159sjh20200300100JU775QX1211JU757DF5020PS1159sjh2Probability Density (×102)τ(s)s(µm/s)AB51005100N202040CB4856PS3125100510051005100510051005100300030003000300030003000300030003000300Cs ×102(µm/s)2 Figure S6: Characterization of orientation statistics across strains. (A) The
mean squared angular displacement of body orientation (black) was fit to a
quadratic function (red) in all strains. (B) The orientation correlation (black)
decays non-exponentially for many strains.
36
QX1211DF5020PS1159sjh2JU775010201.00.50.0JU757τ(s)τ(s)ABPS312CB4856N205010002550050100050100050100050100050100050100050100050100MSAD (rad2)Cψ0255002550025500255002550025500255002550 Figure S7: Characterization of reversal statistics across strains. (A) Distri-
bution of ∆ψ for each strain shows two prominent peaks at 0◦ and 180◦. (B)
Cumulative distributions of the forward and reverse run durations (Tfwd, Trev)
for an individual worm from each strain (black), fit to a biexponential function
(red).
(C) The autocorrelation function of ∆ψ for each strain (black) along
with exponential fit (red). (D, left) The fraction of short runs measured by the
biexponential fits of the transition time distributions (as in B) was inversely
correlated with the average speed of the worm. At low speed, the bearing (and
therefore also ∆ψ, which is used to identify runs) is expected to be dominated
by noise (e.g. pixelation artifacts). (D, right) The fitted time constants for short
forward and reverse intervals were uncorrelated (unlike those for long runs, see
E and also Figure 5A), and typically below the timescale of smoothing filter for
velocity data (1 s), further motivating the exclusion of short intervals in mod-
eling reversal dynamics. (E) τlong, extracted from fits to the transition time
distributions, were correlated with τfwd and τrev, estimated from C∆ψ (panel
C).
37
BAC0100200 Reverse0100200050100% Short RunsForward10-110010110-110010110-110010110210310-1100101102103 10-110010110210310-1100101102103 ED01020010200102001020010200102001020010200102010010-31-P02040S(µm/s)τfwd(s)τrev(s)τfwd,long(s) τrev,long(s)τfwd,short(s)τrev,short(s)τ(s)τfwd(s)τrev(s)0 12∆ψ(o)Prob. Density0180018001800180018001800180018001800.00.51.0N2CB4856JU775QX1211JU757DF5020PS312PS1159sjh210-3100CB4856DF5020JU757JU775N2PS312PS1159QX1211sjh2050050C∆ψ050050050050050050050 Figure S8: Worm run speeds are similar during forward and reverse runs for
individual trajectories (dots, colored by strain). We define run speed as the
top speed during runs (rather than the mean speed, to avoid biases due to run-
length differences). The top speed is computed as the 95th percentile of the
speed distribution (rather than the maximum, to avoid outlier effects).
38
1.61.82.02.22.42.61.61.82.02.22.42.6log10 sfwd(µm/s)log10 srev(µm/s)CB4856DF5020JU757JU775N2PS1159PS312QX1211sjh2 Figure S9:
Cross-correlation analysis of motility dynamics. The cross-
correlation between (top) speed and bearing changes, (middle) speed and veloc-
ity alignment, and (bottom) bearing changes and velocity alignment are shown
for each strain. There is very little cross-correlation among the motility variables
in any of the strains. All cross-correlations were normalized to unit variance by
dividing by the product of the standard deviation (σ) of the two components.
39
N2CB4856JU775QX1211DF5020JU757PS312PS1159sjh2-1001000-1001000-1001000-1001000-1001000-1001000-1001000-1001000-100100010-110-110-1τ(s)Cs,∆ψCs,ψCψ,∆ψ Figure S10: The top behavioral mode effectively captures changes in diffu-
sivity compared to random projections. (A) The effect of variation on the top
behavioral mode (black, as in Figure 5C) compared with a sampling of 100
random modes (red) on the diffusivity of simulated trajectories. For random
modes, the sign of the mode was chosen such that the diffusivity increased with
the projection along the mode. (B) For each random mode we compute the
relative change in diffusivity between mode values ∆Deff = Deff(2)/Deff(−2)
and compare to the same relative diffusivity computed from the top behavioral
mode. The kernel density distribution of the observed change is shown for the
100 samples (ticks). The black line indicates a ratio of 1 (no difference) and
most random projections exhibit less range in ∆Deff.
Figure S11: An Overview of the image processing steps. The video frames
were processed by (1) subtracting the average of 50 frames evenly sampled from
the entire movie and (2) cropping to each of the SDS-enclosed regions. (3) The
largest worm-sized object was identified following several image morphology
operations, and (4) the centroid and image skeleton were measured.
40
Mode 1Random Modes−4−2024103104105106−4−20240.00.20.40.6ProbabilityDeffµm2/slog10∆Deff,random/∆Deff,mode1ABMode projection Figure S12: Comparison of velocity calculation methods. Velocity autocorrela-
tion functions for the three example strains with and without filtering of the data
and without averaging over 100 s windows. The unfiltered velocity (black), esti-
mated using a symmetrized derivative, contained a δ-correlated short-timescale
component in all strains that was particularly prominent in slow-moving strains
such as PS312. The velocity calculated using a 1 s cubic polynomial filter (red)
does not contain this δ-correlated component.
41
050100−5051015N2050100−100102030CB4856 (s)0501000123PS312Symmetrized Deriv.1s Cub. Poly. Deriv.-Corr. ComponentvC (×103µm2/s2)τδ Deff × 102
(µm2/s)
Strain
N2
CB4856
JU775
QX1211
JU757
DF5020
PS312
PS1159
sjh2
Mean
140
429
448
36
210
128
8
81
425
2.5% 97.5%
185
620
558
98
327
255
13
183
553
105
307
360
12
123
55
5
32
314
Table S1: The geometric mean of the effective diffusivity for each strain. For
each trajectory, an effective diffusivity (Deff) was extracted by analysis of mean-
squared displacements and the velocity autocorrelation function.
Strains Mean
N2
77
108
CB4856
112
JU775
40
QX1211
JU757
97
65
DF5020
27
PS312
PS1159
38
159
sjh2
µs
(µm/s)
2.5% 97.5% Mean
1.9
1.8
2.1
0.7
4.2
1.1
0.7
3.1
3.3
85
128
127
66
120
83
32
53
184
68
91
98
26
72
50
23
26
138
τs
(s)
Ds × 102
((µm/s)2/s)
2.5% 97.5% Mean
5.8
4.0
4.3
7.5
3.8
11.7
5.3
0.6
8.8
2.4
2.3
2.7
1.1
5.2
1.3
0.8
6.2
4.7
1.5
1.5
1.6
0.4
3.3
0.8
0.6
1.4
2.5
2.5% 97.5%
7.9
5.2
5.5
11.6
5.0
14.4
7.4
1.1
13.2
4.3
3.1
3.3
4.6
2.8
9.1
3.7
0.3
5.8
Table S2: The model parameters related to the speed dynamics are listed for
each strain. For each worm in a strain, time-averaged parameters were calcu-
lated.
42
kψrms
(rad/s)
(cid:0)rad2/s(cid:1)
Dψ
Strains Mean
0.036
N2
CB4856
0.029
0.038
JU775
0.040
QX1211
0.039
JU757
DF5020
0.037
0.017
PS312
0.023
PS1159
sjh2
0.066
2.5% 97.5% Mean
0.034
0.026
0.018
0.024
0.021
0.026
0.017
0.028
0.036
0.030
0.032
0.033
0.021
0.011
0.009
0.015
0.057
0.090
0.048
0.041
0.053
0.056
0.052
0.042
0.029
0.031
0.077
2.5% 97.5%
0.054
0.017
0.018
0.033
0.030
0.016
0.036
0.009
0.054
0.021
0.026
0.041
0.028
0.014
0.017
0.005
0.065
0.127
Table S3: The model parameters related to the orientation dynamics are listed
for each strain. For each worm in a strain, time-averaged parameters were
calculated.
τfwd
(s)
Strains Mean
N2
23.8
78.6
CB4856
85.3
JU775
26.5
QX1211
32.3
JU757
DF5020
32.7
PS312
5.6
80.8
PS1159
sjh2
155.5
2.5% 97.5% Mean
13.9
4.1
4.3
56.5
8.0
51.5
3.7
12.5
6.4
20.5
15.2
4.8
3.3
4.2
8.8
38.0
75.2
4.9
41.1
109.5
144.0
63.0
50.5
63.2
7.4
174.6
419.9
τrev
(s)
2.5% 97.5%
5.7
6.3
16.3
5.2
10.3
7.2
3.9
16.2
10.8
3.0
2.8
4.2
2.7
4.1
3.2
2.9
5.2
2.3
Table S4: The model parameters related to the reversal state dynamics are
listed for each strain. For each worm in a strain, time-averaged parameters
were calculated.
Parameter Mean
0.50
log10 µs
log10 τs
0.51
-0.19
log10 Ds
0.24
log10 kψ
0.15
log10 Dψ
log10 τf wd
0.50
0.35
log10 τrev
Loading
2.5% 97.5%
0.58
0.25
0.40
0.54
0.08
-0.43
0.43
-0.04
0.40
-0.22
0.36
0.55
0.48
0.15
Table S5: The loadings of each parameter on the top behavioral mode are listed.
43
Strain
N2
CB4856
JU775
QX1211
JU757
DF5020
PS312
PS1159
sjh2
Projection
Mean
-0.26
0.42
0.86
-1.45
0.78
-0.65
-2.42
0.45
1.61
2.5% 97.5%
0.49
-0.48
0.76
-0.35
1.04
-0.06
-1.82
0.32
1.23
-0.07
1.19
-0.70
-0.13
-2.66
-2.64
1.01
2.41
0.80
Table S6: The phenotypic projection along the first behavioral mode is listed
for each strain. For each worm in a strain, a time-averaged projection was
calculated.
44
SI references
[S1] Stephens, G. J, Johnson-Kerner, B, Bialek, W, & Ryu, W. S. (2010) From
modes to movement in the behavior of Caenorhabditis elegans. PLoS One
5, e13914.
[S2] Blaxter, M. L, De Ley, P, Garey, J. R, Liu, L. X, Scheldeman, P, Vier-
straete, a, Vanfleteren, J. R, Mackey, L. Y, Dorris, M, Frisse, L. M, Vida,
J. T, & Thomas, W. K. (1998) A molecular evolutionary framework for
the phylum Nematoda. Nature 392, 71 -- 5.
[S3] Holovachov, O, van Megen, H, Bongers, T, Bakker, J, Helder, J, van den
Elsen, S, Holterman, M, Karssen, G, & Mooyman, P. (2009) A phylo-
genetic tree of nematodes based on about 1200 full-length small subunit
ribosomal DNA sequences. Nematology 11, 927 -- 950.
[S4] Kiontke, K & Fitch, D.
(2005) The phylogenetic relationships of
Caenorhabditis and other rhabditids. WormBook.
[S5] Andersen, E. C, Gerke, J. P, Shapiro, J. A, Crissman, J. R, Ghosh, R,
Bloom, J. S, F´elix, M.-A, & Kruglyak, L. (2012) Chromosome-scale selec-
tive sweeps shape Caenorhabditis elegans genomic diversity. Nat. Genet.
44, 285 -- 90.
[S6] Chiang, J.-T. A, Steciuk, M, Shtonda, B, & Avery, L. (2006) Evolution of
pharyngeal behaviors and neuronal functions in free-living soil nematodes.
J. Exp. Biol. 209, 1859 -- 73.
[S7] Davis, M. W, Somerville, D, Lee, R. Y, Lockery, S, Avery, L, & Fam-
brough, D. M.
(1995) Mutations in the Caenorhabditis elegans Na,K-
ATPase alpha-subunit gene, eat-6, disrupt excitable cell function. J. Neu-
rosci. 15, 8408 -- 18.
[S8] Stiernagle, T. (2006) Maintenance of C. elegans. WormBook.
[S9] Lahl, V, Halama, C, & Schierenberg, E. (2003) Comparative and exper-
imental embryogenesis of Plectidae (Nematoda). Dev. Genes Evol. 213,
18 -- 27.
[S10] Hall, S. E, Beverly, M, Russ, C, Nusbaum, C, & Sengupta, P. (2010) A
cellular memory of developmental history generates phenotypic diversity
in C. elegans. Curr. Biol. 20, 149 -- 55.
[S11] Rechavi, O, Houri-Ze'evi, L, Anava, S, Goh, W. S. S, Kerk, S. Y, Han-
(2014) Starvation-Induced Transgenerational
non, G. J, & Hobert, O.
Inheritance of Small RNAs in C. elegans. Cell 158, 277 -- 87.
[S12] Hart, A. (2006) Behavior. WormBook.
45
[S13] Huang, K.-M, Cosman, P, & Schafer, W. R. (2006) Machine vision based
detection of omega bends and reversals in C. elegans. J. Neurosci. Methods
158, 323 -- 36.
[S14] Kloeden, P. E & Platen, E. (1992) Numerical Solution of Stochastic Dif-
ferential Equations. (Springer Berlin Heidelberg, Berlin, Heidelberg).
[S15] Scott, D. W. (1992) Multivariate density estimation: theory, practice, and
visualization. (John Wiley & Sons).
[S16] Chaikin, P. M., Lubensky, T. C. & Witten, T. A. (1995) Principles of
condensed matter physics.. Cambridge University Press).
46
|
1804.02437 | 1 | 1804 | 2018-03-29T13:41:48 | The impact of electrical couplings on the sequential bursting activity in the ensemble of inhibitory coupled Van der Pol elements | [
"q-bio.NC",
"nlin.CD"
] | The new phenomenological model of the ensemble of three neurons with chemical (synaptic) and electrical couplings has been studied. One neuron is modeled by a single Van der Pol oscillator. The influence of the electrical coupling strength and the frequency mismatch between the elements to the regime of sequential activity is investigated. | q-bio.NC | q-bio |
The impact of electrical couplings on the sequential bursting activity in
the ensemble of inhibitory coupled Van der Pol elements
T. A. Levanova ∗1, A. O. Kazakov1,2, A. G. Korotkov1, and G. V. Osipov1
1Lobachevsky State University of Nizhny Novgorod, Institute of Information Technologies,
Mathematics and Mechanics, 23, Prospekt Gagarina, Nizhny Novgorod, 603950, Russia
2National Research University Higher School of Economics, 25/12 Bolshaya Pecherskaya Ulitsa,
Nizhny Novgorod, 603155, Russia
August 8, 2018
Abstract
The new phenomenological model of the ensemble of three neurons with chemical (synaptic) and
electrical couplings has been studied. One neuron is modeled by a single Van der Pol oscillator. The
influence of the electrical coupling strength and the frequency mismatch between the elements to the
regime of sequential activity is investigated.
1 Introduction
In the last few years, a new field of medicine that is called bioelectronic medicine [1] is actively devel-
oping. The main feature of bioelectronic medicine is the application of electrical stimulus to nervous
system and body's tissues instead of chemical (pharmaceutical) treatment. The main target of electri-
cal impact is nerve fibers, and signals are delivered to them using implants or wearable devices. The
reasons for such interest to bioelectronic medicine are related both to the rapid improvement of technol-
ogy (among the factors we can note the emergence of biocompatible soft electronics, the rapid growth
in computing performance, the small size of the devices [2]), and a limited success of pharmacology in
the treatment of neurological disorders by medication. In this regard, it is also worth noting that in
the coming decades the problem of treating diseases of the nervous system will become more relevant,
because of the tendency of aging of the population and the growing stress in the modern world. For this
reason, a number of scientific and industrial companies have paid attention to bioelectronic medicine.
As recent papers show, [1, 3], this approach can be successfully applied not only to treatment of diseases
of the nervous system, but also in treatment of cardiovascular, inflammatory, metabolic and endocrine
diseases, as evidenced by animal tests and clinical trials. The nervous system is the main regulator of
internal processes in the body. It system affects the processes of thinking, digestion, motor activity,
etc.
[4]. In this connection, there is an increasing interest in the study of electrical couplings in the
nervous system and their role in the generation of various regimes of neuronal activity, as well as the
mechanisms of their formation and suppression. The development of new medical technologies and
their implementation in practical treatment requires much deeper understanding how the peripheral
nervous system is involved in the regulation of various processes in the body.
The main goal of this work is to study the influence of electrical couplings to regimes of sequential
bursting activity in models of neural ensembles with chemical (synaptic) couplings. For this purpose,
there is considered a phenomenological model of the minimal ensemble of three non-identical neurons
which demonstrate the described types of couplings. Each of the neurons is modeled by corresponding
Van der Pol oscillator, but these oscillators have different proper frequencies. In our previous paper [5]
∗[email protected]
1
Figure 1: The topology of the chemical (synaptic) couplings g1 and g2 and electrical couplings d in the
ensemble of neuron-like elements given by the system (1).
the ensemble of identical Van der Pol oscillators only with chemical couplings was studied in details.
In particular, there were studied various dynamical regimes occuring in this ensemble when varying
the strengths of chemical couplings, and scenarios of appearance and disappearance of these regimes
were also investigated. In the papers [6], [7] it was shown that the obtained types of activity and the
mathematical images underlying them, as well as the scenarios of transitions from one type of activity
to another are universal for a wide class of systems. In the present work we investigate the influence
both of electrical couplings and the non-identity of the elements on the dynamics of the neuronal
ensemble, specially focusing on the evolution of sequential bursting activity, since this regime of neural
activity is very important from the point of view of neurodynamics [8]-[10]. We emphasise that results
which are presented in the paper are quite similar, by principal qualitative properties, ro results of real
biological experiments [11].
2 The model
The ensemble of three non-identical neuron-like elements connected to each other by mutual chemical
(synaptic) inhibitory and electrical couplings is modeled by the following system of three Van der Pol
oscillators
(cid:26) xj − µ[λ(xj, xj) − x2
j] xj + ω2
j xj + d(xj+1 − 2xj + xj−1) = 0,
(1)
j = 1, 2, 3.
where the variable xj phenomenologically describe the value of the membrane potential of the j-th
neuron-like element. The electrical couplings between the ensemble elements are given by the expres-
sions d(xj+1 − 2xj + xj−1), where the parameter d is the coefficient of the electrical coupling. The
chemical (synaptic) inhibitory interaction between neuron-like elements in the ensemble is phenomeno-
logically described in the same way as in the paper [5], where:
(cid:16)q
(cid:17) − g2F
(cid:16)q
(cid:17)
λ(xj, xj) = 1 − g1F
j+1 + x2
x2
j+1
j−1 + x2
x2
j−1
,
(2)
Here g1 and g2 are the strengths of the inhibitory couplings directed clockwise and counter-clockwise,
respectively, see Fig. 1. The function F(z) is an activation function with a threshold value z0 that
2
Figure 2: Maps of the largest Lyapunov exponent of the system (1). a - (g1, g2) = (0, 5); b -- (g1, g2) =
(5, 0). Red points denote Λ1 > 0. Green points denote Λ1 = Λ2 = 0. Blue points denote Λ1 = 0. Gray
points denote Λ1 < −0.0005. The upper triangular region on the charts corresponds to the situation when
the trajectories of the system go to infinity.
phenomenologically describes the main principle of the synaptic coupling:
F(z) =
1
1 + exp(−k(z − z0)) .
(3)
With the values k = 100 and z0 = 0.5 of the parameters chosen for modeling, the nonlinear function
F(z) is close to the step function, but remains smooth. When z reaches the threshold value z = z0, what
corresponds to the generation of amplitude oscillations above a certain threshold one by a presynaptic
element, the function F(z) grows jumpwise from 0 to 1 and remains equal to 1 with a further increase
of z. This leads, in turn, to the fact that, in the case of a sufficient coupling strength, the presynaptic
neuron-like element can suppress the activity of the postsynaptic neuron-like element by generating
oscillations of large amplitude. It is known that in real experiments the frequencies for different neurons
and clusters of neurons are different. This allows us to introduce the parameter ∆ into system (1).
Here ω2 = ω1 − ∆, ω3 = ω1 + ∆. The parameter µ << 1 determines the dynamics of a single element,
in which quasi-harmonic oscillations are observed in the absence of couplings [12]. In [5] it was shown
that the regime of sequential bursting activity is observed in system (1) at d = 0 and ∆ = 0 in the
case of strong asymmetry of chemical couplings. To study how the adding of electrical couplings and
the frequency mismatch between the elements affects the evolution of the sequential activity regime,
we built charts of the largest Lyapunov exponent on the parameter plane (d, ∆) (see Fig. 2). In the
red color on the charts the regions corresponding to the positive largest Lyapunov exponent Λ1 > 0
are marked, which indicates the presence of the chaotic dynamics in the system. Such dynamics occurs
when the stable heteroclinic circuit (observed in the system at d = 0) is destroyed by adding electrical
couplings with d 6= 0. The absolute value of Λ1 in this case, as a rule, is not very large. Green color
on the charts denotes regions corresponding to Λ1 = Λ2 = 0. In this case, a two-dimensional torus is
observed in the phase space of the system. The regions corresponding to periodic orbits are marked
with a blue and gray color. Moreover, the blue color corresponds to the regions in which the largest
Lyapnov exponent fluctuates near zero, with minor deviations, of order of numerical error, the positive
3
DD values (0 < Λ1 < 0.0005), and the gray color corresponds to the regions in which −0.0005 < Λ1 < 0 1.
As one can see from Fig. 2, in both cases there is a threshold relation between d and ∆, above which
the trajectories of the system begin to go to infinity (the upper triangular region on both charts). This
fact agrees well with the data of biological experiments, which show that in real biological systems it
is impossible to increase the strength of the couplings without a limitation. Within the framework of
this constraint, periodic as well as quasi-periodic and chaotic regimes are observed in the system for
various ratios of d and ∆. These regimes were not observed in the absence of an electrical couplings
and the frequency mismatch.
3 The evolution of the sequential bursting activity
Recall that the regime of sequential bursting activity is observed in system (1) for the case of strong
asymmetry of the couplings [5]. For example, the value of the coupling parameter g1 can be taken
significantly larger than g2, which is small or equal to zero. The main feature of this regime is the
exponentially increasing in time the length of the burst. The mathematical image of the sequential
activity regime in the phase space of the system (1) is a stable heteroclinic circuit containing the saddle
limit cycles. The phase orbit, which asymptotically approaching this heteroclinic circuit, spends more
and more time in a vicinity of the saddle limit cycles that corresponds to the growth of the burst length.
Let us study analytically how the described heteroclinic circuit evolves in the presence of the
relatively small electrical couplings and the frequency mismatch between elements. We set ω1 = 1,
d = µd1, ∆ = µ∆1. We rewrite the system of differential equations (1) in the following form
1] x1 − µd1(x2 − 2x1 + x3),
x2 + (1 + µ∆1)x2 = µ[λ(x2, x2) − x2
x3 + (1 − µ∆1)x3 = µ[λ(x3, x3) − x2
x1 + x1 = µ[λ(x1, x1) − x2
z1 = [λ(z1, z1) − z1 ¯z1]z1 + id1(z2 − 2z1 + z3),
z2 = [λ(z2, z2) − z2 ¯z2]z2 + id1(z1 − 2z2 + z3) + i∆1z2,
z3 = [λ(z3, z3) − z3 ¯z3]z3 + id1(z1 − 2z3 + z2) − i∆1z3.
2] x2 − µd1(x1 − 2x2 + x3),
3] x3 − µd1(x1 − 2x3 + x2).
Using the Van der Pol method [13] and averaging the system by the period T = 2π we obtain the
following equations for complex averaged amplitudes z1, z2 and z3:
(4)
(5)
(6)
(7)
Now introduce real amplitudes R1, R2, R3 and phases φ1, φ2 and φ3 in the following way
2 e−iφ1 ,
2 e−iφ2 ,
2 e−iφ3 .
After substituting (6) in (5) we obtain the following system:
z2 = R2
z3 = R3
z1 = R1
2
1
R1) − R2
R2) − R2
R3) − R2
4 ]R1 − R2d1 sin(φ1 − φ2) − R3d1 sin(φ1 − φ3),
R1 = [λ(R1,
4 ]R2 − R1d1 sin(φ2 − φ1) − R3d1 sin(φ2 − φ3),
R2 = [λ(R2,
4 ]R3 − R1d1 sin(φ3 − φ1) − R2d1 sin(φ3 − φ2),
R3 = [λ(R3,
R1 φ1 = 2d1R1 − R2d1 cos(φ1 − φ2) − R3d1 cos(φ1 − φ3),
R2 φ2 = 2d1R2 − ∆1R2 − R1d1 cos(φ2 − φ1) − R3d1 cos(φ2 − φ3),
R3 φ3 = 2d1R3 + ∆1R3 − R2d1 cos(φ3 − φ2) − R1d1 cos(φ3 − φ1).
3
that describes the dynamics of averaged amplitudes R1, R2, R3 and phases φ1, φ2 and φ3. In the case
of absence of the electrical couplings, i.e. when d = 0, system (7) splits into two subsystems. The first
subsystem contains equations for averaged amplitudes:
R1 = [λ(R1,
R2 = [λ(R2,
R3 = [λ(R3,
1
R1) − R2
R2) − R2
R3) − R2
4 ]R1,
4 ]R2,
4 ]R3.
2
3
(8)
1This means that, for getting adequate values of Lyapunov exponents, it is required in this case the increasing of the level
of numerical accuracy. However, in both cases, as shown by numerical experiments, such regimes correspond to stable limit
cycles.
4
The second subsystem contains equations for phases:
φ1 = 0,
φ2 = −∆1,
φ3 = +∆1.
(9)
The analytical study of subsystem (8) was carried out before in the paper [5]. System (8) was considered
sequentially in the restrictions to the invariant manifolds R1 = 0, R2 = 0 and R3 = 0. In particular, for
asymmetric couplings it was shown that there are three equilibria on each of the invariant planes [the
unstable node (0, 0), the saddle (2, 0) and the stable node (0, 2)]. Here the equilibria (2, 0) and (0, 2)
are joined by a stable heteroclinic trajectory that goes from the saddle to stable node. Consequently,
three heteroclinic trajectories (corresponding to phase planes (R1, R2), (R1, R3) and (R2, R3) and three
saddles with 2-dimension stable and 1-dimension unstable manifolds form a stable heteroclinic cycle
of system (8). Using equations (8)-(9) it is easy to obtain that this result remains valid also for non-
identical elements (with non-zero frequency mismatch ∆1 6= 0), however, the frequencies of the elements
are different. It should be noted that heteroclinic circuits, as a rule, arise in systems with symmetry,
and when the symmetry breaks down, heteroclinic circuits also disappear. In our case the non-identity
of the elements breaks the symmetry, but the heteroclinic circuit still exists because the amplitude
dynamics described by subsystem (8) does not depend on the phases determined by equations (9).
Now let us study how the adding electrical coupling changes the described above stable heteroclinic
circuit. In order to do this we will study system (7) on the plane R2 = 0:
4 ]R1 − R3d1 sin(φ1 − φ3),
R1 = [λ(R1,
4 ]R3 + R1d1 sin(φ1 − φ3),
R3 = [λ(R3,
R1 φ1 = 2d1R1 − R3d1 cos(φ1 − φ3),
R3 φ3 = 2d1R3 − ∆1R3 − R1d1 cos(φ1 − φ3),
R1 sin(φ1 − φ2) = R3 sin(φ2 − φ3),
R1 cos(φ1 − φ2) = −R3 cos(φ2 − φ3).
R1) − R2
R3) − R2
3
1
(10)
(11)
Using two last expressions we will obtain that φ1 = φ3, so it is possible to rewrite system (10) in the
following form :
R1 = [λ(R1,
R1) − R2
R3) − R2
R3 = [λ(R3,
3d1 = ∆1R1R3 + R2
R2
3
1
4 ]R1,
4 ]R3,
1d1.
It is easy to see that only the equilibria (0, 0) satisfy the relation from (11). Two other equilibria
(2, 0) and (0, 2) of system (8) on the invariant manifold R2 = 0 are no longer equilibria in the system
(11). Thus, introduction of a weak electrical couplings between elements leads to the fact that only the
unstable equilibrium state (0, 0) remains in the system, and the heteroclinic circuit between saddles in
system (7) is destroyed. Consequently, introduction of a weak non-zero electrical couplings between
elements leads to the destruction of the stable heteroclinic circuit existing in the initial system (1).
However, numerical experiments show that for relatively small values of d (corresponding to the red
points on the charts of the largest Lyapunov exponent, see Fig. 3 a,b) in a neighborhood of the circuit
there are many trajectories that sequentially visit regions near the saddle limit cycles (Fig.
ref fig2
b). Since the phase trajectories in this case are not attracted to the destroyed saddle limit cycles, but
only visit their neighborhoods, the burst lengths for all elements are constant, and they are determined
by the values of the electrical coupling and the frequency mismatch between the elements (Fig. 3 a).
The largest Lyapunov exponent is positive in this case, Λ1 > 0. This scenario is similar to the scenario
described in [5], where the destruction of a heteroclinic circuit with the presence of noise was shown.
In the case when the frequency mismatch is relatively small (0 ≤ ∆ ≤ 0.235), with a further
increasing the strength of the electrical coupling d, the dynamics of the system becomes regular,
namely, periodic regimes (Fig. 3 c) and quasiperiodic oscillations (Fig. 3 b) are observed in the
numerical experiment. Upon transition to the region of quasiperiodic oscillations, a two-dimensional
torus appears in the phase space of the system (Fig. 3 d). With a further increase in the strength of
the electrical coupling, the torus collapses, on its place a stable limit cycle is born (Fig. 3 e) as a result
of the supercritical Neimark-Sacker bifurcation.
5
Figure 3: Time series x1, x2, x3 and projections of the trajectories of the system (1) on the 2-dimensional
subspasce (x1, x2) in the case of strong asymmetry of synaptic inhibitory couplings. Parameter values for
synaptic inhibitory couplings: (g1, g2) = (0, 5). Parameter values for electrical couplings and frequency
mismatch: a-b -- d = 0.1, ∆ = 0.02; c-d -- d = 0.1, ∆ = 0.4; f-g -- d = 0.2, ∆ = 0.1.
4 Conclusion
In this paper, the ensemble of neuron-like elements is suggested to be considered as a phenomenological
model of a neural network. This approach has the following advantages: it is possible to investigate
low-dimensional neural models and reproduce the main effects observed in more complex models, for
example, in the biologically realistic Hodgkin-Huxley model [14], as well as in real experiments. In
this study it was shown that the adding of arbitrarily small electrical couplings to the ensemble of Van
der Pol neuron-like elements with chemical (synaptic) couplings leads to the destruction of a stable
heteroclinic circuit between saddle limit cycles. It is also shown that the non-identity of elements (in
the absence of electrical couplings) does not lead to the destruction of this heteroclinic circuit. This
result is not typical for non-symmetric systems. The heteroclinic circuit exists, as a rule, in systems
with symmetry, and when the symmetry breaks, it also disappears. In our case, the heteroclinic circuit
is not destroyed, since the amplitude dynamics of the system does not depend on the phase one. It is
also shown that after the circuit breaks down, a weak chaotic activity appears, but the further increase
of the strength of the electrical couplings leads to a regularization of dynamics in the system.
These results can help to gain a deeper understanding the nature of electrical couplings in the
nervous system. The study of their influence on the evolution of neuronal activity is of interest not
only from the point of view of nonlinear dynamics, but also contributes to the development of the
theoretical basis of bioelectronic medicine and the creation of new methods and approaches for the
treatment of diseases of the nervous system that are not amenable to treatment with the help of
pharmacological agents.
Acknowledgement.
The authors thank S.V. Gonchenko for valuable advices and comments. The analytical results were
obtained with the support of the RFBR grant 16-32-00835. Numerical experiments were performed
within the framework of the RSF grant 14-12-00811. A. Kazakov was supported by the Basic Research
Program at the National Research University Higher School of Economics (HSE) in 2018.
6
References
[1] Birmingham K., et al. Bioelectronic medicines: A research roadmap. Nature Reviews Drug Discover.
2014. Vol. 13. P.399.
[2] Seo, D. et al., Wireless Recording in the Peripheral Nervous System with Ultrasonic Neural Dust,
Neuron, 2016, V. 91, no 3, P. 529.
[3] Sacramento J.F., et al. Bioelectronic modulation of carotid sinus nerve activity in the rat: a poten-
tial therapeutic approach for type 2 diabetes. Diabetologia. 2018. V. 61(3). P. 700.
[4] Afraimovich V.S., et al. On the origin of reproducible sequential activity in neural circuits. Chaos.
2004. V. 14(4). P. 1123.
[5] Levanova T.A., et al. Sequential activity and multistability in an ensemble of coupled Van der Pol
oscillators. Eur. Phys. J. Special Topics. 2013. V. 222. P. 2417.
[6] Mikhaylov A. O., et al. Sequential switching activity in ensembles of inhibitory coupled oscillators.
Europhys. Lett. 2013. V. 101(2). P. 20009.
[7] Levanova T.A., et al. Dynamics of ensemble of inhibitory coupled Rulkov maps. Eur. Phys. J.
Special Topics. 2016. V. 225. P. 147.
[8] Fee M.S. et al. Neural mechanisms of vocal sequence generation in the songbird. Annals NY Acad.
Sci. 2004. V. 1016. P. 153.
[9] Rabinovich M.I. et al. Generation and reshaping of sequences in neural systems. Biol. Cybern. 2006.
V. 95. P. 519.
[10] Varona P. et al. Winnerless competition between sensory neurons generates chaos: A possible
mechanism for molluscan hunting behavior. Chaos. 2002. V. 12. P. 672.
[11] Nicholls J.G., et al. From Neuron to brain. 5th ed. Sinauer Associates, 2011. 621 p.
[12] Andronov A.A., et al. Theory of ascillations. New York: Pergamon Press, 1966.
[13] Zwillinger D. Handbook of Differential Equations, 3rd ed. Boston: Academic Press, 1997.
[14] Komarov M.A., et al. Sequentially activated groups in neural networks. Europhys. Lett. 2009. V.
86. P. 60006.
7
|
1801.04515 | 1 | 1801 | 2018-01-14T05:44:36 | Solving constraint-satisfaction problems with distributed neocortical-like neuronal networks | [
"q-bio.NC"
] | Finding actions that satisfy the constraints imposed by both external inputs and internal representations is central to decision making. We demonstrate that some important classes of constraint satisfaction problems (CSPs) can be solved by networks composed of homogeneous cooperative-competitive modules that have connectivity similar to motifs observed in the superficial layers of neocortex. The winner-take-all modules are sparsely coupled by programming neurons that embed the constraints onto the otherwise homogeneous modular computational substrate. We show rules that embed any instance of the CSPs planar four-color graph coloring, maximum independent set, and Sudoku on this substrate, and provide mathematical proofs that guarantee these graph coloring problems will convergence to a solution. The network is composed of non-saturating linear threshold neurons. Their lack of right saturation allows the overall network to explore the problem space driven through the unstable dynamics generated by recurrent excitation. The direction of exploration is steered by the constraint neurons. While many problems can be solved using only linear inhibitory constraints, network performance on hard problems benefits significantly when these negative constraints are implemented by non-linear multiplicative inhibition. Overall, our results demonstrate the importance of instability rather than stability in network computation, and also offer insight into the computational role of dual inhibitory mechanisms in neural circuits. | q-bio.NC | q-bio |
Solving constraint-satisfaction problems with distributed
neocortical-like neuronal networks 1
Ueli Rutishauser1,2∗, Jean-Jacques Slotine3, Rodney J. Douglas4
1 Computation and Neural Systems, Division of Biology and Biological Engineering, California
Institute of Technology, Pasadena, CA, USA
2 Cedars-Sinai Medical Center, Departments of Neurosurgery, Neurology and Biomedical
Sciences, Los Angeles, CA, USA
3 Nonlinear Systems Laboratory, Department of Mechanical Engineering and Department of
Brain and Cognitive Sciences, Massachusetts Institute of Technology, Cambridge, MA, USA
4 Institute of Neuroinformatics, University and ETH Zurich, Zurich, Switzerland
∗ E-mail: Corresponding author [email protected]
Abstract
Finding actions that satisfy the constraints imposed by both external inputs and internal represen-
tations is central to decision making. We demonstrate that some important classes of constraint
satisfaction problems (CSPs) can be solved by networks composed of homogeneous cooperative-
competitive modules that have connectivity similar to motifs observed in the superficial layers of
neocortex. The winner-take-all modules are sparsely coupled by programming neurons that embed
the constraints onto the otherwise homogeneous modular computational substrate. We show rules
that embed any instance of the CSPs planar four-color graph coloring, maximum independent set,
and Sudoku on this substrate, and provide mathematical proofs that guarantee these graph col-
oring problems will convergence to a solution. The network is composed of non-saturating linear
threshold neurons. Their lack of right saturation allows the overall network to explore the problem
space driven through the unstable dynamics generated by recurrent excitation. The direction of
exploration is steered by the constraint neurons. While many problems can be solved using only
linear inhibitory constraints, network performance on hard problems benefits significantly when
these negative constraints are implemented by non-linear multiplicative inhibition. Overall, our
results demonstrate the importance of instability rather than stability in network computation,
and also offer insight into the computational role of dual inhibitory mechanisms in neural circuits.
1 Introduction
The ability to integrate data from many sources, and to make good decisions for action is essential
for perception and cognition, as well as for industrial problems such as scheduling and routing.
The process of integration and decision is often cast as a constraint satisfaction problem (CSP).
In technological systems CSPs are solved by algorithms that implement strategies such as back-
tracking, constraint propagation, or linear optimization. In contrast to those algorithmic methods,
1This article has been accepted for publication in Neural Computation (MIT Press) on 01/12/18. This is the
final peer-reviewed version that will appear in a future issue of Neural Computation.
1
we explore here the principles whereby neural networks are able to solve some important classes
of CSP directly through their asynchronous parallel distributed dynamics.
Our view of an algorithm here is computational one: The defined sequence of discrete operations
to be carried out on data by a computer to achieve a desired end. By contrast, we view our neural
networks as modeled approximations to physical neuronal networks such as those of the cerebral
cortex, in which processing is not under sequential algorithmic control. Our primary interest in
this paper is to understand how these natural networks can process CSP-like problems.
We explore the behavior of these networks in the context of some reference classes of CSP
that are well understood from an algorithmic point of view: 4-coloring of planar graphs (GC4P);
minimum independent set (MIS); and the game Sudoku (SUD). For each of these classes the
network must decide a suitable color assignment for each node of the graph given a total number
of available colors (which is given a priori). We choose these graph-coloring problems because their
topologies lend themselves to implementation in networks, and because their constraints can be
expressed as simple equal / not-equal relations. Moreover, four-coloring on planar graphs is an
interesting problem because it is computationally hard, and because its solution can be applied
to practical tasks in decision making and cognition (Afek et al., 2011; Dayan, 2008; Koechlin &
Summerfield, 2007). On the other hand, Sudoku is interesting because it involves many constraints
on a dense non-planar graph (Ercsey-Ravasz & Toroczkai, 2012; Rosenhouse & Taalman, 2011).
Also, there are hard input constraints on the values of some SUD nodes, which makes the solution
of SUD significantly more difficult than simple graph coloring, in which any valid coloring is
acceptable.
We show that our neuronal networks can solve arbitrary instances in these three problem
classes. Because many problems in decision making can be reduced to one of these classes (Dayan,
2008) showing that our networks can solve them, implies that they can in principle solve all other
problems that are reducible to these reference classes. These problems can of course be solved also
by algorithmical methods (Karp, 1972; Robertson, Sanders, Seymour, & Thomas, 1996; Appel,
Haken, & Koch, 1977; Kumar, 1992). However, our important contribution here is to explain the
principles of dynamics that allow a network of distributed interacting neurons to achieve the same
effect without relying on the centralized sequential control inherent in these well known algorithms.
Our networks are composed of loosely coupled cooperative-competitive, "winner-take-all" (WTA)
modules (Hahnloser, Sarpeshkar, Mahowald, Douglas, & Seung, 2000; Maass, 2000; Hahnloser,
Douglas, Mahowald, & Hepp, 1999; Douglas & Martin, 2007; Yuille & Geiger, 2003; Ermentrout,
1992; Wersing, Steil, & Ritter, 2001; Abbott, 1994). Each WTA behaves as a special kind of
variable that initially is able to entertain simultaneously many candidate values, but eventually
selects a single best value, according to constraints imposed by the processing evolving at related
variables. This collective selection behavior is driven by the signal gain developed through the re-
current excitatory (positive feedback) connections between the neurons participating in the WTA
circuits, as described below.
WTA circuits are appealing network modules because their pattern of connectivity resembles
the dominant recurrent excitatory and general inhibitory feedback connection motif measured both
physiologically (Douglas, Martin, & Whitteridge, 1991) and anatomically (Binzegger, Douglas, &
Martin, 2004) in the superficial layers of the mammalian neocortex. There is also substantial and
growing evidence for circuit gain and its modulation in the circuits of the neocortex, a finding that
is consistent with the recurrent connectivity required by WTAs (Douglas, Martin, & Whitteridge,
1989; Douglas et al., 1991; Ferster, Chung, & Wheat, 1996; Lien & Scanziani, 2013; Li, Ibrahim,
Liu, Zhang, & Tao, 2013; Carandini & Heeger, 2012; Kami´nski et al., 2017).
Previous studies of WTA networks have focused largely on questions of their stability and
2
convergence (For example (Ben-Yishai, Bar-Or, & Sompolinsky, 1995; Dayan & Abbott, 2001;
Hahnloser et al., 2000; Rutishauser, Douglas, & Slotine, 2011)). However, more recently we have
described the crucial computational implications of the unstable expansion dynamics inherent in
WTA circuits (Rutishauser, Slotine, & Douglas, 2015).
It is these instabilities that drive the
selection processing and therefore the computational process that the network performs. We have
explored this computational instability in networks of linear thresholded neurons (LTN) (Koch,
1998; Douglas, Koch, Mahowald, & Martin, 1999; LeCun, Bengio, & Hinton, 2015) because they
have unbounded positive output. Consequently, networks of LTNs with recurrent excitation must
rely on feedback inhibition rather than output saturation (eg (Hopfield, 1982; Miller & Zucker,
1999; Rosenfeld, Hummel, & Zucker, 1976)) to achieve stability. This also means that networks of
LTNs may change their mode of operation from unstable to stable (Rutishauser et al., 2015).
The principles we develop in this paper depend on concepts we have described previously, in
which we apply contraction theory (Slotine, 2003) to understand the dynamics of collections of
coupled WTAs (Rutishauser et al., 2011, 2015). And so, for convenience, we first summarize briefly
the relevant points of that work. First, note that contraction theory is applicable to dynamics with
discontinuous derivatives such as those introduced by the LTN activation function that we utilize.
This is because in such switched systems the dynamics remain a continuous function of the state
(see results for details). Under suitable parameter regimes, LTN networks can enter unstable
subspaces of their dynamics. The expansion within an unstable subspace of active neurons is
steered and constrained by the negative divergence of their dynamics. This ensures that the
current expanding sub-network will soon switch to a different sub-network, and so to a different
subspace of dynamics (Rutishauser et al., 2015). The new space might be stable or unstable. If
unstable, the network will again switch, and so on, until a stable solution space is found. We
refer to the unstable spaces as 'forbidden' spaces because the network will quickly exit from them.
The exit is guaranteed because the dynamics in forbidden spaces are such that the eigenvectors
associated with the dominant eigenvalue of the system Jacobian are always mixed. This means
that the activity of at least one neuron is falling toward threshold, and will soon pass below it, so
changing the space of network dynamics (see (Rutishauser et al., 2015) for a proof.) Stable spaces,
on the other hand, are said to be be 'permitted'. In these spaces the eigenvalues are all negative
and so the network will converge toward a steady state in that space.
Critically, negative divergence ensures that the dynamics of the space entered next has a lower
dimensionality than the previous space, regardless of whether it is stable or unstable. Thus the
sequence of transitions through the sub-spaces causes the network to compute progressively better
feasible solutions. A further crucial feature of this process is that the direction of expansion
is determined automatically by the eigenvectors of the Jacobian of the currently active neurons
in the network. Thus, the direction of expansion may change according to the particular set
of neurons active in a given forbidden subspace.
In this sense the network is able to actively,
asynchronously, and systematically search through the forbidden spaces until a suitable solution
subspace is encountered. It is this process that constitutes network computation (Rutishauser et
al., 2015).
Now we extend these concepts and show how they can be utilized to construct networks that
solve certain classes of constraint satisfaction problems. We show using new mathematical proofs
and simulations how such problems can be embedded systematically in networks of WTA modules
coupled by negative (inhibitory) and positive (excitatory) constraints. Our overall approach is
to ensure that all dynamical spaces in which a constraint is violated are 'forbidden'. This is
achieved by adding additional neurons that enforce the necessary constraints. Importantly, our
new analytical proofs guarantee that these networks will find a complete and correct solution
3
(provided that such a solution does exist for the problem instance).
We also find that the form of inhibitory mechanism used to implement negative constraints
affects the performance of the network. Two different types of inhibition can be used to implement
negative constraints: linear subtractive and non-linear multiplicative inhibition. While some prob-
lem classes could be solved using only standard subtractive inhibition between modules, we found
that the solution of more difficult problems is greatly facilitated by using multiplicative non-linear
inhibition instead. Recent experimental observations have implicated these two kinds of inhibition
in different modes of cortical computation (Jiang, Wang, Lee, Stornetta, & Zhu, 2013; Pi et al.,
2013), and the work presented here offers a theoretical foundation for the computational roles of
these two types of inhibition also in the neuronal circuits of the neocortex.
2 Network Architecture and Results
2.1 CSP Organization
A CSP consists of a set of variables (concepts, facts) Xi that are assigned discrete or continuous
values and a set of constraints Ci between these variables. The constraints may be unary (restricted
to one variable), binary (a relation between two variables), or higher order. Each constraint Ci
establishes the allowable combinations between values (or relationships between values) of the
variables X. A state (or configuration) of the problem is an assignment of values to some or all of
the variables X. An assignment (or solution) may be complete or partial.
The CSPs are instantiated as neuronal networks by embedding the specific problem in a field of
identical WTA modules. These modules have a standard connection pattern of recurrent excitation
and inhibition that supports the WTA functionality. The CSP is embedded in the network by
coupling the WTA modules via neurons that implement the negative (inhibitory) and positive
(excitatory) constraints. As we will show below, we find that the performance of the CSP network
is affected by the form of negative constraint inhibition onto the WTAs; either linear or non-
linear. We begin by describing the 'standard' WTA (WTAS) and related CSP networks that make
use of linear inhibitory negative constraints; and thereafter describe the extended WTA (WTAE)
networks that make use of non-linear inhibitory negative constraints.
2.2 Standard Winner-Take-All circuit
Each standard WTA (WTAS) consists of N point neurons, N-1 of which are excitatory and the
remaining one (N) is inhibitory (Fig 1A). In the examples below the WTA should express only
a single active unit, and thus the excitatory neurons xi(cid:54)=N receive only self-feedback of excitation
αi. Each neuron may receive also external input Ii. Normally distributed noise with mean µ and
standard deviation sd is added to all these external inputs: Ii = Ii + N (µ, sd)).
The single inhibitory neuron xN sums the β2 weighted input from each of the excitatory neurons,
and returns a common β1 inhibitory signal to all of its excitatory neurons. The dynamics of this
single WTA are
τ xi + Gxi = f (αxi − β1xN − Ti + Ii + N (µ, sd))
τ xN + GxN = f (β2
xj − TN )
N−1(cid:88)
j=1
4
(1)
(2)
Figure 1: Connectivity and network architecture.
(A) Single WTA comprising two excitatory
neurons that encode possible winners, A and B. Top: all connections. Bottom: simplified notation.
(B) Two modules M1 and M2 that implement WTA shown in (A) are connected with one another
by two additional inhibitory cells N C1,2. These cells enforce the negative constraint that the two
WTAs cannot reach the same winner (solution). The constraint inhibition is linear. To maintain
analogy with neurons, the inhibition is shown applied to a dendrite. However, in this point model
neuron, it could as well be applied directly to the soma. See Fig 2 for a simulation of this constraint
problem. (C) The non-linear inhibitory synapse (blue circle) provides on-path inhibition, which
can suppress only dendritic but not somatic excitatory inputs. (D) Example circuit with both (non-
linear) negative, and positive (NC, PC) constraint cells implemented with non-linear inhibitory
synapses.
5
β1β2ABA)B)α1α1I2I1β1Dβ2DAABBβ1DC)M1M2D)"if A then not A"AABB"if B then A"γ1Pα1β1Ddendriticsomaticγ1PICIbiasInhibitoryExcitatoryExcitatory InhibitoryInhibitory synapse (standard)Excitatory synapseInhibitory synapse (non-linear)Branching axonDendriteAxonNC1NC2NC1PC1"if A then not A""if B then not B"β1Dβ2Dγ2Pβ1β2ICIC where f (x) is a non-saturating rectification non-linearity. Here, we take f (x) = max(x, 0), making
our neurons linear threshold neurons (LTNs). Ti ≥ 0 is the activation threshold.
2.3 Constraint satisfaction models implemented on WTAs
The nodes of the graph represent the possible states of the problem at that location, while the
edges of the graph represent the constraints acting between nodes. Each node is a WTA module,
and the patterns of activation of its excitatory neurons represent the allowable states of that node.
Since only one winner is permitted, these WTAs represent as many solution states as they have
excitatory neurons, ie N − 1 states. In the CS problems considered here, all nodes implement the
same set of states. For example, in the case of GC4P, every node of the problem graph is a WTA
with 4 excitatory neurons that represent the four possible color states of that node.
The constraints between WTAs are implemented by additional neurons that add additional
excitatory or inhibitory interactions between the relevant states (corresponding to particular neu-
rons) of the interacting nodes.
In graph-coloring problems, the constraints are negative: The
selection of a particular color neuron at one node, suppresses the selection of the corresponding
color neuron at a neighboring node. However, other problems may require also positive constraints.
For example, if WTA A is in state 1, then WTA B should also be in state 1. The Maximal Inde-
pendent Set (MIS) problem considered below requires such positive constraints. In this section, we
will first describe the dynamics and connectivity of negative-and positive contraint cells whereas
their specific wiring to implement a particular CSP is described later separately for each CSP class
considered.
Negative constraints (NC), or competition, between the states of different WTAs is introduced
through inhibitory feedback by negative constraint cells (NCC) d. In our implementation, each d
provides inhibitory output onto the same set of excitatory cells from which it receives its input
(Fig 1B). In contrast to the inhibitory cells that enforce competition between states within an
WTA, NC cells enforce their competition between (some) specific state neurons across multiple
participating WTAs.
Positive constraints (PC), or hints, are implemented by excitatory positive constraint cells
(PCC) p, each of which receives input from a specific excitatory cell of one WTA and provides
excitation onto specific excitatory neurons(s) on other WTAs.
The dynamics of negative constraint cells d are:
τ di + Gdi = f (β2D
Mi(cid:88)
j=1
xj − T D)
(3)
where Mi are the number of units xj that provide input to di. Note that the xj are members of
different WTAs. Similarly, the dynamics of a positive constraint cell p are:
τ pi + Gpi = f (γ2P
xj − T P )
Ni(cid:88)
j=1
where Ni are the number of units xj that provide input to pi.
The total constraint current, composed of negative and positive components I N C
i
i = − Di(cid:88)
Pi(cid:88)
I cons
i
= I N C
i + I P C
β1Ddk +
γ1P pk
k=1
k=1
6
, and I P C
i
is:
(4)
(5)
where Di and Pi are the total number of negative and positive constraint cells that provide input to
cell i, respectively. β1D > 0 is the strength of the inhibitory synapse made by negative constraint
cell dk onto cell i and γ1P > 0 is the strength of the excitatory synapse made by positive hint cell
pk onto cell i.
Thus, the dynamics of excitatory units in constraint networks composed of standard WTA xi
are:
τ xi + Gxi = fs(αxi − β1xN − Ti + Ii + I bias
i + I cons
i
)
(6)
where I bias
i
are forward inputs that bias the WTA towards solution xi. The initial conditions
for all units are xi(0) = 0, di(0) = 0, pi(0) = 0 throughout this work.
2.4 Dynamics of an example negative constraint network
Before we consider how to choose the parameters of this network and how to analyze its stability
and convergence, consider the example network shown in Fig 1B. This network implements two
negative constraints (NC1 and NC2) between two WTA modules, M1 and M2. Each has only two
possible solutions: A and B. The NC cells enforce the "not same" constraint that both WTAs
should not be in the same state. The dynamics of this circuit (Fig 2) converge to a steady-state
in which the solutions of the two WTAs depend on each other in addition to local constraints (the
inputs). For example, in Fig 2 unit B on both WTAs receive the largest inputs (cyan,red) and
so, independently, the winner on either WTA would be B. However, because of the constraint
dependency only one WTA can express B (node M2), whereas the winner on the other node (M1)
is A despite receiving the lower amplitude input.
2.5 Stability and convergence
Our arguments rely on the concept of the effective Jacobian, which expresses the dynamics of the
currently active subset of neurons (Rutishauser et al., 2015). Consider the network in the form of
x = z(x, t). Here, z(x, t) = f (Wx − Gx), where W is a matrix of weights that describes all the
connections between neurons, f (x) is the non-linearity, and G = diag(G1, . . . , Gn) is a diagonal
matrix containing the dissipative leak terms for each neuron. The effective Jacobian for this system
is
Jef f =
(7)
= ΣW − G
∂z
∂x
where Σ = diag(σ1, . . . , σn) is a diagonal matrix of derivatives of the activation function, evaluated
at the current state xi of each neuron. In our case, all neurons are LTNs, resulting in derivatives
equal to either 0 or 1 depending on whether the state of a neuron is above or below threshold. This
is because the activation function f (x) has slope 1 whenever it is above threshold. Therefore, Σ re-
mains the same as long as no neuron crosses its threshold. Consequently, the effective connectivity
of the network changes whenever a neuron crosses its threshold. 'Effective connectivity' indicates
that connections that arise from inactive (below threshold) neurons cannot influence their target
neurons because they do not provide output. Therefore, the rows of the Jacobian corresponding
to these silent neurons are zeroed out. However, their columns are not zeroed, and so these silent
neurons still receive and process input.
We use Jef f as a mathematical tool to assess and describe network computation. We have
used Jef f previously to show that a single WTA circuit has both permitted (contracting) and
forbidden (expanding) sets of active neurons, and that it is the existence of the forbidden sets that
provides computational power (Rutishauser et al., 2015). This is because forbidden sets are highly
7
Figure 2: Enforcing constraints between WTAs using negative constraint (NC) cells. Simple
example using the two-Node circuit of Fig 1B. (A) Circuit diagram of the network, with two nodes
M1 and M2 with two winners A and B each. The two negative constraint cells NC enforce the
not-same constraint. (B) Weight matrix of the full network. Gray boxes mark nodes M1 and M2.
Connections outside of the boxes correspond to the NCs. (C) The inputs to the network. (D,E)
Activity on nodes M1 and M2. (F) Activity of the two negative constraint cells NC1 and NC2.
8
D)024 0246Node M2 Unit AUnit BInhib012NC cells NC1NC20204060024 M1−AM1−BM2−AM2−BInputs0204060time [tau]time [tau]E)0204060time [tau]0204060time [tau]F)C)Node M1activity [au]activity [au]Unit AUnit BInhibactivity [au]activity [au]A)ABM1NC1ABM2NC2B)Input AInput BInput AInput B24682468unit nr [from]unit nr [to]InhibitoryExcitatory unstable and this drives the network to enter a different set from the one it is currently expressing.
While unstable, the negative feedback ensures that the dynamics during forbidden states steers
network activity towards a suitable solution. A forbidden set must satisfy two conditions:
its
divergence must be negative, and its effective Jacobian Jef f must be positive definite (Rutishauser
et al., 2015). In our WTA networks, these two conditions are enforced by shared inhibition and
excitatory self-recurrence, respectively. Together, they guarantee that an individual WTA will exit
forbidden sets exponentially fast.
The arguments summarized above and the new results derived below rest on contraction theory
(Slotine, 2003), a powerful analytical tool that allows us to systematically reason systematically
about the stability and instability of non-linear networks such as the LTN networks that we use.
Contraction theory is applicable to non-linear networks, such as switched networks, provided that
the dynamics remain a continuous function of the state, and that the contraction metric remains
the same (Lohmiller & Slotine, 2000). It is thus applicable to networks composed of units such
as LTNs that have activation functions whose derivatives are discontinuous. To see why this is
the case, consider the following network:
x + x = max(x, 0). Note that x, which describes the
dynamics of the network, is a continuous function of x despite having discontinuous derivatives
with respect to x. Furthermore, the metric of our networks remains the same throughout their
processing (Rutishauser et al., 2015). Thus, both conditions for the application of contraction
theory are satisfied.
Now we present new proofs that together provide important insights into the operation of this
network. First, we prove that adding NC cells creates new forbidden sets. The NCs are thereby
able to influence the dynamics of network computation. Second, we prove that adding PC cells
creates permitted sets. Third, we prove that networks of WTAs connected by NC/PC cells remain
stable despite the presence of forbidden sets. Together, these new results generalize our previous
findings from individual WTAs (Rutishauser et al., 2015) to networks of WTAs coupled by NC
and PCs that implement specific constraints. Finally, we provide rules that allow all instances of
three classes of CSPs to be implemented in networks of WTAs by installing suitable NC and PC
coupling connections.
2.5.1 Proof 1: Adding negative constraint cells creates forbidden sets
Consider a set of WTAs, with the dynamics of each described by a Jacobian Ji:
0
− l1β1
l2α − G − l2β1
−G
l3β2
Ji =
l1α − G
(cid:34)
0
l3β2
Now, consider a system composed of two copies of the above circuits J1,2:
div(J3) < 0 if div(J1,2) < 0. Thus, combinations of circuits with negative divergence will
always have negative divergence. Next, we add an additional inhibitory neuron ("NC cell") that
enforces additional competition between the two sub-circuits J1 and J2. This new constraint will
create a new forbidden subspace, which must be expanding. This system is described by J4.
(cid:35)
(cid:35)
J3 =
J1 0
0 J2
(cid:34)
J4 =
J3 −kDT
D −G
9
(8)
(9)
(10)
(cid:104)
(cid:104)
V
(cid:105)
(cid:105)
with k = β1
β2 . For example, setting D =
would connect the new in-
hibitory unit such that simultaneously activating the first unit in both sub-circuits J1 and J2 is
forbidden.
β2 0 0 β2 0 0
The individual WTA J1,2 is expanding if V1,2J1,2VT
1 −1 0
V1,2 =
1,2 > 0 with
This V term ensures that both excitatory units on a WTA cannot be simultanously active.
The combined system J4 is expanding if V4J4VT
4 > 0 with
V4 =
0
0
0
V 0
Vj −Vj 0
(cid:104)
(cid:105)
. The term [Vj − Vj] ensures that the first excitatory neurons on both
where Vj =
WTAs cannot be simultaneously active, which is the constraint that the NC cell described by D
embeds. Multiplying out, the result becomes
1 0 0
VJ1VT
0
F4 = V4J4VT
4 =
0
VJ1VT
VjJ1VT
−VjJ2VT
VjJ1VT −VjJ2VT VjJ1VT
j + VjJ2VT
j
which can further be decomposed into
(cid:34)
F4 =
(cid:35)
Q1 BT
B Q2
(11)
(12)
(13)
(14)
(15)
A sufficient condition for above to be positive definite is, Q2 > BT Q−1
1 B (Horn, 1985) (Page 472).
Note that this condition requires Q1,2 to be symmetric. After substituting all variables, the result
is
−2 + 2α
−1 + α
0
F4 =
0
−1 + α
−2 + 2α
1 − α
1 − α −2 + 2α
Thus, Q2 = 2(α− 1), B = [−1 + α, 1− α], and Q1 = 2(α− 1)I. Accordingly, F4 is positive definite
if 2(α − 1) > α − 1. This condition is satisfied if α > 1.
In conclusion, the additional feedback loop introduced by adding the negative constraint cell
creates a forbidden subspace, which is both negatively divergent as well as expanding. This method
can be applied recursively to add arbitrary numbers inhibitory feedback loops to a collection of
WTA circuits.
Positive constraints do not create additional forbidden sets. Instead, a positive constraint of
the kind "if WTA1 is in state 1, WTA2 should be in state 2" reinforces two sets that are already
permitted. Thus, all that is required with regard to positive constraints, is a proof demonstrating
that permitted sets remain permitted (see below).
2.5.2 Proof 2: Stability analysis
In this section, we demonstrate that the addition of positive and negative constraints does not
disturb the overall stability of the system. A key feature of contracting systems is that contraction
10
(cid:34)
τ J =
J3 −kBT
A −G
(cid:105)
(cid:34)
τ J =
J3 BT
A −G
(cid:35)
(cid:35)
(16)
(17)
(cid:105)
is preserved through a variety of combinations of subsystems (Slotine, 2003). Importantly, this
property includes the two kinds of combination that we consider here: the introduction of negative
and positive constraint cells. In the following, we simply outline the grounds for this inclusion.
The detailed proof of preservation of contraction under combination of subsystems is presented in
(Slotine, 2003).
We begin by showing that the addition of negative constraint NC cells does not disturb the
stability of the system. Coupling two WTAs J1 and J2 with NC cells such that the two WTAs
can not have the same winners results in the system Jacobian J3 (see Equation 9).
(cid:104)
where the NC cells have no dynamics apart from the load term G on the diagonal. The connectivity
β2D , B = A. Feedback combinations
of the NC cells is A =
of this form are guaranteed to be contracting as shown in (Slotine, 2003) (section 3.4).
β2D 0 0 β2D 0 0
. With k = β1D
A similar argument holds for positive constraint cells: Adding a positive constraint between
two identical WTAs results in a new system with Jacobian
(cid:105)
(cid:104)
where the positive constraint cell has no dynamics apart from the load term g on the diago-
nal. For the example of a single positive constraint cell that enforces that if WTA 1 is in
state 1, WTA 2 should be in state 2,
and
its connectivity is: A =
γ2P 0 0 0 0 0
(cid:104)
0 0 0 0 γ1P 0
B =
relationship between A and B. However, taking the symmetric part JS = 1
a system in which AS = BS =
form are guaranteed to be contracting (Slotine, 2003) (section 3.4) if
. Unlike the previous negative constraint case, there is no simple
2(J + JT ) results in
. Feedback combinations of this
(cid:104) γ1P +γ2P
0 0 0 γ1P +γ2P
(cid:105)
2
0
2
σ2(AS) < λ(J3)λ(−G)
(18)
where σ is the largest singular value and λ the largest eigenvalue. Both λ(J3) and λ(−G) are
negative by definition, since the systems are both individually contracting.
2 − 1 (see (Rutishauser, Slotine, & Douglas, 2012) section
The contraction rate of a WTA is α
2.3.2). Assuming G = 1, this reduces to
(γP 1 + γP 2)2 < 2 − α
(19)
Therefore, as long as the weights of the positive constraint cells fulfill condition Equation 19, the
system will remain contracting.
2.6 Choice of parameters
The stability and computational power of a WTA depends on the following parameters: excitatory
local recurrence α, inhibitory recurrence β1 and β2 (note that β2 is a superscript index and not a
power), and inhibitory β1,2D and excitatory (γP 1,γP 2) recurrence between WTAs that implements
the constraints. Provided these parameters are set to values within a permitted range, this network
will allow only one winner to emerge, and that solution depends on the pattern of its input I
(Rutishauser & Douglas, 2009). These constraints are:
11
(cid:113)
1 < α < 2
β1β2
1
4
< β1β2 < 1
α2
β1β2 < (1 − 1
2
α
(γP 1 + γP 2)2 < 2 − α
)(β2
1 +
)
(20)
(21)
(22)
(23)
where Equations 20-22 are derived in (Rutishauser et al., 2011) and Equation 23 is derived
above. Note that the constraints on inhibitory feedback loops established by inhibitory neurons
apply to both the local inhibitory neuron of each WTA, as well as the additional inhibitory neurons
that establish inhibitory feedback between WTAs (referred to as β1,2 and β1,2D, respectively).
For all simulations, we chose parameters within the permitted ranges given by Equations 20-23.
Within those restrictions, the parameters were chosen to optimize performance for each problem
class (i.e. GC4P, MIS, and SUD) and network type (W T As and W T Ae). Note that the parameters
used were identical for all instances of a particular problem class and network type (i.e. GC4P
solved with the W T As architecture) and not optimized for a particular problem instance (which
are randomly generated).
2.7 Planar Graph Coloring using only negative constraints (GC4P)
We next applied this architecture to the problem of graph coloring. Here, each node must at
any time express only one of a fixed number of different colors. The selected color represents the
node's current state. The coloring constraint is that nodes that share an edge are forbidden from
having the same color. Finding an assignment of colors to all the graph nodes that respects this
constraint for all their edges solves the graph coloring problem (for example, Fig 3A). The smaller
the number of permitted colors, the harder the problem.
More specifically, we chose to investigate the problem of coloring planar graphs with 4 colors
with an arbitrary number of nodes and undirected edges (Figs 4). A planar graph is one that can
be embedded in the plane, which means that it can be drawn in such a way that no edges cross one
another. Here, we restrict ourselves to planar graphs, because they are guaranteed to be colorable
with 4 colors.
The topology of Graph Coloring can naturally be framed as a topologically distributed con-
straint satisfaction problem, and so implemented by networks of WTA circuits in the manner that
we have described above. The color state of each node is represented by a single WTA, and so the
network implementation requires as many WTA modules as the graph has nodes. In the problems
reported here, the smallest number of colors required to color the graph (its chromatic number)
is constant (here, 4) and given. Thus, each WTA has as many state neurons (possible winners) as
the chromatic number. Because the local competition between these states is unbiased, all WTAs
have the same internal connection architecture.
The edges of the graph are constraints of the kind "not same", and they are implemented using
NC cells (see Methods; Fig 3). At most, the constraints across each edge will require as many
different NC cells as there are node colors. However, a single NC cell can enforce its constraint
across arbitrary numbers of neighbor WTAs. Thus, it is sufficient to add only one NC cell per color
at a given node. This cell is then able to assert the same constraint across all edges connected to
this node (see Methods).
12
Figure 3: Solving graph coloring problems with networks of WTAs. (A) Example 4-node graph,
with one possible color solution computed by a WTA network indicated (colors). (B) Weight matrix
of the 4 module WTA network, implementing the graph shown in (A). Neurons 1 through 20 are
configured as 4 separate WTAs, each with four excitatory neurons that encode the 4 possible
colors of each node, and one global (to that WTA) inhibitory neuron. Neurons 21-24 impose
the inhibitory constraint that no edge may have nodes of the same color. (C) Dynamics of the
network leading to solution in (A). Note the relatively small modulations of constraint neuron
activity required to achieve this solution.
We have shown previously for both symmetrical (Hahnloser et al., 2000) and asymmetrical net-
works (Rutishauser et al., 2015) that the fundamental operation of the WTA modules is an active
selection process whereby the activities of some neurons are driven below threshold by those who
are receiving support from either local or remote excitatory input. Partitions of active neurons
that are inconsistent with stability are forbidden. Such a partition is left exponentially quickly
because the unstably high gain generated by the neurons of forbidden partitions will drive recur-
rent inhibition sufficiently strongly to soon drive a neuron of the set beneath threshold, and so
bring a new partition into being. This process continues until a consistent permitted partition is
found. The previous work was concerned only with the relationship between inhibitory feedback
that is driven by the excitatory members of the local WTA, and the existence of forbidden sets.
We demonstrated there that it is the existence of these forbidden subspaces that provides compu-
tational power (Rutishauser et al., 2015). Now, in these CSP networks the negative constraints
provide an additional source of negative feedback routed via remote WTAs (see above for a proof).
We tested the performance of the network in solving randomly generated planar graphs with
up to 49 four color nodes (Fig. 4). In these cases there were no constraints on the acceptable color
of any specific node. Thus, any solution that satisfies all constraints is acceptable. Consequently
there are many equally valid solutions for each graph (at minimum four, the number of colors).
The goodness of a solution was measured by two metrics: The average time the network took
to converge, and the number of edges that were not satisfied (Number of Errors, in the following)
as a function of time. We found that WTA networks can solve all graphs of up to 25 nodes
correctly within 1500τ . However, larger graphs are incompletely solved and take longer (Fig 4C).
The computational process evolves in such a way that the number of errors decreases exponentially
fast (Fig 4D). This means that networks of a given size solve the majority of random graphs quickly,
with a minority taking much longer. This results in heavy-tailed distributions with respect to the
13
1234A)B)unit nr10205101520123456C)activity [au]unit nr [from]101520unit nr [to]51015205time [tau]InhibitoryExcitatory Figure 4: Performance on solving the planar graph four coloring problem (GC4P). (A) Example
of a GC4P solution. (B) Weight matrix of a network with 505 units (245 for WTAs and 260 NCs).
(C,D) Performance of W T As. (C) Cumulative probability of network convergence as a function of
processing time and network size. (D) Average number of errors (graph edge constraints violated)
as a function of time. (E,F) Performance for W T Ae. Same notation as (C,D). (G) Average time
±s.e. to find a solution as a function of network size and architecture. Time to solution for large
problems was significantly shorter for W T Ae network by comparison with W T As (**, p¡0.01,
kstest). (H) Distribution of times to solution, as a function of network size. (I) Scaling of time
at which solved 50% and 80% of all networks converged as a function of network size. W T As
parameters: α = 1.5, β1 = 3, β2 = 0.3, β1D = 1.5, β2D = 0.15, iid noise of µ = 1.5, σ = 0.15. W T Ae
parameters: same, except α = 1.2, β1D = 3, β2D = 0.3, s = 0.15, o = 0. See section 2.6 for how
the network parameters were chosen. Results are for N=1000 simulations for each network size. A
new random planar graph with 80% density was generated for each simulation.
elapsed time to solution (Fig 4H).
2.8 Maximal Independent Set using both positive and negative con-
straints (MIS)
The use of positive constraint cells is demonstrated by solving a second class of graph coloring
problems: maximal independent sets (MIS)(Fig 5). MIS problems are a second fundamental class
of computational problems (Afek et al., 2011) solvable with the type of network we present here.
MIS is related to graph coloring, but requires different constraints. In this case, each node must
take one of two possible colors A and B. If two nodes are connected (they are neighbors), then they
cannot both be A. Further, any node that takes color B must be connected to at least one other
node that has color A. This problem finds practical application in many distributed algorithms
(Lynch, 1996), where it is used for automatic selection of local leaders.
We translated MIS problems into networks that use both negative and positive constraints. The
first constraint, that two neighbors can not both be color A, is implemented by one NC cell for
each connected pair (see Fig 1D). The second constraint, that if a node has color B it encourages
its neighbors to be of color A, is implemented by positive feedback through two hint cells for each
pair of connected nodes (see Fig 1D). The positive feedback is active conditional on a node being
14
D)A)B)time [tau]C)time [tau]010002000log nr errorstime [tau]H)E)G)time [tau]Probability correcttime [tau]F)−4−202020400200400nr nodes in graph50 %80 %time to solution [tau]020040000.06N=3600.20.4N=900.06N=2501000200000.40.8N=9N=25N=36N=49time [tau]N=9N=25N=36N=49I)9 2536490400800WTAEWTAS******WTAEWTAEunit nr [from]200400unit nr [to]200400cummulative probability010002000010002000log nr errors-20200.40.8cummulative probabilityN=9N=25N=36N=49N=9N=25N=36N=49WTAS Figure 5: Performance on solving the Maximal Independent Set (MIS) problem. (A) Example
maximal independent set (red nodes) on an 8 node graph. Each red node is not connected to any
other red nodes and each green node is connected to at least one red node. (B) Connectivity for
a simple two-node problem. Each node has two possible winners (red, green). NC1 enforces that
not both notes can be red. PC1 and PC2 enforces that if a node is green, the other is red. (C)
Weight matrix for the 8-node graph illustrated in (A). The dashed box indicates the connection
submatrix of the 8 WTAs (3 units each). Remaining entries indicate constraint units and their
connections. (D-E) Performance of W T As (D) W T Ae (E) on random MIS problems of different
size (number of nodes). Graphs were randomly generated planar graphs with 90% density. W T Ae
converges more quickly than W T As for all problem sizes. (F) Performance comparison between
W T As (black) and W T Ae (red). Except for the smallest problem (size N=9), W T Ae converged
significantly more quickly than W T As (** is p¡0.01, ks-test). W T As parameters: α = 1.2, β1 =
3, β2 = 0.3, β1D = 1.5, β2D = 0.15, γ1P = 0.8, γ2P
2 = 0.15, iid noise of µ = 1.5, σ = 0.15. W T Ae
parameters: Same, except γ1P = 1.5.
of color B. Thus, if a node has color A, the positive constraint is inactive. We found that this WTA
network solves MIS in a manner and speed similar to that described above for graph coloring (Fig
5D): The networks solve most MIS problems of large size fairly quickly, however a small number
of large problems remain unsolved even at long times.
2.9 Fully constrained graph coloring problems - Sudoku (SUD)
The standard WTA networks are also able to solve non-planar graph coloring problems in which
the color states of many nodes have a fixed assignment. These initial assignment constraints make
graph coloring significantly more difficult than the case in which any valid solution is acceptable.
A canonical example of this problem class is the popular game Sudoku.
The graph of Sudoku has 81 nodes arranged in a 9x9 lattice (Fig 6A,B). The lattice is composed
of 3x3 boxes, each of which has 3x3 nodes. Each node can take one of nine colors (numbers). The
15
A)B)C)12Node 1NC112Node 2PC1PC2InhibitoryExcitatory02040600204060unit nr [from]unit nr [to]0500100000.20.40.60.81.0cdf P(correct after time t)N=9N=25N=36N=49D)E)0500100000.20.40.60.81.0cdf P(correct after time t) N=9N=25N=36N=49time [tau]9 2536490100200300400F)time to solution [tau]WTAEWTASnr of nodes******WTASWTAEICIC constraints of the problem are that each color can appear only once in each row, in each column,
and in each box. The neural network that implements Sudoku consists of 1052 units. Of those,
810 units (81 nodes, 9 excitatory and 1 inhibitory each) implement the nodes (WTAs) and 243
implement the constraints (9 row, 9 column and 9 box constraints, one for each color; i.e. 27*9).
In addition, there are initial constraints in the form of specified inputs to a subset of the nodes
which describe the specific problem to be solved (Fig 6A).
For the sudoku network, each excitatory cell receives a constant noisy excitatory input from the
network. This input stands for the broader network context in which the particular CS network is
embedded. In addition, each excitatory cell receives several negative constraint inputs. There are
no positive hint constraints. The forward inputs I bias
enforce the fixed color assignments for some
nodes, as required by the specific SUD problem to be solved. Unbiased neurons receive I bias
i = 0.
We found that these standard W T As networks are able to solve many Sudoku problems. But
their performance is poor. The average converge time was 1330τ , and the network was only able
to solve 60% of all networks in the maximal time permitted. As in the case of graph coloring, the
number of violated constraints (errors) decreases exponentially as a function of simulation time
(Fig 6E,F).
i
2.10 Extended WTA (W T Ae) networks
We sought to improve the fraction of correct solutions and rate of reduction in errors by modifying
the network configuration. We reasoned that the exponential form of the network convergence
reflects its exploration of the combinatorial solution space; and that convergence would be more
rapid if the network could be made more sensitive to its constraints.
We noticed that in equation 6 the mechanism of state selection within a WTA is similar to that
of negative constraints; both operate via subtractive inhibition (i.e. the term −β1Ddj is added).
Therefore a negative constraint, rather than only discouraging the selection of a particular state,
could prevent a state from ever being selected:
If the subtractive inhibition through negative
constraints cells is sufficiently strong, this cell will be driven far below the activation threshold and
so become insensitive to positive inputs. We hypothesized that, if the effects of the constraints
scaled multiplicatively rather than being applied subtractively as in equation 6, then the network
would be able to separate the function of state selection at a node from the constraints that biased
that selection, and that this change in architecture might promote more rapid convergence.
The necessary separation of inhibitory constraint functions was achieved by modifying equation
5 to:
Di(cid:88)
Ci(cid:88)
I C
k +
β1Ddk)(
I cons
i
= g(
k=1
k=1
Pi(cid:88)
k=1
γ1Dpk)
(24)
where I C
k is non-specific contextual background excitation (see below), Di is the number of negative
constraint cells synapsing on cell i, Pi the number of positive constraint cells synapsing on cell i,
Ci the number contextual inputs (here, C = 1 throughout), and the function g(z) is an inverse
sigmoid-type non-linearity of the form
g(z) = 1 − [
1
2
(tanh(s(z − o)) + 1)]
(25)
where s is the slope and o the offset. Note that this function takes only values of 0...1. To facilitate
mathematical analysis, we use a linear approximation g(z) = 1 − min(max(sz, 0), 1). When this
function is not in saturation it assumes value sz, where s is the slope with 0 < s < 1.
16
Figure 6: Sudoku, a graph coloring problem, solved by W T Ae. (A) Example "hard" SUD, identi-
cal to the "hard" example used in (Habenschuss et al., 2013). Red values are given. (B) Circuit
implementation of SUD. Each node has 9 possible winners (colors). Row, column and box con-
strains are enforced through negative constraint (NC) cells). The pre-defined (red) winners are
enforced through bias currents to the soma.
(C) Weight matrix of network that implements
SUD. The network consists of 1052 units. Of those, 810 units (81 nodes, 9 excitatory and 1
inhibitory each) implement the nodes (WTAs) and 243 implement the constraints (9 row, 9
(D) Performance of the W T Ae
column and 9 box constraints, one for each color; i.e. 27*9).
(blue) and W T As (gray) network on the sudoku shown in (A). 1000 runs of the same network
with different initial conditions. W T Ae required on average 161 τ to converge, with a maxi-
mal duration of 800τ . (E) Number of violated constraints (errors) decreases exponentially as a
function of simulation time for the simulations shown in (D). (F) Same as in (D), but for sim-
ulations of 50 different sudoku problems of varying difficulty (Project Euler , 2015). For W T Ae
and W T As, average convergence time was 142 τ and 1330τ , respectively. W T Ae parameters:
α = 1.1, β1 = 3, β2 = 0.3, β1D = 3, β2D = 0.3, s = 4, o = 4. All contextual inputs I C = 4, with s.d.
of 1. W T As parameters were identical except α = 1.5, βD1 = 1.5, βD2 = 0.15 (see Fig 4).
17
A)B)C)E)123456789123456789815674239732951486496832715687215943154398672329746851941527368563489127278163594row numbercolumn numberunit nr [from]02004006008001000unit nr [to]02004006008001000InhibitoryExcitatory00.040.08prob correct020040060080000.40.8cdftime [tau]0200400600800050100time [tau]nr of errors0200400600800−10010log nr errorsF)00.10.205001000150000.40.8time [tau]prob correctcdfD)12Node 3,13456789Ibias=112Node 4,13456789Ibias=1NCICICWTAEWTASWTAEWTASWTASWTAE In this extended version of the WTA (W T AE), the inhibitory constraints enter as the argument
of a non-linear function, g. This scales the effect of the two network excitatory sources (the positive
constraint cells pk and the non-specific background excitation I C). Thus, each excitatory unit xi
of the WTAE now receives two kinds of excitatory input: the forward input I bias
, and a constraint
input I cons
is always positive. In contrast, in the
standard WTA this input may also be negative. This has the effect that negative constraints can
never overwrite the forward input.
. Note the critical difference:
in W T AE, I cons
i
i
i
The dynamics of excitatory units in the extended WTA xi are:
τ xi + Gxi = fs(αxi − β1xN − Ti + I bias
i + I cons
i
)
(26)
Note that in Equation 26, the external input Ii no longer appears because it is replaced by the
contextual input I C that is now applied through the non-linearity g(z).
2.11 Enhanced performance of W T Ae networks
We tested the performance of the extended WTA networks on all of the constraint satisfaction
tasks. Unlike the standard networks, the W T Ae networks solve all the problems presented. For
graph coloring, the W T Ae architecture converged significantly faster (Fig 4G) and reduced the
number of errors more rapidly (Fig 4D,F). This difference became more apparent the larger the
problem size. For example, networks with 49 nodes converged on average after 255τ . In contrast,
the same problems required on average 1100τ for W T As (Fig 4G).
Also, for MIS, we found that the W T Ae architecture performed better for the larger problem
sizes, with a speedup of up to 60% (Fig 5E,F).
The greatest advantage of the W T Ae network was for SUD problems. W T Ae solved all Sudoku
problems quickly: convergence was on average 161τ for the hard problem shown in Fig 6A and 142τ
for a set of 50 different sudoku standard problems of varying difficulty (Project Euler , 2015)(Fig
6D-F). By contrast, the W T As network was only able to solve 20% of the instances of the hard
problem (Fig 6D, gray) and 60% of the 50 different problems in (Fig 6F, gray) the same time in
which W T Ae solved 100% of all problems. Note that because the solution times have a heavy tailed
distribution, even for W T Ae a small minority of runs take much longer. For example, whereas
mean convergence is 142τ for the 50 different problems, a few problems required up to 2000τ (Fig
6F).
3 Discussion
Our results indicate that large distributed constraint satisfaction problems can be processed on
a computational substrate composed of many stereotypically connected neuronal WTA modules,
together with a smaller number of more specific 'programming' neurons that interconnect the
WTAs and thereby encode the constraints (or rules) of the particular problem to be solved.
The architecture of these CSP networks is consistent with the strong recurrent excitatory, and
recurrent inhibitory, connection motifs observed in the physiological and anatomical connections
between neurons in superficial layers of mammalian neocortex (Douglas et al., 1991; Binzegger et
al., 2004). Therefore our results are relevant to understanding how these biological circuits might
operate. However, we do not consider that GC4P, MIS, and SUD are per se the problems solved
by the actual neuronal cortical circuits. Instead, we choose these canonical CSP examples because
their properties and applications are well understood in the computational literature, and so they
18
can be used as a basis of comparison for the operation and performance of our WTAs networks,
and then by extrapolation also of neocortical networks.
Graph-coloring CSPs are intriguing computational problems because their structure requires
simultaneous distributed satisfaction of constraints. However, in practice they are solved by se-
quential localized algorithms (Russell & Norvig, 2010; Wu & Hao, 2015). For example, CSPs
can be solved by exhaustive search, in which candidate solutions are systematically generated and
tested for validity. However, this approach does not scale well with problem size (Kumar, 1992).
For our 49 node GC4P and SUD graph, this strategy would require up to approximately 1029 and
1022 configurations to be tested, respectively. Various heuristic algorithms can (but are not guar-
anteed to) improve performance beyond that obtainable by exhaustive search. By contrast, the
neural network we present here solves CSPs efficiently without relying on domain-specific heuris-
tics. However this performance would be a property of the physically realized network, and not of
the algorithmic simulation of the model network that we are obliged to use here. For the moment
our estimates of network performance are in terms of model time steps τ (eg Fig 4F), which stand
proxy for physical performance measurements.
Previous approaches to solving CSPs using artificial neural networks (C. J. Wang & Tsang,
1991; Habenschuss et al., 2013; Jonke, Habenschuss, & Maass, 2016; Hopfield & Tank, 1985;
Mostafa, Muller, & Indiveri, 2013; Mezard & Mora, 2009; McClelland, Mirman, Bolger, & Khaitan,
2014; Rosenfeld et al., 1976; Miller & Zucker, 1999) have relied on the use of saturating neurons to
maintain global stability, and have neglected the important role of instability. Output saturation
is not observed in biological networks, where neurons typically operate at well beneath their
maximum discharge rate. The computational implications of this well-recognized fact have until
recently received little attention. For example, it is now becoming clear that non-saturating 'ReLu'
activation functions are advantageous for deep learning networks (Nair & Hinton, 2010; Maas,
Hannun, & Ng, 2013; LeCun et al., 2015). Here we now show that there are novel principles
of network computation that depend on non-saturating activation.
In this case stability relies
on shared inhibition, which allows transient periods of highly unstable dynamics. This kind of
instability does not exist in networks that utilize saturating neurons (eg (Hopfield & Tank, 1985;
Miller & Zucker, 1999)) where the majority (or even all) neurons are in either positive or negative
saturation. In networks with non-saturating LTNs, the derivative of the activation function (equal
to 1 here) appears in Jef f for all supra-threshold neurons, and hence all currently active neurons
contribute to the expansion or contraction of the network dynamics. Neurons that are in saturation
are of course also active, but because the derivative of their saturated activation function is at
or near zero they contribute little or nothing to the expansion or contraction of the network
dynamics. Analyzing the properties of Jef f is thus a powerful tool to understand the way by
which our networks switch between different states autonomously as driven by their dynamics
during the unstable parts of the dynamics. This tool is analogous to the energy function used in
work pioneered by (Hopfield & Tank, 1985), analysis of which has provided great insight into how
saturating networks compute.
Our approach consists of a set of rules that allows the systematic 'programming' of biologi-
cally plausible networks. Thus, we are able to program the desired computational processes onto
a uniform substrate in a scalable manner (Rutishauser et al., 2015, 2011). This approach has
technological benefits for configuring large scale neuromorphic hardware such as IBM TrueNorth
(Merolla et al., 2014) or the Dynap chip (Qiao et al., 2015), which instantiate a physical network
rather than simulating it as we are doing here. While we deal here only with continously-valued
networks, it has been shown that the types of WTA-networks we use here can also be implemented
using spiking neurons (Neftci et al., 2013; Neftci, Chicca, Indiveri, & Douglas, 2011; W. Wang &
19
Slotine, 2006).
Although we do not claim that our CSP problems are implemented in real cortical networks,
principles such as instability as the driving force of computation (Rutishauser et al., 2015), and
search through forbidden sets, are likely fundamental to spontaneous computation in all types of
LTN networks. In addition, this work provides insight into how analogous hardware should be
engineered. These practical implications are in contrast to more general theoretical frameworks
(e.g. (Heeger, 2017)) that often lack a circuit-level implementation and so cannot make predictions
about the necessary computational roles of cell types such as we do here. Note also that the
computational properties of the networks we describe here are preserved regardless of network
size. This is because all aspects of the network rely on a simple computational motif (that of the
WTA) that can be replicated as many times as needed for a particular problem without having to
make modifications that depend on network size. This scalability is in contrast to other attractor-
based computational approaches, which can be shown to solve small problems but cannot easily
be generalized to larger ones (Afraimovich, Rabinovich, & Varona, 2004).
3.1 Computational properties of network solution of CSP
The fundamental operation of our network involves simultaneous and interactive selection of values
(states) across the WTA modules. The selection process drives the activities of some neurons below
threshold using the signal gain developed by those neurons which receive support from either local
or remote excitatory input. The dynamics of the network forbid partitions of active neurons that
are inconsistent with network stability. These partitions are left exponentially quickly because the
unstably high gain generated by the neurons active in the forbidden partition will increase recurrent
inhibition such that at least one active will be driven beneath its activation threshold (Rutishauser
et al., 2015). As a consequence of this, a new partition of lower divergence is entered. This next
partition may again be forbidden, and so exited; or it might be permitted, and so stable. Exploiting
instability in this manner can be thought of as taking the path of least resistance in state space
(Niemeyer & Slotine, 1997). This exploration of state space continues until a consistent permitted
partition is found (Hahnloser et al., 2000; Rutishauser et al., 2015). In the absence of noise or
other external inputs, any transition between two states results in a reduction of divergence. This
reduction implies that the network cannot return to its previously occupied forbidden state, and
so introduces a form of memory into the network that prevents cycling. In the presence of noise
cycling becomes theoretically possible, but is very unlikely (Rutishauser et al., 2015).
Solving CSPs by sequentially exploring different network subspaces has interesting similarities
with algorithmic linear optimization methods, in particular the simplex and related methods (Miller
& Zucker, 1992, 1999). The critical step in simplex is the Pivot, an algorithmic manipulation that
improves the subset of problem variable(s) to be maximized, while holding all others constant. This
involves a decision, followed by a change of basis, and then maximization along the newly chosen
dimensions. This process is similar to the process whereby the WTA network transitions through
its forbidden sets (a link between neural network operation and simplex similar to that which has
been made through the equivalence between polymatrix games and CSPs in (Miller & Zucker,
1992) for relaxation labeling networks (Rosenfeld et al., 1976) with saturating units). Driven
by the exponentially shrinking volumes implied by negative divergence, the unstable dynamics
rapidly cause a switch to a different state of lower dimensionality of the state space. The direction
in which the expansion proceeds is described by the subset of eigenvectors of the effective Jacobian
Jef f that have positive eigenvalues (Rutishauser et al., 2015). This network step is similar to
maximization of the chosen variable in simplex. Furthermore, whenever the network switches from
20
a forbidden set to another set (which is either forbidden or permitted), the network performs a
'Pivot', changing the basis functions among which the dynamics evolve. In contrast, the mechanism
by which the CSP network implements these search principles is fundamentally different from linear
programming and similar approaches, including the generalization to CSPs based on polymatrix
games based on Lemke's algorithm (Miller & Zucker, 1992). Firstly, our network performs these
steps fully asynchronously and autonomously for every module. Secondly, our network does not
require access to a global cost function, does not require access to the current values of all variables,
and does not depend on an external controller to decide suitable pivots. Instead, our network moves
along the best directions for each WTA, and so the search proceeds for all WTAs in parallel.
The constraint connections affect the mutual selection (value assignment) process of the coupled
WTA variables. These constraints are implemented by directed weighted connections that provide
an immediate and distributed update of all the appropriate values of all affected variables. Within
each subspace, the network behaves as a piecewise linear system that is computationally powerful
in that the partial derivatives of the system update the many interacting variables simultaneously
and consistently as described by the effective Jacobian Jef f .
Importantly, these are updates to possible mixtures of values evolving at each WTA variable,
rather than a replacement of one discrete value by another. But, eventually each WTA variable
will enhance the signal due to one candidate value, while suppressing its competitors. This process
is radically different to the back-tracking / constraint-propagation method implemented by digi-
tal algorithms that procedurally generate and test the consequences of alternative discrete value
assignments to particular variables. Instead, the CSP network approaches its solution by succes-
sive approximation of candidate quality, and so is unlikely to compute towards a false assignment
(unless the CSP has no feasible solution). This optimization process follows from the computa-
tional dynamics of the network: Its computational trajectory follows successively less unfavorable
forbidden subspaces until a permitted subspace is entered.
A further important distinction between algorithmic CS and our network rests in the assignment
of an initial candidate configuration. Algorithmic approaches begin with a candidate configuration
in which every variable is initialized with some legal value. By contrast, the initial CSP network
assignment is effectively null: All xi at all WTA neurons are zero. However, these values are soon
affected by the stochastic context signals I C applied to all xi so that there is almost immediately a
low amplitude mixture of variables across the network xi. The network dynamics then bootstrap
better estimates of these mixtures through the constraints until the network finally converges
towards a complete and consistent assignment. In this way the network offers an novel approach
to CSP that is a dynamic balance between candidate generation and validation through progressive
refinement of a mixture of values at each variable.
3.2 Probabilistic processing
A certain degree of noise is essential for the operation of our networks. Such stochasticity and
bias enters into the CSP network process via inputs from its embedding network. These are the
contextual excitatory inputs IC. For the moment, the IC introduce only randomness and biases
that enable the CSP network to gain access to an otherwise computationally inaccessible solution
subspace, and so provide some degree of innovation in the search process. It is also these inputs that
are modulated by multiplicative inhibition. In a more realistic scenario, I C would be replaced with
input from other parts of the brain to specify priors. This way, the network would be responsive
to constraints set by other parts of the network, such as sensory input or internal states.
We confirmed that the network does find a solution more quickly when biased towards easy
21
Figure 7: Distribution of run times and influence of initial conditions. (A) Probability that a
simulation will find a correct solution after a certain amount of simulation time. N=6000 sim-
ulation runs of random graphs with N=36 nodes, same parameters as in Fig 4H). The majority
of simulations find a solution within 200τ (red line). The data was well fit by the log-normal
(µ = 5.11 − 5.15, σ = 0.72 − 0.75, 95% confidence intervals) and generalized extreme value
(k = 0.50 − 0.56, σ = 74.3 − 78.7, µ = 120.5 − 125.0, 95% confidence intervals) distributions.
Both these distributions are characteristic of heavy-tailed phenomena (Feldman & Taqqu, 1998).
(B) Assessment of fit using a log-log plot. For robustness the y-axis is cumulative rather than log
frequency. The tail of the observed data falls between the two theoretical distributions, indicating
that its tail is heavier than expected by log-normal but less heavy than expected by generalized
extreme value. (C) GC4P for an identical N=25 node graph, but with different initial conditions:
i) random initial conditions (green), ii) partially informative (50% of states are set correctly), and
iii) partially uninformative (50% of states are set incorrectly).
candidates (Fig. 7C). Repetitions of these runs using the same random seed for the initial state
(but not for further processing) confirm that non-deterministic processing nevertheless gives rise
to distinct distributions of solution times for the easier and harder problems (Fig 7C). In the case
of SUD, the prior information is set by the feedforward inputs I bias that implement the required
values of some variables. The larger the number of biases, the harder the problem. But however
hard, the set of problem biases for a given game of Sudoku are a priori known to be compatible with
a solution, and therefore the network can be expected to finally find a complete solution. However,
for more general coloring problems the input biases may only be desirable, and not known to be
compatible with a complete assignment. In such cases the network may find an approximate but
incomplete assignment (Fig 7C).
3.3 Distribution of run times
The network solves the majority of random graphs quickly, with a minority taking much longer
(Fig 7). Thus, the distribution of times required for the WTA networks to solve CSP is heavy-
tailed (Fig 7A,B). The form of the distribution does not depend on the hardness of the problem
(compare Figs 6D and 6F). The heavy-tail persists even if the very same problem is solved multiple
times from the same initial conditions, indicating that probabilistic processing allows the network
to follow different trajectories that may differ substantially in their length (duration) because of
alternative routes through successive forbidden subspaces (Rutishauser et al., 2015). On the other
hand if the network is seeded with initial conditions that favor simple solutions, then the median
processing time is shorter than when the seed is biased towards invalid solutions (Fig 7C. Thus,
the overall distribution of network run times appears to be a composition of the distribution over
22
B)log time [tau]4.55.56.57.5log P(X>x)-8-40Observedlog normalGeneralized extreme valueA)runtime [tau]08001600Prob correct00.0020.004log normalGeneralized extreme value50% valid50% invalid0400800-4-202C)100% randomlog nr errorstime [tau] the hardness problems, as well the distributions over probabilistic trajectories. Such a composition
is a characteristic of heavy-tailed processes (Gomes, Selman, Crato, & Kautz, 2000).
3.4 Computational advantage of multiplicative inhibition
We found that the ability of the network to solve hard problems was improved significantly when
negative constraints were implemented by non-linear multiplicative inhibition (see Fig 4D,F). This
non-linearity improved performance on difficult problems that have both global as well as local
constraints, such as SUD.
Multiple types of inhibition are a prominent feature of nervous systems (Blomfield, 1974-03-29;
Koch, Poggio, & Torre, 1983; Koch, 1998), in particular in neocortex (Isaacson & Scanziani, 2011).
There are both non-linear and linear mechanisms, associated with GABAA chloride-, and GABAB
potassium-mediated inhibition, respectively. Their actions are separable in intracellular recordings
in vivo (Douglas & Martin, 1991; Borg-Graham, Monier, & Fregnac, 1998), but their effects during
processing are mixed (El-Boustani & Sur, 2014; Zhang, Li, Rasch, & Wu, 2013).
Inhibition is mediated by distinct types of neuron encountered in the superficial layers of
neocortex (Rudy, Fishell, Lee, & Hjerling-Leffler, 2011). One large group ( 40%), the basket
(BCs) inhibitory neurons, have horizontally disposed axons that target predominantly the soma
and proximal dendritic segments of pyramidal neurons. The somatic bias of the synapses of these
neurons make them likely candidates for implementing the somatic WTA selective mechanism.
Another large group ( 30%), the bitufted cells or double-bouquet cells (DBCs), have vertically
disposed axons that target predominantly the more distal dendritic segments of pyramidal neurons.
These neurons are candidates for the non-linear NC cells of our model. Although their particular
conductance mechanisms are as yet unknown, non-linear inhibitory effects are considered to play
an important role in the processing of synaptic inputs by dendrites of pyramidal cells (Koch, 1998;
Bar-Ilan, Gidon, & Segev, 2013; Brunel, Hakim, & Richardson, 2014-04; Stuart & Spruston, 2015).
When negative constraints are implemented by direct subtractive inhibition applied to the
somatic compartment, they degrade the local WTA selection process by falsely contributing to
the inhibitory normalization over the WTA xi. We overcame this disadvantage by introducing a
second, dendritic, compartment that receives the positive constraint and contextual input, and
whose output I i
dendrite to the soma is governed by the non-linearity g(z) (Fig 8A). The somatic
compartment receives the standard WTA inputs I bias, I α, and I β1, and the output of the dendritic
compartment. Thus, in addition to local recurrence, each excitatory unit xi of the WTAE receives
two kinds of input: direct somatic input I bias
dendrite. In this configuration,
multiplicative inhibitory constraints quench only the various sources of IC excitation received by
each of the xi, and do not interfere directly with the WTA local decision for the best supported
of the xi. This advantage explains the improved performance of the W T AE networks depicted in
Figs 4D,F,I; 5D,E,F; and 6D,E,F.
and dendritic input I i
i
The g(z) non-linearity provides 'on-path', multiplicative or 'shunting' inhibition previously de-
scribed in biological dendrites (Koch et al., 1983; Gidon & Segev, 2012). This type of inhibition
can veto excitatory input arriving on the the same dendritic branch but has much less influence
on excitatory inputs that arrive on other branches or the soma (Zhang et al., 2013). The g(z) has
two important implications for computation. Firstly, somata that are strongly activated (e.g. by
large I bias) have little or no dendritic sensitivity because the strong feedback inhibition drives g(z)
towards 0 (Fig 8C). They ignore their dendritic excitatory input, thereby reducing the dimension-
ality of the problem. Second, dendritic sensitivity is graded so that when somatic activation varies,
neurons with the low somatic activation are more sensitive to remote excitatory input (Fig 8D).
23
Figure 8: Behavior of dendritic non-linearity that improves performance of network in solving
heavily constrained problems such as sudoku (SUD). (A) Separation of processing into a den-
dritic (top) and somatic (bottom) compartment by non-linearity g(z). Both compartments receive
excitation and inhibition from nearby neurons as well as external inputs. (B) Shape of the non-
linearity g(z), where z is equal to the total inhibitory dendritic input that a dendritic branch
receives. g(z) = 1 − tanh(sz) is plotted for different values of s. The remainder of the fig uses
s = 0.2. (C) Histogram of g(z) values across all dendritic branches in a simulation of sudoku (81
nodes) after a correct solution was found. Note the bimodality (arrows): 64% of all compartments
have g(z) = 0, making them insensitive to dendritic input. This is because their somatic inputs
Ibias are strong. Effectively, this reduces the dimensionality of the problem. (D) Same as (C),
but for a simulation where two different values of Ibias were used (10 and 3). This results in a
tri-modal distribution (arrowheads), with the new mode corresponding to units with non-zero but
weak Ibias. These units thus remain sensitive to dendritic input, but much less so than the units
where Ibias = 0 (right-most mode).
This mechanism provides a weighting of the importance of different dimensions of the problem,
making it easier for solutions with little evidence to switch to alternative solutions, by comparison
with those that have more evidence (ie more somatic activation).
The single dendritic compartment can be generalized to multiple compartments, each governed
by its own nonlinearity, thereby allowing localized inhibitory modulation of specific excitatory
input in the manner of a dendritic tree (Koch, 1998; Tran-Van-Minh et al., 2015). The total
dendritic input I i
dendrite is then the sum of currents provided by all dendritic branches j. There are
∆i branches in total. Each branch receives inhibitory inputs dk from negative constraint cells as
well as two kinds of excitatory inputs: contextual inputs I k
C and inputs from the positive constraint
cells pk. Each branch j receives Ij, Cj, and Pj such inputs, respectively.
∆i(cid:88)
Ij(cid:88)
Cj(cid:88)
Pj(cid:88)
I i
dendrite =
[g(
βD
1 dk)(
I k
C +
γD
1 pk)]
(27)
Future work will explore the potential benefits for network processing of such parallel, or tree-
structured, dendritic structures. For the present we consider only a single dendritic segment.
j
k=1
k=1
k=1
4 Conclusion
We have shown that large distributed constraint satisfaction problems can be processed on a
computational substrate that is composed of stereotypically connected neuronal WTA modules,
and a smaller number of more specific 'programming' neurons which embody the constraints (or
rules) of the particular problem to be solved. The rules of network construction and accompanying
mathematical proofs guarantee that any instance of the three CSP types GC4P, MIS, and SUD
24
dendritic sensitivity g(z)0.51.0total dendritic inhibition z01020slope s=0.1slope s=0.2slope s=0.302000.30.60.9nr of dendritic branches0200β1α1βD1dendriticsomaticγ1ICIbiasA)g(z)B)00dendritic sensitivity g(z)C)D)0.30.60.90dendritic sensitivity g(z)inhibitionexcitationinputinputinhibitionexcitation implemented in the way described will find a solution. Note that the CSPs we considered can be
reduced to graph coloring of planar (GC4P, MIS) and non-planar (SUD) graphs with the number
of colors available given a-priori.
The networks use a combination of unstably high gain and network noise to drive a search
for a consistent assignment values to problem variables. The organization of the network imposes
constraints on the evolving manifold of system dynamics, with the result that the computational
trajectory of the network is steered toward progressive satisfaction of all the problem constraints.
This process takes advantage of the non-saturating nature of the individual neurons, which results
in the effective Jacobian being driven by all neurons that are currently above threshold.
This search performance is greatly improved if the mechanism of value selection at any variable
can reduce its sensitivity to constraints according the confidence of selection. This can be achieved
by using subtractive inhibition for selection, while modulating constraint inputs using multiplica-
tive inhibition. This arrangement allows the constraint satisfaction network to solve more difficult
problems, and to solve all such problems more quickly.
Our findings provide insight into the operation of the neuronal circuits of the neocortex, where
the fundamental patterns of connection amongst superficial neurons is consistent with the WTA
networks described here (Douglas & Martin, 2004; Lee et al., 2016; Rudy et al., 2011; Binzegger
et al., 2004). Our findings are also relevant to the design and construction of hybrid analog digital
neuromorphic processing systems (Liu, Delbruck, Indiveri, Whatley, & Douglas, 2015; Indiveri,
Chicca, & Douglas, 2009; Neftci et al., 2013) because they provide general principles whereby a
physical computational substrate could be engineered and utilized.
5 Methods
5.1 Numerical simulations
All simulations were implemented in MATLAB. Numerical integration of the ODEs is with Euler
integration with δ = 0.01.
The Boost graph library (Siek, Lee, & Lumsdaine, 2002) and its MATLAB interface MatlabBGL
(Gleich, 2009) is used for graph-theoretical algorithms such as confirmation that graphs are planar,
generation of random graphs, Chrobak-Payne Straight Line Drawing etc. The description of the
network is generated automatically based on an XML file that specifies the graph. The XML file
is in JFF format as used by JFLAP (Rodger & Finley, 2006).
5.2 Translating a graph coloring problem into collection of WTAs
The graph is decomposed into fully connected (all-to-all) non-overlapping subgraphs, for each of
which a NC cell ("ring") is added (one for each color, so 4 for each ring). The sum of all βD
1 j
projecting to a particular pyramid cell j is normalized to a constant equal to βD
j . External input
to all pyramid cells (4 for each node) is normally distributed i.i.d. noise (currently µ = 1 and
σ = 0.1, i.e. 10% of the mean).
6 Acknowledgements
We acknowledge the many contributions of Kevan Martin, as well as the inspiration of the Capo
Caccia Workshops on Neuromorphic Engineering. We acknowledge funding by the McKnight
25
Endowment for Neuroscience, the National Science Foundation Award 1554105, and the National
Institute of Health Award R01MH110831.
References
Abbott, L. (1994). Decoding neuronal firing and modelling neural networks. Q. Rev. Biophys.,
27 , 291-331.
Afek, Y., Alon, N., Barad, O., Hornstein, E., Barkai, N., & Bar-Joseph, Z. (2011). A biological
solution to a fundamental distributed computing problem. Science, 331 (6014), 183–185.
Afraimovich, V. S., Rabinovich, M. I., & Varona, P.
ensembles and the winnerless competition principle.
and Chaos, 14 (04), 1195–1208.
(2004). Heteroclinic contours in neural
International Journal of Bifurcation
Appel, K., Haken, W., & Koch, J. (1977). Every planar map is four colorable. Illinois Journal of
Mathematics, 21 , 429-567.
Bar-Ilan, L., Gidon, A., & Segev, I. (2013). The role of dendritic inhibition in shaping the plasticity
of excitatory synapses. Front Neural Circuits, 6 .
Ben-Yishai, R., Bar-Or, R. L., & Sompolinsky, H. (1995). Theory of orientation tuning in visual
cortex. PNAS , 92 (9), 3844–3848.
Binzegger, T., Douglas, R. J., & Martin, K. (2004). A quantitative map of the circuit of cat
primary visual cortex. The Journal of Neuroscience, 24 (39), 8441–8453.
Blomfield, S. (1974-03-29). Arithmetical operations performed by nerve cells. Brain Research,
69 (1), 115–124.
Borg-Graham, L. J., Monier, C., & Fregnac, Y. (1998). Visual input evokes transient and strong
shunting inhibition in visual cortical neurons. Nature, 393 (6683), 369.
Brunel, N., Hakim, V., & Richardson, M. J. (2014-04). Single neuron dynamics and computation.
Current Opinion in Neurobiology, 25 , 149–155.
Carandini, M., & Heeger, D. J. (2012). Normalization as a canonical neural computation. Nat
Rev Neurosci, 13 (1), 51–62.
Dayan, P. (2008). Simple substrates for complex cognition. Frontiers in neuroscience, 2 (2), 255.
Dayan, P., & Abbott, L. (2001). Theoretical neuroscience. Cambridge MA: MIT Press.
Douglas, R. J., Koch, C., Mahowald, M., & Martin, K. (1999). The role of recurrent excitation in
neocortical circuits. In P. Ulinski (Ed.), Models of cortical circuits (pp. 251–282). Springer.
Douglas, R. J., & Martin, K. (1991). A functional microcircuit for cat visual cortex. The Journal
of Physiology, 440 (1), 735–769.
Douglas, R. J., & Martin, K. (2004). Neuronal circuits of the neocortex. Annu. Rev. Neurosci.,
27 , 419–451.
Douglas, R. J., & Martin, K. (2007). Recurrent neuronal circuits in the neocortex. Curr Biol ,
17 (13), R496–R500.
Douglas, R. J., Martin, K., & Whitteridge, D. (1989). A canonical microcircuit for neocortex.
Neural Computation, 1 (4), 480–488.
Douglas, R. J., Martin, K., & Whitteridge, D. (1991). An intracellular analysis of the visual
responses of neurones in cat visual cortex. J Physiol , 440 (1), 659–696.
El-Boustani, S., & Sur, M. (2014). Response-dependent dynamics of cell-specific inhibition in
cortical networks in vivo. Nat Commun, 5 .
Ercsey-Ravasz, M., & Toroczkai, Z. (2012). The chaos within sudoku. Scientific reports, 2 .
26
Ermentrout, B. (1992). Complex dynamics in winner-take-all neural nets with slow inhibition.
Neural Networks, 5 (3), 415-431.
Feldman, R., & Taqqu, M. (1998). A practical guide to heavy tails: statistical techniques and
applications. Springer Science & Business Media.
Ferster, D., Chung, S., & Wheat, H. S. (1996). Orientation selectivity of thalamic inputs to cat
visual cortex. Nature, 380 , 249–252.
Gidon, A., & Segev, I. (2012). Principles governing the operation of synaptic inhibition in dendrites.
Neuron, 75 (2), 330–341.
Gleich, D. F. (2009). Models and algorithms for pagerank sensitivity (Unpublished doctoral dis-
sertation). Stanford University. (Chapter 7 on MatlabBGL)
Gomes, C. P., Selman, B., Crato, N., & Kautz, H. (2000). Heavy-tailed phenomena in satisfiability
and constraint satisfaction problems. Journal of automated reasoning, 24 (1-2), 67–100.
Habenschuss, S., Jonke, Z., & Maass, W. (2013). Stochastic computations in cortical microcircuit
models. PLoS Comput Biol , 9 , e1003311.
Hahnloser, R., Douglas, R. J., Mahowald, M., & Hepp, K. (1999). Feedback interactions between
neuronal pointers and maps for attentional processing. Nat Neurosci, 2 (8), 746-52.
Hahnloser, R., Sarpeshkar, R., Mahowald, M., Douglas, R. J., & Seung, H. S. (2000). Digital
selection and analogue amplification coexist in a cortex-inspired silicon circuit. Nature,
405 (6789), 947-51.
Heeger, D. J. (2017). Theory of cortical function. Proceedings of the National Academy of Sciences,
201619788.
Hopfield, J. (1982). Neural networks and physical systems with emergent collective computational
abilities. Proceedings of the National Academy of Sciences of the United States of America,
79 , 2554-2558.
Hopfield, J., & Tank, D. W. (1985). Neural computation of decisions in optimization problems.
Biological cybernetics, 52 (3), 141–152.
Horn, R. (1985). Matrix analysis. Cambridge University Press.
Indiveri, G., Chicca, E., & Douglas, R. J. (2009). Artificial cognitive systems: From vlsi networks
of spiking neurons to neuromorphic cognition. Cognitive Computation, 1 (2), 119-127.
Isaacson, J., & Scanziani, M. (2011). How inhibition shapes cortical activity. Neuron, 72 (2),
231–243.
Jiang, X., Wang, G., Lee, A. J., Stornetta, R. L., & Zhu, J. J. (2013). The organization of two
new cortical interneuronal circuits. Nature neuroscience, 16 (2), 210–218.
Jonke, Z., Habenschuss, S., & Maass, W. (2016). Solving constraint satisfaction problems with
networks of spiking neurons. Front. Neurosci., 30 March.
Kami´nski, J., Sullivan, S., Chung, J. M., Ross, I. B., Mamelak, A. N., & Rutishauser, U. (2017).
Persistently active neurons in human medial frontal and medial temporal lobe support work-
ing memory. Nature Neuroscience.
Karp, R. M. (1972). Reducibility among combinatorial problems. In R. E. Miller & J. W. Thatcher
(Eds.), Complexity of computer computations (pp. 85–103). New York, USA: Plenum Press.
Koch, C. (1998). Biophysics of Computation: Information Processing in Single Neurons (Compu-
tational Neuroscience). Oxford University Press.
Koch, C., Poggio, T., & Torre, V. (1983). Nonlinear interactions in a dendritic tree: localization,
timing, and role in information processing. Proceedings of the National Academy of Sciences,
80 (9), 2799–2802.
Koechlin, E., & Summerfield, C. (2007). An information theoretical approach to prefrontal exec-
utive function. Trends in cognitive sciences, 11 (6), 229–235.
27
Kumar, V. (1992). Algorithms for constraint-satisfaction problems: A survey. AI magazine, 13 (1),
32.
LeCun, Y., Bengio, Y., & Hinton, G. (2015). Deep learning. Nature, 521 (7553), 436–444.
Lee, W.-C. A., Bonin, V., Reed, M., Graham, B. J., Hood, G., Glattfelder, K., & Reid, R. C. (2016).
Anatomy and function of an excitatory network in the visual cortex. Nature, 532 (7599), 370–
374.
Li, Y.-t., Ibrahim, L. A., Liu, B.-h., Zhang, L. I., & Tao, H. W. (2013). Linear transformation of
thalamocortical input by intracortical excitation. Nat Neurosci, 16 (9), 1324–1330.
Lien, A. D., & Scanziani, M. (2013). Tuned thalamic excitation is amplified by visual cortical
circuits. Nat Neurosci, 16 (9), 1315–1323.
Liu, S.-C., Delbruck, T., Indiveri, G., Whatley, A., & Douglas, R. J. (2015). Event-based neuro-
morphic systems. John Wiley & Sons, Ltd.
Lohmiller, W., & Slotine, J. (2000). Nonlinear process control using contraction theory. AIChE
Journal , 46 (3), 588-596.
Lynch, N. A. (1996). Distributed algorithms. Morgan Kaufmann.
Maas, A. L., Hannun, A. Y., & Ng, A. Y. (2013). Rectifier nonlinearities improve neural network
acoustic models. In Proc. icml (Vol. 30).
Maass, W. (2000). On the computational power of winner-take-all. Neural Computation, 12 ,
2519-2536.
McClelland, J. L., Mirman, D., Bolger, D. J., & Khaitan, P. (2014). Interactive activation and
mutual constraint satisfaction in perception and cognition. Cognitive science, 38 (6), 1139–
1189.
Merolla, P. A., Arthur, J. V., Alvarez-Icaza, R., Cassidy, A. S., Sawada, J., Akopyan, F., . . . others
(2014). A million spiking-neuron integrated circuit with a scalable communication network
and interface. Science, 345 (6197), 668–673.
Mezard, M., & Mora, T. (2009). Constraint satisfaction problems and neural networks: A statistical
physics perspective. Journal of Physiology-Paris, 103 (1), 107–113.
Miller, D. A., & Zucker, S. W. (1992). Efficient simplex-like methods for equilibria of nonsymmetric
analog networks. Neural Computation, 4 (2), 167–190.
Miller, D. A., & Zucker, S. W. (1999). Computing with self-excitatory cliques: a model and an
application to hyperacuity-scale computation in visual cortex. Neural Computation, 11 (1),
21–66.
Mostafa, H., Muller, L. K., & Indiveri, G. (2013). Recurrent networks of coupled winner-take-all
oscillators for solving constraint satisfaction problems. In Advances in neural information
processing systems (NIPS) (Vol. 26, pp. 719–727).
Nair, V., & Hinton, G. E. (2010). Rectified linear units improve restricted boltzmann machines.
In Proceedings of the 27th international conference on machine learning (icml-10) (pp. 807–
814).
Neftci, E., Binas, J., Rutishauser, U., Chicca, E., Indiveri, G., & Douglas, R. J. (2013). Synthe-
sizing cognition in neuromorphic electronic systems. Proceedings of the National Academy of
Sciences, 110 (37), E3468–E3476.
Neftci, E., Chicca, E., Indiveri, G., & Douglas, R. J. (2011). A systematic method for configuring
VLSI networks of spiking neurons. Neural Computation, 23 (10), 2457–2497.
Niemeyer, G., & Slotine, J. (1997). A simple strategy for opening an unknown door. In Robotics
and automation, 1997. proceedings., 1997 ieee international conference on (Vol. 2, pp. 1448–
1453).
28
Pi, H.-J., Hangya, B., Kvitsiani, D., Sanders, J. I., Huang, Z. J., & Kepecs, A. (2013). Cortical
interneurons that specialize in disinhibitory control. Nature, 503 (7477), 521–524.
Project euler. (2015). Retrieved 2015-08-27, from https://projecteuler.net/problem=96
Qiao, N., Mostafa, H., Corradi, F., Osswald, M., Stefanini, F., Sumislawska, D., & Indiveri, G.
(2015). A reconfigurable on-line learning spiking neuromorphic processor comprising 256
neurons and 128k synapses. Frontiers in neuroscience, 9 .
Robertson, N., Sanders, D. P., Seymour, P., & Thomas, R. (1996). Efficiently four-coloring planar
graphs. In Proceedings of the twenty-eighth annual acm symposium on theory of computing
(pp. 571–575).
Rodger, S., & Finley, T. (2006). Jflap - an interactive formal languages and automata package.
Sudbury, MA: Jones and Bartlett. (ISBN 0763738344)
Rosenfeld, A., Hummel, R. A., & Zucker, S. W. (1976). Scene labeling by relaxation operations.
IEEE Transactions on Systems, Man, and Cybernetics, 6 (6), 420–433.
Rosenhouse, J., & Taalman, L. (2011). Taking sudoku seriously: The math behind the worlds most
popular pencil puzzle. Oxford University Press.
Rudy, B., Fishell, G., Lee, S., & Hjerling-Leffler, J. (2011). Three groups of interneurons account
for nearly 100% of neocortical GABAergic neurons. Devel Neurobio, 71 (1), 45–61.
Russell, S. J., & Norvig, P.
(2010). Artificial intelligence: a modern approach (international
edition) (3rd ed.). Prentice Hall.
Rutishauser, U., & Douglas, R. J. (2009). State-dependent computation using coupled recurrent
networks. Neural Computation, 21 (2).
Rutishauser, U., Douglas, R. J., & Slotine, J. (2011). Collective stability of networks of winner-
take-all circuits. Neural computation, 23 (3).
Rutishauser, U., Slotine, J., & Douglas, R. J. (2012). Competition through selective inhibitory
synchrony. Neural Computation, 24 (8), 2033–2052.
Rutishauser, U., Slotine, J., & Douglas, R. J.
(2015). Computation in dynamically bounded
asymmetric systems. PLoS Comput Biol , 11 (1).
Siek, J. G., Lee, L.-Q., & Lumsdaine, A. (2002). The boost graph library: user guide and reference
manual. Boston, MA, USA: Addison-Wesley Longman Publishing Co., Inc.
Slotine, J. (2003). Modular stability tools for distributed computation and control. International
Journal of Adaptive Control and Signal Processing, 17 , 397-416.
Stuart, G. J., & Spruston, N. (2015). Dendritic integration: 60 years of progress. Nature Neuro-
science, 18 (12), 1713–1721.
Tran-Van-Minh, A., Caz´e, R. D., Abrahamsson, T., Cathala, L., Gutkin, B. S., & DiGregorio,
D. A. (2015). Contribution of sublinear and supralinear dendritic integration to neuronal
computations. Front Cell Neurosci, 9 .
Wang, C. J., & Tsang, E. P. (1991). Solving constraint satisfaction problems using neural networks.
In Artificial neural networks, 1991., second international conference on (pp. 295–299).
Wang, W., & Slotine, J.
69 (16), 2320–2326.
(2006). Fast computation with neural oscillators. Neurocomputing,
Wersing, H., Steil, J. J., & Ritter, H. (2001). A competitive-layer model for feature binding and
sensory segmentation. Neural Computation, 13 (2), 357–387.
Wu, Q., & Hao, J.-K. (2015). A review on algorithms for maximum clique problems. European
Journal of Operational Research, 242 (3), 693–709.
Yuille, A., & Geiger, D. (2003). Winner-take-all networks. In M. Arbib (Ed.), The handbook of
brain theory and neural networks (p. 1228-1231). MIT Press.
29
Zhang, D., Li, Y., Rasch, M. J., & Wu, S. (2013). Nonlinear multiplicative dendritic integration
in neuron and network models. Front Comput Neurosci, 7 .
30
|
1004.3598 | 2 | 1004 | 2010-09-23T18:34:15 | An explanation of the distribution of inter-seizure intervals | [
"q-bio.NC",
"math.PR",
"physics.bio-ph"
] | Recently Osorio et al (Eur. J. Neurosci., 30 (2009) 1554) reported that probability distribution of intervals between successive epileptic seizures follows a power law with exponent 1.5. We theoretically explain this finding by modeling epileptic activity as a branching process, which we in turn approximate by a random walk. We confirm the theoretical conclusion by numerical simulation. | q-bio.NC | q-bio | critical branching process and a power law
distribution of avalanches. We have many
neurons that possibly can be induced to fire by
firing of our neuron and probability for each of
them to fire is very small. Thus, the number of
induced firings is Poisson distributed with mean
λ (which is very close to unity). The variance
of the Poisson distribution equals its mean and
therefore it also equals λ. If at given time step
a large number N of neurons are firing then the
number of neurons firing the next time step will
come from a normal distribution with mean
Nλ and variance Nλ . In addition to induced
firings, some neurons will fire spontaneously.
We assume that the number of spontaneously
firing neurons at each time step comes from a
Poisson distribution with mean p. The change
in the number of firing neurons is
An explanation of the distribution of inter-seizure intervals
M.V. Simkin and V.P. Roychowdhury
Department of Electrical Engineering, University of California, Los Angeles, CA 90095-1594
Recently Osorio et al [1] reported that probability distribution of intervals between successive epileptic
seizures follows a power law with exponent 1.5. We theoretically explain this finding by modeling
epileptic activity as a branching process, which we in turn approximate by a random walk. We confirm
the theoretical conclusion by numerical simulation.
Recently Osorio et al [1]
that
reported
probability distribution of epileptic seizure
energies and inter-seizure intervals follow a
power
law with exponents 1.67 and 1.5
respectively. Earlier Beggs and Plenz [2] had
observed spontaneous neuronal avalanches in
neocortical tissues. The distribution of sizes of
these avalanches, computed by summing local
field potentials, followed a power law with
exponent 1.5 as in critical branching process.
The relevance of the branching process to
explanation of seizure energies distribution is
thus obvious. Here we show that in addition the
distribution of inter-seizure intervals is also
consistent with the branching process.
Epileptic seizures result from simultaneous
firing of large number of neurons in the brain.
Anninos and Cyrulnik [3] proposed a neural net
model for epilepsy a simplified version of
which we will use in this study. To start with,
we introduce time discretization. After a neuron
has fired, it cannot fire again for a time interval
known as refractory period. Therefore, the
minimum interval between the beginnings of
two subsequent firings of a neuron is the sum of
spike duration and refractory period. This
interval is few milliseconds [4]. We will use
this interval as our time unit. Suppose that at
given time step N neurons are firing. How many
neurons will be firing next time step? Consider
one of these firing neurons. Its axon connects to
synapses of thousands of neurons. Some of
them are almost ready to fire: their membrane
potential is close to the firing threshold and one
impulse from our neuron will be sufficient to
surpass this threshold. The aforementioned
experiment of Beggs and Plenz [2] suggest that
the average number λof ready to fire neurons
among those to which our neuron is connected
is very close to unity. Only in this case we get a
In the limit of large x, the term, inversely
proportional to x, can be neglected in the first
approximation. When λ is very close to unity
the first term can also be neglected. Equation
21=∆
which means that
(2) reduces to
x
z
where z is a normally distributed random
number with zero mean and unit variance. The
number of firing neurons, N, performs a random
walk, with the size of the step proportional
N performs a simple random walk. A well
know result in random walk theory is that the
to N . Eq.(1) can be simplified by changing
variable from N to
[5], we get
. Using Ito’s formula
(
)
1λ
−
++
pN
1
+×
x
(
λ
−
)
1
1
8
zN
=∆
N
x =
N
x
−
−
z
(2)
(1)
=∆
x
2
p
2
1
2
distribution of first return times to zero (or to
any other chosen point) follows a power law
with exponent 3/21. In the experiment [1],
seizures were counted when electric intensity of
epileptic discharges reached certain threshold.
This in our model is equivalent to N reaching
certain threshold. Then the distribution of first
returns into seizure (inter-seizure intervals) is
the same as the distribution of random walk’s
return times.
Now
let us study Equation (2) without
neglecting any terms. In the particular case
41=p
the second term cancels out and in the
1<λ we get a well studied problem of
case
Brownian motion in a harmonic potential [7]. In
the general p case we get Brownian motion in
the potential
( )
xU
−=
(
λ
−
)
1
4
2
x
+
1
8
−
p
2
×
( )x
ln
.
(3)
One can find probability density of x, by
solving
the
corresponding Fokker-Planck
equation. Alternatively it can be found as a
Boltzmann distribution in the potential given by
Eq.(3) at an appropriate temperature. The result
is:
( )
xP
~
exp
(
−
( )
)
8
xU
=
exp
(
(
λ
2
−
)
1
x
2
)
1
−
41
p
x
(4)
The probability distribution of
immediately found using Eq.(4)
2xN =
can be
(
NP
)
~
exp
(
2
(
−λ
)
1
N
)
1
−
21
pN
(5)
0=p
and
1=λ we get
In the case when
(
)
. This one expects from the theory
1~
N
NP
of branching processes [8]. The survival
probability after N generations for a critical
branching process is N1
, while the expectation
number for the number of individuals is 1. This
means that the average size of surviving family
is N, while the probability is N1
. In the
0>p
case
, branching processes overlap and
this leads to a modified power law exponent.
5.0=p
When
the power law cancels out and
5.0>p
when
the power law exponent becomes
1 One can get this by applying Stirling’s formula to the
result in chapter III.4 of Ref. [6]
positive. Experimentalists did not report the
data which can be compared to Eq.(5) but it is
most likely contained in their data files and they
will be able to compare it with our theory. An
important implication of Eq.(5) is that the
random walk spends less time at higher values
of N. This means that seizures are shorter than
intervals between seizures.
One way to get a critical (or more precisely
slightly subcritical) branching process would be
to use the Self-organized criticality model [9].
The Sandpile model can be easily recast in
neural network terms: accumulation of send
grains corresponds to integration, toppling to
firing, and spontaneous firing to adding sand
grains. In practice it is easier to simulate a
branching process than the generating it SOC
system. Figures 1-3 show results of such
parameters
The
simulation.
were
−
810
−=λ
5.0=p
and
. The seizure threshold
1
810
4 ×
was set at
firing neurons. The seizure
intensities were defined as total number of
neuron firings during seizure. The system was
1110 time steps. Remember that
simulated for
time step is the sum of firing duration and
refractory period. A reasonable estimate for this
is two milliseconds. Thus, our simulation run
corresponds to over six years. The longest inter-
1010
5 ×
time steps (over
seizure interval was
710 time
three years). The longest seizure was
steps (about five hours). The distribution of
inter-seizure intervals is well described by an
inverse power law with exponent 1.47. And the
distribution of seizure intensities - by an inverse
power law with exponent 1.48.
This research gives an answer to the old
question “Do seizures beget seizures?” [10].
After symptoms of seizure end, there still
remains for some time epileptic activity in the
brain. It is easy for this activity to surpass the
threshold again soon. If there was no seizure
recently
this most
likely means
that
the
epileptic activity is minimal or absent and it
will take more time to build up the activity to
pass the threshold. This model also gives an
alternative
explanation
to
power
law
distribution of intervals in other than epileptic
human activity [11].
References
1. Osorio I., Frei M. G., Sornette D. and
Milton J. “Pharmaco-resistant seizures:
self-triggering capacity, scale-free
properties and predictability?” European
Journal of Neuroscience, 30 (2009)
1554.
2. Beggs, J.M. and Plenz, D. Neuronal
avalanches in neocortical circuits. J.
Neurosci., 23 (2003) 11167.
3. Anninos P. A. and Cyrulnik R. “A
Neural Net Model for Epilepsy” J.
theor. Biol. 66 (1977) 695.
4. Arbib M. A. (editor) “Handbook of
Brain Theory and Neural Networks”
(MIT press, Cambridge) 2003.
5. See, for example, J.C. Hull “Options,
futures, and other derivatives” (Prentice
Hall, Upper Saddle River) 1997.
6. Feller W. “An introduction to
probability theory and its applications”
Volume 1 (John Wiley, New York)
1957.
7. Uhlenbeck G. E. and Ornstein L. S.,
“On the Theory of the Brownian
Motion” Phys. Rev. 36 (1930) 823.
8. Harris T. E., “The Theory of Branching
Processes” (Springer, Berlin) 1963.
9. P. Bak “How nature works: the science
of self-organized criticality”
(Copernicus, New York) 1999.
10. Hauser W.A. and Lee J.R. “Do seizures
beget seizures?” Prog. Brain Res., 135
(2002) 215.
11. A.-L. Barabasi “The origin of bursts and
heavy tails in human dynamics” Nature,
435 (2005) 207.
)
s
p
e
t
s
e
m
i
t
n
i
(
h
t
g
n
e
l
1.E+11
1.E+10
1.E+09
1.E+08
1.E+07
1.E+06
1.E+05
1.E+04
1.E+03
1.E+02
1.E+01
1.E+00
s e izu re dura tion
inte rs e izu re in terva l
1
10
100
1000
10000
100000
−
810
−=λ
5.0=p
1
and
. The
Figure 1. Results of numerical simulation of branching epileptic process model with
1110 time steps, which correspond to over six years. The longest inter-seizure interval is over three
simulation was run for
years. The longest seizure is about five hours.
rank
1.E+00
1.E+01
1.E+02
1 .E+03
1.E+04
1.E+05
1.E+06
1 .E+07
1.E+08
1.E+09
1.E+10
interseizure interval (in t ime s teps )
1
0 .1
F
D
C
-
1
0.01
0.001
0.0001
0 .00001
y = 0.884x-0 .4714
Figure 2. Cumulative distribution function (CDF) of inter-seizure intervals. 1 – CDF is fitted by a power law with
47.0−
. This means that probability density function of inter-seizure intervals is a power law with exponent -
exponent
47.1−
.
seizure intens ity (total number of neuron firings )
1.E+10
1.E+11
1.E+12
1.E+13
1.E+14
1.E+15
1.E+16
y = 12209x-0 .4811
1.E+09
1
F
D
C
-
1
0.1
0.01
0.001
0.0001
0.00001
Figure 3. Cumulative distribution function (CDF) of seizure intensities. 1 – CDF is fitted by a power law with exponent
48.1−
48.0−
.
. This means that probability density function of seizure intensities is a power law with exponent
|
1905.12100 | 1 | 1905 | 2019-05-28T21:32:26 | Using local plasticity rules to train recurrent neural networks | [
"q-bio.NC",
"cs.AI"
] | To learn useful dynamics on long time scales, neurons must use plasticity rules that account for long-term, circuit-wide effects of synaptic changes. In other words, neural circuits must solve a credit assignment problem to appropriately assign responsibility for global network behavior to individual circuit components. Furthermore, biological constraints demand that plasticity rules are spatially and temporally local; that is, synaptic changes can depend only on variables accessible to the pre- and postsynaptic neurons. While artificial intelligence offers a computational solution for credit assignment, namely backpropagation through time (BPTT), this solution is wildly biologically implausible. It requires both nonlocal computations and unlimited memory capacity, as any synaptic change is a complicated function of the entire history of network activity. Similar nonlocality issues plague other approaches such as FORCE (Sussillo et al. 2009). Overall, we are still missing a model for learning in recurrent circuits that both works computationally and uses only local updates. Leveraging recent advances in machine learning on approximating gradients for BPTT, we derive biologically plausible plasticity rules that enable recurrent networks to accurately learn long-term dependencies in sequential data. The solution takes the form of neurons with segregated voltage compartments, with several synaptic sub-populations that have different functional properties. The network operates in distinct phases during which each synaptic sub-population is updated by its own local plasticity rule. Our results provide new insights into the potential roles of segregated dendritic compartments, branch-specific inhibition, and global circuit phases in learning. | q-bio.NC | q-bio | Using local plasticity rules to train recurrent neural networks
Owen Marschall, Kyunghyun Cho, and Cristina Savin
Summary: To learn useful dynamics on long time scales, neurons must use plasticity rules that account for long-
term, circuit-wide effects of synaptic changes. In other words, neural circuits must solve a credit assignment
problem to appropriately assign responsibility for global network behavior to individual circuit components. Fur-
thermore, biological constraints demand that plasticity rules are spatially and temporally local; that is, synaptic
changes can depend only on variables accessible to the pre- and postsynaptic neurons. While artificial intelligence
offers a computational solution for credit assignment, namely backpropagation through time (BPTT), this solu-
tion is wildly biologically implausible. It requires both nonlocal computations and unlimited memory capacity,
as any synaptic change is a complicated function of the entire history of network activity. Similar nonlocality
issues plague other approaches such as FORCE [1]. Overall, we are still missing a model for learning in recurrent
circuits that both works computationally and uses only local updates. Leveraging recent advances in machine
learning on approximating gradients for BPTT, we derive biologically plausible plasticity rules that enable re-
current networks to accurately learn long-term dependencies in sequential data. The solution takes the form of
neurons with segregated voltage compartments, with several synaptic sub-populations that have different func-
tional properties. The network operates in distinct phases during which each synaptic sub-population is updated
by its own local plasticity rule. Our results provide new insights into the potential roles of segregated dendritic
compartments, branch-specific inhibition, and global circuit phases in learning.
9
1
0
2
y
a
M
8
2
]
.
C
N
o
i
b
-
q
[
1
v
0
0
1
2
1
.
5
0
9
1
:
v
i
X
r
a
Figure 1: (A) The network architecture consists of just one hidden layer, receiving inputs from an input population
x and projecting to output units y. (B): Basal phase of network activity, during which φ = tanh determines the
firing rate ri from the somatic voltage V S
i . The vectors rS and rA are concatenations of r with the inputs x and
labels y, respectively, with a constant 1 to provide a bias term. (C): Distal phase of network activity, during which
(cid:1). (D): Temporally filtered losses for each set of synapses. Test accuracy
neural firing ri is determined by φ(cid:0)V A
(green) was computed every 500 time steps. Light traces are individual trials, while dark traces are trial-averages.
i
We chose the (4, 6)-back task to quantify learning [2] because it has low-dimensional inputs and outputs, multi-
ple time scales of relevant information, and clear bounds for performance that correspond to learning particular
input-output dependencies. In more detail, the network has to map an i.i.d. temporal sequence of Bernoulli inputs
with px = 0.5 to a Bernoulli output, whose probability depends on the inputs with some lag that can be adjusted
to tune the task difficulty, here lags of 4 and 6 time steps. In particular, the baseline output probability py = 0.5 is
increased (decreased) by 0.5 (0.25) when the input from 4 (6) time steps back is equal to 1.
maxoutput stats.4-back6-backaccuracyW lossJ lossA losstimelearning performance0200040006000800010000120001.00.00.80.60.40.2Doutputsrecurrent networkABCnetwork phase Inetwork phase IIinputs Our model consists of an input layer, a recurrent network, and an output layer (Fig.1A). We define plasticity rules
with the aim of minimizing a loss function that quantifies task performance as the cross entropy between the net-
work outputs and the target distribution. Plasticity rules can be derived by performing stochastic gradient descent
on this loss function, but calculation of the exact gradient requires computations that are nonlocal over space and
time. Instead, we exploit a novel machine learning technique, known as "synthetic gradients," to approximate the
gradient using local computations [3]. Biologically, this approximation manifests as a network of neurons with
multiple compartments that are innervated by functionally distinct sets of synapses W, A and J, one somatic
and two distal. The W synapses are used for solving the actual task, processing inputs and running the primary
network dynamics, whereas the A and J synapses are used for learning. All of these synapses are plastic.
i
(cid:16)
r(t+1)
j
i
i and V A
First, the plasticity rule for a synapse Wij has a very simple form, depending only on presynaptic activity and
the somatic and distal postsynaptic voltages V S
(Fig.1B). Biologically, this corresponds to distal mod-
ulation of synaptic plasticity [4]. Second, the plasticity rule for a synapse Aij depends on presynaptic activity
and postsynaptic distal voltage, gated by the inputs from the J synapses and a top-down error signal δ passed
through fixed feedback synapses WFB (Fig.1C). Lastly, the plasticity rule for Jij follows perceptron-like learning
∆Jij ∝ r(t)
to implement one-step prediction of recurrent dynamics. J is meant to approxi-
mate the Jacobian of the recurrent dynamics ∂r(t+1)/∂r(t). Since all the required signals cannot be simultaneously
represented at the level of voltages, we require two distinct phases for the network dynamics: a "somatic" phase
for the updates of W and J, and a "distal" phase for the updates of A. Biologically, this could be mediated by
targeted inhibition that dynamically gates out unwanted inputs [5].
−(cid:80)
k Jjkr(t)
k
(cid:17)
Each plasticity rule minimizes an implicit loss function w.r.t. its synaptic sub-population: W updates to improve
task performance, A to improve the approximation of credit assignment, and J to approximate the Jacobian.
Fig.1D shows the evolution of the corresponding losses over learning. As expected, the losses decrease and sta-
bilize. Interestingly, the saturation happens fastest for J, driving learning in A, which in turn drives learning in
W, predicting possible differences in timescales of plasticity at distal vs. basal synapses. What does the network
learn? Performance-wise, the network produces the correct output 75% of the time, which is the theoretical bound
given inherent randomness in the task. The 3 blue dashed lines represent cross-entropy bounds for "internalizing"
the different dependencies between the inputs and outputs. The upper-most dashed line represents learning of
the marginal output statistics, i.e. that 62.5% of outputs are 1, while the second and third dashed lines represent
learning of 4- and 6-time-step lags, respectively. On average over many random seeds, our model reliably learns
the 4-time-step lag and is sufficiently close to the next bound to indicate some knowledge of the 6-back compo-
nent. Failing to learn the second dependency is not entirely surprising, because optimal performance w.r.t. cross
entropy requires a perfect calibration of confidence at each time step. Moreover, vanilla recurrent networks are
known to struggle with long-term dependencies even with full BPTT [6]. The fact that our local approximation
of a much more complicated algorithm can learn long-term dependencies at all is exciting.
In summary, we have designed a network model that can learn long-term dependencies using biologically plau-
sible, local learning rules. The required biological features for calculating credit assignment include multi-
compartment neurons, distinct phases for circuit dynamics, and spatial clustering of synapses with similar func-
tion. While functional roles of different compartments and their distinct plasticity properties have received exper-
imental attention, there is relatively little theoretical work on their computational significance [7]. Our work is an
important step in this direction.
References
[1] D. Sussillo and L. F. Abbott, "Generating coherent patterns of activity from chaotic neural networks," Neuron, vol. 63, no. 4,
pp. 544 -- 557, 2009.
[2] S. Pitis, "Recurrent neural networks in tensorflow i." https://r2rt.com/recurrent-neural-networks-in-tensorflow-i.
html, 2016.
[3] M. Jaderberg, W. M. Czarnecki, S. Osindero, O. Vinyals, A. Graves, D. Silver, and K. Kavukcuoglu, "Decoupled neural interfaces
using synthetic gradients," in Proceedings of the 34th International Conference on Machine Learning-Volume 70, pp. 1627 -- 1635,
JMLR. org, 2017.
[4] J. T. Dudman, D. Tsay, and S. A. Siegelbaum, "A role for synaptic inputs at distal dendrites: instructive signals for hippocampal
long-term plasticity," Neuron, vol. 56, no. 5, pp. 866 -- 879, 2007.
[5] P. Somogyi, L. Katona, T. Klausberger, B. Laszt´oczi, and T. J. Viney, "Temporal redistribution of inhibition over neuronal subcellular
domains underlies state-dependent rhythmic change of excitability in the hippocampus," Philosophical Transactions of the Royal
Society B: Biological Sciences, vol. 369, no. 1635, p. 20120518, 2014.
[6] R. Pascanu, T. Mikolov, and Y. Bengio, "On the difficulty of training recurrent neural networks," in International conference on
machine learning, pp. 1310 -- 1318, 2013.
[7] J. Guerguiev, T. P. Lillicrap, and B. A. Richards, "Towards deep learning with segregated dendrites," ELife, vol. 6, p. e22901, 2017.
|
1907.09270 | 1 | 1907 | 2019-07-22T12:28:45 | Spike-timing-dependent plasticity with axonal delay tunes networks of Izhikevich neurons to the edge of synchronization transition with scale-free avalanches | [
"q-bio.NC",
"nlin.AO"
] | Critical brain hypothesis has been intensively studied both in experimental and theoretical neuroscience over the past two decades. However, some important questions still remain: (i) What is the critical point the brain operates at? (ii) What is the regulatory mechanism that brings about and maintains such a critical state? (iii) The critical state is characterized by scale-invariant behavior which is seemingly at odds with definitive brain oscillations? In this work we consider a biologically motivated model of Izhikevich neuronal network with chemical synapses interacting via spike-timingdependent plasticity (STDP) as well as axonal time delay. Under generic and physiologically relevant conditions we show that the system is organized and maintained around a synchronization transition point as opposed to an activity transition point associated with an absorbing state phase transition. However, such a state exhibits experimentally relevant signs of critical dynamics including scale-free avalanches with finite-size scaling as well as branching ratios. While the system displays stochastic oscillations with highly correlated fluctuations, it also displays dominant frequency modes seen as sharp peaks in the power spectrum. The role of STDP as well as time delay is crucial in achieving and maintaining such critical dynamics, while the role of inhibition is not as crucial. In this way we provide definitive answers to all three questions posed above. We also show that one can achieve supercritical or subcritical dynamics if one changes the average time delay associated with axonal conduction. | q-bio.NC | q-bio |
Spike-timing-dependent plasticity with axonal delay tunes networks of Izhikevich
neurons to the edge of synchronization transition with scale-free avalanches
Department of Physics, College of Sciences, Shiraz University, Shiraz 71946-84795, Iran
Mahsa Khoshkhou and Afshin Montakhab∗
(Dated: July 23, 2019)
Critical brain hypothesis has been intensively studied both in experimental and theoretical neuro-
science over the past two decades. However, some important questions still remain: (i) What is the
critical point the brain operates at? (ii) What is the regulatory mechanism that brings about and
maintains such a critical state? (iii) The critical state is characterized by scale-invariant behavior
which is seemingly at odds with definitive brain oscillations? In this work we consider a biologically
motivated model of Izhikevich neuronal network with chemical synapses interacting via spike-timing-
dependent plasticity (STDP) as well as axonal time delay. Under generic and physiologically relevant
conditions we show that the system is organized and maintained around a synchronization transition
point as opposed to an activity transition point associated with an absorbing state phase transition.
However, such a state exhibits experimentally relevant signs of critical dynamics including scale-free
avalanches with finite-size scaling as well as branching ratios. While the system displays stochastic
oscillations with highly correlated fluctuations, it also displays dominant frequency modes seen as
sharp peaks in the power spectrum. The role of STDP as well as time delay is crucial in achieving
and maintaining such critical dynamics, while the role of inhibition is not as crucial. In this way we
provide definitive answers to all three questions posed above. We also show that one can achieve
supercritical or subcritical dynamics if one changes the average time delay associated with axonal
conduction.
PACS numbers: 05.45.Xt, 87.18.Sn, 87.18.Hf, 68.35.Rh
INTRODUCTION
Since its inception nearly two decades ago, the criti-
cal brain hypothesis has gained a considerable amount
of attention in the literature [1 -- 3]. Although it has en-
countered some skepticism at times [4], it has now grown
to a relatively mature field with substantial body of the-
oretical and experimental evidence to support it [4 -- 11].
Brain criticality is thought to underlie many of its fun-
damental properties such as optimal response, learning,
information storage, as well as transfer [12]. The origi-
nal ideas of brain criticality came out of studies of self-
organized criticality, where a threshold dynamics leads to
a balance between slow drive and fast dissipation in open
nonequilibrium systems and thus observation of critical
dynamics [13].
It is now generally believed that long-
term evolution has led to a balance between excitatory
as well as inhibitory tendencies which place the brain "on
the edge", i.e. a critical point. However, this does not
necessarily answer the problem of stability of the critical
state, as some neurophysiological mechanism is needed to
maintain the system near the critical point against many
possible perturbative effects.
It seems like there are some important theoretical is-
sues which have remained open in regards to brain criti-
cality: (i) What exactly is the phase transition which de-
termines the critical point? Traditionally, this has been
assumed to be the absorbing-state phase transition moti-
vated by the studies of self-organized criticality [14, 15].
However, in some recent studies, it has been indicated
that the brain is maintained near a synchronization tran-
sition [10, 16]. We note that some authors have also
argued for the existence of the extended critical region
similar to that of "Griffiths phase" [17 -- 20]. However,
such critical regions also typically occur near the absorb-
ing phase transition where the system transitions from
an inactive phase to an active phase. (ii) What is the
self-organizing mechanism which leads to, and maintains
the system in a critical state? As mentioned above the
balance between excitatory and inhibitory tendencies are
thought to be the long time solution to this question.
However, physiological mechanism such as synaptic plas-
ticity are also considered to be important mechanism
to maintain the nervous system in a balanced state on
shorter time scales. Clearly, extended criticality can also
alleviate such a problem to a certain extend as criticality
is observed for a range of parameter instead of a particu-
lar point. (iii) If the brain is in the critical state with its
associated scale-invariant behavior, how could it also dis-
play definitive rhythmic behavior via brain oscillations?
Brain plasticity is increasingly being recognized as
an important and fundamental property of a healthy
nervous system.
In particular, spike-timing-dependent-
plasticity (STDP) is an important mechanism which
can modify synaptic weights on very short time scales.
Therefore, it seems reasonable to invoke STDP as a self-
organizing mechanism. In a STDP protocol, the strength
of a synapse is modified based on the relative spike-
timing of its corresponding pre- and post-synaptic neu-
rons, i.e., STDP incorporates the causality of pre- and
post-synaptic spikes into the synaptic strength modifica-
tions. If the pre-synaptic neuron spikes first and leads to
the post-synaptic neuron to spike shortly afterward, then
the synapse is potentiated. Reversely, if the pre-synaptic
spike follows the post-synaptic spike the synapse will be
depressed [21 -- 24]. The competition between coupling
and decoupling forces arising from successive potentia-
tion and depression of synapses tunes the neural network
into a balanced dynamical state.
Our work in this paper is motivated by the above con-
siderations. In particular, we propose to study a biolog-
ically plausible model of cortical networks, i.e.
Izhike-
vich neurons, along with neurophysiological regulatory
mechanism such as STDP with suitable axonal conduc-
tion delays in order to answer some of the above posed
questions. Interestingly, we find that our regulatory sys-
tem self-organizes the neuronal network to the "edge of
synchronization" in physiologically meaningful parame-
ter regime. We first establish some of the characteristics
of such a steady state. More importantly, we look for
characteristics of critical dynamics in such a minimally
synchronized steady state. Motivated by various exper-
iments, we look for neuronal avalanches, branching ra-
tios, as well as power spectrum of activity time-series.
We find that such a system on the edge of synchroniza-
tion exhibits significant indications of critical dynamics
including scale-invariant avalanches with finite-size scal-
ing. Our results provide definitive answers to the above
questions in a biologically plausible model of neuronal
networks.
In the following section, we describe the model we use
for our study. Results of our numerical study is presented
in section III, and we close the paper with some conclud-
ing remarks in section IV.
MODEL AND METHODS
The studied cortical networks consist of N spiking
Izhikevich neurons which interact by transition of chemi-
cal synaptic currents with axonal conduction delays. The
dynamics of each neuron is described by a set of two dif-
ferential equations [25]:
dvi
dt
= 0.04v2
i + 5vi + 140 − ui + I DC
i + I syn
i
dui
dt
= a(bvi − ui)
(1)
(2)
with the auxiliary after-spike reset:
if vi≥30,
then vi → c and ui → ui + d
(3)
2
are reset according to Eq.(4). a, b, c and d are four ad-
justable parameters in this model. Tuning these parame-
ters, Izhikevich neuron is capable of reproducing different
intrinsic firing patterns observed in real excitatory and
inhibitory neurons [25]. We set these parameters so that
excitatory neurons spike regularly and inhibitory neurons
produce fast spiking pattern [25 -- 27].
i
The term I DC
is an external current which determines
intrinsic firing rate of uncoupled neurons. Regularly spik-
ing Izhikevich neurons exhibits a Hopf bifurcation at
I DC = 3.78 [28]. We choose values of I DC
randomly
from a Poisson distribution with the mean value 10. The
term I syn
represents the chemical synaptic current deliv-
ered to each post-synaptic neuron i [29]:
i
i
I syn
i =
V0 − vi
Di X
j
gji
exp(− t−(tj +τji)
τs
) − exp(− t−(tj +τji)
τs − τf
τf
)
(4)
Here Di is the in-degree of node i, tj is the instance of
last spike of pre-synaptic neuron j, and τji is the axonal
conduction delay from pre-synaptic neuron j to post-
synaptic neuron i.
If axonal delays are not taken into
account, then τji = 0 for all j 6= i. Axonal delay values
of τji are chosen randomly from a Poisson distribution
with mean value τ = hτjii. τf and τs are the synaptic
fast and slow time constants and V0 is the reversal po-
tential of the synapse. If inhibition is included, then mo-
tivated by the properties of cortical networks [30], we set
population density of inhibitory neurons to twenty per-
cent, i.e. α = 0.2 while the initial strength of inhibitory
synapses are chosen four times the strength of excita-
tory synapses. Therefore, the excitation-inhibition ratio
is balanced. α = 0 indicates that we are only considering
a network of excitatory neurons. gji is the corresponding
element of the adjacency matrix of the network which
denotes the strength of synapse from pre-synaptic neu-
ron j to post-synaptic neuron i. Each type of synapses
are initially static and have equal strength. gji = gs if
neurons j and i are connected and the synapse is excita-
tory, gji = 4gs if neurons j and i are connected and the
synapse is inhibitory, and gji = 0 otherwise. When we
turn the STDP on, strength of excitatory synapses are
modified according to a soft-bound STDP rule [21 -- 24],
while the strength of inhibitory synapses are fixed.
If
pre-synaptic neuron j fires a spike at time t = tpre, then
the strength of synapse is modified to gji → gji + ∆gji,
where:
∆gji =
A+(gmax − gji)e
−
∆t−τji
τ+
−A−(gji − gmin)e
∆t−τji
τ−
if ∆t > τji
if ∆t ≤ τji
(5)
for i = 1, 2, ..., N . Here vi is the membrane potential and
ui is the membrane recovery variable. When vi reaches
its apex (vmax = 30 mV), voltage and recovery variable
Here, ∆t = tpost − tpre is the time difference of last post-
and pre-synaptic spikes, A+ and A− determine the max-
imum synaptic potentiation and depression, τ+ and τ−
TABLE I: Values of constant parameters used in this study.
Izhikevich neuron
Synaptic current
STDP rule
aex = 0.02
τf = 0.2
A+ = 0.05 A− = 0.05
bex = 0.2
τs = 1.7
cex = −65
V0,ex = 0
τ+ = 20
dex = 8
V0,in = −75
τ− = 20
ain = 0.1
bin = 0.2
cin = −65
din = 2
gmin = 0
gmax = 0.6
3
determine the temporal extent of the STDP window for
potentiation and depression, and gmin and gmax are the
lower and upper bounds of synaptic strength. The values
of all parameters for Izhikevich neuron, synaptic current
and STDP rule are listed in Table I.
We consider a temporally shifted STDP window for
which the boundary separating potentiation and depres-
sion does not occur for simultaneous pre- and post-
synaptic spikes, but rather for spikes separated by a small
time interval [31]. We set the value of this shift equal with
the actual axonal delay for each synapse. This rule re-
trieves the conventional STDP rule when no time-delay is
considered, τji = 0. We have plotted the STDP temporal
window function ∆g = f (∆t) and its shift in Fig.1. This
temporal shift causes synchronous or nearly-synchronous
pre- and post-synaptic spikes to induce long-term depres-
sion.
conventional
0.02
0.01
g
0
-0.01
-0.02
-30
We define a time-dependent order parameter:
S(t) =
2
N (N − 1) X
i6=j
cos2(cid:16) φi(t) − φj(t)
2
(cid:17)
(7)
This order parameter measures the collective phase syn-
chronization at time t. S(t) is bounded between 0.5 and
1.
If neurons spike out-of-phase, then S(t)≃0.5, when
they spike completely in-phase S(t)≃1 and for states with
partial synchrony 0.5 < S(t) < 1. The global order pa-
rameter S∗ is the long-time-average of S(t) at the station-
ary state after the influence of STDP (S∗ = hS(t)it). We
note that the intricate details of the model along with
the need to obtain long-time dynamics of the system,
limit our computational abilities. We have therefore per-
formed simulations for 100 < N < 1000. We find that
our general results and conclusions are independent of
the system size used and therefore report most of our re-
sults for N ≈ 500. In the next section we will present a
systematic study of the system above, paying particular
attention to the effect of STDP, time delay, and inhibi-
tion.
shifted
RESULTS
-20
-10
0
t
10
20
30
FIG. 1: Conventional (solid line) and shifted (dashed line)
STDP temporal window function ∆g = f (∆t). Blue parts
denote depression and red parts denote potentiation. Units
of ∆t is ms.
We integrate the dynamical equations using fourth-
order Runge-Kutta method with a time step h = 0.01ms
and obtain vi(t). We typically evolve the entire system
for a long time and make sure that the system has reached
a stationary state. We then perform our measurements
and calculations. We obtain the instants of firings of all
neurons and then assign a phase to each neuron between
each pairs of successive spikes [32]:
φi(t) = 2π
t − tm
i
− tm
i
tm+1
i
(6)
while tm
i
is the time that neuron i emits its mth spike.
Spiking Izhikevich neurons with static
chemical
synapses exhibit a continuous transition to phase syn-
chronization upon increasing synaptic strength, i.e. the
amount of global synchrony depends on the average
synaptic strength [28]. Now, consider the simple case of
an all-to-all network of excitatory neurons without axonal
delays. STDP is off initially. S(t) timeseries for different
values of gs are illustrated in Fig.2(a). It is observed that
S(t) depends on gs as is expected. Next, we turn on the
STDP at t = 5s. Interestingly, it is seen that S(t) time-
series evolve to a common state regardless of their initial
values. Thus, as STDP modifies the synaptic strengths,
neural network organizes into a final state with a spe-
cific global phase synchronization S∗ independent of the
initial synaptic strengths. Our investigations reveal that
this is a generic condition emerging in neural networks
with different underlying structures. We also find that
the amount of S∗ is independent of many parameters in-
cluding the amplitudes and time extents of STDP rule,
and intrinsic firing rate of neurons. However S∗ depends
drastically on the average value of axonal conduction de-
lays. Fig.2(b) shows that increasing τ leads to a phase
transition from strongly synchronized states with S∗ ≃ 1
to asynchronous states with S∗ ≃ 0.5, for neural net-
works with α = 0 and α = 0.2. Fig.2(b) also shows that
inhibition has a secondary role in the amount of steady
state synchronization ,S∗, as compared to axonal delay,
τ . Important to our purposes, it shows that for τ = 10ms
the systems stand at the boundary of phase synchroniza-
tion for both α values.Note the importance of time delay
as it causes STDP to depress (weaken) the synchronous
neurons, thus reducing the amount of S∗ in the system.
(a)
1
)
t
(
S
0.5
0
5
(b)
1
*
S
0.5
0
=0, ji=0
gs=0.5
gs=0.3
gs=0.2
gs=0.1
gs=0.05
10
t
15
20
=0
=0.2
4
8
12
Synchronization and average synaptic weights
4
(a)
1
)
t
(
S
0.5
0
(c)
1
)
t
(
S
0.5
0
(e)
1
)
t
(
S
0.5
0
(g)
1
)
t
(
S
0.5
0
(b)
0.6
)
t
(
G
0.3
= 0, ji = 0
= 0, ji = 0
15
20
0
0
5
15
20
10
t
(d)
0.6
)
t
(
G
= 0, = 10ms
15
20
0
0
(f)
0.6
)
t
(
G
= 0.2, ji = 0
15
20
0
0
= 0.2, = 10ms
(h)
0.6
)
t
(
G
15
20
0
0
= 0, = 10ms
15
20
= 0.2, ji = 0
15
20
= 0.2, = 10ms
15
20
10
t
10
t
10
t
5
5
5
5
5
5
5
10
t
10
t
10
t
10
t
FIG. 2: (a) Effect of STDP on the time evolution of S(t) for
all-to-all networks of excitatory spiking neurons with different
gs. The unit of time axis is in seconds. (b) Dependence of S ∗
on τ for α = 0 and α = 0.2.
FIG. 3: Timeseries of S(t) and G(t) and the influence of
STDP on them. The unit of time axis is in seconds. STDP is
turned on at t = 5s. In each panel different line colors demon-
strate different static synaptic strengths, gs = 0.5 (black),
gs = 0.2 (red) and gs = 0.05 (blue).
In order to further investigate the properties of Izhike-
vich neuronal networks, we consider four different net-
works of N = 500: (1) a network of purely excitatory neu-
rons without time-delay (α = 0, τji = 0), (2) a network
of purely excitatory neurons with axonal conduction de-
lays (α = 0, τ = 10ms), (3) a network of excitatory and
inhibitory neurons without time-delay (α = 0.2, τji = 0),
and (4) a network of excitatory and inhibitory neurons
with axonal conduction delays (α = 0.2, τ = 10ms). We
have studied networks with different τ values, but we
display mostly the results in cases for which all delays
are zero (τji = 0) and S∗ ≫ 0.5 as well as those with
τ = 10ms for which S∗ → 0.5+. We note that while our
results (Fig.2) show that τ = 10ms is an interesting case
of transition point, such an actual value for axonal delay
is experimentally meaningful [33]. We turn on STDP at
t = 5s in a complete network and monitor its influence
on different features of each system.
The influence of STDP on the timeseries S(t) in differ-
ent conditions is illustrated in the left column of Fig.3.
Each panel contains three plots with different values of
gs, i.e. the initial synaptic weights. When STDP is off,
S(t) depends on gs. Turning STDP on, each system
reaches a final state with a specific amount of synchro-
nization S∗, regardless of initial level of order (regardless
of gs). However, S∗ depends on τ and α. Systems (1) and
(3) reach a strongly synchronized states with S∗ ≃ 0.88
and S∗ ≃ 0.75, respectively. Implementation of conduc-
tion delays drive the dynamics toward lower levels of or-
der. Systems (2) and (4) with τ = 10ms lead to states at
the edge of transition with S∗ ≃ 0.509 and S∗ ≃ 0.503,
respectively. The right column of Fig.3 represents the
timeseries of the average strength of excitatory synapses,
for the corresponding system in the left column repre-
sented by, G(t) = 1
NL Pj6=i gji,ex(t), where NL is the
number of existing excitatory links. It is observed that
at the final states G(t) ≃ 0.3 for all the systems. It is in-
teresting that the final average value of synaptic weight
is independent of the amount of inhibition and and/or
axonal delay, as well as initial distribution. However, the
main point here is that the amount of synchronization
in the system is not solely determined by average synap-
tic strength but crucially depends on axonal conduction
delay, and to a lesser degree on inhibition.
Synaptic distributions
(a)
0.3
)
x
e
g
(
P
0
0
0.1
0.2
(c)
0.2
)
x
e
g
(
P
0
0
0.1
0.2
(e)
0.2
)
x
e
g
(
P
0
0
0.1
0.2
(g)
0.15
)
x
e
g
(
P
0
0
0.1
0.2
= 0, ji = 0
0.4
0.5
0.6
= 0, = 10ms
0.4
0.5
0.6
= 0.2, ji = 0
0.4
0.5
0.6
= 0.2, = 10ms
0.4
0.5
0.6
0.3
gex
0.3
gex
0.3
gex
0.3
gex
(b)
0.6
j
i
,
x
e
g
,
i
j
,
x
e
g
0
0
(d)
0.6
j
i
,
x
e
g
,
i
j
,
x
e
g
0
0
(f)
0.6
j
i
,
x
e
g
,
i
j
,
x
e
g
0
0
(h)
0.6
j
i
,
x
e
g
,
i
j
,
x
e
g
0
0
= 0, ji = 0
15
20
= 0, = 10ms
15
20
= 0.2, ji = 0
15
20
= 0.2, = 10ms
15
20
10
t
10
t
10
t
10
t
5
5
5
5
FIG. 4: Distribution of the excitatory synaptic strength at the
stationary state of the systems after the influence of STDP
(left) and time evolution of a pair of reciprocal synapses
(right). The unit of time axis is in seconds.
It is somewhat unexpected that similar average synap-
tic weights would lead to decidedly different synchroniza-
tion patterns. The answer is in the form of the actual
distributions of the weights. In one scenario the average
is the most likely value (unimodal distribution) and in
the other case is the least likely value (bimodal distribu-
tion). The probability distribution function of excitatory
synaptic strengths P (gex) (in the steady state) for each
system is shown in the left column of Fig.4. Also, the
5
FIG. 5: Raster plots of the neural networks with different
values of τ and α at the stationary states after influence of
STDP. The unit of time axis is in seconds.
right column of this figure illustrates time evolution of
strength of a pair of reciprocal synapses. At the absence
of axonal delays, STDP produces a bimodal distribution
of synaptic strengths (Figs.4(a) and 4(e)) which is incom-
patible with the experimentally observed distributions of
synaptic strength. However, addition of time-delays to
the neural network modifies this condition. Simultane-
ous presence of STDP and time-delays produce a uni-
modal distribution of synaptic strengths (Figs.4(c) and
4(g)) resembling those measured in cultured and cortical
networks [34, 35].
Emergence of these different distributions of synaptic
strengths is associated with the amount of phase synchro-
nization in the networks. When neurons interact without
time-delay, the final state of the system is strongly syn-
chronized. Therefore, for each pair of symmetric links,
STDP depresses the link in one direction and potenti-
ates the link in the opposite direction as in Figs.4(b) and
4(f). Thus all symmetric connection would be broken
into unidirectional connections after a while in this case.
This leads to a bimodal distribution of synaptic strengths
whether the network consists of purely excitatory neu-
rons or a mixture of excitatory and inhibitory neurons.
With the inclusion of time-delay in the system the level
of order declines as it also causes to preserve symmetric
connections between each pair of neurons [36]. Although
the strength of synapses fluctuates over time (Figs.4(d)
and 4(h)), both links in opposite directions remain ac-
tive. This leads to a unimodal distribution of synaptic
strengths.
Indicators of criticality
So far we have seen that STDP along with reasonable
time delay (and inhibition) will lead the system on the
edge of synchronization. However, being on the edge of
synchronization could be caused by vastly different spik-
ing patterns [28]. More importantly for the purpose of
the present study, we would like to know whether such a
state of minimal synchronization has any experimentally
relevant indications of criticality. In this subsection we
will address such issues.
Raster plots of neural networks with different values
of α and τ (in the steady state) are displayed in Fig.5.
When time-delay is ignored, neuronal spikes are highly
ordered (Figs.5(a) and 5(c)). This is not the real state
of a healthy nervous system. However, addition of ax-
onal conduction delay modifies the amount of global or-
der in the networks. Simultaneous effect of STDP and
a suitable axonal conduction delays decrease global co-
herence in neural oscillations (See Figs.5(b) and 5(d)).
In Figs.5(c) and 5(d),
inhibitory neurons indexed as
401 − 500, spike with a higher rate as compared to exci-
tatory neurons 1 − 400.
= 0, ji = 0
(a)
300
)
t
(
M
200
100
0
19.5
19.6
19.7
19.8
19.9
20
t
= 0, = 10ms
(c)
)
t
(
M
60
40
20
0
19.5
19.6
19.7
19.8
19.9
20
t
= 0.2, ji = 0
(e)
300
)
t
(
M
200
100
0
19.5
19.6
19.7
19.8
19.9
20
t
(g)
90
)
t
(
M
60
= 0.2, = 10ms
30
19.5
19.6
19.7
19.8
19.9
20
t
(b)
m
u
r
t
c
e
p
s
r
e
w
o
p
2.5
2
1.5
1
0.5
0
20
(d)
m
u
r
t
c
e
p
s
r
e
w
o
p
4
3
2
1
0
20
(f)
m
u
r
t
c
e
p
s
r
e
w
o
p
15
10
5
0
20
(h)
m
u
r
t
c
e
p
s
r
e
w
o
p
8
6
4
2
0
20
106
= 0, ji = 0
21
22
23
24
25
f (Hz)
104
= 0, = 10ms
21
22
23
24
25
f (Hz)
105
= 0.2, ji = 0
21
22
23
24
25
f (Hz)
104
= 0.2, = 10ms
104
10-1
22
22
f (Hz)
21
26
30
23
24
25
FIG. 6: Timeseries of network activity M (t) in the stationary
state after the influence of STDP (left) and their power spec-
trum (right). The unit of time axis is in seconds. The inset
of (h) shows the same on a log-log scale.
The amount of order parameter S∗ and the raster
plots are reasonable evidences indicating the system with
6
τ = 10ms organizes to the edge of synchronization tran-
sition with minimal value of S∗. We now present ex-
perimentally relevant results which indicate that such a
system is in a critical state. We first consider the network
activity timeseries M (t) which is defined as the number
of neuronal spikes at time t, as well as its power spec-
trum. These plots are illustrated in Fig.6. The network
activity oscillates regularly in systems without time-delay
for which phase synchronization is strong (Figs.6(a) and
6(e)). Therefore the power spectrum of these systems
exhibit a sharp peak at f ≃ 23.5Hz (Figs.6(b) and 6(f)).
While neurons are delay-coupled the oscillations of M (t)
are irregular (Figs.6(c) and 6(g)). Despite this deceptive
irregularity, the power spectrum exhibits a large peak at
frequency f ≃ 21.5Hz (Figs.6(d) and 6(h)) along with a
range of other frequencies. This dominant peak reveals
that rhythmic oscillations are still robust at these neu-
ral networks. The inset of Fig.6(h) shows a log-log plot
which indicates that the spectrum has a decaying tail in
the system for which α = 0.2 and τ = 10ms. Note that
the amplitude of oscillations of M (t) depends on the level
of phase synchronization. The stronger the neurons are
synchronized, the larger is the amplitude of M (t) oscil-
lations, i.e. note the scale of the power spectrum on the
y-axis.
Scale-invariant
statistics of neural avalanches
is
thought to be the most important indicator of criti-
cal brain dynamics. Hence, the network displays spon-
taneous activity of various sizes s, known as neural
avalanches, which exhibit scale-free distribution,
i.e.
P (s) ∼ s−β [1]. By monitoring the spiking activity of our
systems, we can identify outbursts of spikes the number
of which is associated with the size of avalanches. An
avalanche begins when the network activity exceeds a
threshold Mth and ends when it turns back below that
threshold. Here, we set the threshold to be equal with the
mean value of activity in the system. s is defined as the
total number of spikes during this avalanche. Critical-
ity is supposed to be indicated by a power-law behavior
P (s) ∼ s−β and a finite-size cut-off which diverges as
system size diverges (N → ∞).
We consider neural networks with α = 0.2 and three
different τ values, i.e. τ = 14ms, τ = 10ms and τ = 8ms.
From the synchronization point of view, Fig. 2(b), these
systems would be subcritical, critical, and supercritical.
Each network is also simulated with different network
sizes N . For any given set of parameters the network is
simulated for a considerably long time, producing a large
number of avalanches. Probability distribution functions
of avalanche sizes for such networks is illustrated in the
left column of Fig.7. For neural networks with τ = 14ms,
P (s) decays with a characteristic scale which is an indica-
tor of subcritical behavior (Fig.7(a)). Note how this scale
saturates as system size increases. For networks with
τ = 8ms, P (s) exhibits a bump for large s which is an
evidence of supercritical behavior (Fig.7(e)). Here, large
avalanches are more likely to occur than intermediate size
avalanches.
Interestingly, in networks with τ = 10ms
P (s) exhibits a power-law behavior P (s) ∼ s−β and a
finite-size cut-off which increases relative to the system
size (Fig.7(c)).
Emergence of power-law behavior in a finite system
does not necessarily prove criticality of the system. To
verify criticality, we perform a finite-size scaling of our
data for different network sizes N (inset of Fig.7(c)). We
observe that indeed we obtain a good collapse for the
system sizes considered in this study. Incidentally, our
finite-size scaling collapse allows us to calculate the β
value more reliably where we obtain the critical exponent
β = 1.4 which is close to the accepted experimental value
β ≈ 1.5 obtained by various studies including the original
neuronal avalanche study of [5].
(a)
100
10-2
)
s
(
P
10-4
10-6
101
= 0.2, = 14ms
102
s
(c)
100
)
s
(
P
10-2
10-4
N
*
)
s
(
P
106
102
10-2
10-2
10-6
101
(e)
100
)
s
(
P
10-2
10-4
101
102
s
102
s
= 0.2, = 8ms
N=125
N=250
N=500
103
N=125
N=250
N=500
103
N=125
N=250
N=500
(b)
3
)
M
(
b
2
()()(b)
1
0
-30
(d)
)
M
(
b
2
1.5
1
0.5
N=125, B=0.93
N=250, B=0.93
N=500, B=0.93
= 0.2, = 14ms
-20
-10
0
M-Ma
10
20
30
N=125, B=1.01
N=250, B=1.00
N=500, B=1.00
= 0.2, = 10ms
0
-30
-20
-10
10
20
30
0
M-Ma
(f)
)
M
(
b
2.5
2
1.5
1
0.5
= 0.2, = 8ms
N=125, B=1.11
N=205, B=1.11
N=500, B=1.12
103
0
-50
0
M-Ma
50
= 0.2, = 10ms
100
10-1
s/N
FIG. 7: Distribution function of size of avalanches, and
activity-dependent branching-ratio b(M ) vs M − Ma, for var-
ious network sizes N : (a) and (b) subcritical, (c) and (d)
critical, (e) and (f) supercritical. Inset of (c) shows the finite-
size-scaling collapse of size of avalanches in the critical sys-
tem for θ = 1.20 and δ = 1.68. Thus the critical exponent is
β = δ
θ = 1.4. On the right column, the average branching-
ratio B for each network with size N is reported in the legend.
Another important quantity to characterize critical dy-
namics is activity-dependent branching ratio [37]. Essen-
tially, this function gives the (relative) expectation value
of the timeseries in the next time step for a given amount
of activity at the present time step. More precisely, it
is defined as, b(M ) = E{ξM /M }. The variable ξM is
the value of the next signal given that the present one
is equal to M , so ξM = {M (t + dt)M (t) = M } [37].
Since a critical system is on the edge and is inherently
7
unpredictable, b(M ) ≈ 1, ∀M . For a finite system one
expects a similar result with the additional consideration
that, with increasing system size, the range of activity
M should increase and that b(M ) should asymptotically
approach 1. Therefore, one expects b(M ) < 1 to gen-
erally indicate subcritical behavior, while b(M ) > 1 to
indicate supercritical behavior. In fact, b(M ) has been
used to ascertain criticality in a wide range of systems
including sandpile models of SOC or solar flares [37] as
well as neural networks [20, 38].
We obtain the activity-dependent branching-ratio
b(M ) using timeseries M (t). The right column of Fig.7
displays b(M ) plots for each one of subcritical, critical
and supercritical systems for different system sizes N
(Figs.7(b), 7(d) and 7(f)). Note that the plots are cen-
tered around their respective average activity Ma. Only
in the critical case (Fig.7(d)) do we observe B(Ma) = 1.
However, more importantly, we see b(M ) increases its
range and decreases its slope (towards zero) with in-
creasing system size, consistent with critical dynam-
ics of the network.
In the two other cases, no such
behavior is observed. For a more common branching
ratio, one calculates the average value of B(M ),
i.e.
b(M )dM . We find B ≃ 1
(τ = 10ms), B ≃ 0.93 (τ = 14ms), and B ≃ 1.1
(τ = 8ms) again indicating critical, subcritical, and su-
percritical dynamics accordingly. The average branching
ratios are reported in the legend of the corresponding
plots in Fig.7.
B = 1/(Mmax − Mmin)R Mmax
Mmin
We have therefore shown how the system with STDP
and physiologically relevant inhibition and axonal delays
will evolve to a unimodal distribution of synaptic weights
starting from a complete uniform network. The resulting
state is a state on the edge of synchronization transition
(not an activity transition) which nevertheless shows ex-
perimentally relevant indicators of critical dynamics in-
cluding power-law avalanches with finite-size scaling as
well as branching ratios. We also show how such indica-
tors of criticality disappear as one moves away from the
edge of synchronization transition via change in average
delay times.
CONCLUDING REMARKS
In this paper we showed that invoking neurophysio-
logical regulatory mechanisms such as temporally shifted
STDP and specific amounts of axonal conduction delays
(τ = 10ms) in a biologically plausible model of corti-
cal networks put the system in a critical state at the
neighborhood of synchronization transition point. In this
state the system exhibits robust rhythmic behavior along
with power-law distributions of avalanche sizes. Further-
more, the behavior of activity-dependent branching-ratio
confirms the criticality of system in this state as well.
However for smaller or larger values of axonal conduc-
tion delays neural networks self-organize into supercriti-
cal or subcritical states, respectively. While the state of
the network is off-critical, neither the statistics of sizes of
avalanches nor branching-ratio exhibit the signs of criti-
cality.
Coexistence of
rhythmic oscillations and scale-
invariant avalanches is important for development of cor-
tical layers [39]. Evidence for this coexistence has been
found in experimental investigations [39, 40]. Also in
theoretical studies, this phenomenon has been reported
to occur as a result of balance between inhibition and
excitation [41], as well as in a periodically driven SOC
model [42]. The neurophysiological mechanisms leading
to this intricate dynamics in the cortex is of fundamen-
tal importance in neuroscience. Here, we revealed that
such intricate dynamics emerges as a result of intrinsic
regulatory mechanisms like STDP and axonal conduction
delays. More strictly, we obtained self-regulated critical-
ity along with coexistence of rhythmic oscillations and
scale invariant activity in a biologically relevant model.
We began this paper by posing three open questions
regarding the critical brain hypothesis. Our results have
provided interesting answers to all three questions. (i)
The critical point and corresponding phase transition
that the brain organizes itself into is not the usual ac-
tivity and/or absorbing phase transition, but the syn-
chronization phase transition.
(ii) The self-organizing
mechanism which tunes and maintains the system around
such critical point is a standard neurophysiological regu-
latory mechanism of a temporally shifted STDP. (iii) The
existence of individual neuronal oscillations which self-
organize to a highly correlated but weakly synchronized
collective state is responsible for a dominate oscillatory
mode in addition to scale-free fluctuations.
We have studied neural networks with different topolo-
gies, various initial conditions, as well as various choices
of STDP parameters and observed that our results are
generally the same upon all such changes. We have also
examined that hard-bound STDP leads to similar results,
except for the distribution function of synaptic strengths
that would be bimodal regardless of all conditions imple-
mented in the neural network.
ACKNOWLEDGEMENTS
Support from Shiraz University research council
is
kindly acknowledged. This work has been supported in
part by a grant from the Cognitive Sciences and Tech-
nologies Council.
∗ Electronic address: [email protected]
8
[1] Plenz, D. (2014). Criticality in Neural Systems. Wiley
VCH. New York.
[2] Legenstein, R., and Maass, W. (2008). New Directions
in Statistical Signal Processing: From Systems to Brain.
MIT Press.
[3] Chialvo, D. R. (2010). Emergent complex neural dynam-
ics. Nature Phys. 6, 744.
[4] Beggs, J. M., and Timme, N. (2012). Being critical of
criticality in the brain. Front. Physiol. 3, 163.
[5] Beggs, J. M., and Plenz, D. (2003). Neuronal avalanches
in neocortical circuits. J. Neurosci. 23, 11167.
[6] Beggs, J. M., and Plenz, D. (2004). Neuronal avalanches
are diverse and precise activity patterns that are stable
for many hours in cortical slice cultures. J. Neurosci. 24,
5216.
[7] Haimovici, A., Taglizucchi, E., Balenzuela, P., and
Chialvo, D. R. (2013). Brain organization into resting
state networks emerges at criticality on a model of the
human connectome. Phys. Rev. Lett. 110, 178101.
[8] Plenz, D., and Thiagarajan, T. C. (2007). The organizing
principles of neuronal avalanches: cell assemblies in the
cortex? Trends Neurosci. 30, 101.
[9] Friedman, N., Ito, S., Brinkman, B. A. W., Shimono,
M., LeeDeVille, R. E., Dahmen, K. A., Beggs, J. M., and
Butler, T.C. (2012). Universal critical dynamics in high
resolution neuronal avalanche data. Phys. Rev. Lett. 108,
208102.
[10] Fontenele, A. J., de Vasconcelos, N. A. P., Feliciano, T.,
Aguiar, L. A. A., Soares-Cunha, C., Coimbra, B., Porta,
L. D., Ribeiro, S., Rodrigues, A. J., Sousa, N., Carelli, P.
V., and Copelli, M. (2019). Criticality between cortical
states. Phys. Rev. Lett. 122, 208101.
[11] Levina, A., Herrmann, J. M., and Geisel, T. (2007). Dy-
namical synapses causing self-organized criticality in neu-
ral networks. Nature Phys. 3, 857.
[12] Hesse, J., and Gross, T. (2014). Self-organized criticality
as a fundamental property of neural systems. Front. Syst.
Neuro. 8, 166.
[13] Bak, P., Tang, C., and Wiesenfeld, K., (1987). Self-
organized criticality: An explanation of the 1/f noise.
Phys. Rev. Lett. 59, 381.
[14] Montakhab, A., and Carlson, J. M. (1998). Avalanches,
transport, and local equilibrium in self-organized criti-
cality. Phys. Rev. E 58, 5608.
[15] Vespignani, A., Dickman, R. Munoz, M. A., Zapperi, S.
(2000). Absorbing-state phase transitions in fixed-energy
sandpiles. Phys. Rev. E 62, 4564.
[16] di Santo, S., Villegas, P., Burioni, R., and Munoz, M.
A., (2018). Landau Ginzburg theory of cortex dynamics:
Scale-free avalanches emerge at the edge of synchroniza-
tion . PNAS 115(7), 1356-1365.
[17] Moretti, P., and Munoz, M. A. (2013). Griffiths phases
and the stretching of criticality in brain networks. Nat.
Commun. 4, 2521.
[18] Munoz, M. A., Juhasz, R., Castellano, C., and Odor, G.
(2010). Griffiths phases on complex networks. Phys. Rev.
Lett. 105(12), 128701.
[19] Odor, G., Dickman, R., Odor, G. (2015). Griffiths phases
and localization in hierarchical modular networks. Sci.
Rep. 5, 14451.
[20] Moosavi, S. A., Montakhab, A., and Valizadeh, A.
(2017). Refractory period in network models of excitable
nodes: self-sustaining stable dynamics, extended scaling
region and oscillatory behavior. Sci. Rep. 7, 7107.
[21] Markram, H., Gerstner, W., and Sjostrom, P. J.,
(2012). Spike-timing-dependent plasticity: a comprehen-
sive overview. Front. Synaptic Neurosci. 4, 1.
[22] Song, S., Miller, K. D., and Abbott, L. F. (2000).
Competitive Hebbian learning through spike-timing-
dependent synaptic plasticity. Nat. Neurosic. 3, 919-926.
[23] Bi G. Q. and Poo, M. (2001). Synaptic modification by
correlated activity: Hebb's postulate revisited. Annu.
Rev. Neurosci. 24, 139-166.
[24] Sjostrom, J., and Gerstner, W., (2010). Spike-timing de-
pendent plasticity. Scholarpedia 5, 1362.
[25] Izhikevich, E. M., (2003). Simple model of spiking neu-
rons. IEEE Trans. Neural Netw. 14, 1569-1572.
[26] Izhikevich, E. M., (2007). Dynamical Systems in Neuro-
science. MIT Press, Cambridge.
[27] Izhikevich, E. M., (2006). Polychronization: computation
with spikes. Neural Comput. 18, 245-282.
[28] Khoshkhou, M., and Montakhab, A. (2018). Beta-
rhythm oscillations and synchronization transition in net-
work models of Izhikevich neurons: effect of topology and
synaptic type. Front. Comput. Neurosci. 12, 59.
[29] Roth, A., and van Rossum, M. (2009). Computational
Modeling Methods for Neuroscientists. ed E De Schutter.
Cambridge, MA: MIT Press.
[30] DeFelipe, J.
(1993). Neocortical neuronal diversity:
chemical heterogeneity revealed by colocalization stud-
ies of classic neurotransmitters, neuropeptides, calcium-
binding proteins, and cell surface molecules. Cerebral
Cortex 3, 273289.
[31] Babadi, B., and Abbott, L. F., (2010). Intrinsic stability
of temporally shifted spike-timing dependent plasticity.
PLoS Comput. Biol. 6(11): e10000961.
[32] Pikovsky, A., Rosenblum, M., and Osipov, G., (1997).
Phase synchronization of chaotic oscillators by external
driving. Physica D 104, 219.
[33] Swadlow, H. A., and Waxman, S. S. (2012). Axonal con-
duction delays. Scholarpedia 7(6),1451.
9
[34] Turrigiano, G., Leslie, K., Desai, N., Rutherford, L., and
Nelson, S. (1998). Activity-dependent scaling of quantal
amplitude in neocortical neurons. Nature 391 892895.
[35] Song, S., Sjostrom, P., Reigl, M., Nelson, S., and
Chklovskii, D. (2005). Highly nonrandom features of
synaptic connectivity in local cortical circuits. PLoS Biol.
3 e68.
[36] Asl, M. M., Valizadeh, A., and Tass, P. A., (2017). Den-
dritic and axonal propagation delays determine emergent
structures of neuronal networks with plastic synapses.
Sci. Rep. 7,39682.
[37] Martin, E., Shrein, A., and Paczuski, M. (2010).
Activity-dependent branching ratios in stocks, solar x-
ray flux, and the Bak-Tang-Wiesenfeld sandpile model.
Phys. Rev. E 81, 016109.
[38] Moosavi S. A., and Montakhab, A. (2014). Mean-field be-
havior as a result of noisy local dynamics in self-organized
criticality: Neuroscience implications. Phys. Rev. E 89,
052139.
[39] Gireesh, E. D.,
and Plenz. D.
(2008). Neuronal
avalanches organize as nested theta-and beta/gamma-
oscillations during development of cortical
layer 2/3.
PNAS 105, 7576.
[40] Yang, H., Shew, W. L., Roy, R., and Plenz, D. (2012).
Maximal variability of phase synchrony in cortical net-
works with neuronal avalanches. J Neurosci. 32(27),
10611072.
[41] Poil, S. S., Hardstone, R., Mansvelder, H. D., and
Linkenkaer-Hansen, K. (2012). Critical-state dynamics of
avalanches and oscillations jointly emerge from balanced
excitation/inhibition in neuronal networks. J. Neurosci.
32.29, 9817.
[42] Moosavi, S. A., Montakhab, A., and Valizadeh, A.
(2018). Coexistence of scale-invariant and rhythmic be-
havior in self-organized criticality. Phy. Rev. E 98(2),
022304.
|
1905.04283 | 1 | 1905 | 2019-05-10T17:50:21 | Visuospatial short-term memory and dorsal visual gray matter volume | [
"q-bio.NC"
] | Visual short-term memory (VSTM) is an important cognitive capacity that varies across the healthy adult population and is affected in several neurodevelopmental disorders. It has been suggested that neuroanatomy places limits on this capacity through a map architecture that creates competition for cortical space. This suggestion has been supported by the finding that primary visual (V1) gray matter volume (GMV) is positively associated with VSTM capacity. However, evidence from neurodevelopmental disorders suggests that the dorsal visual stream more broadly is vulnerable and atypical volumes of other map-containing regions may therefore play a role. For example, Turner syndrome is associated with concomitantly reduced volume of the right intraparietal sulcus (IPS) and deficits in VSTM. As posterior IPS regions (IPS0-2) contains topographic maps, together this suggests that posterior IPS volumes may also associate with VSTM. In this study, we assessed VSTM using two tasks, as well as a composite score, and used voxel-based morphometry of T1-weighted magnetic resonance images to assess GMV in V1 and right IPS0-2 in 32 healthy young adults (16 female). For comparison with previous work, we also assessed associations between VSTM and voxel-wise GMV on a whole-brain basis. We found that total brain volume (TBV) significantly correlated with VSTM, and that correlations between VSTM and regional GMV were substantially reduced in strength when controlling for TBV. In our whole-brain analysis, we found that VSTM was associated with GMV of clusters centered around the right putamen and left Rolandic operculum, though only when TBV was not controlled for. Our results suggest that VSTM ability is unlikely to be accounted for by the volume of an individual cortical region and may instead rely on distributed structural properties. | q-bio.NC | q-bio | Visuospatial short-term memory and dorsal visual gray matter volume
Dennis Dimonda,b,c, Rebecca Perryb,d, Giuseppe Iariab,e,f, Signe Brayb,c,d,g
a Department of Neuroscience, Cumming School of Medicine, University of Calgary, Calgary,
AB, Canada
b Alberta Children's Hospital Research Institute, University of Calgary, Calgary, AB, Canada.
c Child and Adolescent Imaging Research Program, University of Calgary, Calgary, AB, Canada.
d Department of Paediatrics, Cumming School of Medicine, University of Calgary, Calgary, AB,
Canada.
e Department of Psychology, University of Calgary, Calgary, AB, Canada.
f Hotchkiss Brain Institute, Cumming School of Medicine, University of Calgary, Calgary, AB,
Canada
g Department of Radiology, Cumming School of Medicine, University of Calgary, Calgary, AB,
Canada
Correspondence: * indicates corresponding author
Dennis Dimond*; [email protected] ; Alberta Children's Hospital, 2888 Shaganappi
Trail NW, Calgary, Alberta, Canada, T2B 6A8.
DOI: https://doi.org/10.1016/j.cortex.2018.12.007
© 2019. This manuscript version is made available under the CC-BY-NC-ND 4.0
license http://creativecommons.org/licenses/by-nc-nd/4.0/
Abstract
Visual short-term memory (VSTM) is an important cognitive capacity that varies across the
healthy adult population and is affected in several neurodevelopmental disorders. It has been
suggested that neuroanatomy places limits on this capacity through a map architecture that
creates competition for cortical space. This suggestion has been supported by the finding that
primary visual (V1) gray matter volume (GMV) is positively associated with VSTM capacity.
However, evidence from neurodevelopmental disorders suggests that the dorsal visual stream
more broadly is vulnerable and atypical volumes of other map-containing regions may therefore
play a role. For example, Turner syndrome is associated with concomitantly reduced volume of
the right intraparietal sulcus (IPS) and deficits in VSTM. As posterior IPS regions (IPS0-2)
contains topographic maps, together this suggests that posterior IPS volumes may also associate
with VSTM. In this study, we assessed VSTM using two tasks, as well as a composite score, and
used voxel-based morphometry of T1-weighted magnetic resonance images to assess GMV in
V1 and right IPS0-2 in 32 healthy young adults (16 female). For comparison with previous work,
we also assessed associations between VSTM and voxel-wise GMV on a whole-brain basis. We
found that total brain volume (TBV) significantly correlated with VSTM, and that correlations
between VSTM and regional GMV were substantially reduced in strength when controlling for
TBV. In our whole-brain analysis, we found that VSTM was associated with GMV of clusters
centered around the right putamen and left Rolandic operculum, though only when TBV was not
controlled for. Our results suggest that VSTM ability is unlikely to be accounted for by the
volume of an individual cortical region and may instead rely on distributed structural properties.
Keywords: attention, visuospatial short-term memory, intraparietal sulcus, gray matter volume
Abbreviations: Intraparietal sulcus (IPS); visual short-term memory (VSTM); gray matter
volume (GMV); total brain volume (TBV).
1. Introduction
Visual short-term memory (VSTM) is the ability to hold visual information in memory for
several seconds (Baddeley, 2003). VSTM provides an essential gate between perception and
cognition, allowing the retention of sensory information for higher cognitive processing
(Baddeley, 2003). Although the precise limits on VSTM capacity are debated (Cowan,
2001)(and appended commentaries), it is clear that capacity varies between individuals (Conway
et al., 2001; Vogel & Machizawa, 2004; Vogel et al., 2005).
VSTM changes across development (Park et al., 2002; Riggs et al., 2011) and is affected in many
neurodevelopmental (Attout et al., 2017; Costanzo et al., 2013) and mental health conditions
(Girard et al., 2018). Understanding the neuro-anatomical basis of VSTM has the potential to
inform our basic understanding of cognitive capacity and has clinical relevance to disorders
affecting VSTM and the potential for remediation. The map theory of VSTM capacity postulates
that topographic maps found in early sensory (Sereno et al., 1995) and higher cognitive (Silver &
Kastner, 2009) regions may place limits on cognitive capacity, as items must compete for actual,
'bounded', space (Franconeri et al., 2013). According to this model, one would predict a linear
relationship between VSTM capacity and gray matter volume (GMV) of specific cortical
regions. Such a relationship has indeed been reported by two recent neuroimaging studies; one
reporting a positive association between VSTM and GMV of the primary visual cortex (V1)
(Bergmann et al., 2016), the other reporting findings in the left inferior parietal lobule
(Konstantinou et al., 2017). These findings provide support for the map-based theory of VSTM
capacity and raise the possibility that other cortical areas might also play a role in limiting
VSTM.
The intraparietal sulcus (IPS) is a cortical region that plays an important role in visuospatial
attention and memory (Corbetta & Shulman, 2002). The IPS is topographically organized into 6
distinct regions, each containing a topographic 'map' of the contralateral visual field (Silver &
Kastner, 2009). The posterior IPS (IPS regions 0-2) makes white matter connections to early
visual areas such as V1 (Bray et al., 2013; Greenberg et al., 2012), and is believed to send top-
down signals to influence sensory salience in the visual cortex (Lauritzen, D'Esposito, Heeger, &
Silver, 2009). These brain regions (V1 and IPS0-2) are part of the dorsal visual stream, which
has been suggested to be particularly vulnerable to neurodevelopmental abnormalities (Braddick
et al., 2003). Deficits in dorsal stream function are present in several neurodevelopmental
disorders (Atkinson et al., 2003) including patients with Williams and Turner syndrome, who
also show deficits in VSTM (Attout et al., 2017; Cornoldi et al., 2001; Costanzo et al., 2013). In
the case of Turner syndrome, VSTM deficits are concurrent with reduced GMV in posterior
parietal and occipital regions (Zhao & Gong, 2017), as well as abnormal shape of the right
posterior IPS (Molko, Cachia, Riviere et al., 2003). Taken together, the known involvement of
IPS0-2 and V1 in visuospatial processing, along with concurrent findings of VSTM impairment
and right posterior IPS and visual cortex abnormalities in TS, suggests that these regions may
play a central role in limiting VSTM capacity and related deficits.
Taken together, a map-based theory of VSTM capacity, in combination with findings in
neurodevelopmental populations, suggests that GMV of the right posterior IPS regions 0-2, in
addition to early visual regions such as V1, may place constraints on VSTM capacity. In the
present study we used voxel-based morphometry of T1-weighted magnetic resonance images to
extract GMVs of V1 and IPS0-2 regions of interest (ROIs) using a probabilistic atlas (Wang et
al., 2014), in 32 healthy young adults. We characterized VSTM using two assessments and
correlated volumetric with cognitive measures. For comparison with previous findings, we also
assessed associations between voxel-wise GMV and VSTM on a whole-brain basis. We found
that total brain volume predicted VSTM, and individual ROIs did not add predictive value. At
the whole-brain level, VSTM associated with GMV in the left parietal/temporal lobes and right
putamen, though only when total brain volume was not accounted for. Our results suggest that
VSTM is unlikely to rely on individual brain regions and may be more accurately associated
with distributed representations.
2. Methods
2.1 Participants
All participants provided informed written consent, and the study was approved by the Conjoint
Health Research Ethics Board at the University of Calgary. Inclusion criteria for the study were
having normal or normal-corrected vision and right handedness. Participants were recruited from
the University of Calgary student community and consisted of 32 young adults (aged 18.1 -- 36.1,
mean = 22.2, SD = 3.69, M/F = 16/16) having no history of neurological or psychiatric
conditions. Participants received financial compensation for participating.
2.2 VSTM measurement
VSTM was assessed using a set of computer-based tasks programmed in the PsychToolbox
(http://psychtoolbox.org/) in MATLAB (The Mathworks Inc., Natick, MA, USA). VSTM was
assessed in a session that preceded MRI acquisition by several hours up to one month. Two
assessments of VSTM were administered; the dot memory task, an assessment of location
VSTM, and the modified span board task, which assesses location VSTM, but also includes a
cognitive control component as it requires remembering both the location and order in which
stimuli were presented.
In the dot memory task (Miyake et al., 2001), participants viewed a 5x5 grid and were asked to
remember the location of items that flashed in the grid. Throughout the task a grayscale
background was shown. At the start of each trial they were shown black gridlines for 1s followed
by a set of black circles that appeared in specific locations in the grid for 750ms. The circle
diameter was equivalent to the height and width of the grid cells. The dots disappeared, and
participants were instructed to click in the grid cells to indicate where the dots were shown.
There were no time limits within trials. The number of locations to remember increased from 3-7
across a total of 25 trials, with 5 trials for each number of locations to remember. Between each
level there was a 2s pause during which a fixation cross was shown. The task was scored as the
total number of correctly remembered locations out of a maximum score of 125. Participants
started with 3 practice trials (3, 4 and 5 dots shown) to familiarize them with the task before the
scored trials started. No performance criteria were imposed for the practice trials.
In the modified span board task (Westerberg et al., 2004), participants were asked to recall the
spatial position of stimuli and also the order. A grayscale background was shown throughout this
task. Each trial began with the presentation of a number in the center of the screen, indicating the
number of squares that would be illuminated on the trial. This number was presented for 1s
during the intertrial interval. On each trial, participants were shown a set of 9 small black squares
irregularly spaced on the screen for 500ms. A set of squares were illuminated in white for 650ms
one after the other with no temporal gap. Participants were then asked to click the squares in the
order they were illuminated. There were no time limits within trials. Once clicked, the squares
remained illuminated. The number of illuminated squares increased from 4-8 across a total of 25
trials, with 5 trials for each number of illuminated squares. The task was scored as the total
number of squares recalled in the correct order, out of a maximum of 150. Participants
completed two practice trials to become familiar with the task. No performance criteria were
used for the practice trials.
Since combining tasks assessing the same cognitive construct can reduce measurement error and
potentially provide a more accurate assessment, we also calculated a composite VSTM score by
averaging Z-scores of the dot memory and span board tasks.
2.3 MRI data collection and processing
T1-weighted anatomical MR images (3D SPGR, 180 slices, FOV = 25.6cm, voxel size = 1mm
isotropic, flip angle = 12) were collected on a 3T GE Signa scanner with an 8-channel head coil
at the Seaman Family MR Research Center at the University of Calgary. We assessed gray
matter volume (GMV) with voxel-based morphometry (Ashburner & Friston, 2000) in SPM12,
using standard processing procedures. In brief, T1-weighted images were segmented, images
were co-registered to a study specific template generated using the Dartel toolbox (Ashburner,
2007), and subsequently normalized to standardized Montreal Neurological Institute (MNI)
template space (Mazziotta, Toga, Evans, Fox, & Lancaster, 1995). Motivated by concurrent
findings of impaired VSTM and structural abnormalities in Turner syndrome, we took an ROI-
based approach to assess GMV of the bilateral primary visual cortex, and right posterior IPS
regions 0-2. GMV was averaged within a set of 4 ROIs; V1 (left and right combined), right IPS0,
right IPS1 and right IPS2. All ROIs were defined using a probabilistic atlas (Wang et al., 2014).
2.4 ROI Statistical analyses
We calculated correlations between the dot memory, span board and composite VSTM scores
and total brain volume (TBV), in both bivariate Pearson's correlations and partial correlations
controlling for sex. Next, we used partial correlations to assess associations between the dot
memory, span board, and composite scores and ROI GMV, controlling for sex. For comparison
with previous work, these partial correlations were conducted with and without controlling for
TBV. ROI analyses were false discovery rate corrected for multiple comparisons, and statistical
significance was set at p<0.05.
2.5 Whole-brain statistics
To test our hypothesis that VSTM is related to GMV of specific regions of the visual cortex and
right posterior IPS, we elected to utilize an ROI-based approach. However, considering this
approach might miss associations in brain areas outside of our hypothesized regions, which could
be of interest in comparison to previous work (Konstantinou et al., 2017), we utilized voxel-
based morphometry to assess associations between VSTM task performance and GMV
throughout the whole-brain. Images were smoothed using an 8mm full-width at half-maximum
Gaussian smoothing kernel and entered into whole-brain general linear models, in which task
performance on the VSTM tasks were regressed against voxel-wise GMV while controlling for
sex, with and without controlling for TBV. Initial voxel-wise threshold was set at p<0.001,
followed by family-wise error multiple comparison correction with cluster-wise thresholding at
p<0.05.
3 Results
3.1 VSTM task performance
Raw scores on the span board task ranged from 75 to 138 (mean=111.1, std=18.8), and the dot
memory task from 88 to 125 (mean=114.3, std=8.8). The two tasks were significantly correlated
(r=0.75, p<0.001).
3.2 Association between ROI GMV and VSTM
Before testing individual ROIs, we assessed associations between VSTM scores and TBV. We
found that the dot memory task and composite score were significantly correlated with TBV
before (r=0.459 and 0.404 respectively, p<0.05) and after (r=0.467 and 0.368 respectively,
p<0.05) controlling for sex. For the dot memory task, with exclusion of 3 lower limit outliers this
correlation fell below significance in Pearson's (r=0.171, p=0.375) and partial correlations
controlling for sex (r=0.179, p=0.362). We calculated associations both with and without
controlling for TBV in partial correlations with ROI GMV. Before controlling for TBV, we
found a trendline correlation between dot memory task performance and V1 GMV (r=0.374,
uncorrected p<0.05), however this fell below trendline significance with exclusion of outliers
(r=0.185, p=0.346). No significant associations were found after controlling for TBV, though
interestingly, correlation coefficients were substantially reduced for all tasks when TBV was
included in the model (Figure 1).
Figure 1. Regional GMV-VSTM task correlations. Correlations between gray matter volume
and VSTM task (dot memory, span board, and composite score) performance for the four regions
of interest are shown before and after correcting for total brain volume. V1 = primary visual
cortex; rIPS0-2 = right intraparietal sulcus 0-2; TBV = total brain volume.
3.3 Association between voxel-wise GMV and VSTM
In whole-brain voxel-based morphometry analyses without controlling for TBV, we found dot
memory scores correlated significantly with two clusters; one encompassing the right basal
ganglia, limbic structures, and the hippocampus, with a peak value in the right putamen (MNI
x,y,z coordinates: 19.5,9,-4.5; cluster size = 1398 voxels), the other extending from the left
parietal lobe towards the superior temporal gyrus, with a peak value in the left Rolandic
operculum (MNI x,y,z coordinates: -63,-12,12; cluster size=1168 voxels) (Figure 2). These
clusters were no longer significant when TBV was controlled for, or when outliers in the dot
memory score were excluded. No significant associations were found with the span board or
composite VSTM scores.
Figure 2: Assocation between dot memory task and voxel-wise GMV. Correlations between
dot memory task performance and voxel-wise GMV in the whole-brain voxel-based
morphometry analysis. Significant clusters are visualized as red overlays on a smoothed 3D
image of the MNI template brain (p<0.05).
4. Discussion
Motivated by predictions about dorsal stream vulnerability and map theories of VSTM capacity,
this study assessed correlations between VSTM performance and GMV in V1 and posterior IPS
regions. For comparison with previous studies, we also assessed associations between voxel-wise
GMV and VSTM at the whole-brain level. We found that TBV, but not individual regions, was a
significant predictor of performance on the dot memory and VSTM composite scores in a sample
of healthy young adults. At the whole-brain level, we found that GMV of clusters centered
around the right putamen and left Rolandic operculum was associated with dot memory scores,
though only when TBV was not accounted for. These results suggest that associations between
VSTM and regional GMV are weak when TBV is accounted for.
The findings in this study are in contrast to recent work showing a significant association
between V1 volume and VSTM capacity (Bergmann et al., 2016). We note however that this
study did not control for TBV in their analyses. Konstantinou et al., (2017) conducted a whole
brain analysis to identify neuroanatomical correlates of location VSTM while controlling for
TBV. They found that GMV in the left inferior parietal lobule, but not in V1, associated with
VSTM. These findings are similar to that of our whole-brain analysis, wherein we found that
performance on the dot memory task associated with GMV of left parietal and temporal regions,
including a portion of the left inferior parietal lobule. Interestingly however, this association was
only significant when TBV was not controlled for. Taken together these two studies and the one
reported here suggest that regional GMV-VSTM associations may be relatively weak when TBV
is considered, particularly in V1.
It is worth noting that discrepancies between our results and those of previous studies (Bergmann
et al., 2016; Konstantinou et al., 2017) could be due to differences in the tasks utilized to assess
VSTM. Konstantinou et al., (2017) showed that object and location VSTM tasks are dependent
on the structure of different cortical regions. This might mean that our findings are more
comparable to that of Konstantinou et al., (2017), who similarly utilized a location VSTM task,
as compared to that of Bergmann et al., (2016), who utilized an object VSTM task. While this
might limit the comparability of our results to that of Bergmann et al., (2016), our reported
associations between dot memory performance and V1 GMV (r=0.374) suggests that our studies
may be similarly sensitive to GMV-VSTM relationships, though we additionally found that the
specificity of regional associations was reduced when TBV was accounted for (r=0.185).
Another thing to consider in comparing tasks across studies is the extent to which executive
functions play a role. It could be argued that tasks involving recollection of both location and
order of stimuli, as in our span board task, might recruit cognitive control (Marshuetz & Smith,
2006), making direct comparisons more difficult.
Despite these considerations, we feel that an important contribution of the present work is to
demonstrate the importance of considering TBV when calculating brain volume -- behavior
correlations. We report correlation sizes before and after this correction so that future studies
investigating this question may be adequately powered. We further suggest that our data do not
necessarily rule out the map account of VSTM capacity (Franconeri et al., 2013), but that
perhaps the VSTM-regional GMV relationship is more subtle than previously suggested, or that
a distributed set of regions beyond V1 and the posterior IPS may play a role in mediating inter-
individual capacity differences.
A limitation of this work is relatively small sample size. Although the study is adequately
powered to detect effects of the size reported in previous work (i.e. 86% power to detect r=0.44
anatomical V1 volume association in Bergmann et al.; calculated using gPower), correlations in
small samples are prone to inaccurate estimates (Rousselet, basic statistics blog, 2018) and
should be interpreted with caution. Considering this, and for the sake of transparency, we report
correlations with the dot memory task with and without three lower limit outlier scores included.
While removal of these outliers did affect the results, it is important to point out that the three
participants in question had a poor performance on both the dot memory and span board tasks,
suggesting that these scores accurately reflect the participant's VSTM capacity, and are not due
to measurement error.
A second limitation of this work is the use of probabilistic/anatomically defined regions of
interest, rather than functionally localized regions. The main reason for this choice is that
preliminary analysis of functional localizer data collected on this sample (Bray et al., 2015)
showed significant associations between performance on the functional localizer task (essentially
a vigilance task) and blood-oxygen-level-dependent (BOLD) response (unpublished). That is,
participants who were perhaps more fatigued and therefore less attentive showed lower BOLD
amplitude, which could introduce bias. Indeed, studies on visual attention and working memory
have shown associations between concurrent performance and BOLD response (Pessoa et al.,
2002; Roalf et al., 2014). We therefore chose to use probabilistic ROIs (Wang et al., 2014),
which would no suffer from this issue. However, there is a trade-off, as the mapping to
individual neuroanatomy is less accurate. Though Bergmann et al., (2016) found similar results
in their study when using functionally vs anatomically defined regions, suggesting this may not
have influenced our results.
VSTM is an important cognitive capacity and understanding it's neuroanatomical correlates has
both theoretical importance and potential clinical utility. In the present work, we found that
posterior IPS regions did not associate with VSTM, and relationships between regional GMV
and VSTM was weakened when TBV was controlled for. Our results suggest that regional
GMV-VSTM associations, particularly in V1, may be weaker than previously suggested.
Moreover, we found a significant association between VSTM and TBV, suggesting VSTM may
be more accurately described by distributed representations, in line with findings that VSTM
tasks engage distributed cortical regions (e.g. Dotson et al., 2018; Pessoa et al., 2002).
Declaration of interests
The authors declare that they have no competing interests.
Funding
This work was supported by an NSERC Postdoctoral Fellowship awarded to SB, an NSERC
CGSD award and AIHS Graduate Studentship to DD, and the Hotchkiss Brain Institute. The
funding sources were not involved in any aspect of the study, including; study design, data
collection, analysis, interpretation of results, writing of the manuscript or decision to publish.
Acknowledgements
We gratefully acknowledge the participation of all volunteers, and staff at the Seaman Family
MR Research Centre; Dr. Mounir Nour and Eileen Pyra.
References
Ashburner, J., & Friston, K. J. (2000) Voxel-based morphometry--the methods Neuroimage.
Neuroimage, 11(6 Pt 1), 805-21.
Ashburner, J. (2007) A fast diffeomorphic image registration algorithm Neuroimage.
Neuroimage, 38(1), 95-113.
Atkinson, J., Braddick, O., Anker, S., Curran, W., Andrew, R., Wattam-Bell, J., & Braddick, F.
(2003) Neurobiological models of visuospatial cognition in children with Williams syndrome:
measures of dorsal-stream and frontal function Dev Neuropsychol. Dev Neuropsychol, 23(1-2),
139-72.
Attout, L., Noël, M. P., Nassogne, M. C., & Rousselle, L. (2017) The role of short-term memory
and visuo-spatial skills in numerical magnitude processing: Evidence from Turner syndrome
PLoS ONE. PLoS ONE, 12(2), e0171454.
Baddeley, A. (2003) Working memory: looking back and looking forward Nat. Rev. Neurosci.
Nat. Rev. Neurosci, 4(10), 829-39.
Bergmann, J., Genç, E., Kohler, A., Singer, W., & Pearson, J. (2016) Neural Anatomy of
Primary Visual Cortex Limits Visual Working Memory Cereb. Cortex. Cereb. Cortex, 26(1), 43-
50.
Blair, J. R., & Spreen, O. (1989) Predicting premorbid IQ: a revision of the National Adult
Reading Test The Clinical Neuropsychologist. The Clinical Neuropsychologist.
Braddick, O., Atkinson, J., & Wattam-Bell, J. (2003) Normal and anomalous development of
visual motion processing: motion coherence and 'dorsal-stream vulnerability' Neuropsychologia.
Neuropsychologia, 41(13), 1769-84.
Bray, S., Almas, R., Arnold, A. E., Iaria, G., & MacQueen, G. (2015) Intraparietal sulcus
activity and functional connectivity supporting spatial working memory manipulation Cereb.
Cortex. Cereb. Cortex, 25(5), 1252-64.
Bray, S., Arnold, A. E., Iaria, G., & MacQueen, G. (2013) Structural connectivity of visuotopic
intraparietal sulcus Neuroimage. Neuroimage, 82, 137-45.
Bray, S., Dunkin, B., Hong, D. S., & Reiss, A. L. (2011) Reduced functional connectivity during
working memory in Turner syndrome Cereb. Cortex. Cereb. Cortex, 21(11), 2471-81.
Buchanan, L., Pavlovic, J., & Rovet, J. (1998) A reexamination of the visuospatial deficit in
turner syndrome: Developmental Neuropsychology. Developmental Neuropsychology, 14(2),
341 -- 367.
Conway, A. R., Cowan, N., & Bunting, M. F. (2001) The cocktail party phenomenon revisited:
the importance of working memory capacity Psychon Bull Rev. Psychon Bull Rev, 8(2), 331-5.
Corbetta, M., & Shulman, G. L. (2002). Control of goal-directed and stimulus-driven attention in
the brain. Nat Rev Neurosci, 3(3), 201-215
Cornoldi, C., Marconi, F., & Vecchi, T. (2001) Visuospatial working memory in Turner's
syndrome Brain And Cognition. Brain And Cognition, 46(1-2), 90-94.
Costanzo, F., Varuzza, C., Menghini, D., Addona, F., Gianesini, T., & Vicari, S. (2013)
Executive functions in intellectual disabilities: a comparison between Williams syndrome and
Down syndrome Res Dev Disabil. Res Dev Disabil, 34(5), 1770-80.
Cowan, N. (2001) The magical number 4 in short-term memory: a reconsideration of mental
storage capacity Behav Brain Sci. Behav Brain Sci, 24(1), 87-114; discussion 114-85.
Dotson, N. M., Hoffman, S. J., Goodell, B., & Gray, C. M. (2018) Feature-Based Visual Short-
Term Memory Is Widely Distributed and Hierarchically Organized Neuron. Neuron.
Franconeri, S. L., Alvarez, G. A., & Cavanagh, P. (2013) Flexible cognitive resources:
competitive content maps for attention and memory Trends Cogn. Sci. (Regul. Ed.). Trends
Cogn. Sci. (Regul. Ed.), 17(3), 134-41.
Girard, T. A., Wilkins, L. K., Lyons, K. M., Yang, L., & Christensen, B. K. (2018) Traditional
test administration and proactive interference undermine visual-spatial working memory
performance in schizophrenia-spectrum disorders Cogn Neuropsychiatry. Cogn Neuropsychiatry,
1-12.
Greenberg, A. S., Verstynen, T., Chiu, Y. C., Yantis, S., Schneider, W., & Behrmann, M. (2012)
Visuotopic cortical connectivity underlying attention revealed with white-matter tractography J.
Neurosci. J. Neurosci, 32(8), 2773-82.
Konstantinou, N., Constantinidou, F., & Kanai, R. (2017) Discrete capacity limits and
neuroanatomical correlates of visual short-term memory for objects and spatial locations Hum
Brain Mapp. Hum Brain Mapp, 38(2), 767-778.
Lauritzen, T. Z., D'Esposito, M., Heeger, D. J., & Silver, M. A. (2009). Top-down flow of visual
spatial attention signals from parietal to occipital cortex. J Vis, 9(13), 18.11-14.
Marshuetz, C., & Smith, E. E. (2006). Working memory for order information: multiple
cognitive and neural mechanisms. Neuroscience, 139(1), 195-200.
Mazziotta, J. C., Toga, A. W., Evans, A., Fox, P., & Lancaster, J. (1995). A probabilistic atlas of
the human brain: theory and rationale for its development. The International Consortium for
Brain Mapping (ICBM). Neuroimage, 2(2), 89-101.
Miyake, A., Friedman, N. P., Rettinger, D. A., Shah, P., & Hegarty, P. (2001) How are
visuospatial working memory, executive functioning, and spatial abilities related? A latent-
variable analysis Journal Of Experimental Psychology-general. Journal Of Experimental
Psychology-general, 130(4), 621-640.
Molko, N., Cachia, A., Riviere, D., Mangin, J. F., Bruandet, M., Le Bihan, D., Cohen, L., &
Dehaene, S. (2003) Functional and structural alterations of the intraparietal sulcus in a
developmental dyscalculia of genetic origin Neuron. Neuron, 40(4), 847-58.
Molko, N., Cachia, A., Rivière, D., Mangin, J. F., Bruandet, M., Le Bihan, D., Cohen, L., &
Dehaene, S. (2003) Functional and structural alterations of the intraparietal sulcus in a
developmental dyscalculia of genetic origin Neuron. Neuron, 40(4), 847-58.
Nelson, H. E., & Willison, J. (1991) National Adult Reading Test (NART).
Park, D. C., Lautenschlager, G., Hedden, T., Davidson, N. S., Smith, A. D., & Smith, P. K.
(2002) Models of visuospatial and verbal memory across the adult life span Psychol Aging.
Psychol Aging, 17(2), 299-320.
Pessoa, L., Gutierrez, E., Bandettini, P., & Ungerleider, L. (2002) Neural correlates of visual
working memory: fMRI amplitude predicts task performance Neuron. Neuron, 35(5), 975-87.
Riggs, K. J., Simpson, A., & Potts, T. (2011) The development of visual short-term memory for
multifeature items during middle childhood J Exp Child Psychol. J Exp Child Psychol, 108(4),
802-9.
Roalf, D. R., Ruparel, K., Gur, R. E., Bilker, W., Gerraty, R., Elliott, M. A., Gallagher, R. S.,
Almasy, L., Pogue-Geile, M. F., Prasad, K., Wood, J., Nimgaonkar, V. L., & Nimgaonkar, V. L.
(2014) Neuroimaging predictors of cognitive performance across a standardized neurocognitive
battery Neuropsychology. Neuropsychology, 28(2), 161-76.
Rousselet G. (2018) basic statistics blog: Small n correlations cannot be trusted;
https://garstats.wordpress.com/2018/06/01/smallncorr/; Accessed July 11 2018.
Sereno, M. I., Dale, A. M., Reppas, J. B., Kwong, K. K., Belliveau, J. W., Brady, T. J., Rosen, B.
R., & Tootell, R. B. (1995) Borders of multiple visual areas in humans revealed by functional
magnetic resonance imaging Science. Science, 268(5212), 889-93.
Silver, M. A., & Kastner, S. (2009) Topographic maps in human frontal and parietal cortex
Trends Cogn. Sci. (Regul. Ed.). Trends Cogn. Sci. (Regul. Ed.), 13(11), 488-95.
Vogel, E. K., & Machizawa, M. G. (2004) Neural activity predicts individual differences in
visual working memory capacity Nature. Nature, 428(6984), 748-751.
Vogel, E. K., McCollough, A. W., & Machizawa, M. G. (2005) Neural measures reveal
individual differences in controlling access to working memory Nature. Nature, 438(7067), 500-
3.
Wang, L., Mruczek, R. E., Arcaro, M. J., & Kastner, S. (2014) Probabilistic Maps of Visual
Topography in Human Cortex Cereb. Cortex. Cereb. Cortex.
Westerberg, H., Hirvikoski, T., Forssberg, H., & Klingberg, T. (2004) Visuo-spatial working
memory span: A sensitive measure of cognitive deficits in children with ADHD Child
Neuropsychology. Child Neuropsychology, 10(3), 155-161.
Zhao, C., & Gong, G. (2017) Mapping the effect of the X chromosome on the human brain:
Neuroimaging evidence from Turner syndrome Neurosci Biobehav Rev. Neurosci Biobehav Rev,
80, 263-275.
|
1903.09103 | 1 | 1903 | 2019-03-21T16:38:49 | Quantum entanglement in physical and cognitive systems: a conceptual analysis and a general representation | [
"q-bio.NC",
"quant-ph"
] | We provide a general description of the phenomenon of entanglement in bipartite systems, as it manifests in micro and macro physical systems, as well as in human cognitive processes. We do so by observing that when genuine coincidence measurements are considered, the violation of the 'marginal laws', in addition to the Bell-CHSH inequality, is also to be expected. The situation can be described in the quantum formalism by considering the presence of entanglement not only at the level of the states, but also at the level of the measurements. However, at the "local'" level of a specific joint measurement, a description where entanglement is only incorporated in the state remains always possible, by adopting a fine-tuned tensor product representation. But contextual tensor product representations should only be considered when there are good reasons to describe the outcome-states as (non-entangled) product states. This will not in general be true, hence, the entangement resource will have to generally be allocated both in the states and in the measurements. In view of the numerous violations of the marginal laws observed in physics' laboratories, it remains unclear to date if entanglement in micro-physical systems is to be understood only as an 'entanglement of the states', or also as an 'entanglement of the measurements'. But even if measurements would also be entangled, the corresponding violation of the marginal laws (no-signaling conditions) would not for this imply that a superluminal communication would be possible. | q-bio.NC | q-bio |
Quantum entanglement in physical and cognitive systems:
a conceptual analysis and a general representation
Diederik Aerts∗, Jonito Aerts Arguelles∗, Lester Beltran∗, Suzette Geriente∗
Massimiliano Sassoli de Bianchi∗† Sandro Sozzo‡ and Tomas Veloz§∗
Abstract
We provide a general description of the phenomenon of entanglement in bipartite systems, as it manifests
in micro and macro physical systems, as well as in human cognitive processes. We do so by observing that
when genuine coincidence measurements are considered, the violation of the 'marginal laws', in addition
to the Bell-CHSH inequality, is also to be expected. The situation can be described in the quantum
formalism by considering the presence of entanglement not only at the level of the states, but also at the
level of the measurements. However, at the "local" level of a specific joint measurement, a description
where entanglement is only incorporated in the state remains always possible, by adopting a fine-tuned
tensor product representation. But contextual tensor product representations should only be considered
when there are good reasons to describe the outcome-states as (non-entangled) product states. This
will not in general be true, hence, the entangement resource will have to generally be allocated both in
the states and in the measurements. In view of the numerous violations of the marginal laws observed
in physics' laboratories, it remains unclear to date if entanglement in micro-physical systems is to be
understood only as an 'entanglement of the states', or also as an 'entanglement of the measurements'.
But even if measurements would also be entangled, the corresponding violation of the marginal laws (no-
signaling conditions) would not for this imply that a superluminal communication would be possible.
Keywords: Human cognition; quantum structures; quantum measurements; Bell inequalities; entangle-
ment; superposition; marginal laws; no-signaling conditions; marginal selectivity.
1
Introduction
The term 'entanglement' was firstly officially introduced by Schrodinger, in the thirties of the last century.
He described it as a situation of two systems whose states are initially known, which following a temporary
interaction enter into a state where one has a complete knowledge of the state of the bipartite system formed
by their combination, but not anymore (apparently at least) a complete knwowledge of their individual
states, and this even though they may become widely separated in space and therefore, one would expect, in
a condition such that they should possess well-defined individual properties. Schrodinger famoulsy asserted
that he did not consider entanglement as "one but rather the characteristic trait of quantum mechanics,
the one that enforces its entire departure from classical lines of thought" (Schrodinger, 1935).
∗Center Leo Apostel for Interdisciplinary Studies, Brussels Free University, Krijgskundestraat 33, 1160 Brussels (Belgium).
Email addresses: [email protected],[email protected],[email protected],[email protected],[email protected]
†Laboratorio di Autoricerca di Base, via Cadepiano 18, 6917 Barbengo (Switzerland). Email address:[email protected]
‡School of Business and Research Centre IQSCS, University Road, LE1 7RH Leicester (United Kingdom). Email address:
[email protected]
§Universidad Andres Bello, Departamento Ciencias Biol´ogicas, Facultad Ciencias de la Vida, 8370146 Santiago (Chile),
Instituto de Filosof´ıa y Ciencias de la Complejidad IFICC, Los Alerces 3024, Nunoa, Santiago (Chile). Email address:
[email protected]
1
Thirty years later, John Bell derived the famous inequalities that today bear his name, which are able
to test the presence of entanglement in bipartite systems (Bell, 1964). When he did so, there was still an
ongoing and widespread debate in the physics' community regarding the validity of quantum mechanics
as a fundamental theory of our physical reality. Bell himself did not believe that in actual experiments,
like those that would be realized by Aspect's and other experimental groups in the following decades, his
inequalities would be violated (Aspect et al., 1982a; Aspect, 1983; Aspect et al., 1982b; Tittel et al., 1998;
Weihs et al., 1998; Giustina et al., 2013; Christensen et al., 2013; Hensen et al., 2016). But nowadays the
predictions of quantum theory are no longer put into question, not only as regards entanglement, which
has been shown to be preservable over distances of more than a thousand of kilometers (Yin et al., 2017),
but also with respect to many other effects predicted by the theory, like the delocalization of large organic
molecules (Gerlich et al., 2011), just to cite one. On the other hand, the debate about the profound meaning
of the theory never stopped, and in fact has constantly renewed and expanded over the years, so much so
that one can envisage this will produce in the end a Copernican-like revolution in the way we understand
the nature of our physical reality (Deutsch, 1998; Stapp, 2011; Kastner, 2013; Fuchs, 2017; Aerts et al.,
2018). Such debate, however, has not remained confined to physicists or philosophers of science, but also
reached new fields of investigation, in particular that of psychology, due to the development of that research
domain called 'quantum cognition', which saw its beginnings in the nineties of the last century (Aerts &
Aerts, 1995; Aerts et al., 1999; Khrennikov, 1999; Gabora & Aerts, 2002; Atmanspacher et al., 2002; Aerts
& Czachor, 2004; Aerts & Gabora, 2005a,b) and borrowed ideas from quantum physics to develop new
promising models for a variety of cognitive phenomena, also providing in return interesting insights as
regards our understanding of physical systems (Khrennikov, 2010; Busemeyer & Bruza, 2012; Haven &
Khrennikov, 2013; Wendt, 2015).
That said, it is worth observing that since the days of Schrodinger, one of the main elements of dissat-
isfaction was the presence of an irreducible (irremovable) probability in quantum theory. This famously
led Einstein to say, in a letter that he wrote to Max Born in 1926, that God "is not playing at dice"
(Born, 1971). Following the research that was carried out during the last four decades by our group, first
in Geneva and then in Brussels, we can say in retrospect that we do agree with him, as we are today
confident in asserting that quantum mechanics is not at odds with determinism, if the latter is understood
at the global level (Aerts, 1999; Aerts & Sassoli de Bianchi, 2014, 2016). God does not play dice, indeed,
because s/he does not have to, as the irreducible quantum probabilities come from the fact that, as we
will explain in the article, as humans we are forced to play dice when we perform quantum measurements,
being the latter much like (weighted) symmetry breaking processes (the actual breaking the symmetry of
the potential), integrating in their very protocol the presence of fluctuations that cannot be eliminated
without at the same time altering the very nature of what is being measured (Sassoli de Bianchi, 2015).
The view that quantum entities might not always have well-defined values for certain observables, and
this not because we would be ignorant about these values, but because, literally, there would be no actual
values, was considered to be problematic by many, as against the view of realism, although this is truly
a matter of concern only for those adhering to the (we think false) prejudice that our physical reality
should be fully contained in space (or spacetime). However, considering all that we learned from quantum
physics and relativity, this is very likely to be a wrong (or incomplete) view, space being instead only a
very particular theater staging a small portion of our multidimensional reality, typically that populated by
the classic macroscopic entities (Aerts, 1999; Aerts & Sassoli de Bianchi, 2014, 2017).
The presumed antirealism of quantum theory has brought people to investigate whether it would be
possible to substitute quantum theory by so called hidden-variable theories, aiming at explaining the
quantum probabilities as resulting from our lack of knowledge of an underlying (pre-empirical, pre-spatial,
non-spatial) deterministic reality. Bell's work, in the sixties of last century (Bell, 1964, 1966, 1987), fits
into this context of searching for a hidden-variable theory (Genovese, 2005), and was deeply inspired by a
2
situation that was described in 1935 by Einstein, Podolsky and Rosen (EPR), still called today the 'EPR
paradox' (Einstein et al., 1935), although there are no more paradoxes, the situation having been clarified
not only from the experimental point of view (Aspect et al., 1982a; Aspect, 1983; Aspect et al., 1982b;
Tittel et al., 1998; Weihs et al., 1998; Giustina et al., 2013; Christensen et al., 2013; Hensen et al., 2016),
but also from the logical one (Aerts, 1984; Sassoli de Bianchi, 2019a).
The situation put forward in the EPR article was later reformulated by David Bohm, using the clearer
example of two spins in an entangled spin state (Bohm, 1951), which is today considered to be the archetypi-
cal quantum entanglement situation, and we will also consider such example in our discussion in this article.
Regarding how the Bohm model situation is described by many working physicists, we observe that there
is still a disturbing "schizophrenia" about how the two entangled spin entities in a Bohm-EPR set up are
viewed. On the one hand, there is agreement in acknowledging that two spins, even though separated
in spatial terms, nevertheless form a single and whole system. On the other hand, there is difficulty in
accepting the consequence of such statement, implying that our spatial theater, as we said already, can
only be viewed as the tip of a much vaster non-spatial reality, which cannot be fully represented in space,
hence, cannot be understood only in terms of spatio-temporal phenomena akin to localized particles or
extended waves and fields.
In another letter to Max Born, Einstein wrote in 1947 (Born, 1971): "I admit, of course, that there is
a considerable amount of validity in the statistical approach which you were the first to recognise clearly
as necessary given the framework of the existing formalism. I cannot seriously believe in it because the
theory cannot be reconciled with the idea that physics should represent a reality in time and space, free
from spooky actions at a distance. [··· ] I am quite convinced that someone will eventually come up with a
theory whose objects, connected by laws, are not probabilities but considered facts, as used to be taken for
granted until quite recently." In a later commentary, Born wrote that the decisive sentence in Einstein's
letter (Born, 1971): "[··· ] is the one where he says: 'that physics should represent a reality in time and
space, free from spooky actions at a distance'. I too had considered this postulate to be one which could
claim absolute validity. But the realities of physical experience had taught me that this postulate is not an
a priori principle but a time-dependent rule which must be, and can be, replaced by a more general one."
This more general rule, mentioned by Born, asks us to abandon the 'space contains reality hypoth-
esis' and to accept what follows from the quantum formalism and its numerous successful tests in the
laboratories, i.e., that (Aerts, 1999): "reality is not contained within space. Space is a momentaneous
crystallization of a theatre for reality where the motions and interactions of the macroscopic material and
energetic entities take place. But other entities -- like quantum entities for example -- 'take place' outside
space, or -- and this would be another way of saying the same thing -- within a space that is not the
three-dimensional Euclidean space."
Ironically enough, Einstein's reality criterion (Einstein et al., 1935) provides one of the strong argu-
ments that intimate us to accept the non-spatiality of the quantum micro-entities. Indeed, Heisenberg's
uncertainty principle prevents us to simultaneously define both the position and momentum of a quan-
tum entity like an electron. Therefore, one cannot determine, not even in principle, how the position and
momentum of the entity will vary in time, and consequently one cannot predict with certainty, not even
in principle, its future locations. Following Einstein's reality criterion, we have then to conclude that the
entity in question doesn't possess the property of being somewhere in space, hence it would be a non-spatial
entity, which does not mean, however, that it would be an unreal entity (Sassoli de Bianchi, 2011).
Many additional arguments can be brought forward in support of the thesis that quantum entities should
be considered to be non-spatial, like those following from a study of their temporal behaviors, by means
of the notion of sojourn time (Sassoli de Bianchi, 2012), or from an analysis of spins greater than one-half,
which cannot be associated with any specific spatial direction (Aerts & Sassoli de Bianchi, 2017), and of
course, there are also the many no-go theorems, in particular those of Kochen & Specker (1967), which if
3
taken seriously tell us just that: that quantum entities cannot be depicted as the factual objects connected
by laws of Einstein's desiderata, being instead more like entities having an unexpected "conceptual nature,"
being able to manifest in states having a varying degree of abstractness or concreteness, the more concrete
ones being those we usually describe as the classical spatio-temporal objects of our ordinary experience
(Aerts et al., 2018).
It is certainly not the purpose of the present paper to enter into a comprehensive discussion of the
non-spatial and conceptual behavior of quantum entities (see also, in that respect, the perspective offered
by Kastner's possibilistic transactional interpretation of quantum mechanics (Kastner, 2013)). We only
want here to emphasize that it would be wrong to consider that a physical entity, to be real, has to exist in
space. If we let go such "classical prejudice," then, when studying the phenomenon of entanglement, one
is not forced any more, as Einstein considered to be, to understand two entities that are separated in space
as two entities that would be necessarily disconnected to one another. If their state is non-spatial, then
the nature of their possible connection will simply be non-spatial as well, i.e., non-manifest as a connection
through space. If this is so, then there is no need to speak of "spooky actions at a distance," as there
would be no phantom-like action of one entity over the other, during a Bell-test experiment. More simply,
Alice and Bob,1 with their instruments, localized in different regions of space, would both act on a same
entity, which forms a whole, and not on two independent entities.
In other words, it would be wrong to conceive a bipartite entity in an entangled state as two fully
separated entities, just because they can respond to different instruments placed at distant locations, and
that in between these locations the probability of a detection tends to zero, once the entangled entities
have been emitted by the source and have propagated away. If outcomes can be actualized in a coincident
and correlated way, in distant detectors, this is because the two apparent spatially divided fragments are
the tip of an undivided non-spatial entity, having some well-defined degree of availability in interacting
with different spatial measuring instruments, and by doing so acquiring spatial properties (for instance by
leaving a trace in a detection screen). If this is correct, then there is no reason to speak in terms of an
action, or influence, of Alice's measurement on Bob's measurement, and vice versa, as they would both
operate on a same undivided entity, at the same time. There is no influence of one measurement on the
other, only a single measurement jointly performed by Alice and Bob, on a whole entity. The latter can
then remain whole, or possibly disentangle, following their joint action, depending on the entity's nature
and on the experimental operations. When joint measurements on entangled entities are understood in
this way, there is also no reasons to require that Alice's statistics of outcomes would be independent of the
choice of actions operated by Bob, i.e., that the system would necessarily obey the so-called 'no-signaling
conditions' (also called 'marginal laws', or 'marginal selectivity').
These no-signaling conditions are implicitly assumed to be valid in the standard quantum formalism,
when joint measurements are represented by tensor product observables, defined with respect to a same
given tensor product decomposition of the Hilbert space. However, it remains today unclear if this is
the correct way to model certain experimental situations, considering that significant violations of the
no-signaling conditions have been evidenced in the physics' laboratories (Adenier & Khrennikov, 2007; De
Raedt et al., 2012, 2013; Adenier & Khrennikov, 2017; Bednorz, 2017; Kupczynski, 2017). These violations
are totally unexpected if the adopted view is that in which Alice and Bob perform independent (although
coincident in time) measurements, instead of a bigger unique joint measurement. In other words, if we think
of Alice's and Bob's measurements as two distinct interrogative contexts, asking different and independent
questions, then there is indeed no reasons to expect that the statistics of answers collected by Alice could
be influenced by the questions that are asked by Bob.
1Alice and Bob are the two archetypical fictional characters used to described, in physics and cryptology, the joint ex-
perimental actions and data collections resulting from their sharing of a same bipartite system, typically in an entangled
state.
4
The usual understanding of quantum entanglement in physics is indeed that of a situation where we have
a bipartite entity emitted by a source, like the two spin one-half fermions of Bohm's archetypical model,
such that when they are flying apart, they are assumed to become fully independent. If this would be true,
it would be natural to think that the choice of measurement performed by Alice on its sub-system could
not be influenced by the choice of measurement performed by Bob, and vice versa. If the two fermionic
entities are fully separated (spatially and experimentally) this is indeed something to be expected, which
then translates in the constraints of the no-signaling conditions. The same reasoning would however lead
one to also expect that no correlations of the kind able to violate Bell's inequality should be observed
between Alice's and Bob's answers, which is why Einstein described, and many physicists nowadays still
describe, the situation as a "spooky action at a distance." The action is considered to be "spooky" because
no force field seems to be involved in it, but would just happen as a consequence of the linear structure of
the quantum state space.
But it seems that there is a limit in the "spookiness" that many physicists are ready to digest: it can
be spooky enough to violate Bell's inequalities, but not so spooky to also violate the marginal laws. Why?
Because these laws are the conditions that guarantee that no faster than light communications can arise
between Alice and Bob. Still, the very idea of Alice sending a signal to Bob, or vice versa, relies on Alice
and Bob selectively acting only on their sub-entities, which in turn presupposes some sort of separateness
of these two sub-entities, which is only reasonable to assume if they would be genuine spatial entities,
as then their separation in space would be sufficient to also produce a "separation in substance." As we
emphasized, this is however an untenable assumption, hence not only the violation of Bell's inequalities is
to be expected, but also a violation of the no-signaling conditions should be expected, unless the system
possesses some remarkable symmetries (which can indeed be the case for the quantum micro-systems, in
case the observed violations of the no-signaling conditions would only be the result of experimental errors).
This does not imply, however, that there would be the possibility to exploit these violations in order to
produce a superluminal communication between Alice and Bob. As we will explain in this article, we
believe that a subtle logical error has been made in the standard analysis of the no-signaling conditions.
It is usually considered that since the produced correlations in a typical Bell-test situation are spacelike
separated, if they could be used to send signals, then such signals would travel faster than light, hence they
would violate relativity theory. The error in question consists in not carefully distinguishing between the
'origin of the correlations' and the 'mechanism of signaling by means of such correlations'. Indeed, even if
correlations are associated with spacelike separated events, the mechanism of using them for signaling will
not necessarily lead to a faster than light propagation.
What above described leads us to consider another domain of investigation, that of human cognition,
where it was also observed that experimental situations can be created where Bell's inequalities are violated
(Bruza et al., 2008, 2009; Aerts & Sozzo, 2011, 2014a; Bruza et al., 2015; Gronchi & Strambini, 2017; Aerts
et al., 2018a,b; Beltran & Geriente, 2019; Aerts et al., 2019). This is so because there can be more or less
strong (non-spatial) connections between different conceptual entities, depending on how much meaning
they share. In other words, meaning connections in the conceptual realm are the equivalent of the (non-
spatial) coherence-connection shared by entangled micro-physical entities, and can be exploited to create
correlations in well-designed psychological experiments. Generally, these experiments will also violate the
marginal laws, hence, in their modeling one cannot simply use the standard quantum representation, with
a single tensor product representation for all considered joint measurements.
This caused some authors to doubt that a genuine form of entanglement is at play in cognitive systems
(Dzhafarov & Kujala, 2013; Aerts, 2014; Aerts et al., 2018b). For instance, in the abstract of Cervantes
& Dzhafarov (2018), the authors write: "All previous attempts to find contextuality in a psychological
experiment were unsuccessful because of the gross violations of marginal selectivity in behavioral data,
making the traditional mathematical tests developed in quantum mechanics inapplicable." In our opinion,
5
statements of these kind rest on the aprioristic view that entanglement should be caused by some sort of
("spooky") information flow from Alice to Bob, or vice versa, rather than by a process where Alice and
Bob are able to jointly co-create information/meaning, by acting at once on a same whole entity, the latter
being of course a process that is expected to also generally violate the marginal laws. So, we could say
that, in a sense, the same kind of misunderstanding seems to be at play in the analysis of the entanglement
phenomenon both in the physical and psychological laboratories, when some physicists and psychologists
try to figure out what could be able to generate the observed correlations.
It is the purpose of the present paper to bring some clarity to all this, presenting the phenomenon of
entanglement and its quantum modeling, both in physical and cognitive situations, under the light of a
unified and general perspective, resulting from the work that our group has carried out in the last decades,
aimed at understanding the foundations of physical theories and of our human cognitive processes, always
bringing particular attention to the aspects that unite these two domains. To do so, in Sec. 2, we start by
describing three paradigmatic examples of bipartite systems violating the CHSH version of Bell's inequality,
which we will use to exemplify our analysis throughout the paper. The first one (Sec. 2.1) is Bohm's spin
model of two spin- 1
2 entities in a singlet state. The second one (Sec. 2.2) is a variation of Aerts' "vessels of
water" macroscopic model, which involves an elastic entity formed by multiple bands. The third example
(Sec. 2.3) is a 'psychological model', exploiting the meaning connection characterizing a given conceptual
combination.
In Sec. 3, we then provide a probabilistic definition of sub-measurements, and make explicit the marginal
laws (or no-signaling conditions), showing that they are obeyed in Bohm's spin model example, because of
the assumed product (non-entangled) structure of the measurements within the customary tensor product
description of the situation, but violated in the elastic and the psychological models. In Sec. 4, we also
define the notions of compatibility and separability, showing that when all joint measurements are formed
by separate sub-measurements, both the CHSH inequality and the marginal laws are necessarily obeyed. In
Sec. 5, we continue our analysis by introducing and explaining the important distinction between correla-
tions of the first and second kind, showing that only the latter, which operate at the (non-local) level of the
whole entangled entity, are able to violate the CHSH inequality, without for this implying a superluminal
influence traveling from Alice to Bob, or vice versa, i.e., with no "spooky action at a distance" and no
violations of relativistic principles, even if the marginal laws are also violated.
In Sec. 6, we show that the standard quantum formalism allows one to model the joint measurements
even in situations where the marginal laws are disobeyed, emphasizing that a tensor product structure
is always relative to the choice of a specific isomorphism and that when the marginal laws are violated
a single isomorphism is insufficient to introduce a tensor product structure for all joint measurements.
Hence, some will necessary appear to be entangled measurements. In Sec. 7, a specific modeling example
is provided, and in Sec. 8 the issue of how and when to introduce isomorphisms in order to tensorialize
specific joint measurements is discussed, emphasizing that it should be limited to those situations where
there is evidence that the outcome-states are product states, i.e., describe a disentanglement of the system
as produced by the measurements.
In Sec. 9, we lean on the problem of the definition of sub-measurements when also the change of state
induced by them, as described by the quantum projection postulate, is taken into due consideration, em-
phasizing that in quantum mechanics sub-measurements do not arise as a simple procedure of identification
of certain outcomes. Finally, in Sec. 9, we offer a number of concluding remarks, trying to bring attention
to some important points that emerged from our analysis.
6
2 Bipartite systems and joint measurements
Our discussion in this article will be limited to a particular class of systems called 'bipartite systems', or
'bipartite entities'. As the name indicates, these are systems (or entities, we will use these two terms inter-
changeably in the article) of a composite nature, i.e., in which two parts, called sub-systems, or sub-entities,
can be identified at some level. When considering bipartite entities, one can adopt a "deconstructivist view-
point," where the starting point is a single whole entity and one considers to which extent such entity can
be understood as a system formed by two parts, or a complementary "constructivist viewpoint," where one
starts from two clearly distinguishable entities that are brought together in some way, and one considers
to which extent they can be understood as the sub-entities forming a bigger emerging composite system.
Independently of the viewpoint considered, what is important for our analysis is that the system
possesses some properties that are characteristics of a bipartite structure. For example, two electrons, even
when in a singlet (entangled) state, form a bipartite system because even though we cannot attach individual
vector-states to each one of the electrons (one can nevertheless attach individual density operator-states
(Aerts & Sassoli de Bianchi, 2016)), there are properties that remain always actual and characterize the
bipartiteness of the system, like the fact that we are in the presence of two electronic masses and two
electric charges, that the system can produce two distinct detection events on spatially separated screens,
instead of a single one, etc. In the present discussion, we will limit our analysis to three different kinds
of bipartite systems, which will help us to illustrate the different aspects of our approach to quantum
entanglement and its modeling. The first system is David Bohm's archetypical example of two half-spins
in an entangled state. The second system is a macroscopic elastic structure presenting two distinguishable
ends, on which it is possible to act simultaneously, and the third one is a combination of two abstract
concepts, subjected to participants in a psychological experiment.
2.1 Bohm's spin model
In the historical discussion of their paradoxical situation, EPR considered the position and momentum
observables. As we mentioned in the Introduction, David Bohm subsequently proposed a simpler situation,
which expresses the situation equally well: that of two spin- 1
2 entities (two fermions) in a rotationally
invariant entangled state (a so-called singlet state), the wave function of which can be written as:
ψ(r1, r2)s(cid:105) =
1√
2
ψ(r1)ψ(r2)(+(cid:105) ⊗ −(cid:105) − −(cid:105) ⊗ +(cid:105),
(1)
where ψ(r1) and ψ(r2) are the spatial components of the wave functions of the two fermions, and +(cid:105) and
−(cid:105) are the "up" and "down" eigenstates of the spin operators relative to some given spatial direction.
The spatial component ψ(r1, r2) is of course important in order to describe the evolution of the composite
entity in relation to space. However, we will focus here only on the spinorial component, being understood
that the spatial factor describes two entities emitted by a source that propagate away from each other,
with some average velocity, in opposite directions, towards some distant Stern-Gerlach apparatuses and
the associated detection screens. So, we will more simply consider that the state of the bipartite system
is:
(2)
s(cid:105) =
1√
2
(+(cid:105) ⊗ −(cid:105) − −(cid:105) ⊗ +(cid:105).
It is worth observing that the above spin vector-state, being rotationally invariant, cannot be associated
with well-defined individual spin properties prior to the measurements (we will come back to this point
later in our discussion).
In the typical experimental setting of a Bell-test experiment, one has four joint measurements, which
we simply denote AB, AB(cid:48), A(cid:48)B and A(cid:48)B(cid:48). In the joint measurement AB, the spin entity moving to the
7
left (let us call it the spin measured by Alice) is subjected to a Stern-Gerlach apparatus oriented along the
A-axis, whereas the spin entity moving to the right (let us call it the spin measured by Bob) is subjected
to a Stern-Gerlach apparatus whose magnet is oriented along the B-axis. The same holds for the other
three joint measurements, which use Stern-Gerlach apparatuses also oriented along the A(cid:48) and B(cid:48) axes.
If α is the angle between the A and B axes, then according to the quantum formalism the probabilities
p(A1, B1) and p(A2, B2) that Alice and Bob jointly obtain a spin up outcome, respectively jointly obtain
a spin down outcome, are:
On the other hand, the probabilities p(A1, B2) and p(A2, B1) of Alice finding a spin up and Bob a spin
down, respectively Alice finding a spin down and Bob a spin up, are given by:
p(A1, B1) = p(A2, B2) = 1
2 sin2 α
2 .
(3)
p(A1, B2) = p(A2, B1) = 1
2 sin2 π−α
2 = 1
2 cos2 α
2 .
(4)
If an angle α = π
4 is considered, one then finds:
p(A1, B1) = p(A2, B2) = 1
8 (2 − √
(5)
An optimal choice for the A(cid:48) and B(cid:48) measurements, maximizing the correlations, is to also consider an
4 between the B and A(cid:48) axes, with angles of π
angle of 3π
between A and A(cid:48) and between B and B(cid:48); see Fig. 1. This gives the additional probabilities:
4 between the A and B(cid:48) axes, and an angle of π
p(A1, B2) = p(A2, B1) = 1
8 (2 +
2),
2).
2
√
Figure 1: The four orientations of the Stern-Gerlach apparatuses used by Alice and Bob in their joint measurements, which
maximize the violation of the CHSH inequality.
p(A1, B(cid:48)
p(A1, B(cid:48)
2) = p(A2, B(cid:48)
1) = p(A2, B(cid:48)
1) = p(A(cid:48)
2) = p(A(cid:48)
1, B1) = p(A(cid:48)
1, B2) = p(A(cid:48)
2, B2) = p(A(cid:48)
2, B1) = p(A(cid:48)
1, B(cid:48)
1, B(cid:48)
1) = p(A(cid:48)
2) = p(A(cid:48)
2, B(cid:48)
2, B(cid:48)
(6)
By attributing the value +1 to the situation where both spins are either up or down, and the value −1
to the situation where one spin is up and the other spin is down, we can calculate the following 'expectation
values' (also called 'correlation functions') of the four joint measurements:
8 (2 +
2).
2) = 1
1) = 1
8 (2 − √
2),
√
E(A, B) = p(A1, B1) − p(A1, B2) − p(A2, B1) + p(A2, B2) = − 1√
E(A, B(cid:48)) = p(A1, B(cid:48)
2) = 1√
2, B2) = − 1√
E(A(cid:48), B) = p(A(cid:48)
2) = − 1√
E(A(cid:48), B(cid:48)) = p(A(cid:48)
2, B(cid:48)
1, B2) − p(A(cid:48)
2) − p(A(cid:48)
1, B(cid:48)
1, B1) − p(A(cid:48)
1) − p(A(cid:48)
1, B(cid:48)
2, B1) + p(A(cid:48)
2, B(cid:48)
1) + p(A(cid:48)
1) − p(A1, B(cid:48)
2) − p(A2, B(cid:48)
1) + p(A2, B(cid:48)
2
2
2
2
.
(7)
8
The Clauser Horne Shimony Holt (CHSH) version of Bells inequality then says that the quantity (Clauser
et al., 1969):
CHSH ≡ E(A, B) − E(A, B(cid:48)) + E(A(cid:48), B) + E(A(cid:48), B(cid:48)),
(8)
or similar expressions obtained by interchanging the roles of A and A(cid:48) and/or the roles of B and B(cid:48), is
bounded by:
−2 ≤ CHSH ≤ 2.
√
= −2
(9)
Since (7) gives CHSH = − 4√
2 < −2, the CHSH inequality (9) is clearly violated by Bohm's spin
model. The violation corresponds here to the value known as Tsirelson's bound, which is a maximal value
for the quantum correlations, for as long as the no-signaling conditions are fulfilled (see Sec. 3), which
will always be the case if measurements are represented as product observables, as we will discuss in more
detail later in the article.
2
2.2 An elastic band model
The second kind of bipartite system that we intend to consider is a macroscopic physical entity. Note
that since the eighties of the last century, it was observed that the quantum laws are not the exclusive
prerogative of the micro-entities, or of the very low temperature regimes, being possible to describe ide-
alized macroscopic physical entities having a genuine quantum-like behavior, resulting from how certain
non-standard experiments, perfectly well-defined in operational terms, can be performed on them, giving
rise to non-Kolmogorovian probability models (Aerts, 1986; Aerts et al., 1997; Aerts, 1998, 1999; Sassoli
de Bianchi, 2013a,b). Some of these "quantum machine" models were also studied in order to better un-
derstand the phenomenon of entanglement, as it is possible to conceive classical laboratory situations able
to violate Bell's inequalities, thus throwing some light on the possible mechanisms at the origin of the
observed correlations (Aerts, 1982, 1984, 1991; Aerts et al., 2000; Aerts, 2005; Sassoli de Bianchi, 2013b;
Aerts & Sassoli de Bianchi, 2016; Aerts et al., 2018b). Here we consider a model using breakable elastic
bands, which is a variation of previous similar models (Aerts, 2005; Sassoli de Bianchi, 2013b).
More precisely, the entity we consider, subjected to different joint measurements, is formed by n + 1
uniform elastic bands of same length d, one of which is black, and all the others are white; see Fig. 2.
It can be considered as a bipartite entity because all the elastics are aligned, parallel to each other, thus
Figure 2: A schematic representation of the 'bundle of elastic bands' entity on which Alice and Bob perform the four AB,
A(cid:48)B, AB(cid:48) and A(cid:48)B(cid:48) coincidence measurements.
presenting all their left ends to Alice and all their right ends to Bob. The four joint measurements AB,
AB(cid:48), A(cid:48)B and A(cid:48)B(cid:48) are then defined as follows. Measurement AB consists in Alice and Bob pulling with
force, at the same predetermined time, the black elastic. Outcomes (A1, B2) and (A2, B1) consist in Alice
collecting in this way a fragment of length greater than d/2 and Bob collecting a fragment of length less
than d/2, respectively Alice collecting a fragment of length lesser than d/2 and Bob of length greater than
d/2. Clearly, the elastics being assumed to be uniform, the associated probabilities are the same for Alice
and Bob and equal to one-half. Also, considering that outcome (A1, B1), where Alice and Bob both collect
9
a fragment of length greater than d/2, and outcome (A2, B2), where Alice and Bob both collect a fragment
of length lesser than d/2, cannot be observed, we have:
p(A1, B2) = p(A2, B1) = 1
2 ,
(10)
Measurement AB(cid:48) (A(cid:48)B) is performed in the same way, but this time Bob (Alice) has to select in a
random way the elastic to be pulled (for instance, keeping the eyes closed). This means that it can now
either be the black one or one of the n whites, and therefore Alice and Bob will not necessarily pull the
same elastic. If they do not, they will simply collect the entire elastic, which therefore will be of length
d > d/2. So, this time we have the probabilities:
p(A1, B1) = p(A2, B2) = 0.
1, B1) = p(A1, B(cid:48)
p(A(cid:48)
p(A(cid:48)
1, B2) = p(A2, B(cid:48)
2, B2) = p(A2, B(cid:48)
p(A(cid:48)
1) = n
1) = p(A1, B(cid:48)
2) = p(A(cid:48)
2, B1) = 1
2
n+1 ,
2) = 0,
1
n+1 .
(11)
Note that for the calculation of p(A(cid:48)
1, B1), we observed that the probability for Alice not to grab the same
elastic grabbed by Bob, which is the black one, is given by the number n of white elastics divided by the
total number n + 1 of elastics, i.e., n
Finally, measurement A(cid:48)B(cid:48) consists in Alice and Bob both pulling a random elastic, hence the proba-
n+1 .
bilities are:
p(A(cid:48)
1, B(cid:48)
p(A(cid:48)
2, B(cid:48)
p(A(cid:48)
1, B(cid:48)
2) = p(A(cid:48)
2, B(cid:48)
1
n+1 .
2) = 0,
1, B(cid:48)
1) = 1
2
1) = n
n+1 ,
Note that for the calculation of p(A(cid:48)
1), one has to reason as follows. There are (n + 1)2 ways to grab the
left and right ends of n + 1 elastics. Among these (n + 1)2 ways, n + 1 of them consist in grabbing the left
and right ends of the same elastic. These events have to be excluded, as they cannot produce the outcome
(A(cid:48)
1) outcome,
are (n + 1)2 − (n + 1) = n(n + 1). Dividing this number by the total number (n + 1)2 of possible events,
1, B(cid:48)
one obtains n(n+1)
2)
is only possible for the n + 1 cases where Alice and Bob grab the same elastic, and when this happens the
probability is 1
1). So, the events where Alice and Bob grab a different elastic, which yield the (A(cid:48)
2), one can reason as follows. Outcome (A(cid:48)
n+1 . For the calculation of p(A(cid:48)
n+1 , and the same holds of course for p(A(cid:48)
(n+1)2 = n
2 , hence p(A(cid:48)
2, B(cid:48)
1).
2) = 1
2
1, B(cid:48)
1, B(cid:48)
1, B(cid:48)
1, B(cid:48)
(12)
1
We thus obtain the expectation values:
E(A, B) = −1, E(A, B(cid:48)) = E(A(cid:48), B) = E(A(cid:48), B(cid:48)) = n−1
n+1 .
(13)
Therefore, considering the quantity obtained by interchanging the roles of B and B(cid:48) in (8), i.e., CHSH ≡
−E(A, B) + E(A, B(cid:48)) + E(A(cid:48), B) + E(A(cid:48), B(cid:48)), we find:
n − 1
n + 1
n − 1
2
n + 1
CHSH = 1 + 3
(14)
For n = 0, we have CHSH = −2, so there is no violation, as in this case all four joint measurements are
the same measurement. For n = 1, 2, we have CHSH = 1, 2, respectively, hence there is still no violation.
But for n = 3, 4, 5, 6, . . . , we have CHSH = 5
7 , . . . , respectively, so the inequality is violated for
n > 2, and will be maximally violated (CHSH = 4) in the n → ∞ limit (in this limit, the model becomes
equivalent to the 'vessels of water model' introduced by one of us almost forty years ago (Aerts, 1982)).
5 , 3, 22
2 , 14
= 4
.
2.3 A psychological model
Entanglement has also been identified and extensively investigated in human cognitive processes (Bruza et
al., 2008, 2009; Aerts & Sozzo, 2011, 2014a; Bruza et al., 2015; Gronchi & Strambini, 2017; Aerts et al.,
10
2018a,b; Beltran & Geriente, 2019; Aerts et al., 2019). So, as a third paradigmatic situation, we consider
joint measurements performed by human participants in a psychological experiment. The system is formed
by two distinct concepts (Aerts & Sozzo, 2014a): Animal and Acts, which are combined in a specific
sentence: The Animal Acts. This sentence carries a certain meaning, which corresponds to a specification
of the state of the bipartite system formed by the two concepts Animal and Acts. A different choice of a
sentence containing the two concepts, like for instance The Animal Acts in a Strange Way, would correspond
to a different state, producing different probabilities when performing joint measurements like those we
are now going to describe. For a review of the Brussels' operational-realistic approach to cognition, where
concepts are considered to be entities that can be in different states and be subjected to measurements
performed by cognitive entities sensitive to their meaning, like human minds, we refer the reader to Aerts
et al. (2016), and the references cited therein.
The four joint measurements AB, AB(cid:48), A(cid:48)B and A(cid:48)B(cid:48) are defined as follows. Measurement AB consists
in participants jointly selecting a good example for the concept Animal, from the two possibilities A1 =
Horse and A2 = Bear, and a good example for the concept Acts, from the two possibilities B1 = Growls
and B2 Whinnies. Hence, the four outcomes of AB are: The Horse Growls, The Horse Whinnies, The
Bear Growls and The Bear Whinnies; see Fig. 3. The experimental probabilities obtained in an experiment
Figure 3: A symbolic representation of a participant (in the appearance of David Bohm) performing the joint measurement
AB, selecting one among the four possible combinations of exemplars, here producing a transition from the pre-measurement
state The Animal Acts to the outcome-state The Bear Growls.
performed in 2011 (Aerts & Sozzo, 2011) give:
81 , p(A1, B2) = 51
81 , p(A2, B1) = 21
p(A1, B1) = 4
(15)
The joint measurement AB(cid:48) is carried over in the same way, but considering for the exemplars of Acts the
two possibilities B(cid:48)
2 = Meows. Hence, the four outcomes of AB(cid:48) are: The Horse Snorts,
The Horse Meows, The Bear Snorts and The Bear Meows, and the experimental probabilities obtained in
the 2011 experiment are:
1 = Snorts and B(cid:48)
81 , p(A2, B2) = 5
81 .
p(A1, B(cid:48)
1) = 48
81 , p(A1, B(cid:48)
2) = 2
81 , p(A2, B(cid:48)
1) = 24
81 , p(A2, B(cid:48)
(16)
Similarly, the joint measurement A(cid:48)B considers Growls and Whinnies for the two exemplars of Acts, as in
the AB measurement, but this time A(cid:48)
2 = Cat, for the two possible choices for Animal.
1 = Tiger and A(cid:48)
2) = 7
81 .
11
Hence, the four outcomes of A(cid:48)B are: The Tiger Growls, The Tiger Whinnies, The Cat Growls and The
Cat Whinnies, and the obtained experimental probabilities were in this case:
p(A(cid:48)
(17)
Finally, in joint measurement A(cid:48)B(cid:48) the four outcomes are The Tiger Snorts, The Tiger Meows, The Cat
Snorts and The Cat Meows, and the experimental probabilities obtained in the 2011 experiment are:
2, B2) = 4
81 .
1, B1) = 63
2, B1) = 7
81 , p(A(cid:48)
1, B2) = 7
81 , p(A(cid:48)
81 , p(A(cid:48)
p(A(cid:48)
1, B(cid:48)
1) = 12
81 , p(A(cid:48)
1, B(cid:48)
2) = 7
81 , p(A(cid:48)
2, B(cid:48)
1) = 8
81 , p(A(cid:48)
2, B(cid:48)
2) = 54
81 .
(18)
Calculating the expectation values (7), using the above experimental probabilities, we obtain:
E(A, B) = − 31
(19)
Therefore, considering the quantity CHSH ≡ −E(A, B) + E(A, B(cid:48)) + E(A(cid:48), B) + E(A(cid:48), B(cid:48)), we have the
violation:
81 , E(A, B(cid:48)) = 29
81 , E(A(cid:48), B) = 53
81 , E(A(cid:48), B(cid:48)) = 51
81 .
CHSH =
+
+
+
= 2 +
> 2.
(20)
63
81
29
81
53
81
51
81
=
196
81
34
81
3 Sub-measurements and marginal selectivity
In the above three paradigmatic examples of bipartite systems violating the CHSH inequality, we have each
time defined four joint measurements AB, AB(cid:48), A(cid:48)B and A(cid:48)B(cid:48). These joint measurements are performed
as a whole, in a pure coincidental way. However, one can also consider in more specific terms the sub-
measurements which they are the joint of. As regards the outcome probabilities of these sub-measurements,
they can be obtained by simply identifying certain outcomes of the joint measurements of which they are
part. Consider the case of AB, whose outcomes are defined by the four couples (A1, B1), (A1, B2), (A2, B1)
and (A2, B2). Let us denote by A and B the associated sub-measurements (we will sometimes say that
these are the measurements executed by Alice and Bob, although, as we will further discuss in Sec. 9, there
are some subtleties involved when defining Alice's and Bob's sub-measurements in a fully operational way,
according to the quantum formalism).
Sub-measurement A can be defined as follows:
it is the measurement having the two outcomes A1
and A2, such that A1 is actualized each time that either (A1, B1) or (A1, B2) are actualized, and A2 is
actualized each time that either (A2, B1) or (A2, B2) are actualized. Similarly, sub-measurement B can be
defined as the measurement having the two outcomes B1 and B2, such that B1 is actualized each time that
either (A1, B1) or (A2, B1) are actualized, and B2 is actualized each time that either (A1, B2) or (A2, B2)
are actualized. In probabilistic terms, this means that:
pB(A1) = p(A1, B1) + p(A1, B2),
pA(B1) = p(A1, B1) + p(A2, B1),
pB(A2) = p(A2, B1) + p(A2, B2),
pA(B2) = p(A1, B2) + p(A2, B2).
(21)
Of course, we can do the same with the others three joint measurements, defining the corresponding
(marginal) sub-measurement probabilities. For AB(cid:48) we have sub-measurement A jointly performed with
sub-measurement B(cid:48):
pB(cid:48)(A1) = p(A1, B(cid:48)
1) = p(A1, B(cid:48)
pA(B(cid:48)
1) + p(A1, B(cid:48)
2),
1) + p(A2, B(cid:48)
1),
pB(cid:48)(A2) = p(A2, B(cid:48)
2) = p(A1, B(cid:48)
pA(B(cid:48)
1) + p(A2, B(cid:48)
2),
2) + p(A2, B(cid:48)
2).
(22)
For the joint measurement A(cid:48)B, we have sub-measurement A(cid:48) jointly performed with sub-measurement B:
pB(A(cid:48)
1) = p(A(cid:48)
pA(cid:48)(B1) = p(A(cid:48)
1, B1) + p(A(cid:48)
1, B1) + p(A(cid:48)
1, B2),
2, B1),
pB(A(cid:48)
2) = p(A(cid:48)
pA(cid:48)(B2) = p(A(cid:48)
2, B1) + p(A(cid:48)
1, B2) + p(A(cid:48)
2, B2),
2, B2).
(23)
12
Finally, A(cid:48)B(cid:48) is the joining of the two sub-measurement A(cid:48) and B(cid:48), with marginal outcome probabilities:
pB(cid:48)(A(cid:48)
pA(cid:48)(B(cid:48)
1) = p(A(cid:48)
1) = p(A(cid:48)
1, B(cid:48)
1, B(cid:48)
1) + p(A(cid:48)
1) + p(A(cid:48)
1, B(cid:48)
2),
2, B(cid:48)
1),
pB(cid:48)(A(cid:48)
pA(cid:48)(B(cid:48)
2) = p(A(cid:48)
2) = p(A(cid:48)
2, B(cid:48)
1, B(cid:48)
1) + p(A(cid:48)
2) + p(A(cid:48)
2, B(cid:48)
2),
2, B(cid:48)
2).
(24)
The marginal laws (also called marginal selectivity, or no-signaling conditions) are then the require-
ments that the probabilities of the different sub-measurements do not change when one changes the sub-
measurement with which they are jointly measured. More precisely, the conditions are:
pB(A1) = pB(cid:48)(A1),
pA(B1) = pA(cid:48)(B1),
pB(A2) = pB(cid:48)(A2),
pA(B2) = pA(cid:48)(B2),
pB(A(cid:48)
pA(B(cid:48)
1) = pB(cid:48)(A(cid:48)
1),
1) = pA(cid:48)(B(cid:48)
1),
pB(A(cid:48)
pA(B(cid:48)
2) = pB(cid:48)(A(cid:48)
2),
2) = pA(cid:48)(B(cid:48)
2).
(25)
In the example of Sec. 2.1, all the marginals are equal to 1
2 , hence they do obey the above marginal
laws. The reason for that is that the quantum joint measurements AB, AB(cid:48), A(cid:48)B and A(cid:48)B(cid:48) are associated
with (tensor) product (Pauli matrix) operators σA ⊗ σB, σA(cid:48) ⊗ σB, σA ⊗ σB(cid:48) and σA(cid:48) ⊗ σB(cid:48), which by
construction have to obey the marginal laws, as it is easy to verify by a direct calculation. Consider for
instance the probability pB(A1). Writing σA = A1(cid:105)(cid:104)A1 − A2(cid:105)(cid:104)A2, and σB = B1(cid:105)(cid:104)B1 − B2(cid:105)(cid:104)B2, we
have:
pB(A1) = p(A1, B1) + p(A1, B2) = (cid:104)s (A1(cid:105)(cid:104)A1 ⊗ B1(cid:105)(cid:104)B1)s(cid:105) + (cid:104)s (A1(cid:105)(cid:104)A1 ⊗ B2(cid:105)(cid:104)B2)s(cid:105)
= (cid:104)s (A1(cid:105)(cid:104)A1 ⊗ (B1(cid:105)(cid:104)B1 + B2(cid:105)(cid:104)B2))s(cid:105) = (cid:104)s (A1(cid:105)(cid:104)A1 ⊗ I)s(cid:105),
(26)
which is clearly independent of B, hence pB(A1) = pB(cid:48)(A1), and same for the other marginal probabilities.
What about the elastic example of Section 2.2, does it obey the marginal laws? We have:
pB(A1) = 1
2 ,
pB(A(cid:48)
1) =
pA(B1) = 1
2 ,
pA(B(cid:48)
1) =
n+ 1
n+1 = pB(cid:48)(A1),
2
n+1 = pB(cid:48)(A(cid:48)
n+ 1
1),
2
n+ 1
n+1 = pA(cid:48)(B1),
2
n+1 = pA(cid:48)(B(cid:48)
n+ 1
1),
2
1
n+1 = pB(cid:48)(A2),
2
1
pB(A2) = 1
pB(A(cid:48)
2) = 1
2
pA(B2) = 1
pA(B(cid:48)
2) = 1
2
2 , 1
n+1 = pB(cid:48)(A(cid:48)
2),
2 , 1
n+1 = pA(cid:48)(B(cid:48)
2).
1
1
n+1 = pA(cid:48)(B2),
2
(27)
Clearly, for the trivial case n = 0, all marginal laws are obeyed, but for n > 0 some of them are violated.
Note that for n = 1, 2, the CHSH inequality is not violated, but the marginal laws are violated, hence the
two violations are not perfectly correlated in the model. Consider however that, for example:
pB(cid:48)(A1) − pB(A1) =
n+ 1
2
n+1 − 1
2 = 1 − 1
n+1 .
(28)
Hence, the degrees of the violation of the marginal laws and CHSH inequality jointly increase as n increases.
Let us also consider the psychological measurement of Sec. 2.3. We have:
pB(A1) = 55
pB(A(cid:48)
1) = 70
pA(B1) = 25
pA(B(cid:48)
1) = 72
81 (cid:54)= 50
81 (cid:54)= 19
81 (cid:54)= 70
81 (cid:54)= 20
81 = pB(cid:48)(A1),
81 = pB(cid:48)(A(cid:48)
1),
81 = pA(cid:48)(B1),
81 = pA(cid:48)(B(cid:48)
1),
pB(A2) = 26
pB(A(cid:48)
2) = 11
pA(B2) = 56
pA(B(cid:48)
2) = 9
81 (cid:54)= 31
81 (cid:54)= 62
81 (cid:54)= 11
81 (cid:54)= 61
81 = pB(cid:48)(A2),
81 = pB(cid:48)(A(cid:48)
2),
81 = pA(cid:48)(B2),
81 = pA(cid:48)(B(cid:48)
2).
(29)
In this example we thus have even less symmetry than in the model with the elastic bands, as all the
marginal equalities are manifestly violated.
13
4 Compatibility and separability
In the previous section we have defined the outcome probabilities of Alice's sub-measurements A, A(cid:48),
and Bob's sub-measurements B and B(cid:48), by identifying some of the outcomes of the corresponding joint
measurements, as per (21)-(24). These relations express the 'compatibility' of the sub-measurements, in
the sense that the sub-measurements behave, at least probabilistically, as if they could be substituted by
one big experiment, after an identification of the corresponding outcomes. An important notion in our
discussion is that of 'separability'. The two compatible sub-measurements A and B of a joint measurement
AB are said to be separated, with respect to a given pre-measurement state, if the following factoring
relationships holds (with the factors being independent from one another):
p(A1, B1) = p(A1)p(B1),
p(A2, B1) = p(A2)p(B1),
p(A1, B2) = p(A1)p(B2),
p(A2, B2) = p(A2)p(B2).
Combining (30) with (21), one finds the necessary and sufficient condition for separability:
p(A1, B2)p(A2, B1) = p(A1, B1)p(A2, B2).
(30)
(31)
It is easy to check that none of the three experimental situations considered in Secs. 2.1-2.3, describe sep-
arate sub-measurements. In other words, in all these situations the joint measurements reveal correlations
that violate (30) or (31).
Note that if all four joint measurements AB, AB(cid:48), A(cid:48)B and A(cid:48)B(cid:48) are formed by separate sub-measurements,
then both the CHSH inequality and the marginal laws are necessarily obeyed. This is so because then the
marginal probabilities, like for instance:
pB(A1) = p(A1, B1) + p(A1, B2) = p(A1)p(B1) + p(A1)p(B2)
= p(A1)(pA(B1) + pA(B2)) = p(A1),
(32)
do not depend anymore on the measurement with which they are jointly executed. It is also easy to show
that the CHSH is necessarily obeyed in this case. Indeed, we have:
E(A, B) = p(A1, B1) − p(A1, B2) − p(A2, B1) + p(A2, B2)
= p(A1)p(B1) − p(A1)p(B2) − p(A2)p(B1) + p(A2)p(B2)
= [p(A1) − p(A2)][p(B1) − p(B2)] = E(A)E(B),
and similarly E(A, B(cid:48)) = E(A)E(B(cid:48)), E(A(cid:48), B) = E(A(cid:48))E(B) and E(A(cid:48), B(cid:48)) = E(A(cid:48))E(B(cid:48)). Hence:
CHSH = E(A, B) − E(A, B(cid:48)) + E(A(cid:48), B) + E(A(cid:48), B(cid:48))
= E(A)E(B) − E(A)E(B(cid:48)) + E(A(cid:48))E(B) + E(A(cid:48))E(B(cid:48))
= E(A)(E(B) − E(B(cid:48)) + E(A(cid:48))(E(B) − E(B(cid:48))
≤ E(A)E(B) − E(B(cid:48)) + E(A(cid:48))E(B) − E(B(cid:48))
≤ E(B) − E(B(cid:48)) + E(B) − E(B(cid:48)),
(33)
(34)
(35)
where for the penultimate inequality we have used the triangle inequality, and for the last inequality the
fact that E(A) = p(A1) − p(A2) = 2p(A1) − 1 ≤ 1, and similarly E(A(cid:48)) ≤ 1. The last expression in
(35) contains numbers that are within the interval −1 ≤ E(B), E(B(cid:48)) ≤ 1, hence it is necessarily bounded
by 2, i.e., CHSH ≤ 2.
14
5 Correlations of the first and second kind
An important aspect in the discussion of the phenomenon of entanglement (rarely taken into due account)
is the distinction between 'correlations of the first kind' and 'correlations of the second kind' (Aerts, 1990).
Consider the joint measurement AB of Sec. 2.2, where Alice and Bob jointly pull their respective ends
of the black elastic band. When they do so, the elastic (assumed to be uniform) will break with equal
probability in one of its points, which neither Alice nor Bob can predict in advance. If, say, the elastic
breaks at a distance λ from Alice's end, Alice's collected fragment will be of length LA = λ, whereas Bob's
collected fragment will be of length LB = d− λ. The remarkable property of these two lengths is that their
sum LA + LB is independent of λ and is always equal to the initial (unstretched) length d of the elastic. In
other words, the two lengths LA and LB are perfectly correlated: independently of their actual value, their
sum is necessarily equal to the total length d of the initially unbroken elastic. So, the joint measurement
AB 'creates a correlations' or, to say it in more precise terms, 'actualizes one among an infinite number of
potential correlations'. Indeed, each value of λ corresponds to a different couple of correlated lengths LA
and LB, and at each run of the AB measurement, because of the inevitable fluctuations in the interaction
between Alice's and Bob's hands and the elastic, a different value for λ will be obtained. Correlations that
are created by a joint measurement are called 'of the second kind', whereas if they were present prior to
the measurement, hence are only discovered and not created by a joint measurement, they are called 'of
the first kind'. Consider the situation where the black elastic would be already broken, but we do not know
where exactly it is broken, i.e., we do not know the exact value of λ. We only know that such "hidden
variable" has a well-defined value prior to the joint measurement to be executed. We can for instance model
our lack of knowledge by means of a uniform probability distribution, so that we still have in this case, as it
2 , so that E(A, B) = −1, and we
was the case for the unbroken black elastic, that p(A1, B2) = p(A2, B1) = 1
also have E(A(cid:48), B(cid:48)) = n−1
n+1 . What now changes are the probabilities for the A(cid:48)B and AB(cid:48) measurements.
We have:
p(A(cid:48)
1, B1) = 1
2
n
n+1 ,
p(A(cid:48)
1, B2) = 1
2 ,
p(A(cid:48)
2, B1) = 1
2
1
n+1 ,
p(A(cid:48)
2, B2) = 0.
(36)
and similarly for the AB(cid:48) measurement, so that E(A(cid:48), B) = E(A, B(cid:48)) = − 1
n+1 . For the quantity CHSH =
−E(A, B)+E(A, B(cid:48))+E(A(cid:48), B)+E(A(cid:48), B(cid:48)), we therefore obtain: CHSH = 1− 2
n+1 . Clearly,
for all n we have CHSH ≤ 2. Considering instead CHSH = E(A, B) − E(A, B(cid:48)) + E(A(cid:48), B) + E(A(cid:48), B(cid:48)),
n+1 , so again for all n we have CHSH ≤ 2. In other
we obtain: CHSH = −1 + 1
words, we find that when correlations of the second kind are replaced by correlations of the first kind, the
CHSH inequality is not anymore violated.
n+1 = 2 n−1
n+1 + n−1
n+1 − 1
n+1 + n−1
n+1 = − 2
The above example shows the crucial difference between an experimental situations where the lack
of knowledge is about correlations (of the first kind) that pre-exist the measurement processes, which
cannot give rise to a violation of the CHSH inequality, and an experimental situations where the lack
of knowledge is about correlations (of the second kind) that do not pre-exist the measurements, but are
created by them, during their executions. The former situation is that of so-called hidden-variable theories,
introducing 'elements of reality' describing those past factors that would have determined the correlations
in the bipartite system, already existing prior to the measurement. These correlations of the first kind
are precisely those that are filtered by Bell's inequalities, like the CHSH inequality. The latter situation
is instead that of so-called hidden measurement "interactions," which are assumed to be responsible for
the actualization of potential properties in quantum measurements. When measurements are of the joint
kind, i.e., they are performed on bipartite entities in a coincident way, at once, they can clearly also create
correlations, when the bipartite entity forms an undivided whole.
The existence of a general mechanism based on hidden measurement-interactions, responsible for quan-
It is impor-
tum indeterminism, remains of course a hypothesis in need of experimental confirmation.
15
tant however to emphasize that such possibility is inherent in the very geometry of Hilbert space, if one
adopts its 'extended Bloch representation (EBR)', in which the elements of reality describing the available
measurement-interactions are represented by specific simplex-structures inscribed in the Bloch sphere, al-
lowing for a simple (non-circular) derivation of the Born rule (Aerts, 1986; Aerts & Sassoli de Bianchi,
2014, 2016). In other words, not only the standard quantum formalism allows, but in fact also suggests,
the process of actualization of potential properties, indicated by the projection postulate, to result from
the presence of a hidden (non-local, non-spatial) level of interaction between the measuring system and the
measured entity, which in the case of entangled systems would be responsible for the creation of correlations
at the level of the whole entangled entity.
Of course, in the elastic model, the presence of these hidden measurement-interactions is self-evident
and there are no mysteries there in how the CHSH inequality and the marginal laws are violated. Different
"ways of stretching the elastic" give rise to different non-local interactions between Alice's and Bob's hands
and the elastic, which in turn will produce different breaking points, thus creating different correlations.
It is important to observe that the joint action of Alice and Bob is exerted at the level of the entire
elastic entity. There are no direct or indirect influences exerted by Alice's sub-measurements on Bob's
sub-measurements, or vice versa, there are just global joint measurements, operated at the level of the
whole bipartite structure, at once, in a non decompositional way, hence the absence of factorizability of the
joint probabilities evidenced by the violation of the CHSH inequality and the additional possible violation
of the marginal laws. To put it in a different way, there is nothing traveling from Alice to Bob, or from
Bob to Alice, no "spooky action at a distance," no superluminal communication of one part of the elastic
instructing the other part how to behave.
The situation described in the elastic model, and also subtended by the quantum formalism (when the
projection postulate is taken seriously and the further elucidation provided by the EBR is considered), is
also what seems to happen, mutatis mutandis, in human minds participating in a psychological experiments,
like the one described in Sec. 2.3, where two concepts -- Animal and Acts -- are combined to form a bipartite
entity in a given meaning state, described by the specific sentence in which the two concepts are combined,
here the very simple sentence: The Animal Acts. Even though the joint measurements are about associating
different exemplars to the two conceptual entities, this operation is not performed by Alice (assuming Alice
would be responsible to select the exemplar for the Animal sub-entity) in a way that would be separated
from the operation performed by Bob (assuming Bob would be responsible to select an exemplar for the
Acts sub-entity), so much so that one would be allowed to introduce notions like "the direct or indirect
influence of Alice's sub-measurement on Bob's sub-measurements, and vice versa." Indeed, each participant
in the experiment performs the joint measurements as whole measurements, at once, considering the entire
meaning of the The Animal Acts sentence in relation to its four possible outcome-states, which in the case
of the AB measurement are the outcomes The Horse Growls, The Horse Whinnies, The Bear Growls and
The Bear Whinnies. This is very similar to the two hands jointly pulling the two ends of the black elastic
band: they are not influencing or communicating with one another, they are simply "working in unison."
Again, there is no Alice's measurement influencing (directly or indirectly) Bob's measurement, there
is no action at a distance between the two sub-entities, or sub-measurements: there is just a single whole
entity with a bipartite structure which is acted upon by a single measurement process -- for clarity, let
us call it David's measurement -- which in a sense can be understood as the joining of Alice and Bob
sub-measurements, because it implies a choice to be jointly made on Animal and Acts, simultaneously,
in the same way that for the model of Sec. 2.2, to perform the AB measurement, the black elastic needs
to be jointly and simultaneously pulled from both sides. But the effective action is on the totality of
the elastic entity, in the same way that David, in the psychological measurement, acts on the totality of
the The Animal Acts conceptual entity, and there would be no sense in trying to frame the experimental
process in terms of possible mutual influences of Alice's action on Bob's action, and vice versa. Because,
16
in ultimate analysis, there are no separated processes to be associated with Alice (corresponding to the
choice of an exemplar for Animal ) and Bob (corresponding to the choice of an exemplar for Acts), there
is only a 'single mind process', that of David, describing the collective participant in the psychological
measurement, selecting at once both exemplars (fore the notion of 'collective participant', see Aerts et al.
(2019)).
Coming back to the joint measurement AB on the elastic band entity, note that from the local viewpoint
of Alice and Bob, outcomes become available to them at the same time, and if the elastic is long enough,
these two events will be spacelike separated. However, nothing travels at speed greater than light, as
it requires time for the two elastic fragments to reach Alice's and Bob's hands (or David's hands, if
the experiment is performed by a single individual). The same can be expected to be the case also
with micro-physical entangled entities, like electrons and photons: even if the final detection events are
spacelike separated, this does not necessarily imply that there is the propagation of a superluminal influence.
Therefore, the observed violations of the marginal laws should not necessarily raise concern regarding
possible violations of Einsteinian relativity. Independently of the above, the situation in the psychology
laboratory is of course also different from that of a physics laboratory, as is clear that there are no principles
equivalent to the relativistic one that would make one expect marginal selectivity to generally apply. This
also because, when two conceptual entities are combined, and their possible exemplars are jointly selected,
this only happens in the mind of a single subject, not in a spatial environment equipped with specific
symmetries (see also the additional discussion in Sec. 10).
6 Hilbertian modeling
When some of the marginal laws are violated, in addition to the CHSH inequality, we must proceed more
carefully in the Hilbert space modeling of the situation. Indeed, as we showed it explicitly in the Bohm
example, if the joint measurements are represented by 'product operators' in the tensor product of the
Hilbert spaces of the sub-entities, the marginal laws will be satisfied (see Sec. 3). However, we should not
forget that in addition to introducing a tensor product starting from the state spaces of the sub-entities, it
is also possible to introduce it with respect to their operator spaces, allowing in this way for a much more
general tensor product construction. It can be proven that the set of linear operators on the tensor product
of the state space of the joint entity is isomorphic to the tensor product of the sets of linear operators on
the state spaces of the sub-entities. This means that from a mathematical perspective, when confronted
with the presence of entanglement due to a violation of the CHSH inequality, there is no reason to prefer
to model it in terms of an entangled state, i.e., of an element of the tensor product of the state spaces
of the sub-entities, or in terms of an entangled measurement, i.e., of an element of the tensor product of
the operator spaces of the sub-entities. Indeed, as we will show in the following of the article, and as we
put forward in Aerts & Sozzo (2014a,b,c), when in addition to the violation of the CHSH inequality the
marginal laws are also violated, this indicates that entanglement can also be present in the measurements,
hence be mathematically modeled by entangled self-adjoint operators, rather than only in the state, where
it can be mathematically modeled by an entangled state.
But let us analyse the situation carefully, starting from the more common way of introducing the tensor
product at the level of the state spaces of the sub-systems. We can also then see very explicitly that it is
indeed possible to model the situation by considering a larger class of measurements of the non-product
kind, which we have called 'entangled measurements' (Aerts & Sozzo, 2014a,b,c). In other words, when the
marginal laws are violated, one cannot model entanglement only at the level of the state of the bipartite
system and, as we will analyse in detail in the following, in many experimental situations some of the
measurements are also to be considered as entangled.
Let us denote by H the complex Hilbert space describing the states of the bipartite system under study.
17
For the three examples that we considered in Sections 2.1-2.3, H is typically a 4-dimensional Hilbert space,
which we will now assume to be the case, i.e., a Hilbert space whose orthonormal bases have four elements.
More precisely, let us denote x1, x2, x3 and x4 the four states forming one of the basis of H, which from
now on we will represent as orthonormal kets {x1(cid:105), . . . ,x4(cid:105)}, (cid:104)xjxj(cid:105) = δij, i, j = 1, ..., 4, using Dirac's
notation. Let us then also consider a state p, represented by the normalized ket p(cid:105) ∈ H, (cid:104)pp(cid:105) = 1. Is
it an entangled state or a product state? A question of this kind cannot be answered on itself, but only
with reference to a tensorial representation of H. More precisely, one has first to introduce an isomorphism
I between H and another (isomorphic) Hilbert space K, having the tensorial structure: K = HA ⊗ HB.
Then, the previous question can be more precisely stated as follows: Is the state p, represented by the ket
p(cid:105), an entangled state with respect to the isomorphism I?
Before continuing, let us recall that an isomorphism I between two Hilbert spaces H and K is a surjective
linear map, I : H → K, preserving the inner product, i.e., (cid:104)pq(cid:105)H = (cid:104)I(p)I(q)(cid:105)K, where (cid:104)··(cid:105)H and (cid:104)··(cid:105)K
denote the scalar products in H and K, respectively. Note that this automatically implies that I is also
injective, i.e., that I is a bijection.2 So, the answer to the above question is that p is an 'entangled state
with respect to I' if it is not a 'product state with respect to I', that is, if there are no two states represented
by the kets pA(cid:105) ∈ HA and pB(cid:105) ∈ HB, such that one could write: Ip(cid:105) = pA(cid:105) ⊗ pB(cid:105). Similarly, one can
ask if a joint measurement AB is an 'entangled measurement', and again the question has to be addressed
in relation to a specific isomorphism I. So, the answer is that a joint measurement AB, represented by a
given (here bounded) self-adjoint linear operator EAB ∈ L(H), is an 'entangled measurement with respect
to I' if it is not a 'product measurement with respect to I', that is, if there are no two sub-measurements
A and B, represented by the self-adjoint linear operators EA ∈ L(HA) and EB ∈ L(HB), respectively, such
that one could write: IEABI−1 = EA ⊗ EB. An important observation is that, given a joint measurement
AB, one can always find a tailor-made isomorphism IAB : H → HA ⊗ HB, such that, with respect to IAB,
AB is a product measurement. This is easy to prove by introducing two observables EA ∈ L(HA) and
EB ∈ L(HB), with spectral decompositions EA = a1PA1 + a2PA2 and EB = b1PB1 + b2PB2, respectively,
where a1 and a2 are the eigenvalues of EA, and PA1 = A1(cid:105)(cid:104)A1 and PA2 = A2(cid:105)(cid:104)A2 are the associated
one-dimensional projection operators, and similarly for EB. We thus have:
EA ⊗ EB = (a1PA1 + a2PA2) ⊗ (b1PB1 + b2PB2)
= a1b1PA1 ⊗ PB1 + a1b2PA1 ⊗ PB2 + a2b1PA2 ⊗ PB1 + a2b2PA2 ⊗ PB2.
(37)
Introducing also the spectral decomposition
EAB = ab11PA1B1 + ab12PA1B2 + ab21PA2B1 + ab22PA2B2,
(38)
where the abij and the PAiBj = AiBj(cid:105)(cid:104)AiBj, i, j = 1, 2, are the eigenvalues and associated one-dimensional
projection operators of EAB ∈ L(H), one can then define the isomorphism IAB by simply specifying its
action on the four eigenstates of EAB, or the associated projection operators:
IABAiBj(cid:105) = Ai(cid:105) ⊗ Bj(cid:105),
IABPAiBj I−1
AB = PAi ⊗ PBj ,
i, j = 1, 2.
Eq. (37) can then be rewritten as:
I−1
ABEA ⊗ EBIAB = a1b1PAiB1 + a1b2PA1B2 + a2b1PA2B1 + a2b2PA2B2 = EAB,
(39)
(40)
where the last equality follows from the fact that given four real numbers abij, i, j = 1, 2, one can always
write them as products abij = ajbj, for well chosen ai and bj, i, j = 1, 2.
2Note that by definition I is an isometric operator and that for finite-dimensional Hilbert spaces, which will be the case in
our discussion, I is also necessarily a unitary operator.
18
Consider now two joint measurements AB and AB(cid:48), represented by the self-adjoint operators EAB,EAB(cid:48) ∈
L(H). Let IAB be the isomorphism allowing to represent AB as a product measurement, i.e., EAB =
I−1
ABEA ⊗ EBIAB. Can we also have EAB(cid:48) = I−1
ABEA ⊗ EB(cid:48)IAB, i.e., can we represent both measurements AB
and AB(cid:48) as product measurements, with respect to a same isomorphism? This is what is usually consid-
ered to be the case in typical quantum situations, like in the Bohm's spin model, where a unique tensorial
representation is introduced from the beginning, and all joint observables are constructed to be product
observables with respect to that same tensorial representation. As we have shown already in (26), when all
joint measurements are product measurements, the marginal laws have to be satisfied. This means that if
they are not, we cannot find a same isomorphism I = IAB = IAB(cid:48) that would allow to represent both AB
and AB(cid:48) as product measurements. Indeed:
p(A1, B1) + p(A1, B2) = (cid:104)p(PA1B1 + PA1B2)p(cid:105) = (cid:104)pI−1I(PA1B1 + PA1B2)I−1Ip(cid:105)
= (cid:104)pI−1(PA1 ⊗ PB1 + PA1 ⊗ PB2)Ip(cid:105) = (cid:104)pI−1PA1 ⊗ (PB1 + PB2)Ip(cid:105)
= (cid:104)pI−1PA1 ⊗ (PB(cid:48)
)Ip(cid:105)
= (cid:104)pI−1I(PA1B(cid:48)
= p(A1, B(cid:48)
)Ip(cid:105) = (cid:104)pI−1(PA1 ⊗ PB(cid:48)
)I−1Ip(cid:105) = (cid:104)p(PA1B(cid:48)
+ PB(cid:48)
+ PA1B(cid:48)
1) + p(A1, B(cid:48)
2),
+ PA1 ⊗ PB(cid:48)
)p(cid:105)
+ PA1B(cid:48)
2
1
2
2
1
2
1
1
(41)
A(cid:48)BEA(cid:48) ⊗ EBIA(cid:48)B, EAB(cid:48) = I−1
AB(cid:48)EA ⊗ EB(cid:48)IAB(cid:48) and EA(cid:48)B(cid:48) = I−1
i.e, the existence of such isomorphism requires the marginal laws to be satisfied. And since they are generally
not, one is forced to adopt a more general representation where not all measurements are necessarily of
the product form.
In general, one can introduce four different isomorphisms, IAB, IA(cid:48)B, IAB(cid:48) and IA(cid:48)B(cid:48), associated with
the four joint measurements AB, A(cid:48)B, AB(cid:48) and A(cid:48)B(cid:48), respectively, such that EAB = I−1
ABEA ⊗ EBIAB,
EA(cid:48)B = I−1
A(cid:48)B(cid:48)EA(cid:48) ⊗ EB(cid:48)IA(cid:48)B(cid:48). In other words, AB
is a product measurement with respect to IAB, A(cid:48)B is a product measurement with respect to IA(cid:48)B, AB(cid:48)
is a product measurement with respect to IAB(cid:48), and A(cid:48)B(cid:48) is a product measurement with respect to IA(cid:48)B(cid:48).
However, AB will not in general be a product measurement with respect to IA(cid:48)B, IAB(cid:48) and IA(cid:48)B(cid:48); A(cid:48)B will
not in general be a product measurement with respect to IAB, IAB(cid:48) and IA(cid:48)B(cid:48); AB(cid:48) will not in general be
a product measurement with respect to IAB, IA(cid:48)B and IA(cid:48)B(cid:48); and A(cid:48)B(cid:48) will not in general be a product
measurement with respect to IAB, IA(cid:48)B and IAB(cid:48). To put it differently, if one wants to force a tensor
product representation for all the four joint measurements AB, A(cid:48)B, AB(cid:48) and A(cid:48)B(cid:48), this is possible, but
the price to be paid is that the representation will then become contextual, in the sense that entanglement
will have to be incorporated into states that are in general different for each one of the joint measurements,
i.e., the states IABp(cid:105), IA(cid:48)Bp(cid:105), IAB(cid:48)p(cid:105) and IA(cid:48)B(cid:48)p(cid:105), respectively.
7 A simple modeling example
It is instructive at this point to provide an example of an explicit Hilbertian modeling. For simplicity, we
only consider the elastic band model presented of Sec. 2.2, in the limit n → ∞, of maximal violation of
the CHSH inequality. For an explicit Hilbertian representation of the psychological model of Sec. 2.3, we
refer the reader to Aerts & Sozzo (2014a); see also the Hilbert space modeling provided in Aerts et al.
(2019), relative to a recently performed test on the co-occurrences of two concepts and their combination
in retrieval processes on specific corpuses of documents.
So, the marginal laws being violated, we cannot describe all joint measurements as product measure-
ments with respect to a same isomorphism. Let us choose the eigenvectors of EAB to be the vectors of the
canonical basis of H = C4, that is:
A1B1(cid:105) = 1, 0, 0, 0(cid:105), A1B2(cid:105) = 0, 1, 0, 0(cid:105), A2B1(cid:105) = 0, 0, 1, 0(cid:105), A2B2(cid:105) = 0, 0, 0, 1(cid:105).
(42)
19
Then, we can write the pre-measurement state as the superposition state:
p(cid:105) =
1√
2
0, 1, 1, 0(cid:105) =
1√
2
(0, 1, 0, 0(cid:105) + 0, 0, 1, 0(cid:105)),
which clearly gives the correct probabilities:
p(A1, B1) = (cid:104)1, 0, 0, 0p(cid:105)2 = 0,
p(A2, B1) = (cid:104)0, 0, 1, 0p(cid:105)2 = 1
2 ,
p(A1, B2) = (cid:104)0, 1, 0, 0p(cid:105)2 = 1
2 ,
p(A2, B2) = (cid:104)0, 0, 0, 1p(cid:105)2 = 0.
Considering the isomorphism IAB and the observable EAB, defined as:
IAB1, 0, 0, 0(cid:105) = 1, 0(cid:105) ⊗ 1, 0(cid:105),
IAB0, 0, 1, 0(cid:105) = 0, 1(cid:105) ⊗ 1, 0(cid:105),
i,j=1 abijAiBj(cid:105)(cid:104)AiBj,
EAB =(cid:80)2
IAB0, 1, 0, 0(cid:105) = 1, 0(cid:105) ⊗ 0, 1(cid:105),
IAB0, 0, 0, 1(cid:105) = 0, 1(cid:105) ⊗ 0, 1(cid:105),
δ1 = 1, δ2 = −1,
abij = δiδj,
such that E(A, B) = (cid:104)pEABp(cid:105), we can write:
(43)
(44)
(45)
IABEABI−1
AB = IAB(1, 0, 0, 0(cid:105)(cid:104)1, 0, 0, 0 − 0, 1, 0, 0(cid:105)(cid:104)0, 1, 0, 0
−0, 0, 1, 0(cid:105)(cid:104)0, 0, 1, 0 + 0, 0, 0, 1(cid:105)(cid:104)0, 0, 0, 1)I−1
AB
= 1, 0(cid:105) ⊗ 1, 0(cid:105)(cid:104)1, 0 ⊗ (cid:104)1, 0 − 1, 0(cid:105) ⊗ 0, 1(cid:105)(cid:104)1, 0 ⊗ (cid:104)0, 1
−0, 1(cid:105) ⊗ 1, 0(cid:105)(cid:104)0, 1 ⊗ (cid:104)1, 0 + 0, 1(cid:105) ⊗ 0, 1(cid:105)(cid:104)0, 1 ⊗ (cid:104)0, 1
= (1, 0(cid:105)(cid:104)1, 0 − 0, 1(cid:105)(cid:104)0, 1) ⊗ (1, 0(cid:105)(cid:104)1, 0 − 0, 1(cid:105)(cid:104)0, 1) = σz ⊗ σz,
(46)
where σz is Pauli's z-matrix, such that σz1, 0(cid:105) = 1, 0(cid:105) and σz0, 1(cid:105) = −0, 1(cid:105). So, the two sub-measurements
A and B can be represented, with respect to IAB, using the same operator EA = EB = σz, and we have the
quantum average:
E(A, B) = (cid:104)pEABp(cid:105) = (cid:104)pI−1
ABIABEABI−1
((cid:104)1, 0 ⊗ (cid:104)0, 1 + (cid:104)0, 1 ⊗ (cid:104)1, 0)σz ⊗ σz(1, 0(cid:105) ⊗ 0, 1(cid:105) + 0, 1(cid:105) ⊗ 1, 0(cid:105))
ABIABp(cid:105) = (cid:104)pI−1
ABσz ⊗ σzIABp(cid:105)
1
=
2
= − 1
2
= −1.
− 1
2
Let us now also model measurement AB(cid:48). We can consider in this case the eigenvectors:
A1B(cid:48)
0, 1,−1, 0(cid:105), A2B(cid:48)
1(cid:105) = 1, 0, 0, 0(cid:105), A2B(cid:48)
0, 1, 1, 0(cid:105), A1B(cid:48)
2(cid:105) =
1(cid:105) =
2(cid:105) = 0, 0, 0, 1(cid:105),
1√
2
1√
2
which clearly give the correct probabilities:
(47)
(48)
(49)
(50)
p(A1, B(cid:48)
1) = 1√
p(A2, B(cid:48)
2
p(A1, B(cid:48)
2) = 1√
2
(cid:104)0, 1,−1, 0p(cid:105)2 = 0,
p(A2, B(cid:48)
2) = (cid:104)0, 0, 0, 1p(cid:105)2 = 0.
For the self-adjoint operator EAB(cid:48), describing AB(cid:48), we have the spectral decomposition:
(cid:104)0, 1, 1, 0p(cid:105)2 = (cid:104)pp(cid:105)2 = 1,
1) = (cid:104)1, 0, 0, 0p(cid:105)2 = 0,
2(cid:88)
ijAiB(cid:48)
ab(cid:48)
j,
j(cid:105)(cid:104)AiB(cid:48)
i,j=1
20
EAB(cid:48) =
ab(cid:48)
ij = δiδj,
δ1 = 1, δ2 = −1,
and of course E(A, B(cid:48)) = (cid:104)pEAB(cid:48)p(cid:105) = 1. However, the isomorphism IAB cannot be used now to also write
2(cid:105) are entangled vectors
EAB(cid:48) in a tensor product form, as is clear that the eigenvectors A1B(cid:48)
with respect to IAB.
The operators EA(cid:48)B and EA(cid:48)B(cid:48) can be taken to be identical to EAB(cid:48), so that the Bell operator:
1(cid:105) and A1B(cid:48)
ECHSH = −EAB + EAB(cid:48) + EA(cid:48)B + EA(cid:48)B(cid:48) = −EAB + 3EAB(cid:48)
(51)
(52)
(53)
can be explicitly written in the canonical basis:
ECHSH =
0
0
0
If we calculate the expectation value CHSH = (cid:104) pECHSHp(cid:105), we thus find:
0 1 0
1 0 0
0 0 1
0
0
0
−1 0 0
0
1 0
0
0 1
0
0 0 −1
0
0
0
+ 3
−1 0 0 0
=
−4 0 0 0
.
(cid:0)0 1 1 0(cid:1)−4 0 0 0
1√
0
= 4,
1 3 0
3 1 0
0 0 2
2
CHSH =
1√
2
in accordance with (14), as n → ∞.
0
0
0
1 3 0
3 1 0
0 0 2
1
1
0
8 Product or entangled measurements?
In the previous sections, we have seen that entanglement, both for states and measurements, has to be de-
fined relative to a given isomorphism, specifying a tensorial representation. This means that entanglement
is a contextual property that depends on the way sub-structures are identified in a composite system. In
some experimental circumstances, the way such identification has to be carried out can be dictated by the
very geometry of the setting. For instance, it appears natural in a typical Bohm model situation, where
Alice and Bob operate two distinct Stern-Gerlach apparatuses separated by an arbitrarily large spatial
distance, to introduce a product representation for their joint measurements. However, what appears to
be natural might not necessarily be correct. Indeed, in view of the numerous observed violations of the
no-signaling conditions in Bell-test experiments (Adenier & Khrennikov, 2007; De Raedt et al., 2012, 2013;
Adenier & Khrennikov, 2017; Bednorz, 2017; Kupczynski, 2017), one may wonder to which extent the
fact that the overall measuring apparatus is formed by two spatially separated instruments, producing
outcomes in a coincident way, would be sufficient to characterize all the considered joint measurements as
product measurements with respect to a same tensor product representation, i.e., a same isomorphism and
therefore also a same pre-measurement state.
When entanglement was initially theorized, physicists were not expecting that correlations could man-
ifest independently of the distance separating Alice's and Bob's measuring apparatuses. This means that
when a bipartite entity is in an entangled state, even if a joint measurement is represented by means of a
product operator EA ⊗ EB, this does not mean that the corresponding sub-measurements would be sepa-
rated, i.e., would obey (30), this being the case only when the system is in a product state. In other words,
product measurements are not separated measurements, when entanglement is present in the system. Now,
what we observed in Sec. 6, is that there is some freedom in the way such non-separability of the joint
measurements (reflecting the non-separability of the bipartite entity) can be modeled within the Hilbert
space formalism. At the "local level" of a given joint measurement, one can always contextually consider a
specific isomorphism that will be able to push all the entanglement resource in the state. However, this will
21
generally only work contextually for that specific joint measurement, and not for all joint measurements
that can be defined and performed on the system.
So, contextually, by considering a suitable isomorphism, one can always assert that 'entanglement is
just in the state', and not also in the joint measurement, i.e., that the notion of 'entangled states' is
sufficient to describe the fact that a bipartite system forms an interconnected whole, so that one does not
need to also consider a notion of 'entangled measurements', to properly describe the situation. As we are
going to explain, this way of proceeding remains consistent only if one can consider the outcome-states to
be product states, which however is something that might not be true for all experimental situations. But
let us for a moment assume that the outcome-states are correctly described as product states. It becomes
then possible to support the following view. First, one observes that the presence of entanglement in the
system manifests at the level of the probabilities in such a way that the joint probabilities cannot be written
as the products (30). Of course, this is a consequence of entanglement, i.e., a necessary condition for it,
not a sufficient one, and it is precisely one of the merits of John Bell to have identified conditions able to
characterize the presence of entanglement, via his inequalities, combining the probabilistic data obtained
from different joint measurements, so as to demarcate correlations of the first kind from those of the second
kind. But of course, entanglement will manifest its presence in each joint measurement, as the source of
the non-factorizability of the corresponding joint probabilities.
Now, the very notion of entanglement implicitly contains the idea that we are in the presence of a
bipartite system, i.e., a system formed by two parts that, precisely, have been entangled and therefore do
form a whole. Such bipartite structure can then be implemented in the formalism by describing a given joint
measurement performed on the entity as a product measurement. This has the consequence that outcome-
states will be modeled as product states, i.e., that the measurement process will have to be understood
as a process of disentanglement of the previously entangled sub-entities (as implied by the projection
postulate). So, the tensor product is to be viewed as a mathematical procedure of recognition of the
bipartite structure of the system, in the context of a given joint measurement that takes the interconnected
system and produces its disconnection, creating in this way the correlations. When we only focus on the
situation of 'one state and one joint measurement', this 'tensor product procedure' presents no specific
problems, as is clear that one can always find a well-defined isomorphism that can do the job of pushing
all the entanglement in the state (see Sec. 6). However, when more than a single joint measurement is
considered, and these measurements do not obey the marginal laws, we have seen that the associated
isomorphisms, implementing for each of them the tensor product structure, will not coincide. When they
do coincide, i.e., when the marginal laws are obeyed, one can of course forget about the different possible
ways of introducing tensor product structures, and just start from the beginning with a given tensor
product representation for the state space, as is usually done in quantum mechanics, where the validity
of the no-signaling conditions is taken for granted. But when marginal selectivity is not satisfied, one is
forced to conclude that the tensor product structure can only be locally applied to each one of the joint
measurements, and not globally applied to all joint measurements.
On could be tempted to consider the above as a shortcoming of the quantum Hilbertian formalism,
however, another possibility is to simply consider this as the signature of a more complex underlying reality.
Consider the example of a curved surface imbedded in R3. Each point of it can be locally associated with
a two-dimensional tangent plane, but one cannot define a global tangent plane, i.e., a plane that would be
tangent to all points of the curved surface. The situation could be here analogous: each isomorphism can
locally produce a simpler tensor product structure, which however cannot be applied to all measurement
situations, because the system would not possess enough symmetries for this to be possible. We could say
that when the marginal laws are obeyed, the system (understood here as the entity plus its collection of
relevant measurements) is as close as possible to the situation of two genuinely separated systems,3 where
3Note that because of the superposition principle, quantum mechanics does not allow for the description of fully separated
22
the presence of the sub-systems is recognizable in a much stronger way than when marginal selectivity is
not satisfied. In other words, the idea is that when marginal selectivity is not satisfied, the tensor product
procedure of recognition of sub-systems would not be given once for all, but needs to be fine-tuned with
respect to each joint measurement considered. Otherwise, if one forces a single tensor product structure onto
the system, say via the isomorphism IAB, the other joint measurements will become, from the perspective
of that IAB, entangled measurements. However, if they produce transitions that disentangle the system
(like when by pulling an elastic band it gets broken and separated into two fragments), such description by
means of a single isomorphism is to be considered as non-optimal, as it makes certain measurements appear
as if they would be entangled, whereas in practice they would produce a factual disentanglement of the
previously entangled system. But again, their "entangled appearance" would only be due to the specific
isomorphism used, and when reverting to a better fine-tuned isomorphism, each measurement can always
be "locally" conveniently transformed into a product measurement, giving rise to product outcome-states
(relative to that specific isomorphism).
The above discussion is pertinent only in situations such that there is a way to ascertain that the
considered joint measurements give rise to product outcome-states, i.e., states that cannot be used any-
more to violate the CHSH inequality, which therefore should be properly represented as tensor product
vectors. However, there certainly are situations where the outcome-states of a joint measurement are to be
considered to remain entangled states.4 Consider the cognitive model of Sec. 2.3. Take the four outcome-
states of the joint measurement AB, described by the conceptual combinations: The Horse Growls, The
Horse Whinnies, The Bear Growls and The Bear Whinnies. These are more concrete states than the
pre-measurement state, The Animal Acts, but can still be considered to be entangled states of the compos-
ite entity formed by the joining of the two individual entities Animal and Acts. Indeed, in the cognitive
domain, what the quantum structure of entanglement captures, not only mathematically but also conceptu-
ally, is the presence of a 'meaning connection' between the different concepts. The two concepts Horse and
Whinnies are strongly meaning connected, hence one could make the case that their combination should
still be represented as an entangled state, and not as a product state. The meaning connection between
Bear and Whinnies is instead much more feeble, hence The Bear Whinnies is to be considered to be a
less entangled state than The Horse Whinnies. Now, every time there is a meaning connection between
concepts, in a given conceptual combination, one can certainly conceive joint measurements extracting
correlations from it and by doing so violate the CHSH inequality.
Therefore, for cognitive measurements the situation is that 'entangled measurements' would be the
default way to represent the situation of joint measurements that can truly preserve the wholeness (or part
of the wholeness) of the measured conceptual entity, and not mere mathematical artefacts resulting from an
inappropriate choice of the isomorphism introducing a tensor product structure.5 In this kind of situation,
there is of course a lot of freedom in the choice of the isomorphism, as different criteria can certainly be
adopted. Typically, one will consider for the pre-measurement state a maximally symmetric and entangled
state, like a singlet state, so that the outcome-states of the different entangled measurements can reflect
in a natural way the higher or lower meaning connection that is carried by the different combination of
exemplars, which in turn translates in the fact that these outcome-states will be more or less entangled
(Aerts & Sozzo, 2014a; Aerts et al., 2019). So, one should not expect to find within the quantum formalism
systems (see Sassoli de Bianchi (2019a) and the references cited therein), which is the reason why we say "as close as possible,"
i.e., within the limits of what the Hilbertian formalism allows to model in terms of separation.
4An example of entangled measurements is the so-called Bell-state measurements, in quantum teleportation, where two
given qubits are collapsed onto one of the four entangled states known as Bell states (Bennett et al., 1993).
5In the domain of macroscopic models, an example of joint measurements preserving the entanglement connection was
described in Sassoli de Bianchi (2014), with an idealized experimental situation of two prisms connected through a rigid
rod, glued on their opposed polygon-faces, which were jointly rolled in order to create correlations able to violate the CHSH
inequality, however preserving their interconnection through the rigid rod.
23
a 'unique recognition procedure' in the description of the structure of a bipartite system subjected to joint
measurements, i.e., a unique recipe for choosing one or multiple isomorphisms, with respect to which entan-
glement at the level of the states and/or measurements can be defined. Each situation requires an attentive
analysis and interpretation, based on the nature of the outcome-states of the different measurements, from
which a suitable representation can then be adopted.
9 More on sub-measurements
In Sec. 3, we defined the two sub-measurements A and B of a joint measurement AB by a simple procedure
of identification of certain outcomes. For instance, we defined sub-measurement A as the measurement
having Ai, i = 1, 2, as its outcomes, where Ai is actualized each time that either (Ai, B1) or (Ai, B2)
are actualized, when performing AB, so that pB(Ai) = p(Ai, B1) + p(Ai, B2), i = 1, 2, and same for the
outcomes of B; see (21). In standard quantum mechanics, however, this is not how the sub-measurements of
a joint measurement are usually described. Indeed, if A would be just obtained by performing the "bigger"
measurement AB, then simply identifying two by two its outcomes, A should collapse the pre-measurement
state to exactly the same set of outcome-states as AB does, as operationally speaking A and AB would be
executed in exactly the same way. However, using the same notation as in Sec. 6, the standard quantum
formalism tells us to represent the two sub-measurements A and B by the following operators:
EA = a1(PA1B1 + PA1B2) + a2(PA2B1 + PA2B2),
EB = b1(PA1B1 + PA2B1) + b2(PA1B2 + PA2B2).
(54)
The "tilded notation" is here used to distinguish EA, EB ∈ L(H), which act on the whole Hilbert space H,
from the operators EA ∈ L(HA) and EB ∈ L(HB), introduced in Sec. 6, which only act on the sub-spaces
HA and HB, in a given tensorial decomposition of H.
So, according to the projection postulate, if p(cid:105) is the pre-measurement state, when performing the
sub-measurement A, represented by operator EA, the two possible outcome-states are:
(PA1B1 + PA1B2)p(cid:105)
(cid:107)(PA1B1 + PA1B2)p(cid:105)(cid:107) ,
(PA2B1 + PA2B2)p(cid:105)
(cid:107)(PA2B1 + PA2B2)p(cid:105)(cid:107) .
Similarly, the two possible outcome-states of sub-measurement EB are:
(PA1B1 + PA2B1)p(cid:105)
(cid:107)(PA1B1 + PA2B1)p(cid:105)(cid:107) ,
(PA1B2 + PA2B2)p(cid:105)
(cid:107)(PA1B2 + PA2B2)p(cid:105)(cid:107) .
Clearly, these four outcome-states are superpositions of the four outcome-states of AB, which are:
PA1B1p(cid:105)
(cid:107)PA1B1p(cid:105)(cid:107) ,
PA1B2p(cid:105)
(cid:107)PA1B2p(cid:105)(cid:107) ,
PA2B1p(cid:105)
(cid:107)PA2B1p(cid:105)(cid:107) ,
PA2B2p(cid:105)
(cid:107)PA2B2p(cid:105)(cid:107) .
(55)
(56)
(57)
Note however that since (see also (37)):
EAB = EA EB = EB EA = a1b1PA1B1 + a1b2PA1B2 + a2b1PA2B1 + a2b2PA2B2,
(58)
even though EA and EB project onto different outcome-states than those of EAB, when considered individ-
ually, if their processes are combined the same set of outcome-states (57) will be consistently obtained.
This can be easily seen by considering a situation where it makes sense to consider that A and B are per-
formed in a sequence, one after the other (the sub-measurements being compatible, the order is irrelevant).
Imagine that after performing first sub-measurement A, the outcome-state associated with the eigenvalue
24
a1 has been obtained (the first vector-state in (55)). If this outcome-state becomes the pre-measurement
state for the subsequent measurement B, inserting it at the place of p(cid:105) in (56), then observing that
(PA1B1 + PA2B1)(PA1B1 + PA1B2) = PA1B1,
(PA1B2 + PA2B2)(PA1B1 + PA1B2) = PA1B2,
(59)
we clearly obtain the first two vectors in (57) as possible outcome-states. Similarly, if the outcome-state
of A is that associated with the eigenvalue a2 (the second vector-state in (55)), observing that
(PA1B1 + PA2B1)(PA2B1 + PA2B2) = PA2B1,
(PA1B2 + PA2B2)(PA2B1 + PA2B2) = PA2B2,
(60)
we now obtain the last two vectors in (57) as possible outcome-states.
Let us now consider the example given in Sec. 2.3, of the psychological measurements with the bipartite
conceptual entity formed by the two concepts Animal and Acts, in the state described by the conceptual
combination The Animal Acts. Can we specify the two sub-measurements A and B in more specific terms,
also describing the way they individually change the initial state, so as to obtain a description in full
compliance with how sub-measurements are defined in quantum mechanics? The answer is affirmative,
and to show how we have to come back to the figure of David that we introduced in Sec. 7, to express
the fact that the joint measurements resulted from a single mind process (David being understood as the
collective participant acting at once on the totality of the bipartite conceptual entity), and not from the
combination of two separate mind processes, those of Alice and Bob.
In other words, to see how the
two sub-measurements A and B can be described, in a way that remains consistent with the statistics of
outcomes generated by David, we have to consider some additional processes that transform the outcome-
states (57) of AB into the superposition states (55) and (56) corresponding to the outcome-states of A and
B, respectively.
Starting from David's measurement AB, we now have to disjoin it into two sub-measurements, A and
B, associated with the figures of Alice and Bob. Clearly, the latter will have to be associated with the
marginal probabilities only, and will therefore contain less information, when considered individually, than
their join, hence a process of erasure must be considered to go from the whole joint measurement to the
associated sub-measurements, and it is precisely this erasure (of information) process that can explain why
the sub-measurements' outcome-states have to be described as superposition states (in the same way in
delayed choice quantum eraser experiments one recovers the interference pattern when the "which-path"
information is erased (Yoon-Ho Kim et al., 2000; Walborn et al., 2002; Aerts, 2009b)). So, consider the
situation of measurement AB, with David collecting the four outcomes (A1, B1), (A1, B2), (A2, B1) and
(A2, B2), corresponding to the four conceptual combinations The Horse Growls, The Horse Whinnies, The
Bear Growls and The Bear Whinnies. To define the sub-measurements A and B, we have to consider two
additional processes, where David sends information about the outcomes to Alice and Bob, but does so
taking care to erase information about the chosen exemplar of Acts, when informing Alice, and about the
chosen exemplar for Animal, when informing Bob. More precisely, assuming that the outcome was, say,
The Horse Growls, David will communicate the following to Alice:
An Exemplar For The Combination "The Animal Acts" Has Been Selected Among The Four
Following Possibilities: "The Horse Growls," "The Horse Whinnies," "The Bear Growls" and
"The Bear Whinnies," And It Contained "Horse."
Mutatis mutandis, the same kind of communication, with removed information about what was the selection
for the Animal concept, will be given to Bob, so that Alice and Bob will be able to deduce the correct
marginal probabilities, pB(A1), pB(A2) and pA(B1), pA(B2), respectively.
Now, the above information that Alice receives from David, corresponds to an articulate conceptual
combination that defines a specific state of the bipartite conceptual entity (we recall that in our Brus-
sel's operational-realistic approach to cognition the different conceptual combinations are associated with
25
different states for the conceptual entity under consideration; see Aerts et al. (2016)). When such state
is represented using a Hilbertian vector-state, it has to incorporate not only the lack of knowledge about
the choice for Acts, but also the fact that Whinnies has greater probability to be chosen (in the human
conceptual realm) than Growls, in association with Horse. Hence, this will correspond to a superposition
state represented by the first vector in (55), and of course one can reason in the same way for the other
outcome A2, and for the outcomes collected by Bob, who will receive from David a communication with
the Animal information that has been erased.
Note that a conceptual combination expressing a situation of lack of knowledge is to be associated
with a genuine new element of reality in the human conceptual realm, considering that it is human minds
that constitute the measurement instruments and that the latter are sensitive to the meaning content
of such combinations. So, the outcome-states associated with Alice's and Bob's sub-measurements have
to be described by superposition states, and not by classical mixtures. Indeed, the additional action of
David communicating the obtained outcomes to Alice and Bob, in the way described above, erasing part
of the information, is equivalent to a change of state of the conceptual entity under consideration, able to
produce interference effects, when subjected to additional measurements, so it truly has to be modeled as
a superposition state, in accordance with the quantum projection postulate (Aerts, 2009a).
It is interesting to also note that in the above scheme there is no influence exerted by Alice on Bob,
and vice versa: they both simply receive some incomplete information about the outcomes of the joint
measurement creating correlations, and since the information sent to Alice is consistent with that sent to
Bob, in the sense that the erased part for Alice corresponds to the non-erased part for Bob, and vice versa,
and that Alice (Bob) also receive specific information on the possible outcomes of Bob's (Alice's) sub-
measurements, there is no mystery as to why the marginal conditions (25) can be violated, when different
joint measurements are considered, and of course this has no implications as regards the involvement of
superluminal communication processes, even though David communicates simultaneously to Alice and Bob,
sending them information by means of signals traveling, say, at the light speed (assuming that Alice and
Bob would be separated by some large spatial distance).
10 Concluding remarks
Let us conclude our analysis by recapitulating our findings and offering a few additional comments and
contextualization. A central element and crucial insight was the recognition that in its essence the quantum
entanglement phenomenon, both in physical and conceptual systems, is first of all a "name giving" to non-
product structures that appear in a procedure of identification of what are the possible sub-entities forming
a composite system. Such "name giving" procedure will depend on the specific choice of the isomorphism
considered to implement a given bipartite tensor product structure for the space of states, and consequently
also for the linear operators acting on it. A consequence of that, is that entanglement can naturally appear
not only in the states, but also in the observables, i.e., in the measurements that the different observables
represent. It is of course a known fact that the quantum formalism allows for the entanglement resource
to be totally or partially shifted from the state to the observables, so that depending on the considered
factorization a quantum state can appear either entangled or separable (Thirring et al., 2011; Harshman &
Ranade, 2011; Harshman, 2012). Our more specific point is that the interpretative freedom offered by the
possibility of choosing one or multiple isomorphisms, implementing different tensor product factorizations,
becomes crucial when dealing with experimental situations violating the marginal laws. According to our
analysis, it is licit to affirm that in situations where the measurements' outcome-states can be assumed to
remain entangled, the (non-spatial) entanglement resource should be attributed not only to the states, but
also to the accessible interactions (i.e., to the measurements), operating at the level of the overall bipartite
entity.
26
Thinking of entanglement as being present also at the level of the measurements might seem like a
very drastic perspective, compared to the standard situation where it is only attributed to the state of
the bipartite entity, particularly in those experimental situations where there is a clear spatial separation
between the measurement apparatuses working in a coincident way. However, if the measured entity forms
a whole, it is to be expected that also the measurements can become entangled, precisely through the very
wholeness of the measured entity, because their action on the latter would occur simultaneously and not
sequentially.6 In other words, the notions of locality and separability, usually intended as 'spatial locality'
and 'spatial separability', need here be replaced by the more general notions of 'sub-system locality' and
'sub-system separability'. This because among the salient properties of physical and conceptual systems,
there is precisely that of non-spatiality, and therefore 'separation in space' is not anymore a sufficient
criterion for characterizing a separation of two sub-systems and corresponding joint measurements.
We have however also emphasized that it is always possible to also adopt each time, for each joint
measurement, a specific tailor-made entanglement identification. Then, all the entanglement can be pushed,
for each joint measurement, into the state only, with the joint measurement being described as usual as
a product measurement. In this way, everything becomes explicitly contextual, in the sense that for each
coincidence experiment a different state has to be used to represent the compound entity, which can only
be justified when the effect of the measurement is that of disentangling the previously entangled entity.
In Sec. 5, we already emphasized that a violation of the marginal laws does not necessarily imply a
violation of the Einsteinian no-signaling principle.
It is instructive to recall here the typical reasoning
that makes one believe that this is instead to be expected (Ballentine & Jarrett, 1987). One assumes that
Alice and Bob have their laboratories located at great distance from one another, and that they succeeded
sharing a very large number of identically prepared entangled bipartite systems, and that they are also
able to jointly experiment on all of them in a parallel way, so obtaining an entire statistics of outcomes in
a negligible time. Then, if all these extraordinary things can be done with great efficiency, one can imagine
that Alice and Bob could have arranged things in such a way that the choice of which sub-measurement
to perform, say on the part of Bob, is the expression of a code they use to communicate. And since
Bob's choice of sub-measurements can be distinguished by Alice in her statistics of outcomes, because
of the violation of marginal selectivity, one might conclude in this way that some kind of supraluminal
communication could arise between them.
Now, there is a loophole in the above reasoning which is usually not taken into consideration. There is
no doubt that Alice has a means to infer the sub-measurement performed by Bob, if the marginal laws are
violated, but what is the total duration of their overall communication? Certainly, one has to include into
that duration also the time required to prepare the shared entangled entities, on which their numerous joint
measurements have to be performed in parallel. These need to propagate from the source towards the two
interlocutors, who we can assume to be equidistant from the latter. So, if d is the distance separating Alice
from Bob, in the best scenario they can collect all these shared entangled states, sent in parallel, in a time
equal to d
2c . If we assume that their measurements are then associated with instantaneous collapses, one
would be tempted to conclude that their communication can arise at an effective speed of twice the light
speed. What one is here forgetting, however, is to also include the time needed for this whole process to be
initiated. Indeed, the communication does not start when Alice and Bob perform their joint measurements
on a statistical ensemble of entangled entities, but at the moment when they decide to activate the source
in order to use it to communicate. This means that, assuming that the communication is started by Bob,
he will have to send a signal to activate it, which will require at best an additional time d
2c . Hence, the
controlled transfer of information between Alice and Bob cannot happen with effective speed greater than
6It is of course possible to also imagine experimental situations where the sub-measurements would be genuinely sequential,
with mixed orders of execution.
In this case, one can show that the violation of the marginal laws could result from the
possible incompatibility of Alice's and Bob's experimental procedures, giving rise to order effects (Sassoli de Bianchi, 2019b).
27
the speed of light. In other words, if we intend a communication as a process such that the sender of the
message decides when to initiate it, and does so on purpose, then even when the marginal laws are violated
one cannot exhibit a contradiction with special relativity.
In fact, additionally to the theoretical reasoning presented above, there is an even easier way to see the
loophole, which is the following. Consider again the second example we put forward in this article, that of
the elastics violating both the CHSH inequality and the marginal laws. We can easily make an experimental
arrangement such that the events where Alice and Bob gather the outcomes of their joint measurements
are spacelike separated, relativistically speaking. If the spacelike separation would be the criterion which
makes it possible to use the violation of the marginal laws to send signals faster than light, according to
the method sketched above, this should also work with our example of the elastics, which obviously is not
the case, exactly because of the mentioned problem of the extra time needed for a complete execution of
the communication. In other words, if quantum micro-entities in entangled states violating the marginal
laws would lend themselves to signaling, elastic bands would be equally suitable for this purpose, which is
obviously not the case.
One might disagree with the part of the reasoning above saying that a communication between Bob
and Alice can be said to have started only when one of the two has sent a signal to activate the source,
but the fact remains that an effective signaling resulting from correlations of the second kind would be
of a completely different nature than a causal influence of the "spooky action at a distance" kind, which
is usually imagined to happen in situations where entanglement is at play. All sorts of correlated events
happen in our physical (and cognitive) reality, which are spacelike separated, as a result of the fact that
there are common causes at their origin, and nobody would of course dare to say that relativity should
forbid spacelike correlated events based on common causes. Now, our proposed description of quantum
measurements as processes resulting from some 'hidden-measurement interactions' (Aerts & Sassoli de
Bianchi, 2014, 2016), is also of the kind where a common cause would be responsible for the actualization
of two correlated outcomes, which can be spacelike separated, with the only (important) difference that the
common cause in question would be present at the (non-spatial) potential level and would be actualized
each time in a different way, at each run of a joint measurement. From that perspective, even when the
marginal laws are violated, there would be no signaling in the strict sense of the notion, i.e., in the sense
of having signals propagating in space with a velocity greater than that of light (see also the discussion in
Aerts (2014); Sassoli de Bianchi (2019b)), so, no abandonment of the relativity principles is necessary.
Having said that, we conclude by coming back to the criticism expressed by Dzhafarov & Kujala
(2013), which we already mentioned in the Introduction, according to which our analysis of the conceptual
model of Sec. 2.3 would not be revealing of the presence of a genuine form of entanglement, because the
data not only violate the CHSH inequality, but also the marginal laws.7 Such criticism is based on their
Contextuality-by-Default (CbD) approach, which they use to derive a modified CHSH inequality where
the usual expression is corrected by subtracting terms that are non-zero if the marginal laws are violated
(Dzhafarov & Kujala, 2016). This means that according to their modified inequality, a violation of the
marginal laws would generally reduce the amount of entanglement (which they simply call 'contextuality'
in their approach) that a bipartite system is truly manifesting. This goes completely countercurrent with
the view that emerges from our analysis, according to which a violation of the marginal laws, in addition to
a violation of the CHSH inequality, would instead be the signature of a stronger presence of entanglement,
requiring a modeling also at the level of the measurements.8
The reasons for this discrepancy between our interpretation and that of Dzhafarov and Kujala, is that
their definition of contextuality is too restrictive to capture the overall interconnectedness that permeates
7For some previous responses to these criticisms, see also Aerts (2014); Aerts et al. (2018b).
8This is so not only when the marginal laws are disobeyed, but also when they are obeyed but the CHSH inequality is
violated beyond Tsirelson's bound (Aerts & Sozzo, 2013).
28
both the micro-physical and cognitive realms, which give rise to the phenomenon of entanglement and
which in the case of conceptual entities can be generally understood as a 'connection through meaning'.
Indeed, their way of looking at the experimental situation is based on the same kind of prejudice that led
physicists to describe entanglement as a "spooky action at a distance." For Dzhafarov and Kujala, the
joint measurements described in Sec. 2.3, always involve two distinct interrogative contexts, one associated
with Alice, who is assumed to "select an animal," and the other associated with Bob, who is assumed to
"select an act," and these two processes are meant not to influence one another. This is of course a very
unsatisfactory way of depicting the actual experimental situation, which as we explained is the result of a
unique interrogative context that we described using David's fictitious character, whose cognitive process,
when selecting an exemplar for The Animal Acts, cannot be decomposed into two separate processes.
Indeed, it is a process operating at the level of the non-decomposable meaning of the entire sentence, in
the same way that two hands pulling an elastic band operate at the level of the entire non-decomposable
unity of the elastic. And of course, this non-decompositionality of the process will generally be able to
produce a violation of the CHSH inequality and of the marginal laws, the latter being obeyed only if some
remarkable symmetries would be present both in the system and measurement processes.
In other words, Dzhafarov and Kujala's disagreement is an expression of the same prejudice that we
believe is still in force today among some physicists, when they think that a 'spatial separation' should also
imply an 'experimental separation', thus failing to recognize that our physical reality is mostly non-spatial.
That a same kind of classical prejudice would be in force also among some cognitive scientists is a bit more
surprising, considering that in the conceptual domain, because of the all-pervading meaning connections,
the claim that an overall cognitive process should decompose into separate sub-processes is not a very
natural one, for instance because a conceptual combination cannot be modelled using a joint probability
distribution with its variables corresponding to the interpretation of the individual concepts. Dzhafarov
and Kujala's objection is based on the observation that the choice of an exemplar for Animal would be
influenced not only by the options that are offered for that choice, but also by the options that are offered
for the choice of an exemplar for Acts. But speaking in terms of mutual influences (as is systematically
done for instance in the general mathematical theory of 'selective influences'; see Schweickert et al. (2012)),
means that one is already presupposing a possible separation of the two processes, i.e., a splitting of David's
mind into two separate Alice's and Bob's minds, whereas in reality there is only a single-mind cognitive
process or, to put it in a different way, David's mind cannot be understood as a 'juxtaposition' of Alice's
and Bob's sub-minds, but as their 'superposition', as it also emerged from our discussion of Sec. 9.
References
Adenier, G. & Khrennikov, A. (2007). Is the fair sampling assumption supported by EPR experiments? J.
Phys. B: Atomic, Molecular and Optical Physics 40, 131 -- 141.
Adenier, G. & Khrennikov, A. (2017). Test of the no-signaling principle in the Hensen loophole-free CHSH
experiment. Fortschritte der Physik (Progress in Physics) 65, 1600096.
Aerts, D. (1982). Example of a macroscopical situation that violates Bell inequalities. Lettere al Nuovo
Cimento, 34, 107111.
Aerts, D. (1984). The missing elements of reality in the description of quantum mechanics of the EPR
paradox situation. Helvetica Physica Acta 57, 421 -- 428.
Aerts, D. (1986). A possible explanation for the probabilities of quantum mechanics. Journal of Mathe-
matical Physics 27, 202 -- 210.
Aerts, D. (1990). An attempt to imagine parts of the reality of the micro-world. In: J. Mizerski et al.
(Eds.), Problems in Quantum Physics II; Gdansk '89. World Scientific Publishing Company, Singapore.
29
√
Aerts, D. (1991). A mechanistic classical laboratory situation violating the Bell inequalities with 2
2,
exactly 'in the same way' as its violations by the EPR experiments. Helvetica Physica Acta 64, 1 -- 23.
Aerts, D. (1998). "The entity and modern physics: the creation discovery view of reality," in: E. Castellani
(Ed.), Interpreting Bodies: Classical and Quantum Objects in Modern Physics (pp. 223 -- 257). Princeton:
Princeton Unversity Press.
Aerts, D. (1999). "The stuff the world is made of: Physics and reality," in: D. Aerts, J. Broekaert and
E. Mathijs (Eds.), Einstein Meets Magritte: An Interdisciplinary Reflection (pp. 129 -- 183). Dordrecht:
Springer Netherlands.
Aerts, D. (2009a). Quantum structure in cognition. Journal of Mathematical Psychology 53, 314 -- 348.
Aerts, D. (2009b). Quantum particles as conceptual entities: A possible explanatory framework for quantum
theory. Foundations of Science 14, 361 -- 411.
Aerts, S. (2005). A realistic device that simulates the non-local PR box without communication.
arXiv:quant-ph/0504171.
Aerts, D. (2014). Quantum and Concept Combination, Entangled Measurements, and Prototype Theory.
Topics in Cognitive Science 6, 129 -- 137; doi: 10.1111/tops.12073.
Aerts, D. & Aerts, S. (1995). Applications of quantum statistics in psychological studies of decision pro-
cesses. Foundations of Science 1, 85 -- 97.
Aerts, D. and Czachor, M. (2004). Quantum aspects of semantic analysis and symbolic artificial intelligence.
Journal of Physics A: Mathematical and Theoretical 37, L123 -- L132.
Aerts, D. & Gabora, L. (2005). A theory of concepts and their combinations I: The structure of the sets of
contexts and properties. Kybernetes 34, 167 -- 191.
Aerts, D. & Gabora, L. (2005). A theory of concepts and their combinations II: A Hilbert space represen-
tation. Kybernetes 34, 192 -- 221.
Aerts, D. & Sassoli de Bianchi M. (2014). The Extended Bloch Representation of Quantum Mechanics and
the Hidden-Measurement Solution to the Measurement Problem. Annals of Physics 351, 975 -- 1025.
Aerts, D. and Sassoli de Bianchi, M. (2016). "The Extended Bloch Representation of Quantum Mechanics.
Explaining Superposition, Interference and Entanglement," J. Math. Phys. 57, 122110.
Aerts, D. & Sassoli de Bianchi, M. (2017). Do spins have directions? Soft Computing 21, 1483 -- 1504.
Aerts, D., & Sozzo, S. (2011). Quantum structure in cognition. Why and how concepts are entangled. In:
Quantum interaction 2011. Lecture notes in computer science, 7052 (116 -- 127). Berlin: Springer.
Aerts, D. & Sozzo, S. (2013). General Quantum Hilbert Space Modeling Scheme for Entanglement. In:
Proceedings of the Seventh International Conference on Quantum, Nano and Micro Technologies (pp.
25-31), Eds. V. Ovchinnikov and P. Dini, IARIA, 2013.
Aerts, D., & Sozzo, S. (2014a). Quantum entanglement in conceptual combinations. International Journal
of Theoretical Physics 53, 3587 -- 360.
Aerts D., Sozzo S. (2014b) Entanglement Zoo I: Foundational and Structural Aspects. In: Atmanspacher
H., Haven E., Kitto K., Raine D. (eds) Quantum Interaction. QI 2013. Lecture Notes in Computer
Science, vol 8369. Springer, Berlin, Heidelberg, pp. 84 -- 96.
Aerts D., Sozzo S. (2014c) Entanglement Zoo II: Examples in Physics and Cognition. In: Atmanspacher H.,
Haven E., Kitto K., Raine D. (eds) Quantum Interaction. QI 2013. Lecture Notes in Computer Science,
vol 8369. Springer, Berlin, Heidelberg, pp. 97 -- 109.
Aerts, D., Coecke, B., Durt, T. & Valckenborgh, F. (1997). Quantum, classical and intermediate I: a model
on the Poincar´e sphere. Tatra Mountains Mathematical Publications 10, 225.
Aerts, D., Aerts, S., Broekaert, J. & Gabora, L. (2000). The violation of Bell inequalities in the macroworld.
Foundations of Physics, 30, 1387 -- 1414.
Aerts, D., Aerts Arguelles, J., Beltran, L., Geriente, S., Sassoli de Bianchi, M., Sozzo, S & Veloz, T.
(2018a). Spin and wind directions I: Identifying entanglement in nature and cognition. Foundations of
Science 23, 323 -- 335.
Aerts, D., Aerts Arguelles, J., Beltran, L., Geriente, S., Sassoli de Bianchi, M., Sozzo, S & Veloz, T.
30
(2018b). Spin and wind directions II: A Bell State quantum model. Foundations of Science 23, 337 -- 365.
Aerts, D., Beltran, L., Geriente, S. & Sozzo, S. (2019). Quantum-theoretic Modeling in Computer Science.
A complex Hilbert space model for entangled concepts in corpuses of documents; arXiv:1901.04299v1
[cs.CL].
Aerts, D., Broekaert, J. & Smets, S. (1999). A quantum structure description of the liar paradox. Interna-
tional Journal of Theoretical Physics 38, 3231 -- 3239.
Aerts, D., Sassoli de Bianchi, M. & Sozzo, S. (2016). On the foundations of the Brussels operational-realistic
approach to cognition. Frontiers in Physics 4. Doi: 10.3389/fphy.2016.00017.
Aerts, D., Sassoli de Bianchi, M., Sozzo, S. & Veloz, M. (2018). On the Conceptuality interpretation of
Quantum and Relativity Theories. Foundations of Science. https://doi.org/10.1007/s10699-018-9557-z.
Aerts, D., Sassoli de Bianchi, M., Sozzo, S. & Veloz, M. (2019). Quantum cognition goes beyond-quantum:
modeling the collective participant in psychological measurements. In: Probing the Meaning of Quantum
Mechanics - Information, Contextuality, Relationalism and Entanglement. World Scientific, pp. 355 -- 382.
Doi: 10.1142/9789813276895 0017.
Aspect, A., Grangier, P. & Roger, G. (1982a). Experimental realization of Einstein-Podolsky-Rosen-Bohm
Gedankenexperiment: A new violation of Bell's Inequalities. Physical Review Letters 49, 91 -- 94.
Aspect, A., Dalibard, J., & Roger, G. (1982b). Experimental test of Bell's Inequalities using time-varying
analyzers. Physical Review Letters 49, 1804 -- 1807.
Aspect, A. (1983). Trois tests exp´erimentaux des in´egalit´es de Bell par mesure de corr´elation de polarization
de photons. Orsay: Th`ese d'Etat.
Atmanspacher, H., Romer, H. and Walach, H. (2002). Weak quantum theory: Complementarity and
entanglement in physics and beyond. Foundations of Physics 32, 379 -- 406.
Ballentine, L. E. & Jarrett, J. P. (1987). Bell's theorem: Does quantum mechanics contradict relativity?
American Journal of Physics 55, 696 -- 701; doi: 10.1119/1.15059.
Bednorz A. (2017). Analysis of assumptions of recent tests of local realism. Phys. Rev. A 95, 042118.
Bell, J. (1964). On the Einstein Podolsky Rosen paradox. Physics 1, 195 -- 200.
Bell, J. S. (1966). On the problem of hidden variables in quantum mechanics. Review of Modern Physics
38, 447.
Bell, J. (1987). Speakable and Unspeakable in Quantum Mechanics. Cambridge: Cambridge University
Press.
Beltran, L. & Geriente, S. (2018). Quantum entanglement in corpuses of documents. Foundations of Science.
Doi: 10.1007/s10699-018-9570-2.
Bennett, C. H., Brassard, G. Cr´epeau, C., Jozsa, R., Peres A. & Wootters W. K. (1993). Teleporting an
unknown quantum state via dual classical and Einstein-Podolsky-Rosen channels. Phys. Rev. Lett. 70,
1895 -- 1899.
Bohm, D. (1951). Quantum Theory. New-York: Prentice-Hall.
Born, M. (1971). The Born-Einstein Letters 1916-1955. Macmillan Press, New York.
Bruza, P. D., Kitto, K., McEvoy, D. & McEvoy, C. (2008). Entangling words and meaning. In Proceedings
of the second quantum interaction symposium (118 -- 124). Oxford: Oxford University Press.
Bruza, P. D., Kitto, K., Nelson, D. & McEvoy, C. (2009). Extracting spooky-activation-at-a-distance from
considerations of entanglement. In Quantum interaction 2009. Lecture notes in computer science, 5494
(71 -- 83). Berlin: Springer.
Bruza, P., Kitto, K., Ramm, B. & Sitbon, L. (2015). A probabilistic framework for analysing the compo-
sitionality of conceptual combinations. Journal of Mathematical Psychology 67, 26 -- 38.
Busemeyer, J. R. & Bruza, P. D. (2012). Quantum Models of Cognition and Decision, Cambridge University
Press, Cambridge.
Cervantes, V. H. & Dzhafarov, E. N. (2018). Snow Queen is Evil and Beautiful: Experimental Evidence for
Probabilistic Contextuality in Human Choice. To be published in: Decision, arXiv:1711.00418 [q-bio.NC].
Christensen, B. G., McCusker, K.T., Altepeter, J., Calkins, B., Gerrits, T., Lita, A., Miller, A., Shalm, L.
31
K., Zhang, Y., Nam, S. W., Brunner, N., Lim, C. C. W., Gisin, N., & Kwiat, P. G. (2013). Detection-
loophole-free test of quantum nonlocality, and applications. Physical Review Letters 111, 1304 -- 1306.
Clauser, J. F., Horne, M. A., Shimony, A. & Holt, R.A. (1969). Proposed experiment to test local hidden-
variable theories. Physical Review Letters 23, 880 -- 884.
De Raedt, H., Michielsen, K. & Jin, F. (2012). Einstein-Podolsky-Rosen-Bohm laboratory experiments:
Data analysis and simulation. AIP Conf. Proc. 1424, 55 -- 66.
De Raedt H., Jin, F. & Michielsen, K. (2013). Data analysis of Einstein-Podolsky-Rosen-Bohm labora-
tory experiments. Proc. of SPIE 8832, The Nature of Light: What are Photons? V, 88321N; doi:
10.1117/12.2021860.
Deutsch, D. (1998). The Fabric of Reality. Penguin Books, London.
Dzhafarov, E. N. & Kujala, J. V. (2013). On selective influences, marginal selectivity, and Bell/CHSH
inequalities. Topics in Cognitive Science 6, 121 -- 128.
Dzhafarov, E. N. & Kujala, J. V. (2016). Context-content systems of random variables: The contextuality-
by-default theory. Journal of Mathematical Psychology 74, 11 -- 33.
Einstein, A., Podolsky, B. & Rosen, N. (1935). Can Quantum-Mechanical Description of Physical Reality
Be Considered Complete? Physical Review 47, 777 -- 780.
Fuchs, C. A. (2017). On Participatory Realism. In: Durham I., Rickles D. (eds). Information and Interac-
tion. The Frontiers Collection. Springer, Cham.
Gabora, L. & Aerts, D. (2002). Contextualizing concepts using a mathematical generalization of the quan-
tum formalism. Journal of Experimental & Theoretical Artificial Intelligence 14, 327 -- 358.
Genovese, M. (2005). Research on hidden variable theories. A review of recent progresses. Physics Reports
413, 319 -- 396.
Gerlich, S., Eibenberger, S., Tomandl, M., Nimmrichter, S., Hornberger, K., Fagan, P. J., Tuxen, J., Mayor,
M. & Arndt, M. (2011). Quantum interference of large organic molecules. Nature Communications 2,
263.
Giustina, M., Mech, A., Ramelow, S., Wittmann, B., Kofler, J., Beyer, J., Lita, A., Calkins, B., Gerrits,
T., Woo Nam, S., Ursin, R., & Zeilinger, A. (2013). Bell violation using entangled photons without the
fair-sampling assumption. Nature 497, 227 -- 230.
Gronchi, G. & Strambini, E. (2017). Quantum cognition and Bell's inequality: A model for probabilistic
judgment bias. Journal of Mathematical Psychology 78, 65 -- 75.
Harshman, N. L. & Ranade K. S. (2011). Observables can be tailored to change the entanglement of any
pure state, Phys. Rev. A 84, 012303.
Harshman, N. L. (2012). Observables and Entanglement in the Two-Body System. AIP Conference Pro-
ceedings 1508, 386 -- 390.
Haven, E. & Khrennikov, A.Y. (2013). Quantum Social Science, Cambridge University Press, Cambridge.
Hensen, B., Bernien, H., Dr´eau, A. E., Reiserer, A., Kalb, N., Blok, M. S., Ruitenberg, J., Vermeulen, R. F.
L., Schouten, R. N., Abell´an, C., Amaya, W., Pruneri, V., Mitchell, M. W., Markham, M., Twitchen, D.
J., Elkouss, D., Wehner, S., Taminiau T. H., & Hanson R. (2016). Loophole-free Bell inequality violation
using electron spins separated by 1.3 kilometres. Nature, 526, 682 -- 686.
Kastner, R. E. (2013). The transactional Interpretation of Quantum Mechanics: The Reality of Possibility.
Cambridge University Press, New York.
Kochen, S. & Specker, E. P. (1967). The problem of hidden variables in quantum mechanics. J. Math.
Mech. 17, 59 -- 87.
Khrennikov, A. (1999). Classical and quantum mechanics on information spaces with applications to cog-
nitive, psychological, social and anomalous phenomena. Foundations of Physics 29, 1065 -- 1098.
Khrennikov, A. Y. (2010). Ubiquitous Quantum Structure. Berlin: Springer.
Kupczynski, M. (2017). Is Einsteinian no-signalling violated in Bell tests? Open Phys. 15, 739 -- 753, doi:
10.1515/phys-2017-0087.
Sassoli de Bianchi, M. (2011). Ephemeral properties and the illusion of microscopic particles. Foundations
32
of Science 16, 393 -- 409.
Sassoli de Bianchi, M. (2012). From permanence to total availability: a quantum conceptual upgrade.
Foundations of Science 17, 223 -- 244.
Sassoli de Bianchi, M. (2013a). The delta-quantum machine, the k-model, and the non-ordinary spatiality
of quantum entities. Foundations of Science 18, 11 -- 41.
Sassoli de Bianchi, M. (2013b). Using simple elastic bands to explain quantum mechanics: a conceptual
review of two of Aerts' machine-models. Central European Journal of Physics 11, 147 -- 161.
Sassoli de Bianchi, M. (2014). A remark on the role of indeterminism and non-locality in the violation of
Bell's inequality. Annals of Physics 342, 133 -- 142.
Sassoli de Bianchi, M. (2015). "God may not play dice, but human observers surely do," Foundations of
Science 1, 77 -- 105.
Sassoli de Bianchi, M. (2019a). On Aerts' overlooked solution to the EPR paradox. In: Probing the Meaning
of Quantum Mechanics -- Information, Contextuality, Relationalism and Entanglement, World Scientific,
pp. 185 -- 201. Doi: 10.1142/9789813276895 0010.
Sassoli de Bianchi, M. (2019b). An upper bound for quantum correlations of mixed sequential measurements
with order effects. In preparation.
Schrodinger, E. (1935). Discussion of Probability Relations between Separated Systems. Mathematical
Proceedings of the Cambridge Philosophical Society 31, 555 -- 563; doi:10.1017/S0305004100013554.
Schweickert, R., Fisher, D. L., & Sung, K. (2012). Discovering Cognitive Architecture by Selectively Influ-
encing Mental Processes. New Jersey: World Scientific.
Stapp H. P. (2011). Mindful Universe, The Frontiers Collection, Springer-Verlag, Berlin Heidelberg.
Thirring, W., Bertlmann, R. A., Kohler, P. & Narnhofer, H. (2011). Entanglement or separabil-
ity: The choice of how to factorize the algebra of a density matrix. Eur. Phys. J. D 64, 181 -- 196.
https://doi.org/10.1140/epjd/e2011-20452-1.
Tittel, W., Brendel, J., Zbinden, H. & Gisin N. (1998). Violation of Bell's inequalities by photons more
than 10 km apart. Physical Review Letters 81, 3563 -- 3566.
Walborn, S. P., Terra Cunha, M. O., P´adua, S. & Monken, C. H. (2002). Double-slit quantum eraser.
Physical Review A 65, 033818.
Wendt, A. (2015). Quantum mind and social science, Cambridge University Press; Cambridge.
Weihs, G., Jennewein, T., Simon, C., Weinfurter, H. & Zeilinger, A. (1998). Violation of Bell's inequality
under strict Einstein locality condition. Physical Review Letters 81, 5039 -- 5043.
Yin, J. et al. (2017). Satellite-based entanglement distribution over 1200 kilometers. Science 16, 1140 -- 1144.
DOI: 10.1126/science.aan3211.
Yoon-Ho Kim, R. Yu, S. P. Kulik, Y. H. Shih & Scully, M. O. (2000). Delayed "choice" quantum eraser.
Physical Review Letters 84, 1 -- 5.
33
|
1601.01358 | 2 | 1601 | 2017-08-15T14:46:56 | Fast, invariant representation for human action in the visual system | [
"q-bio.NC"
] | Humans can effortlessly recognize others' actions in the presence of complex transformations, such as changes in viewpoint. Several studies have located the regions in the brain involved in invariant action recognition, however, the underlying neural computations remain poorly understood. We use magnetoencephalography (MEG) decoding and a dataset of well-controlled, naturalistic videos of five actions (run, walk, jump, eat, drink) performed by different actors at different viewpoints to study the computational steps used to recognize actions across complex transformations. In particular, we ask when the brain discounts changes in 3D viewpoint relative to when it initially discriminates between actions. We measure the latency difference between invariant and non-invariant action decoding when subjects view full videos as well as form-depleted and motion-depleted stimuli. Our results show no difference in decoding latency or temporal profile between invariant and non-invariant action recognition in full videos. However, when either form or motion information is removed from the stimulus set, we observe a decrease and delay in invariant action decoding. Our results suggest that the brain recognizes actions and builds invariance to complex transformations at the same time, and that both form and motion information are crucial for fast, invariant action recognition. | q-bio.NC | q-bio | A fast, invariant representation for human action in the visual system
Leyla Isik*, Andrea Tacchetti*, and Tomaso Poggio
Center for Brains, Minds, and Machines, MIT
Corresponding author: Leyla Isik
77 Massachusetts Avenue
Bldg 46-4141B
Cambridge, MA 02139
617.258.6933
[email protected]
Running title: Invariant neural representations for human action
* These authors contributed equally to this work
1
Abstract
Humans can effortlessly recognize others' actions in the presence of complex transformations,
such as changes in viewpoint. Several studies have located the regions in the brain involved in
invariant action recognition, however, the underlying neural computations remain poorly
understood. We use magnetoencephalography (MEG) decoding and a dataset of well-
controlled, naturalistic videos of five actions (run, walk, jump, eat, drink) performed by different
actors at different viewpoints to study the computational steps used to recognize actions across
complex transformations. In particular, we ask when the brain discounts changes in 3D
viewpoint relative to when it initially discriminates between actions. We measure the latency
difference between invariant and non-invariant action decoding when subjects view full videos
as well as form-depleted and motion-depleted stimuli. Our results show no difference in
decoding latency or temporal profile between invariant and non-invariant action recognition in
full videos. However, when either form or motion information is removed from the stimulus set,
we observe a decrease and delay in invariant action decoding. Our results suggest that the
brain recognizes actions and builds invariance to complex transformations at the same time,
and that both form and motion information are crucial for fast, invariant action recognition.
Keywords
Action recognition, Magnetoencephalography, Neural decoding, Vision
2
As a social species, humans rely on recognizing the actions of others in their everyday lives. We
quickly and effortlessly extract action information from rich dynamic stimuli, despite variations in
the visual appearance of action sequences, due to transformations such as changes in size,
position, actor, and viewpoint (e.g., is this person running or walking towards me, regardless of
which direction they are coming from). The ability to recognize actions, the middle ground
between action primitives (e.g., raise the left foot and move it forward) and activities (e.g.,
playing basketball) (Moeslund and Granum 2001), is paramount to humans' social interactions
and even survival. The computations driving this process, however, are poorly understood. This
lack of computational understanding is evidenced by the fact that even state of the art computer
vision algorithms, convolutional neural networks, which match human performance on object
recognition tasks (He et al. 2015), still drastically underperform humans on action recognition
tasks (Le et al. 2011; Karpathy et al. 2014). In particular, what makes action and other visual
recognition problems challenging are transformations (such as changes in scale, position and
3D viewpoint) that alter the visual appearance of actions, but are orthogonal to the recognition
task (DiCarlo and Cox 2007).
Several studies have attempted to locate the regions in the brain involved in processing
actions, and in some cases, locate regions in the brain containing viewpoint-invariant
representations. In humans and nonhuman primates, the extrastriate body area (EBA) has been
implicated in recognizing human form and action (Michels et al. n.d.; Downing et al. 2001;
Lingnau and Downing 2015), and the superior temporal sulcus (STS) has been implicated in
recognizing biological motion and action (Perrett et al. 1985; Oram and Perrett 1996; Grossman
et al. 2000; Vaina et al. 2001; Grossman and Blake 2002; Beauchamp et al. 2003; Peelen and
Downing 2005; Vangeneugden et al. 2009). The posterior portion of the STS (pSTS) represents
particular types of biological motion data in a viewpoint invariant manner (Grossman et al. 2010;
Vangeneugden et al. 2014). Beyond visual cortex, action representations have been found in
3
human parietal and premotor cortex when people perform and view certain actions, particularly
hand grasping and goal-directed behavior (analogous to monkey "mirror neuron" system)
(Hamilton and Grafton 2006; Dinstein, Gardner, et al. 2008; Dinstein, Thomas, et al. 2008;
Oosterhof et al. 2010, 2012, 2013; Freeman et al. 2013). However, recent work suggests that
these "mirror neuron" regions do not code the abstract, invariant representations of actions,
which are coded in occipitotemporal regions (Wurm et al. 2015, 2016).
Here we investigate the neural dynamics of action processing, rather than the particular
brain regions
involved,
in order
to elucidate
the underlying computations. We use
magnetoencephalography (MEG) decoding to understand when action information is present
and how the brain computes representations that are invariant to complex, non-affine
transformations such as changes
in viewpoint. Timing
information can constrain
the
computations underlying visual recognition by informing when different visual representations
are computed. For example, recent successes in MEG decoding have revealed interesting
properties about invariant object recognition in humans, mainly that it is fast and highly dynamic,
and that varying levels of abstract categorization and invariance increase over the first 200ms
following image onset (Carlson et al. 2011, 2013; Cichy et al. 2014; Isik et al. 2014).
Specifically, we use timing data to distinguish between two competing hypotheses of
how the brain computes invariance to complex, non-affine transformations. The first hypothesis
is that an initial recognition module identifies which action is being performed, and then
invariance to various transformations is computed at a later processing stage by discounting
specific poses or actor identity information. This hypothesis is consistent with previous data for
invariant face and object recognition where shapes can first be discriminated without any
generalization and then invariance develops at later stages in the visual processing hierarchy
(Logothetis and Sheinberg 1996; Freiwald and Tsao 2010), and is also the basis for viewpoint-
invariant computational models of vision (Leibo et al. 2017). Alternatively, the brain may carry
4
out recognition and invariance computations simultaneously, in which case actions could be
decoded at the same time within and across different transformations.
Our results show no difference in decoding latency or temporal profile between invariant
and non-invariant action recognition, supporting the hypothesis that the brain performs
recognition and invariance in the same processing stage. We further show that two types of
action information, form (as tested with static images) and motion (as tested with point light
figures), both contribute to these immediately view-invariant representations; when either form
or motion information is removed, view-invariant decoding is no longer computed at the same
time as non-invariant decoding. These results suggest that features that are rich in form and
motion content drive the fast, invariant representation of the actions in the human brain.
Materials and Methods
Action recognition dataset
To study the effect of changes in view on action recognition, we used a dataset of five actors
performing five different actions (drink, eat, jump, run and walk) on a treadmill from two different
views (0 and 90 degrees from the front of the actor/treadmill; the treadmill rather than the
camera was rotated in place to film from different viewpoints) [Figure 1] (Tacchetti et al. 2016).
These actions were selected to be highly familiar, and thus something subjects would have
experienced under many viewing conditions, to include both reaching-oriented (eat and drink)
and leg-oriented (jump, run, walk) actions, as well as to span both coarse (eat and drink versus
run and walk) and fine (eat versus drink and run versus walk) action distinctions. Every video
was filmed on the same background, and the same objects were present in each video,
regardless of action. Each action-actor-view combination was filmed for at least 52-seconds.
The videos were then cut into two-second clips that each included at least one cycle of each
action, and started at random points in the cycle (for example, a jump may start midair or on the
5
ground). This dataset allows testing of actor and view invariant action recognition, with few low-
level confounds.
To explore the roles of form and motion in invariant action representations, we extended
this video dataset with two additional components: a form only dataset, consisting of
representative single frames for each action, and a motion-only dataset, consisting of point light
figures performing the same actions. For the form dataset, the authors selected one frame per
video making sure that the selected frames were unambiguous for action identity (special
attention was paid to the actions eat and drink to ensure the food or drink was near the mouth,
and occluded views to ensure there was some visual information about action). For the motion
point light dataset, the videos were put on Amazon Mechanical Turk and workers were asked to
label 15 landmarks in every single frame: center of head, shoulders, elbows, hands, torso, hips,
knees, and ankles. Three workers labeled each video frame. We used the spatial median of the
three independent labels for each frame and landmark to increase the signal to noise ratio, and
independently low-pass filtered the time series (Gaussian Filter with a 30 frames aperture and
normalized convolution) for each of the 15 points to reduce the high frequency artifacts
introduced by single-frame labeling.
Participants
Three separate MEG experiments were conducted (see below). Ten subjects (5 female)
participated in experiment one, ten subjects (7 female) participated experiment two, and ten
subjects (7 female) participated in experiment three. All subjects had normal or corrected to
normal vision. The MIT Committee on the Use of Humans as Experimental Subjects approved
the experimental protocol. Subjects provided informed written consent before the experiment.
Experimental procedure
6
In the first experiment, we assessed if we could read out different actions both within viewpoint
(training and testing on videos at 0 degrees or 90 degrees, without any generalization) and
across viewpoint, by training and testing on two different views (0 and 90 degrees). In this
experiment ten subjects were shown 50 two-second video clips (one for each of five actors,
actions, and two views, 0 and 90 degrees), each presented 20 times.
To examine whether form and motion information were necessary to construct invariant
action representations, in the second and third experiments we showed subjects limited "form"
(static image) or "motion" (point-light walkers) datasets. Specifically, in the second experiment,
ten subjects were shown 50 static images (one for each of five actors, actions, and two views, 0
and 90 degrees), which were single frames from the videos in Experiment 1, for 2 seconds
presented 20 times each. In the third experiment, ten subjects were shown 10 two-second video
clips, which consisted of point-light walkers traced along one actor's videos from two views in
experiment one (labelled by Mechanical Turk workers as described above), presented 100
times each.
In each experiment, subjects performed an action recognition task, where they were
asked after a random subset of videos or images (in a randomly interspersed 10% of the trials
for each video or image) what action was portrayed in the previous image or video. The purpose
of this behavioral task was to ensure subjects were attentive and assess behavioral
performance on the various datasets. The button order for each action was randomized across
trials to avoid systematic motor confounds in the decoding. Subjects were instructed to fixate
centrally. The videos were presented using Psychtoolbox to ensure accurate timing of stimulus
onset. Each video had a duration of 2s and a 2s inter-stimulus interval. The videos were shown
in grayscale at 3 x 5.4 degrees of visual angle on a projector with a 48 cm × 36 cm display, 140
cm away from the subject.
7
MEG data acquisition and preprocessing
The MEG data were collected using an Elekta Neuromag Triux scanner with 306 sensors, 102
magnetometers at 204 planar gradiometers, and were sampled at 1000 Hz. First the signals
were filtered using temporal Signal Space Separation (tSSS) with Elekta Neuromag software.
Next, Signal Space Projection (SSP) (Tesche et al. 1995) was applied to correct for movement
and sensor contamination. The MEG data were divided into epochs from -500 - 3500 ms,
relative to video onset, with the mean baseline activity removed from each epoch. The signals
were band-pass filtered from 0.1–100 Hz to remove external and irrelevant biological noise
(Acunzo et al. 2012; Rousselet 2012). The convolution between signals and bandpass filter was
implemented by wrapping signals in a way that may introduce edge effects at the beginning and
end of each trial. We mitigated this issue by using a large epoch window (-500-3500 ms) and
testing significance in a manner that takes into account temporal biases in the data (see
significance testing below). The above pre-processing steps were all implemented using the
Brainstorm software (Tadel et al. 2011).
General MEG decoding methods
MEG decoding analyses were performed with the Neural Decoding Toolbox (Meyers 2013), a
Matlab package implementing neural population decoding methods. In this decoding procedure,
a pattern classifier was trained to associate the patterns of MEG data with the identity of the
action in the presented image or video. The stimulus information in the MEG signal was
evaluated by testing the accuracy of the classifier on a separate set of test data. This procedure
was conducted separately for each subject and multiple re-splits of the data into training and
test data were utilized.
The time series data of the magnetic field measured in each sensor (including both the
magnetometers and gradiometers) were used as classifier features. We averaged the data in
8
each sensor into 100 ms overlapping bins with a 10 ms step size, and performed decoding
independently at each time point. Decoding analysis was performed using cross validation,
where the dataset was randomly divided into five cross validation splits. The classifier was then
trained on data from four splits (80% of the data), and tested on the fifth, held out split (20% of
the data) to assess the classifier's decoding accuracy.
Decoding - feature pre-processing
To improve signal to noise, we averaged together the different trials for each semantic
class (e.g. videos of run) in a given cross validation split so there was one data point per
stimulus per cross validation split. We next Z-score normalized that data by calculating the
mean and variance for each sensor using only the training data. We then performed sensor
selection by applying a five-way ANOVA to each sensor's training data to test if the sensor was
selective for the different actions. We use sensors that were selective for action identity, i.e.,
show a significantly greater variation across class than within class, with p<0.05 significance
based on a F-test (if no sensors were deemed significant, the one with the lowest p-value is
selected). The selected sensors were then fixed and used for testing. Each sensor (including
both magnetometers and gradiometers) was considered as an independent sensor input into
this algorithm, and the feature selection, like the other decoding steps is performed separately
at each 100ms time bin, and thus a different number of sensorswas selected for each subject at
each time bin. The average number of sensors selected for each subject across all significant
decoding time bins is shown in Supplemental Table 2. These pre-processing parameters have
been shown to empirically improve MEG decoding signal to noise in a previous MEG decoding
study (Isik et al. 2014), however as we did not use absolute decoding performance (rather
significantly above chance decoding) as a metric for when information is present in the MEG
signals, we did not further optimize decoding performance with the present data.
9
Decoding - classification
The pre-processed MEG data was then input into the classifier. Decoding analyses were
performed using a maximum correlation coefficient classifier, which computed the correlation
between each test vector and a mean training vector that is created from taking the mean of the
training data from a given class. Each test point was assigned the label of the class of the
training data with which it was maximally correlated. When we refer to classifier "training" this
could alternatively be thought of as learning to discriminate patterns of electrode activity
between the different classes in the training data, rather than a more involved training procedure
with a more complex classifier. We intentionally chose a very simple algorithm to see in the
simplest terms what information is coded in the MEG data. Prior work has also shown
empirically that results with a correlation coefficient classifier are very similar to standard linear
classifiers like support vector machines (SVMs) or regularized least squares (RLS) (Isik et al.
2014).
We repeated the above decoding procedure over 50 cross validation splits, at each time
bin to assess the decoding accuracy versus time. We measured decoding accuracy as the
average percent correct of the test set data across all cross-validation splits, and reported
decoding results for the average of ten subjects in each experiment. Plots and latency
measures were centered at the median value of each of the 100ms time bins.
For more details on these decoding methods see (Isik et al. 2014).
Decoding invariant information
To see if information in the MEG signals could generalize across a given transformation,
we trained the classifier on data from subjects viewing the stimuli under one condition (e.g. 0-
degree view) and tested the classifier on data from subjects viewing the stimuli under a
10
separate, held out condition (e.g. 90-degree view). This provided a strong test of invariance to a
given transformation. In all three experiments, we compared the within and across view
decoding. For the "within" view case, the classifier was trained on 80% of data from one view,
and tested on the remaining 20% of data from the same view. For the "across" view case, the
classifier was trained on 80% of data from one view, and tested on 20% of data from the
opposite view, so the same amount of training and test data was evaluated in each case.
Significance testing
We assessed action decoding significance using a permutation test. We ran the decoding
analysis for each subject with the labels randomly shuffled to create a null distribution. Shuffling
the labels breaks the relationship between the experimental conditions that occurred. We
repeated the procedure of shuffling the labels and running the decoding analysis 100 times to
create a null distribution, and reported p-values as the percentage rank of the actual decoding
performance within the null distribution.
For each experiment and decoding condition, we averaged the null decoding data
across ten subjects and determined when the mean decoding across subjects was above the
mean null distribution. We define the decoding "onset time" as the first time the subject-
averaged decoding accuracy was greater than the subject-averaged null distribution, with p <
0.05. This provided a measure of when significant decodable information was first present in the
MEG signals, and is a standard metric to compare latencies between different conditions (Isik et
al. 2013; Cichy et al. 2016). Time of peak decoding accuracy for each condition, an alternative
established measure of decoding latency, was found to be much more variable (with 95% CI
that were on average over 400 ms larger than onset times), we therefore restricted ourselves to
using onset latency only.
11
To compare onset latencies for different decoding conditions (e.g. within view versus
across view decoding), we performed 1000 bootstrap resamples of subjects and use the
resulting distribution to compute empirical 95%-confidence intervals for the onset latency of ech
condition to estimate the temporal sensitivity of our measure (Hoenig and Heisey 2001), as well
as for the difference in onset latency between the two conditions. Specifically, in each bootstrap
run, we randomly selected a different subset of ten subjects with replacement, computed onset
latencies for each condition (as outlined above) and calculated the difference in onset latency
between the invariant and non-invariant conditions. We defined the onset latencies for invariant
and non-invariant decoding significantly different with p<0.05 if the empirical 95% interval for
their difference did not include 0 (Cichy et al. 2016).
Temporal Cross Training
Beyond decoding latency, we sought to examine the dynamics of the MEG decoding
using temporal-cross-training analysis (Meyers et al. 2008; Meyers 2013; Isik et al. 2014; King
and Dehaene 2014). In this analysis, rather than training and testing the classifier on the same
time point, a classifier was trained with data from one time point and then tested on data from all
other time points. Otherwise the decoding methods (including feature pre-processing, cross
validation and classification) were identical to the procedure outlined above. This method
yielded a matrix of decoding accuracies for each training and test time point, where the rows of
the matrix indicate the times when the classifier was trained, and the columns indicate the times
when the classifier was tested. The diagonal entries of this matrix contained the results from
when the classifier was trained and tested on data from the same time point (identical to the
procedure described above).
Results
Readout of actions from MEG data is early and invariant
12
Ten subjects viewed 2-second videos of five actions performed by five actors at two views (0
degrees and 90 degrees) (Figure 1, top row) while their neural activity was recorded in the
MEG. We then trained our decoding classifier on only on one view (0 degrees or 90 degrees),
and tested it on the second view (0 degrees or 90 degrees). Action can be read out from the
subjects' MEG data in the case without any invariance ("within view" condition) at, on average,
250 ms (210-330 ms) (mean decoding onset latency across subjects based on p<0.05
permutation test, 95% CI reported throughout in parentheses, see Methods) [Figure 3, blue
trace, Table 1].
The onset latency was much less than the 2s long stimulus and the cycle of most actions
(drink, eat, jump which were filmed to each last approximately 2s) in the dataset. Each video
began at a random point in a given action sequence, suggesting that the brain can compute this
representation from different partial sequences of each action. We also observed a significant
rise in decoding after the video offset, consistent with offset responses that have been observed
in MEG decoding of static images (Carlson et al. 2011).
We assessed if the MEG signals are invariant to changes in viewpoint by training the
classifier on data from subjects viewing actions performed at one view and testing on a second
held out view. This invariant "across-view" decoding arose on average at 230 ms (220-270ms).
The onset latencies between the within and across were not significantly different at the group
level (p = 0.13), suggesting that the early action recognition signals are immediately view
invariant.
To ensure that the lack of latency difference between the within and between view
conditions is not due to the fact that we are using 100ms overlapping time bins, we re-ran the
decoding 10ms time bins and 10ms step size (non-overlapping time bins) [Supplemental Figure
3]. Although the overall decoding accuracy is lower, the within and across view decoding onsets
were still not significantly different (p = 0.62).
13
The dynamics of invariant action recognition
Given that the within- and across-view action decoding conditions ha similar onset latencies, we
further compared the temporal profiles of the two conditions by asking if the neural codes for
each condition are stable over time. To test this, we trained our classifier with data at one time
point, and tested the classifier at all other time points. This yielded a matrix of decoding
accuracies for different train times by test times, referred to as a temporal cross training (TCT)
matrix (Meyers et al. 2008; Carlson et al. 2013; Meyers 2013; Isik et al. 2014). The diagonal of
this matrix shows when the classifier is trained and tested with data at the same time point, just
as the line plots in Figure 3.
The within-view and across-view TCTs show that the representations for actions, both
with and without view, are highly dynamic as there is little off-diagonal decoding that is
significantly above chance (Figure 4a-b, Supplemental Figure 4a-b). The window of significantly
above chance decoding performance from 200-400 ms, in particular, is highly dynamic and
decoding only within a 50-100 ms window is significantly above chance. At later time points, the
above chance decoding extends to a larger window that spans 300ms, suggesting the late
representations for action are more stable across time than the early representations. Further,
we find that significant decoding for the within and across view conditions are largely
overlapping (Supplemental Figure 4C) showing that information for both conditions are
represented at the same time scale in the MEG data.
Coarse and fine action discrimination
We examined how the five different actions are decoded in both the within and across
decoding conditions. By analyzing the confusion matrices for the within- and across-view
decoding, we see that not only are coarse action distinctions made (e.g., between run/walk and
14
eat/drink), but so are fine action distinctions (e.g., between eat and drink) even at the earliest
decoding of 250 ms (Supplemental Figure 2). Further, actions performed in a familiar context
(i.e. run and walk on a treadmill) are not better classified than those performed in an unfamiliar
context (i.e. eat and drink on a treadmill).
To more fully examine the difference between coarse and fine action distinctions, we
performed decoding on four action pairs - two action pairs that spanned body parts and can be
done based on coarse discrimination (drink/run and eat/walk), and two action pairs involving the
same body parts and require
fine discrimination (run/walk and eat/drink), (Figure 4,
Supplemental Figure 5). This decoding is above chance in three of the four pairs (excluding,
eat/drink), and as expected, the coarse action distinctions are better decoded than the fine
action distinctions. Importantly, in all four pairs there is no significant difference between the
latencies for the within and across view decoding (Table 1), strengthening the notion that the
representation of action sequence computed by the human brain is immediately invariant to
changes in 3D viewpoint.
The roles of form and motion in invariant action recognition
To study the roles of two information streams, form and motion, in action recognition, subjects
viewed two limited stimulus-sets in the MEG. The first 'Form' stimulus set consisted of one static
frame from each video (containing no motion information). The second 'Motion' stimulus set,
consisted of point light figures that are comprised of dots on each actor's head, arm joints, torso,
and leg joints and move with the actor's joints (containing limited form information) (Johansson
1973).
Ten subjects viewed each of the form and motion datasets in the MEG. We could
decode action from both datasets in the within view case without any invariance (Figure 5). The
early view-invariant decoding that was observed with full movies, however, was impaired for
15
both the form or motion datasets. In the form-only experiment, within view could be read out at
410 ms and across view at 510ms. The decoding for both the within and across view conditions,
however, were not reliably above chance as the lower limit of the 95% CI for both the within and
across view decoding fell below the significance threshold. In the motion-only experiment, within
view action information could be read out significantly earlier than across view information: 210
ms (180-260 ms) versus 300 ms (300-510 ms), and was significantly different between the two
conditions (p = 0.013).
Discussion
We investigated the dynamics of invariant action recognition in the human brain and
found that action can be decoded from MEG signals around 200 ms post video onset. This is
extremely fast, particularly given that the duration of each video and most action cycles (e.g.,
one drink from a water bottle) was 2s. These results are consistent with a recent MEG decoding
study that classified two actions, reaching and grasping, slightly after 200ms post video onset
(Tucciarelli et al. 2015). Crucially, we showed that these early neural signals are selective to a
variety of full-body actions as well as invariant to changes in 3-D viewpoint.
The fact that invariant and non-invariant action representations are computed at the
same time may represent a key difference between the neural computations underlying object
and action recognition. Invariant object information increases along subsequent layers of the
ventral stream (Logothetis and Sheinberg 1996; Rust and Dicarlo 2010) causing a delay in
invariant decoding relative to non-invariant decoding (Isik et al. 2014). Further, physiology data
(Freiwald and Tsao 2010) and computational models (Leibo et al. 2017) of static face
recognition have shown that invariance to 3D viewpoint, in particular, arises at a later
processing stage than initial face recognition. One possible account of this discrepancy is that
even non-invariant ("within view") action representations rely on higher-level visual features (that
16
carry some degree of viewpoint invariant information), than those used in basic object
representations.
We characterized the dynamics of action representations using temporal cross training
and found that the decoding windows for within and across view decoding are largely
overlapping (Supplemental Figure 4C), suggesting that the beyond onset latencies, the overall
dynamics of decoding are similar for non-invariant and view-invariant action representations. It
has been suggested that visual recognition, as studied with static object recognition, has a
canonical temporal representation that is demonstrated by highly diagonal TCT matrices (King
and Dehaene 2014). Our action results seem to follow this pattern (Figure 4). Representations
for action are highly dynamic, but they are more stable over time than previously reported for
object decoding (Carlson et al. 2013a; Cichy et al. 2014; Isik et al. 2014).
As shown previously, we find that people can recognize actions with either biological
motion or form information removed from the stimulus (Johansson 1973; Schindler and van
Gool 2008; Singer and Sheinberg 2010), and that decoding actions within-view is largely intact
when form or motion cues are removed. This is likely due to the fact that within-view decoding,
unlike the across-view condition, requires little generalization and can thus be performed using
low-level cues in the form or motion stimuli. The across-view decoding, on the other hand,
requires substantially more generalization and cannot be performed as well, or as quickly with
form or motion depleted stimuli. While our datasets are a best attempt to isolate form and
motion information, it is important to note that static images contain implied motion and that
point light figures contain some form information and have less motion information than full
movies. Nevertheless, the low-accuracy and delayed decoding with either limited stimulus set
suggests that both form and motion information are necessary to build a robust action
representation.
17
Importantly these invariant action representations cannot be explained by low-level
stimulus features, such as motion energy (Tacchetti et al. 2016). While we cannot fully rule out
the effects of eye movements or shifts in covert attention, eye movement patterns cannot be
accounting for our early MEG decoding accuracy, because we do not observe a significant shift
in the eye positions between different actions until after 600 ms post video onset and further the
same decoder applied to MEG signals does not successfully decode action information using
raw eye position data (Supplemental Figure 1).
The five actions tested in this study comprise only a small subset of the wide variety of
familiar actions we recognize in our daily lives. The five-way classification shows similar
decoding across between all five actions, including both coarse and fine action distinctions
(Supplemental figure 2). The binary action decoding results show that, while coarse action
discriminations are more easily decoded than fine action discriminations, there is no difference
in the latency of within and across view decoding for any action pair, suggesting that invariant
and non-invariant action signals are indeed computed at the same stage (Figure 4,
Supplemental Figure 5).
These five actions were selected to be highly familiar, and thus we do not know to what
extent familiarity is necessary for the immediate invariance we observed. Indeed, modeling and
theoretical work suggest that in order to build templates to be invariant to non-affine
transformations such as changes in 3-D viewpoint, one must learn templates from different
views of each given category (Leibo et al. 2015). It remains an open question how this
invariance would translate to unfamiliar actions and how many examples would be needed to
learn invariant representations of new actions.
Finally, the longer latency and greater cross-temporal stability of action decoding raises
the question of whether recurrent and feedback connections are used to form invariant action
representations. This is difficult to test explicitly without high spatiotemporal resolution data.
18
Beyond 100ms post-stimulus onset it is likely that feedback and recurrent connections occur
(Lamme and Roelfsema 2000), however further studies have shown that this data is well
approximated by feedforward hierarchical models (Tacchetti et al. 2016).
Taken as a whole, our results show that the brain computes action selective
representations remarkably quickly and, unlike in the recognition of static faces and objects, at
the same time that it computes invariance to complex transformations that are orthogonal to the
recognition task. This contrast represents a qualitative difference between action and object
visual processing. Moreover, our findings suggest that both form and motion information are
necessary to construct these fast invariant representations of human action sequences. The
methods and results presented here provide a framework to study the dynamic neural
representations evoked by natural videos, and open the door to probing neural representations
for higher level visual and social information conveyed by video stimuli.
Acknowledgements
This material is based upon work supported by the Center for Brains, Minds and Machines
(CBMM), funded by NSF STC award CCF-1231216. We thank the McGovern Institute for Brain
Research at MIT and the Athinoula A. Martinos Imaging Center at the McGovern Institute for
Brain Research at MIT for supporting this research. We would like to thank Patrick Winston,
Gabriel Kreiman, Martin Giese, Charles Jennings, Heuihan Jhuang, and Cheson Tan for their
feedback on this work.
19
References
Acunzo DJ, Mackenzie G, van Rossum MCW. 2012. Systematic biases in early ERP and ERF
components as a result of high-pass filtering. J Neurosci Methods. 209:212–218.
Beauchamp MS, Lee KE, Haxby J V, Martin A. 2003. FMRI responses to video and point-light
displays of moving humans and manipulable objects. J Cogn Neurosci. 15:991–1001.
Carlson TA, Hogendoorn H, Kanai R, Mesik J, Turret J. 2011. High temporal resolution
decoding of object position and category. J Vis. 11.
Carlson T, Tovar DA, Alink A, Kriegeskorte N. 2013. Representational dynamics of object vision:
the first 1000 ms. J Vis. 13.
Cichy RM, Pantazis D, Oliva A. 2014. Resolving human object recognition in space and time.
Nat Neurosci. 17:455–462.
Cichy RM, Pantazis D, Oliva A. 2016. Similarity-Based Fusion of MEG and fMRI Reveals
Spatio-Temporal Dynamics in Human Cortex During Visual Object Recognition. Cereb
Cortex. bhw135.
DiCarlo JJ, Cox DD. 2007. Untangling invariant object recognition. Trends Cogn Sci. 11:333–
341.
Dinstein I, Gardner JL, Jazayeri M, Heeger DJ. 2008. Executed and Observed Movements
Have Different Distributed Representations in Human aIPS. J Neurosci. 28:11231–11239.
Dinstein I, Thomas C, Behrmann M, Heeger DJ. 2008. A mirror up to nature. Curr Biol. 18:R13-
8.
Downing PE, Jiang Y, Shuman M, Kanwisher N, Downing PE, Jiang Y, Jiang Y, Shuman M,
Shuman M, Kanwisher N, Kanwisher N. 2001. A cortical area selective for visual
20
processing of the human body. Science. 293:2470–2473.
Freeman J, Ziemba CM, Heeger DJ, Simoncelli EP, Movshon JA. 2013. A functional and
perceptual signature of the second visual area in primates. Nat Neurosci. 16:974–981.
Freiwald WA, Tsao DY. 2010. Functional compartmentalization and viewpoint generalization
within the macaque face-processing system. Science. 330:845–851.
Grossman E, Donnelly M, Price R, Pickens D, Morgan V, Neighbor G, Blake R. 2000. Brain
Areas Involved in Perception of Biological Motion. J Cogn Neurosci. 12:711–720.
Grossman ED, Blake R. 2002. Brain Areas Active during Visual Perception of Biological Motion.
Neuron. 35:1167–1175.
Grossman ED, Jardine NL, Pyles JA. 2010. fMR-Adaptation Reveals Invariant Coding of
Biological Motion on the Human STS. Front Hum Neurosci. 4:15.
Hamilton AF d. C, Grafton ST. 2006. Goal Representation in Human Anterior Intraparietal
Sulcus. J Neurosci. 26:1133–1137.
He K, Zhang X, Ren S, Sun J. 2015. Delving Deep into Rectifiers: Surpassing Human-Level
Performance on ImageNet Classification. arXiv:1502.01852.
Hoenig JM, Heisey DM. 2001. The Abuse of Power. Am Stat. 55:19–24.
Isik L, Meyers EM, Leibo JZ, Poggio T. 2014. The dynamics of invariant object recognition in the
human visual system. J Neurophysiol. 111:91–102.
Johansson G. 1973. Visual perception of biological motion and a model for its analysis. Percept
Psychophys. 14:201–211.
Karpathy A, Toderici G, Shetty S, Leung T, Sukthankar R, Fei-Fei L. 2014. Large-Scale Video
Classification with Convolutional Neural Networks. In: 2014 IEEE Conference on Computer
21
Vision and Pattern Recognition. IEEE. p. 1725–1732.
King J-R, Dehaene S. 2014. Characterizing the dynamics of mental representations: the
temporal generalization method. Trends Cogn Sci. 18:203–210.
Lamme VAF, Roelfsema PR. 2000. The distinct modes of vision offered by feedforward and
recurrent processing. Trends Neurosci. 23:571–579.
Le Q V., Zou WY, Yeung SY, Ng AY. 2011. Learning hierarchical invariant spatio-temporal
features for action recognition with independent subspace analysis. In: CVPR 2011. IEEE.
p. 3361–3368.
Leibo JZ, Liao Q, Anselmi F, Freiwald WA, Poggio T. 2017. View-Tolerant Face Recognition
and Hebbian Learning Imply Mirror-Symmetric Neural Tuning to Head Orientation. Curr
Biol. 27:62–67.
Leibo JZ, Liao Q, Anselmi F, Poggio TA. 2015. The Invariance Hypothesis Implies Domain-
Specific Regions in Visual Cortex. PLOS Comput Biol. 11:e1004390.
Lingnau A, Downing PE. 2015. The lateral occipitotemporal cortex in action. Trends Cogn Sci.
Logothetis NK, Sheinberg DL. 1996. Visual object recognition. Annu Rev Neurosci. 19:577–621.
Meyers EM. 2013. The neural decoding toolbox. Front Neuroinform. 7.
Meyers EM, Freedman DJ, Kreiman G, Miller EK, Poggio T. 2008. Dynamic population coding
of category information in inferior temporal and prefrontal cortex. J Neurophysiol.
100:1407–1419.
Michels L, Lappe M, Vaina LM. n.d. Visual areas involved in the perception of human movement
from dynamic form analysis. Neuroreport. 16:1037–1041.
Moeslund TB, Granum E. 2001. A Survey of Computer Vision-Based Human Motion Capture.
22
Comput Vis Image Underst. 81:231–268.
Oosterhof NN, Tipper SP, Downing PE. 2012. Viewpoint (in)dependence of action
representations: an MVPA study. J Cogn Neurosci. 24:975–989.
Oosterhof NN, Tipper SP, Downing PE. 2013. Crossmodal and action-specific: neuroimaging
the human mirror neuron system. Trends Cogn Sci. 17:311–318.
Oosterhof NN, Wiggett AJ, Diedrichsen J, Tipper SP, Downing PE. 2010. Surface-Based
Information Mapping Reveals Crossmodal Vision–Action Representations in Human
Parietal and Occipitotemporal Cortex. J Neurophysiol. 104.
Oram MW, Perrett DI. 1996. Integration of form and motion in the anterior superior temporal
polysensory area (STPa) of the macaque monkey. J Neurophysiol. 76:109–129.
Peelen M V, Downing PE. 2005. Selectivity for the human body in the fusiform gyrus. J
Neurophysiol. 93:603–608.
Perrett DI, Smith PAJ, Mistlin AJ, Chitty AJ, Head AS, Potter DD, Broennimann R, Milner AD,
Jeeves MA. 1985. Visual analysis of body movements by neurones in the temporal cortex
of the macaque monkey: A preliminary report. Behav Brain Res. 16:153–170.
Rousselet GA. 2012. Does Filtering Preclude Us from Studying ERP Time-Courses? Front
Psychol. 3:131.
Rust NC, Dicarlo JJ. 2010. Selectivity and tolerance ("invariance") both increase as visual
information propagates from cortical area V4 to IT. J Neurosci. 30:12978–12995.
Schindler K, van Gool L. 2008. Action snippets: How many frames does human action
recognition require? In: 2008 IEEE Conference on Computer Vision and Pattern
Recognition. IEEE. p. 1–8.
Singer JM, Sheinberg DL. 2010. Temporal cortex neurons encode articulated actions as slow
23
sequences of integrated poses. J Neurosci. 30:3133–3145.
Tacchetti A, Isik L, Poggio T. 2016. Spatio-temporal convolutional neural networks explain
human neural representations of action recognition. arXiv:1606.04698.
Tadel F, Baillet S, Mosher JC, Pantazis D, Leahy RM. 2011. Brainstorm: a user-friendly
application for MEG/EEG analysis. Comput Intell Neurosci. 2011:879716.
Tesche CD, Uusitalo MA, Ilmoniemi RJ, Huotilainen M, Kajola M, Salonen O. 1995. Signal-
space projections of MEG data characterize both distributed and well-localized neuronal
sources. Electroencephalogr Clin Neurophysiol. 95:189–200.
Tucciarelli R, Turella L, Oosterhof NN, Weisz N, Lingnau A. 2015. MEG Multivariate Analysis
Reveals Early Abstract Action Representations in the Lateral Occipitotemporal Cortex. J
Neurosci. 35:16034–16045.
Vaina LM, Solomon J, Chowdhury S, Sinha P, Belliveau JW. 2001. Functional neuroanatomy of
biological motion perception in humans. Proc Natl Acad Sci U S A. 98:11656–11661.
Vangeneugden J, Peelen M V, Tadin D, Battelli L. 2014. Distinct neural mechanisms for body
form and body motion discriminations. J Neurosci. 34:574–585.
Vangeneugden J, Pollick F, Vogels R. 2009. Functional differentiation of macaque visual
temporal cortical neurons using a parametric action space. Cereb Cortex. 19:593–611.
Wurm MF, Ariani G, Greenlee MW, Lingnau A. 2016. Decoding Concrete and Abstract Action
Representations During Explicit and Implicit Conceptual Processing. Cereb Cortex.
26:3390–3401.
Wurm MF, Lingnau A, Wurm XMF, Lingnau A. 2015. Decoding Actions at Different Levels of
Abstraction. J Neurosci. 35:7727–7735.
24
Figures
Figure 1 – Action recognition dataset. A) We used a dataset of two-second videos depicting
five actors performing five actions from five viewpoints. Frames from one example walk video at
90 degrees (top) and one example drink video at 0 degrees (bottom) are shown. We extended
this dataset to B) a "Form only" dataset, containing single (action informative) frames from each
two-second movie, and C) a "Motion only" dataset of point light videos created by labeling joints
on actors in each video (A, bottom).
25
Figure 2 – Within and across view action decoding. We can decode action by training and
testing a simple on the same view ('within-view' condition), or, to assess viewpoint invariance,
training on one view (0 degrees or 90 degrees) and testing on second view ('across view'
condition). Results are from the average of ten subjects. Error bars represent standard error
across subjects. Horizontal line indicates chance decoding accuracy (see Supplementary
Materials). Line at bottom of plot indicates group-level significance with p<0.05 permutation test,
for the average null distribution across the ten subjects. Inset shows a zoom of decoding time
courses from 175-525 ms post-video onset.
26
Figure 3 – Dynamics of action representations. A temporal cross training matrix showing the
decoding results for training a classifier at each point in time (y-axis) and testing the classifier at
all other times (x-axis), zoomed in to the time period from 0-1500ms post-video onset, for (a)
within-view decoding, and (b) across-view decoding for subjects watching the 2-view video
dataset (Experiment 1). Colorbar indicates mean decoding accuracy for ten subjects. Black dots
indicate points when decoding is significantly above chance at group level based on p<0.05
significance test. Results along the diagonal for the within and across view decoding are the
same as shown in the line plots in Figure 3. For TCT results from -500-3500 ms post-video
onset, see Supplemental Figure 4.
27
Figure 4 – Coarse and fine binary action decoding. We can decode action in a pair of
dissimilar actions (across body part, coarse discrimination) or a pair of similar actions (within
body part, fine discrimination) both 'within-view' and 'across-view'. A) decoding Drink versus
Run (coarse discrimination) and b) Run versus Walk (fine discrimination) are shown. Results
are from the average of ten subjects watching the 2-view videos dataset (Experiment 1). Error
bars represent standard error across subjects. Horizontal line indicates chance decoding
accuracy (see Supplementary Materials). Line at bottom of plot indicates group-level
significance with p<0.05 permutation test, for the average null distribution across the ten
subjects.
28
Figure 5 – The effects of form and motion on invariant action recognition. (a) Action can
also be decoded invariantly to view from form information alone (static images) (b) Action can
be decoded from biological motion only (point light walker stimuli). Results are each from the
average of ten subjects. Error bars represent standard error across subjects. Horizontal line
indicates chance decoding (20%). Line at bottom of plot indicates group-level significance with
p<0.05 permutation test, for the average null distribution across the ten subjects.
29
98
Experiment
Within view onset time
Across-view onset time
mean (95% CI)
mean (95% CI)
Experiment 1 (videos) – 100 ms
250 ms (210-330)
230 ms (220-270)
Experiment 1 – 10 ms
210 ms (190 – 300)
250 ms (230-300)
Experiment 1 – Run/Drink
230 ms (200-240)
210 ms (200-260)
Experiment 1 – Run/Eat
230 (200-240)
240 ms (220-270)
Experiment 1 – Run/Walk
370 ms (330-430)
320 ms (280-430)
Experiment 1 – Eat/Drink
N/A
N/A
Experiment 2 (frames)
410 ms (320 --)
510 ms (430--)
Experiment 3 (point lights)
210 ms (180-260)
300 ms (290-420)
Table 1 – Onset times and 95% CI for each experiment. The onset time for the average
across all 10 subjects' data for each experiment, and the 95% CI for each onset time as
calculated with a bootstrapping procedure (see Methods).
30
|
1809.01020 | 3 | 1809 | 2019-09-27T19:26:48 | Self-sustained activity of low firing rate in balanced networks | [
"q-bio.NC",
"physics.bio-ph"
] | Self-sustained activity in the brain is observed in the absence of external stimuli and contributes to signal propagation, neural coding, and dynamic stability. It also plays an important role in cognitive processes. In this work, by means of studying intracellular recordings from CA1 neurons in rats and results from numerical simulations, we demonstrate that self-sustained activity presents high variability of patterns, such as low neural firing rates and activity in the form of small-bursts in distinct neurons. In our numerical simulations, we consider random networks composed of coupled, adaptive exponential integrate-and-fire neurons. The neural dynamics in the random networks simulate regular spiking (excitatory) and fast-spiking (inhibitory) neurons. We show that both the connection probability and network size are fundamental properties that give rise to self-sustained activity in qualitative agreement with our experimental results. Finally, we provide a more detailed description of the self-sustained activity in terms of lifetime distributions, synaptic conductances, and synaptic currents. | q-bio.NC | q-bio |
Self-sustained activity of low firing rate in balanced
networks
F. S. Borges1, P. R. Protachevicz2, R. F. O. Pena3, E. L. Lameu4,5, G. S. V.
Higa1, A. H. Kihara1, F. S. Matias6,7, C. G. Antonopoulos8, R. de
Pasquale9, A. C. Roque3, K. C. Iarosz10, P. Ji11,12, A. M. Batista2,13
1Center for Mathematics, Computation, and Cognition, Federal University of ABC, Sao
Bernardo do Campo, SP, Brazil.
2Graduate in Science Program - Physics, State University of Ponta Grossa, PR, Brazil.
3Laboratory of Neural Systems, Department of Physics, University of Sao Paulo,
Ribeirao Preto, SP, Brazil.
4National Institute for Space Research, Sao Jos´e dos Campos, SP, Brazil.
5Department of Physics, Humboldt University, Berlin, Germany.
6Institute of Physics, Federal University of Alagoas, Macei´o, AL, Brazil.
7Cognitive Neuroimaging Unit, CEA DRF/I2BM, INSERM, Universit´e Paris-Sud,
Universit´e Paris-Saclay, F-91191 Gif/Yvette, France
8Department of Mathematical Sciences, University of Essex, Wivenhoe Park, UK.
9Department of Physiology and Biophysics, ICB, University of Sao Paulo, Sao Paulo,
SP, Brazil.
10Institute of Physics, University of Sao Paulo, Sao Paulo, SP, Brazil.
11Key Laboratory of Computational Neuroscience and Brain-Inspired Intelligence (Fudan
University), Ministry of Education, China.
12Institute of Science and Technology for Brain-Inspired Intelligence, Fudan University,
Shanghai, China.
13Department of Mathematics and Statistics, State University of Ponta Grossa, Ponta
Grossa, PR, Brazil.
Abstract
Self-sustained activity in the brain is observed in the absence of external
∗Corresponding author: [email protected], [email protected]
Preprint submitted to Elsevier
October 1, 2019
stimuli and contributes to signal propagation, neural coding, and dynamic
stability. It also plays an important role in cognitive processes. In this work,
by means of studying intracellular recordings from CA1 neurons in rats and
results from numerical simulations, we demonstrate that self-sustained activ-
ity presents high variability of patterns, such as low neural firing rates and
activity in the form of small-bursts in distinct neurons.
In our numerical
simulations, we consider random networks composed of coupled, adaptive
exponential integrate-and-fire neurons. The neural dynamics in the random
networks simulates regular spiking (excitatory) and fast spiking (inhibitory)
neurons. We show that both the connection probability and network size are
fundamental properties that give rise to self-sustained activity in qualitative
agreement with our experimental results. Finally, we provide a more de-
tailed description of self-sustained activity in terms of lifetime distributions,
synaptic conductances, and synaptic currents.
Keywords: Spontaneous activity, neural networks, whole-cell recordings,
asynchronous irregular activity
1. Introduction
Self-sustained activity (SSA), where neurons display persistent activity
even in the absence of external stimuli [1, 2], is observed in diverse situations
such as in in vitro cortical cultures and in slice preparations [3, 4, 5], in in
vivo cortical preparations [6], in slow-wave sleep [7], in anesthesia [8], and
in the resting state [9, 10]. Electrophysiological recordings of SSA states
show irregular neural spiking, typically with low average frequencies of a few
Hertz, obeying long-tailed distributions [11, 12, 13].
Many works have modelled neural networks with SSA by using ran-
dom networks composed of excitatory and inhibitory leaky, integrate-and-fire
(LIF) neurons with external background input [14, 15, 16, 17, 18, 19]. Other
studies have considered networks with non-random architectures, composed
of LIF neurons [20, 21, 22, 23, 24] or nonlinear, two-dimensional integrate-
and-fire neuron models [25, 26, 27, 28, 29, 30]. Both the architecture and
neuron types that comprise the network play an important role in SSA states.
Such states are generated and maintained by recurrent interactions within
networks of excitatory and inhibitory neurons. A stable SSA state is related
to strong recurrent excitation within the neural network, which is restrained
by inhibition to prevent runaway excitation. The balance between excitation
2
and inhibition in neural networks is considered critical to maintain a SSA
state [5, 14, 31, 32, 33, 34]. Kumar et al.
[17] studied the effects of net-
work size on SSA states. Barak and Tsodyks [35] shown that combinations
of synaptic depression and facilitation result in different network dynam-
ics. Triplett et al.
[36] analysed spontaneous activity in developing neural
networks and shown that networks of binary threshold neurons can form
structured patterns of neural activity.
In this work, we verify the existence of SSA in a random network com-
posed of neurons with different intrinsic firing patterns (the so-called electro-
physiological classes [37]). We consider the adaptive, exponential integrate-
and-fire (AdEx) [38] model with cortical neurons modelled as regular spiking
(RS) cells with spike frequency adaptation and fast-spiking (FS) cells with a
negligible level of adaptation. In such networks, depending on the excitatory
synaptic strength, neurons can exhibit a transition from spiking to bursting
synchronisation [39, 40] and bistable firing patterns [41]. We find conditions
in which unstructured, sparsely connected random networks of AdEx neurons
can display low frequency, self-sustained activity.
In particular, we show that not only the balance between excitation and
inhibition but also the connection density and network size are both im-
portant factors for low frequency SSA. In balanced networks, high mean
node-degree connectivity is necessary to give rise to low mean neural firing
rates, and for such high values, large networks are necessary to support SSA
states. In our computer simulations, we obtain qualitatively similar results to
the ones we observed in our experimental recordings, where we use CA1 neu-
rons whole-cell recordings in rats to demonstrate the possible variability of
firing rate patterns observed in the brain. Our intracellular recordings show
high variability of spontaneous activity patterns including low and irregular
neural firing rates of approximately 1 Hz and spike-train power spectra with
slow fluctuations [23], and small-bursts activity in distinct recorded neurons.
Interestingly, we show that these results can be reproduced qualitatively by
our model with cortical neurons modelled as regular spiking (RS) cells with
spike frequency adaptation and fast-spiking (FS) cells with a negligible level
of adaptation.
The paper is organised as follows: in Sec. 2, we introduce the neural net-
work model and in Sec. 3, the various quantities used to study self-sustained
activity in our numerical simulations and the details of our electrophysiolog-
ical experiments. In Sec. 4, we present our results on self-sustained activity
in the numerical simulations and experimental data, and in the last section
3
we discuss them and draw our conclusions.
2. Neural network model
We start by building a random neural network of N AdEx neurons by
connecting them with probability p, where p is the probability that any two
neurons in the network are connected, excluding autapses (i.e. neurons con-
nected to themselves, thus self-loops are not allowed). The N neurons are
split into excitatory and inhibitory neurons according to the ratio 4:1 (mean-
ing that 80% are excitatory and 20% are inhibitory), following [42]. The
connection probability p and the mean connection degree K are associated
by means of the relation
.
(1)
p =
K
N − 1
The dynamics of each AdEx neuron i = 1, . . . , N in the network is given
by the system of coupled equations [27]
C
dVi
dt
τw
dwi
dt
∆T (cid:19)
= −gL(Vi − EL) + gL∆T exp(cid:18)Vi − VT
gij(Vi − Ej) + Γi! ,
S wi +
−
1
N
Xj=1
= a(Vi − EL) − wi,
(2)
(3)
where Vi and wi are, respectively, the membrane potential and adaptation
current of neuron i, gij the synaptic conductance of the synapse from neuron
j to neuron i, and Γi the external perturbation applied to neuron i. The
synaptic conductance gij has exponential decay with synaptic time-constant
τs. The parameter values in Eqs. (2) and (3) are given in Table 1. These val-
ues have been chosen so that the system can reproduce the spiking character-
istics of RS (excitatory) and FS (inhibitory) neurons observed in experiments
with real neurons [27].
When the membrane potential of neuron i is above a threshold potential
(Vi(t) > Vthres = −30 mV), the neuron is assumed to generate a spike and
the following update conditions are applied
Vi → Vr = −60 mV,
wi → wi + b,
gji → gji + gs,
4
(4)
(5)
(6)
where Vr is the reset potential. Parameters b and gs have different values
depending on whether the neuron is excitatory or inhibitory: For excitatory
neurons, b = 0.01 nA and gs = gex, and for inhibitory neurons, b = 0 and gs =
gin. Updates have synaptic delays of 1.5 ms and 0.8 ms for excitatory and
inhibitory synapses, respectively. After the update, gji decays exponentially
with a fixed time-constant τs (5 ms for excitatory and 10 ms for inhibitory
synapses [22]). We define the relative inhibitory conductance g by
g = gin/gex
(7)
as the parameter we will use in the investigation of network dynamics. In each
simulation, we apply external stimuli Γ to 5% of the N neurons (randomly
chosen) for 50 ms to initiate network activity, and then stop the external
stimuli to observe the activity triggered, which can be persistent (SSA) or
transient. For each neuron i, the external stimulus Γi has the same charac-
teristics: it consists of excitatory current pulses with synaptic conductances
that rise instantaneously to 0.01 µS and decay exponentially afterwards with
a decay time of 5 ms, generated by a homogeneous Poisson process with rate
400 Hz.
All numerical simulations were implemented in C and the ordinary differ-
ential equations were integrated using the fourth order Runge-Kutta method
with a fixed time step of h = 0.01 ms.
3. Quantities for the study of self-sustained activity in numerical
simulations and details of electrophysiological experiments
In this work, we use the following mathematical quantities to study SSA
based either on time intervals or specific spike timings.
3.1. Coefficient of variation
The coefficient of variation is defined based on time intervals and exploits
the interspike interval (ISI) where the mth interval is defined as the difference
between two consecutive spike-timings tm
of neuron i, namely ISIi =
tm+1
i − tm
i > 0. The coefficient of variation of the ith neuron is then given by
i and tm+1
i
CVi =
σISIi
ISIi
,
(8)
where σISIi is the standard deviation of the interspike intervals of neuron i.
5
Table 1: Parameter values used in Eqs. (2) and (3). Values for excitatory neurons are
indicated by • and for inhibitory by ⋆. The ranges, where applicable, are indicated by
square brackets. These values have been chosen so that the system of Eqs.
(2) and
(3) reproduces the spiking characteristics of RS (excitatory) and FS (inhibitory) neurons
observed in experiments with real neurons [27].
Parameter
Symbol
Value
Membrane capacitance
Resting leak conductance
Resting potential
Slope factor
Spike threshold
Membrane area
Refractory time period
Adaptation intensity
Adaptation time constant
Integration time step
Time
Reversal potential
Synaptic time constant
Synaptic conductance
Synaptic delay
C
gL
EL
∆T
VT
S
tr
a
τw
h
t
Ej
τs
1 µF/cm2
0.05 mS/cm2
-60 mV
2.5 mV
-50 mV
20,000 µm2
2.5 ms
0.001 µS
600 ms
0.01 ms
[0,10] s
0 mV •
-80 mV ⋆
5 ms •
10 ms ⋆
gs = gex
gs = gin
d
[0,12] nS •
[0,240] nS ⋆
1.5 ms •
0.8 ms ⋆
3.2. Time-varying and mean network firing rates
We define the spike-train of neuron i as the sum of delta functions [43]
xi(t) =X[tm
i ]
δ(t − tm
i ),
(9)
where [tm
i ] is the set of all spike-timings of neuron i for t ∈ [0, T ] and T the
final integration time. Based on Eq. (9), we calculate the mean firing rate
of neuron i over the time interval [0, T ] as
xi(t)dt.
(10)
¯Fi =
1
T Z T
0
6
For all neurons in the network, we define the time-varying network firing rate
(in Hz) in intervals of ψ = 1 ms as
F (t) =
1
ψN
N
Xi=1 (cid:18)Z t+ψ
t
δ(t − tm
i )dt(cid:19) ,
and the mean network firing rate over the time interval [0, T ] as
¯F =
1
hISIi
,
where hISIi is the mean ISI of the network given by
hISIi =
1
N
N
Xi=1
hISIii.
(11)
(12)
(13)
3.3. Power spectrum, Fano factors and mean firing rate
Based on the definition of spike-trains xi(t) in Eq.
(9), we define the
power spectrum of neuron i as
Sxx
i (f ) =
hxi(f )x∗
i (f )i
T
,
(14)
where h·i indicates ensemble average, T is the final integration time, xi(f ) is
the Fourier transform of neuron i, given by
xi(f ) =
T
Z0
e2πif txi(t)dt,
(15)
and x∗
i (f ) is the complex conjugate of xi(f ). The power spectrum of a set of
M (M ≤ N) neurons, ¯Sxx(f ) is then defined as the average power spectrum
of the M neurons
M
Xi=1
7
¯Sxx(f ) =
1
M
Sxx
i (f ).
(16)
To obtain consistent results over different simulations (see Fig. 5), we have
chosen M = 5×104, independently of N and K. In this work, we will use two
quantities related to ¯Sxx to describe spike-train characteristics [44, 45, 46, 47].
The first quantity is the Fano factor F F = h(n−hni)2i/hni = h(∆n)2i/hni,
which is defined as the ratio of the variance of n to the mean of the averaged
spike count of the M neurons over the time window [0, T ], namely
n =
1
M
M
Xi=1 Z T
0
xi(t)dt.
(17)
Its relation to ¯Sxx is meaningful in the vanishing frequency limit of ¯Sxx(f ), i.e.
¯Sxx(f ) = ¯F · F F . The Fano factor, F F , is a standard measure of neural
lim
f →0
variability (F F = 1 corresponds to a Poisson process) and is related to the
k=1 rk),
CV of the ISIs of the M neurons [48] by lim
f →0
where rk is the serial correlation coefficient between ISIs that are lagged by
k.
¯Sxx(f ) = ¯F ·CV2 (1 + 2P∞
The second quantity is the mean firing rate of the M neurons, ¯FM (defined
as in Eq. (12) but dividing by M), which is related to ¯Sxx(f ) by
¯FM = lim
f →∞
¯Sxx(f ).
(18)
3.4. Synaptic input
The instantaneous synaptic conductance of neuron i due to all excitatory
and inhibitory synapses arriving to it, is denoted by
Gi(t) =Xj∈E
gij(t) +Xj∈I
gij(t) = Gex,i(t) + Gin,i(t),
(19)
where E and I are the sets of all excitatory and inhibitory neurons in the
network, respectively.
The instantaneous synaptic input to neuron i is then defined by
Isyn,i(t) = Xj∈E
gij(Ej − Vi) +Xj∈I
= I E
syn,i(t) + I I
syn,i(t),
gij(Ej − Vi)
(20)
and the mean synaptic input ¯Isyn over all neurons in the network and over
the simulation time in [0, T ] by
¯Isyn =
1
N
N
Xi=1
1
T Z T
t=0(cid:0)I E
syn,i(t) + I I
syn,i(t)(cid:1) dt.
8
(21)
3.5. Mean network decay-time
Finally, for a given neuron i, we define the time of its last spike as the
maximum time in its spike train,
tlast
i = max {tm
i }.
(22)
The network decay-time (DT) is then defined by the maximum of the
last spikes of all neurons in the network during the simulation time, i.e. the
maximum tlast
of all neurons in the network,
i
DT = max {tlast
i
, i = 1, . . . , N}.
(23)
Finally, we define the mean network decay-time DT as the mean DT over
103 simulations with different random initial conditions.
3.6. Details of electrophysiological experiments
3.6.1. Animals
Electrophysiological experiments were conducted using male Wistar rats
with 20-25 postnatal days. All animals were kept in an animal facility in a
12:12h light-dark cycle at a temperature of 23◦C ± 2◦ with free access to
food and water. All procedures were approved by the Institutional Animal
Care Committee of the Institute of Biomedical Sciences, University of Sao
Paulo (CEUA ICB/USP n. 090, fls. 1◦).
3.6.2. Preparation of brain slices
After animals were deeply anesthetised through isoflurane inhalation (A-
Errane; Baxter Pharmaceuticals), they were decapitated and the brain was
quickly removed and submerged in cooled (0◦C) oxygenated (5% CO2-95%
O2) cutting solution (in mM): 206 sucrose, 25 NaHCO3, 2.5 KCl, 10 MgSO4,
1.25 NaH2PO4, 0.5 CaCl2, and 11 D-glucose. After removing the cerebellum,
brain hemispheres were separated by a single sagittal cut. Both brain hemi-
spheres were trimmed up and glued in a metal platform and sectioned using
a vibratome (Leica - VT1200). 350-400 µm brain slices were obtained by ad-
vancing the vibratome blade from anterior-posterior orientation. Slices were
rapidly transferred to a holding chamber containing artificial cerebrospinal
fluid (ACSF; in mM): 125 NaCl, 25 NaHCO3, 3 KCl, 1.25 NaH2PO4, 1 MgCl2,
2 CaCl2, and 25 D-glucose. Slices were kept oxygenated at room tempera-
ture (20-25◦) for at least one hour before proceeding with electrophysiological
recordings.
9
3.6.3. Electrophysiological recordings
Brain slices containing the hippocampal formation were placed in a sub-
mersion-type recording chamber upon a modified microscope stage and main-
tained at 30◦C with constant perfusion of oxygenated ACSF (5% CO2-95%
O2). Whole-cell recordings were made from neurons located in the pyra-
midal layer of CA1. Recording pipettes were fabricated from borosilicate
glass (Garner Glass) with input resistances of about 4-6 MΩ and were filled
with intracellular solution (in mM): 135 K-gluconate, 7 NaCl, 10 HEPES,
2 Na2ATP, 0.3 Na3GTP, 2 MgCl2; at a pH of 7.3 obtained with KOH and
osmolality of 290 mOsm. All experiments were performed using a visualised
slice setup under a differential interference contrast-equipped Nikon Eclipse
E600FN microscope. Recordings were made using a Multiclamp 700B am-
plifier and pClamp software (Axon Instruments). Only recordings from cells
that presented spontaneous activity with membrane potentials lower than
−60 mV, access resistance lower than 20 MΩ, and input resistance higher
than 100 MΩ and lower than 1000 MΩ, were included in our data. We in-
jected depolarising currents to identify regular, tonic, or bursting spike pat-
terns and neural spontaneous activity was assessed by 10 min of continuous
recordings in current clamp mode.
4. Self-sustained activity
4.1. Self-sustained activity in electrophysiological experiments
SSA assessed by electrophysiological recordings in physiological brain
states is characterised by irregular neural spiking, normally with low av-
erage frequency where rates follow a long-tailed distribution across neurons
[23]. Interestingly, some brain regions are able to produce spontaneous net-
work activity after a slicing procedure, including hippocampal sharp waves
[49, 50, 51]. Once hippocampal slices present SSA manifested by spontaneous
activity, we use CA1 neurons whole-cell recordings to demonstrate the possi-
ble variability of firing rate patterns observed in the brain. Our intracellular
recordings show a high variability of spontaneous activity patterns including
low neural firing rates and small-bursts activity in distinct recorded neurons.
Figure 1 shows a representation of traces obtained by whole-cell patch
clamp from CA1 neurons during a period of 20 s. The raster plot of the 150
traces is shown in Fig. 1(a), and five traces of membrane potential are shown
in Fig. 1(b) for the same time window. The mean firing rate and interspike
interval (ISI) distributions are plotted in Figs. 1(c) and 1(d), respectively.
10
Recorded neurons present distinct firing patterns, including very low firing
rate and small bursts of spikes, as shown in Fig. 1(b). In the whole record,
the mean firing rate over all neurons is approximately equal to 1.172 Hz.
Similar firing patterns, with high variability and low neural firing rates, are
found in different recordings in the hippocampus and other cortices of the
rat brain during slow-wave sleep [52], as well as in recordings from human
middle temporal gyrus during sleep [53].
(a)
1
50
i
100
150
(b)
(c)
100
10-1
10-2
f
10-3
0
1
2
3
Fi (Hz)
4
5
(d)
100
10-1
10-2
f
10-3
0
2
4
6
ISI (s)
8
20
Figure 1: Results from electrophysiological experimental data.
(a) Raster plot of 150
traces. (b) Twenty seconds of recordings are shown (horizontal black bar denotes 1 s and
vertical black bar denotes 50 mV). (c) Mean firing rate and (d) interspike interval (ISI)
distributions. Note that the vertical axes in (c) and (d) are in logarithmic scale and that
the label f represents the fraction of values in the bins in the horizontal axes.
4.2. Self-sustained activity in numerical simulations
Here, we use the results from numerical simulations of the a random net-
work model of N adaptive exponential integrate-and-fire neurons [54] (Eqs.
(2) and (3)). We focus on reproducing three aspects of the previous subsec-
tion (electrophysiological experimental data): i) mean firing rates around 1
Hz, ii) small-bursts activity in distinct neurons (Fig. 1(b)) and iii) data with
variability in firing rates where most of the neurons have frequencies between
0 and 1 Hz (Fig. 1(c)).
11
The random network model of N adaptive, exponential integrate-and-fire
neurons [54] we use here is composed of 80% excitatory and 20% inhibitory
neurons (i.e. a ratio of 4:1) following [42]. The synaptic conductance shows
an exponential decay with synaptic delays of 1.5 ms and 0.8 ms for excitatory
and inhibitory synapses, respectively. In each simulation, we apply external
stimuli Γ to 5% of the N neurons (randomly chosen) for 50 ms to initiate
network activity. Then, we stop the external stimulation to observe the
activity triggered, which can be persistent (i.e. SSA) or transient.
To reproduce SSA firing patterns with low neural firing rates, we analyse
the parameter space gex × g for a neural, random, network of N = 104
AdEx neurons with connection probability p = 0.02. We focus on an area
of the parameter space where there is a balanced regime of excitation and
inhibition. In this area, we do not observe SSA for gex < 0.004 µS, i.e.
in
the weak coupling region. Examples of time-dependent network firing rates
in this region are shown in Fig. 2(a), for gex = 0.0035 µS and g = 16. The
initial stimulus applied in the first 50 ms, generates short-lived activity in
the network. The mean synaptic input, ¯Isyn, is negative (left-hand inset in
Fig. 2(a)) and the network decay-time DT is always less than 3 s (right-hand
inset in Fig. 2(a)). Examples of time-dependent network firing rates in the
region of SSA are shown in Fig. 2(b), corresponding to gex = 0.008 µS and
g = 16. The mean synaptic input is approximately balanced ( ¯Isyn ≈ 0) and
the decay-time is higher than the maximum time used in our simulations (i.e.
DT ≥ 5 s).
For N = 104 neurons, the lowest mean network firing rates are approx-
imately 2 Hz. The region where these rates occur is shown in purple in
Fig. 2(c). A particular case for gex < 0.0035 µS is indicated by a turquoise
square, where the activity is not self-sustained. For the random network and
parameter space considered, the region with SSA is roughly determined by
gex > 0.004 µS (indicated in yellow in Fig. 2(e)). The lowest mean network
firing rate of an SSA state is around 4 Hz (red region in Fig. 2(c)). Within
this region, one can identify the region of excitation/inhibition balance by
considering the region with Isyn ≈ 0 (see Eq. (21)). This is shown in red in
Fig. 2(d). The black and yellow regions in Fig. 2(d) correspond to slightly
predominant inhibitory and excitatory mean synaptic input, respectively. An
SSA case with ¯F ≈ 4 Hz and excitation/inhibition balance is indicated by
the green circle in Figs. 2(c)-(e) and corresponds to gex = 0.008 µS and
g = 16.
To study the effect of the size of the network on SSA, we considered
12
(a)
60
0.2
40
f
0.1
0.2
0.1
0
-0.1
-0.075
Isyn (nA)
-0.05
0
0
1.5
DT (s)
3.0
)
z
H
(
F
20
0
60
(b)
0.2
40
f
0.1
g
g
(c)
(d)
(e)
20
15
10
5
20
15
10
5
20
15
10
_
F
(Hz)
8.0
6.0
4.0
2.0
0
0.02
0.01
0
Isyn
(nA)
-0.01
-0.02
5.0
4.0
3.0
2.0
1.0
DT
(s)
1.0
0.5
)
z
H
(
F
20
0
-0.025
0
Isyn (nA)
0.025
0
0
2.5
DT (s)
5.0
g
0
0
0.5
1.0
time (s)
1.5
2.0
5
0
0
0.003 0.006 0.009 0.012
gex
Figure 2: Results from the numerically simulated data. Time-dependent network firing
rate (F ) in two distinct regimes: (a) non-SSA and (b) SSA. Each case corresponds to
a different simulation. The two neural network models use the same set of parameters
except that the excitatory conductance is gex = 0.0035 µS in (a) and gex = 0.008 µS in
(b). The insets show the probability distribution of the mean synaptic inputs ( ¯Isyn) over
103 simulations and decay-times (DT) in each case. (c) Mean network firing rate ( ¯F ), (d)
mean synaptic input ( ¯Isyn) and (e) Mean decay-time (DT) as a function of excitatory (gex)
and relative inhibitory synaptic coupling (g). The neural network has N = 104 neurons
and the connection probability is p = 0.02. In (a) and (b), we fixed g = 16. The turquoise
squares correspond to the parameters in (a) and the green circles to those in (b). Note
that the label f in the insets in (a) and (b) represents the fraction of values in the bins in
the horizontal axes.
two networks of different sizes with the same mean degree. The first has
N = 104 and p = 0.04 and the second N = 2 × 104 and p = 0.02, both having
K = 400. The parameters (gex = 0.005 µS and g = 8) are the same for the
two networks, and put them close to the balanced state. In Fig. 3(a), we
show the firing rate evolution of the two networks. For N = 104 (turquoise
line), the activity decays before 2 s, and for N = 2 × 104 (green line) SSA
is observed. The distributions of ¯Isyn for the two cases (see the insets in
Fig. 3(a)) have similar (positive) average values but the one for N = 104
is broader and left-skewed. The decay-time distribution (Figs. 3(b), (c))
clearly shows that the networks with N = 2 × 104 have SSA, while the ones
13
(a)
60
0.2
f
0.1
)
z
H
(
F
30
0
0
(b)
0.10
f
0.05
0.3
0.2
0.1
0
-0.05
0
Isyn(nA)
0.05
0
-0.05
0
Isyn(nA)
0.05
1.0
time (s)
(c)
1.0
0.5
(d)
(f)
(h)
20
15
10
5
20
15
10
5
20
15
10
g
g
2.0
g
(e)
(g)
(i)
_
F
(Hz)
8.0
6.0
4.0
2.0
0
0.02
0.01
0
(nA)
Isyn
-0.01
-0.02
5.0
4.0
3.0
2.0
1.0
DT
(s)
0
0
2.5
DT (s)
5.0
0
0
2.5
DT (s)
5.0
5
0
0.003 0.006 0.009 0.012
0
0
0.003 0.006 0.009 0.012
gex
gex
Figure 3: Results from the numerically simulated data. (a) Time-dependent network firing
rate (F ) for N = 104 (turquoise line) and N = 2 × 104 (green line). The insets show the
probability distribution of the mean synaptic input ( ¯Isyn) over 103 simulations, for each
case. (b) and (c) show the decay-time (DT) distribution (also for 103 simulations) for
N = 104 and N = 2 × 104, respectively (colours follow the same convention as in (a)). We
considered the excitatory conductance gex = 0.005 µS, mean node-degree K = 400, and
relative inhibitory synaptic coupling g = 8. The mean network firing rates (d) and (e),
mean synaptic inputs (f) and (g), and mean decay-times (h) and (i) are shown as a function
of gex and g for N = 104 (left) and N = 2×104 (right), respectively. The turquoise squares
correspond to N = 104 and p = 0.04 and the green circles to N = 2 × 104 and p = 0.02.
Note that the label f in the insets in (a) and (b) represents the fraction of values in the
bins in the horizontal axes.
with N = 104 are predominantly short-lived. A comparison of networks with
the two sizes in the gex × g parameter space is shown in Figs. 3(d)-(i). The
mean network firing rate and mean synaptic input display similar behaviour
in the parameter space for the two network sizes (Figs. 3(d)-(g)). However,
this similarity is not seen in the diagram for DT (Figs. 3(h)-(i)). The region
corresponding to SSA (yellow) is larger for N = 2 × 104 than for N = 104.
Moreover, the shape of this region for N = 2 × 104 discloses almost absent
sensitivity of SSA duration to the relative inhibitory synaptic conductance
g. On the other hand, for N = 104, the SSA lifetime is sensitive to g for
0.006 µS . gex . 0.008 µS, and only for strong coupling (gex & 0.008 µS), it
becomes insensitive to g. In Fig. 3, we observe large decay time DT for high
14
frequencies.
Next, we investigate the influence of N, p and K (see Eq. (1)) on the
network firing rate of SSA states. We fix gex = 0.008 µS and g = 16 to
keep the dynamics around an excitatory/inhibitory balance. Figure 4 shows
the mean network firing rate of SSA states (colour scale) in the parameter
space N × p. The white area in the parameter space corresponds to non-SSA
states. We can see that the mean network firing rate of SSA states depends
on both N and p. In particular, low firing rate SSA states appear when the
network size N increases. The black solid line in Fig. 4 represents networks
with mean node-degree K = 1500, in which case there are SSA states for
N ≥ 1.5 × 105 neurons. The inset in Fig. 4(a) shows the dependence of
the mean network firing rate ¯F of SSA states on the mean connection node-
degree K for constant N. We can observe that lower rates are obtained for
increased K. However, very low firing rates (≈ 1 Hz) are present only for
very large network sizes. In the inset in Fig. 4(b), we can observe that the
network size does not alter the mean network firing rate when K is fixed.
Therefore, large K plays an important role in the occurrence of low network
firing rates, and for such low firing rates, large-size networks are necessary
to support SSA states.
Figure 5 shows results for three cases of network sizes and mean node-
degrees K, namely for N = 105 and K = 1300, for N = 3×105 and K = 1500
and for N = 5 × 105 and K = 1700. The time-varying network firing rate is
non-periodic and its mean value decreases as both N and K increase (Fig.
5(a)).
In Figs. 5(c)-(e), we observe that most ISIs are distributed in the
interval [0, 2] s. These results agree with the experimental results in Fig.
1(d), where more than 60% of ISIs are in a similar interval.
In the three
cases, the spiking variability, characterised by the ISI distribution, is very well
described by a Poisson distribution (Figs. 5 (c)-(e)). This is characteristic of
irregular neural firing [55, 56, 57, 19]. Moreover, the CV of the ISIs slowly
converges to the one of a Poisson process (= 1) as both N and K increase.
As discussed in Sec. 3, the power spectrum of a set of M neurons is
related to different quantities that characterise network firing. In our numer-
ical simulations, we can observe these relations by taking the limits of the
power spectrum as f → ∞ or f → 0. The mean firing rate ¯FM is associated
with the infinity frequency limit, namely limf →∞ ¯Sxx(f ) = ¯FM (see also Eq.
(18)). In Fig. 5(b), one can see that these mean firing rates are very low
(1 Hz . ¯FM . 1.5 Hz) and that the smaller rates occur for larger N and
K. On the other hand, irregularity is associated with vanishing frequency,
15
0.03
0.025
0.02
0.015
0.01
p
16
12
)
z
H
(
F
8
4
0
0
6
(a)
)
z
H
(
F
4
2
(b)
0
0
N=100000
N=300000
400
800
K
1200
1600
K=500
K=1500
100000 200000 300000 400000 500000
N
0.005
5000
100000
N
200000
300000
8
7
6
5
4
3
2
1
--
F
(Hz)
Figure 4: Results from the numerically simulated data. Each point in the coloured part
of the plot gives ¯F calculated for a simulation with the corresponding parameter values
of N and p. The white area corresponds to non-SSA states. The black line represents
networks with mean node-degree K = 1500 (according to Eq. (1)). Upper inset: mean
network firing rate ¯F as a function of K, keeping N fixed, where the red squares and
black circles represent N = 105 and N = 3 × 105 neurons, respectively. Lower inset: mean
network firing rate ¯F as a function of N , showing independence of ¯F on N for constant
connection degree K. The curves represent two selected K values: K = 500 (red squares)
and K = 1500 (black circles). Note that we have used gex = 0.008 µS and g = 16 for all
simulations in this figure.
¯Sxx(f ) = ¯F .F F , where F F is the Fano factor (see Subsec. 3.3).
namely lim
f →0
Assuming approximately the same ¯F for the three networks in Fig. 5(b), say
¯F ≈ 1.3, the f → 0 limit of ¯Sxx(f ) shows that F F > 1 for all of them. This
suggests a process more irregular than the Poisson process [58], e.g. bursting
[56]. Moreover, as the network size N increases, F F decreases, indicating
tendency to converge to a Poisson process and is linked to a standard irregu-
larity measure and CV. We see that in the limit of very small frequencies, the
value of the power spectra decreases as both K and N increases, confirming
the behaviour observed by CV, i.e. the spiking times are becoming more reg-
ular. The power spectra display slow fluctuations, which can be explained
by the low neural firing rates and bursting spiking patterns (as discussed
16
(a)
)
z
H
(
F
(b)
)
z
H
(
x
x
S
4
3
2
1
0
4
3
2
1
0
4
3
2
1
0
5
4
3
2
1
10-1
10-2
10-3
f
10-4
10-5
10-2
10-3
f
10-4
10-5
10-2
10-3
f
10-4
10-5
0
(c)
CV = 1.37 ± 0.38
(d)
CV = 1.35 ± 0.38
(e)
CV = 1.32 ± 0.40
1
2
ISI (s)
3
6
7
time (s)
8
9
10
N=100000 and K=1300
N=300000 and K=1500
N=500000 and K=1700
100
101
Frequency (Hz)
102
103
Figure 5: Results from the numerically simulated data. (a) From top to bottom, we show
F as N and K increase (for the N and K values, see legend in (b)). (b) Power spectra
averaged over 5 × 104 neurons in the network and 10 s of simulation. Note that the
horizontal axis is logarithmic and the vertical linear. (c)-(e) ISI distributions for the same
three networks with the vertical axes in logarithmic and horizontal in linear scales. The
values of the corresponding CVs are shown in the plots. In the three cases, the parameter
values used are gex = 0.008 µS and g = 16. Note that the vertical axes in (c), (d) and (e)
are in logarithmic scale and that the label f represents the fraction of values in the bins
in the horizontal axes.
below). Slow power spectrum fluctuations are also features of spontaneous
activity in cortical networks [10, 46, 59].
In Fig. 6, we show characteristics of the SSA firing patterns with low rate
exhibited by the network (a particular case with N = 5 × 105 and K = 1700
is shown as a representative example). The patterns are characterised by
sparse and non-synchronous activity (Fig. 6(a)), akin to what has been
termed the heterogeneous variant [19] of the asynchronous irregular (AI)
regime [14, 60].
In the homogeneous AI regime, all neurons fire with the
same mean rate, but in the heterogeneous AI regime, the mean firing rates
fluctuate in time and across neurons. In some cases, neurons can even exhibit
bursting periods (as can be seen in Fig. 6(b)). Neurons have low firing
rates, with a right-skewed distribution that peaks around 1 Hz (Fig. 6(d)).
17
(a)
1
i
250
500
(b)
(c)
)
S
µ
(
2
1
i
G
100
10-1
10-2
f
(d)
10-3
0
1
4
5
2
3
Fi (Hz)
(e)
0.05
0.10
0.15
0
2
4
6
(f)
(g)
8
10
(h)
0.08
f
0.04
0
0
0.08
excitatory
inhibitory
f
0.04
0
1.0
1.25
1.5
time (s)
1.75
2.0
0
0
1.0
0.5
Gi (µS)
1.5
0
-2
-8
-10
-6
-4
Isyn,i (nA)
Figure 6: Results from the numerically simulated data. All plots refer to a single simulation
of a network with N = 5 × 105, K = 1700, gex = 0.008 µS and g = 16. (a) Raster plot of 2
s of simulation time (only 500 randomly chosen neurons from the network are shown). (b)
membrane potential of three randomly chosen neurons from the network (the horizontal
black bar denotes 0.2 s and the vertical black bar denotes 50 mV). (c) Time-varying
synaptic conductances (Gex,i(t) in green and Gin,i(t) in cyan) of neurons represented by
the blue line in (b). (d) distribution of mean firing rates ¯Fi (Eq. (10)) of all neurons
in the network. Distribution of excitatory (Gex,i) (e) and inhibitory (Gin,i) (f) synaptic
conductances of all neurons at the end of the simulation. Distribution of excitatory (I E
syn,i)
(g) and inhibitory (I I
syn,i) (h) synaptic inputs of all network neurons at the end of the
simulation. Note that the vertical axis in (d) is in logarithmic scale and that the label f
in (d), (e), (g), (f) and (h) represents the fraction of values in the bins in the horizontal
axes.
Irregularity is confirmed by the CV of ISIs and power spectra analysis (Fig.
5). The similarity between the firing regime observed in our simulations and
the heterogeneous AI state can be further verified by a comparison of the
respective spike-train power spectra. Spectra for networks in heterogeneous
AI have been calculated elsewhere [47] and display slow fluctuations as shown
here (Fig. 5(b)).
The distributions of the excitatory (Gex) and inhibitory (Gin) synaptic
conductances over neurons (panels (e) and (f) in Fig. 6) have nearly sym-
metric shapes, with the Gin distribution slightly right skewed, in agreement
with experimental evidence [61], even though the conductance values per se
18
syn) and inhibitory (I I
are higher than in the experimental recordings. The distributions of excita-
tory (I E
syn) synaptic inputs are nearly identical and vary
over nearly the same range of absolute values, though the upper end of the
range is slightly higher for the inhibitory synaptic inputs (see panels (g) and
(h) in Fig. 6). This is a hallmark of a balanced state, and the neural spikes
occur due to the large synaptic conductance variability, which allows for the
necessary conditions that can support the appearance of SSA.
Finally, it is worthwhile the fact that in Figs. 6(a) (raster plot) and 6(b)
(membrane potential), there are several striking similarities, namely: non-
synchronous activity and, the coexistence of spike and bursting neural activ-
ity. More importantly though, qualitatively similar activities and behaviours
are present in the data obtained from hippocampal slices (electrophysiologi-
cal experiments) as demonstrated in Fig. 1(a)-(b) and in the results from the
numerical simulations data. In both experimental and numerically simulated
data, we observe high variability in neural firing rates where most of the neu-
rons fire with frequencies in the interval [0, 1] Hz, with similar distribution of
mean firing rates in the interval [1, 5] Hz (compare Fig. 1(c) and Fig. 6(d)).
5. Conclusions and discussion
In this paper, motivated by self-sustained activity observed in the brain in
the absence of external stimuli, we sought to study necessary conditions for
which irregular and low-frequency self-sustained dynamics emerge in mod-
els of neural networks with random connectivity. We build neural network
models of 80% excitatory and 20% inhibitory neurons randomly connected
and mathematically described by the adaptive, exponential integrate-and-fire
model. This model mimics the behaviour of biological neurons, exhibiting
spiking and bursting patterns of activity. We studied network features by
varying: (i) the balance between excitation and inhibition in the system and
(ii) the topological characteristics of the random network by means of vary-
ing the mean node-degree. Results were obtained by running a large number
of numerical simulations to compute firing rates and related quantities, and
compared them with results from our electrophysiological experiments using
whole-cell patch clamp from CA1 rat neurons.
Our numerical results allowed us to observe neural activity with slow fluc-
tuations in the absence of external perturbation. We found that the patterns
of low firing-rate self-sustained activity is asynchronous and irregular as de-
picted in raster plots. We shown that the irregular spikes with low firing
19
rate depend on the mean node degrees of the neurons, and that low-rate self-
sustained activity occurs for a tight balance between inhibition/excitation
and large network sizes. We found that for fixed mean connection degree,
the mean network firing rate does not change with the network size. In an
excitation/inhibition balanced random network, if the network size is fixed,
the mean network firing rate decreases when the mean node-degree increases.
However, there is a maximum mean node-degree value for which it is possi-
ble to maintain self-sustained activity states, and this value is proportional
to the network size. Therefore, large networks are necessary to give rise to
self-sustained activity states when the mean node-degree is large. Moreover,
the occurrence of spikes is due to the synaptic-conductance variability and
the inhibitory synaptic input is slightly higher than the excitatory.
More importantly, we found that our results from the numerical simula-
tions resemble those obtained from our experimental recordings and analysis,
such as irregular firing with firing rates of about 1 Hz. In addition, we have
demonstrated that our model is able to reproduce the coexistence of spike
and bursting neurons and the distribution of mean firing rates between 0 and
5 Hz obtained by our electrophysiological experiments using whole-cell patch
clamp from CA1 rat neurons.
Concluding, we have been able to demonstrate the existence of low neu-
ral firing rate self-sustained activity in excitation/inhibition balanced ran-
dom networks of adaptive, exponential integrate-and-fire neurons. This phe-
nomenon is characterised by irregular neural oscillations and shows that low
frequency self-sustained activity can be found in large balanced random net-
works, a result that is qualitatively similar to those obtained from our elec-
trophysiological experiments using whole-cell patch clamp from CA1 rat neu-
rons.
Acknowledgements
We wish to acknowledge the support from the following grants and sche-
mes: CAPES (88881.120309/2016-01), CNPq (154705/2016-0, 306251/2014-
0, 311467/2014-8, 432429/2016-6), FAPEAL, International Visiting Fellow-
ships Scheme of the University of Essex, and FAPESP (2011/19296-1, 2013/
07699-0, 2013/25667-8, 2015/50122-0, 2015/ 07311-7, 2016/23398-8, 2017/
13502-5, 2017/18977-1, 2017/20920-8, 2017/26439-0, and 2015/50122-0).
20
References
[1] M.D. Greicius, B. Krasnow, A.L. Reiss, V. Menon, Functional connec-
tivity in the resting brain: a network analysis of the default mode hy-
pothesis, Proceedings of the National Academy of Sciences of the United
States of America 100 (2003) 253.
[2] M.D. Fox, A.Z. Snyder, J.L. Vincent, M. Corbetta, D.C. Van Essen,
M.E. Raichle, The human brain is intrinsically organized into dynamic,
anticorrelated functional networks, Proceeding of the National Academy
of Science of the United States of America 102 (2005) 9673.
[3] D. Plenz, A. Aertsen, Neural dynamics in cortex-striatum co-cultures II,
Spatiotemporal characteristics of neural activity, Neuroscience 70 (1996)
893.
[4] M.V Sanchez-Vives, D.A McCormick, Cellular and network mechanisms
of rhythmic recurrent activity in neocortex, Nature Neuroscience 10
(2000) 1027.
[5] Y. Shu, A. Hasenstaub, D.A. McCormick, Turning on and off recurrent
balanced cortical activity, Nature 432 (2003) 288.
[6] I. Timofeev, F. Grenier, M. Bazhenov, T.J. Sejnowski, Steriade M, Ori-
gin of slow cortical oscillations in deafferented cortical slabs, Cerebral
Cortex 10 (2000) 1185.
[7] M. Steriade, I. Timofeev, F. Grenier, Natural waking and sleep states:
a view from inside neocortical neurons, Journal of Neurophysiology 85
(2001) 1969.
[8] M. Steriade, A. Nunez, F. Amzica, A novel slow (< 1 Hz) oscillation of
neocortical neurons in vivo: depolarizing and hyperpolarizing compo-
nents, Journal of Neuroscience 13 (1993) 3252.
[9] A. Arieli, D. Shoham, R. Hildesheim, A. Grinvald, Coherent spatiotem-
poral patterns of ongoing activity revealed by real-time optical imaging
coupled with single-unit recording in the cat visual cortex, Journal of
Neurophysiology 73 (1995) 2072.
21
[10] D. Mantini, M.G. Perrucci, C. Del Gratta, G.L. Romani, M. Corbetta,
Electrophysiological signatures of resting state networks in the human
brain, Proceeding of the National Academy of Sciences of the United
States of America 104 (2007) 13170.
[11] T. Hrom´adka, M.R. DeWeese, A.M. Zador, Sparse representation of
sounds in the unanesthetized auditory cortex, PLoS Biology 6 (2008)
0124.
[12] D.H. O'Connor, S.P. Peron, D. Huber, K. Svoboda, Neural activity
in barrel cortex underlying vibrissa-based object localization in mice,
Neuron 67 (2010) 1048.
[13] G. Buzs´aki, K. Mizuseki, The log-dynamic brain: how skewed distribu-
tions affect network operations, Nature Reviews Neuroscience 15 (2014)
264.
[14] N. Brunel, Dynamics of sparsely connected networks of excitatory and
inhibitory spiking neurons, Journal of Computational Neuroscience 8
(2000) 183.
[15] T.P. Vogels, L.F. Abbott, Signal propagation and logic gating in net-
works of integrate-and-fire neurons, Journal of Neuroscience 25 (2005)
10786.
[16] N. Parga, L.F. Abbott, Network model of spontaneous activity exhibit-
ing synchronous transitions between up and down states, Frontiers in
Neuroscience 1 (2007) 57.
[17] A. Kumar, S. Schrader, A. Aertsen, S. Rotter, The high-conductance
state of cortical networks, Neural Computation 20 (2008) 1.
[18] B. Kriener, H. Enger, T. Tetzlaff, H.E. Plesser, M.-O. Gewaltig, G.T.
Einevoll, Dynamics of self-sustained asynchronous irregular activity in
random networks of spiking neurons with strong synapses, Frontiers in
Computational Neuroscience 8 (2014) 1.
[19] S. Ostojic, Two types of asynchronous activity in networks of excitatory
and inhibitory spiking neurons, Nature Neuroscience 17 (2014) 594.
22
[20] A. Renart, R. Moreno-Bote, X.-J. Wang, N. Parga, Mean-driven and
fluctuation-driven persistent activity in recurrent networks, Neural
Computation 19 (2007) 1.
[21] M. Kaiser, C.C. Hilgetag, Optimal hierarchical modular topologies for
producing limited sustained activation of neural networks, Frontiers in
Neuroinformatics 4 (2010) 1.
[22] S.-J. Wang, C.C. Hilgetag, C. Zhou, Sustained activity in hierarchi-
cal modular neural networks: self-organized criticality and oscillations,
Frontiers in Computational Neuroscience 5 (2011) 1.
[23] A. Litwin-Kumar, B. Doiron, Slow dynamics and high variability in bal-
anced cortical networks with clustered connections, Nature Neuroscience
15 (2012) 1498.
[24] T.C. Potjans, M. Diesmann, The cell-type specific cortical microcircuit:
relating structure and activity in a full-scale spiking network model,
Cerebral Cortex 24 (2014) 785.
[25] A. Compte, Computational and in vitro studies of persistent activity:
edging towards cellular and synaptic mechanisms of working memory,
Neuroscience 139 (2006) 135.
[26] E.M. Izhikevich, G.M. Edelman, Large-scale model of mammalian tha-
lamocortical systems, Procedings of the National Academy of Science of
the States of America, 105 (2008) 3593.
[27] A. Destexhe, Self-sustained asynchronous irregular states and up-down
states in thalamic, cortical and thalamocortical networks of nonlinear
integrate-and-fire neurons, Journal of Computational Neuroscience 27
(2009) 493.
[28] P. Stratton, J. Wiles, Self-sustained non-periodic activity in networks of
spiking neurons: The contribution of local and long-range connections
and dynamic synapses, Neuroimage 52 (2010) 1070.
[29] P. Tomov, R.F. Pena, M.A. Zaks, A.C. Roque, Sustained oscillations,
irregular firing, and chaotic dynamics in hierarchical modular networks
with mixtures of electrophysiological cell types, Frontiers in Computa-
tional Neuroscience 8 (2014) 1.
23
[30] P. Tomov, R.F. Pena, A.C. Roque, M.A. Zaks, Mechanisms of self-
sustained oscillatory states in hierarchical modular networks with mix-
tures of electrophysiological cell types, Frontiers in Computational Neu-
roscience 10 (2016) 1.
[31] C.A. van Vreeswijk, H. Sompolinsky, Chaos in neural networks with
balanced excitatory and inhibitory activity, Science 274 (1996) 1724.
[32] B. Haider, A. Duque, A.R. Hasenstaub, D.A. McCormick, Neocortical
network activity in vivo is generated through a dynamic balance of ex-
citation and inhibition, Journal of Neuroscience 26 (2006) 4535.
[33] A.H. Taub, Y. Katz, I. Lampl, Cortical balance of excitation and inhibi-
tion is regulated by the rate of synaptic activity, Journal of Neuroscience
33 (2013) 14359.
[34] E. Nanou, A. Lee, W. A. Catterall, Control of Excitation/Inhibition
Balance in a Hippocampal Circuit by Calcium Sensor Protein Regulation
of Presynaptic Calcium Channels, Journal of Neuroscience 38 (2018)
4430.
[35] O. Barak, M. Tsodyks, Persistent activity in neural networks with dy-
namic synapses, PLoS Computational Biology 3 (2007) e35.
[36] M.A. Triplett, L. Avitan, G.J. Goodhill, Emergence of spontaneous as-
sembly activity in developing neural networks without afferent input,
PLoS Computational Biology 14 (2018) e1006421.
[37] D. Contreras, Electrophysiological classes of neocortical neurons, Neural
Networks 17 (2004) 633.
[38] R. Brette, W. Gerstner, adaptive, exponential integrate-and-fire model
as an effective description of neural activity, Journal of Neurophysiology
94 (2005) 3637.
[39] F.S. Borges, P.R. Protachevicz, E.L. Lameu, R.C. Bonetti, K.C. Iarosz,
I.L. Caldas, M.S. Baptista, A.M. Batista, Synchronised firing patterns
in a random network of adaptive, exponential integrate-and-fire neuron
model, Neural Networks 90 (2017) 1.
24
[40] P.R. Protachevicz, R.R. Borges, A.S. Reis, F.S. Borges, K.C. Iarosz,
I.L. Caldas, E.L. Lameu, E.E.M. Macau, R.L. Viana, I.M. Sokolov,
F.A.S. Ferrari, J. Kurths, A.M. Batista, Synchronous behavior in net-
work model based on human cortico-cortical connections, Physiological
Measurements 39 (2018) 1.
[41] P.R. Protachevicz, F.S. Borges, E.L. Lameu, P. Ji, K.C. Iarosz, A. H.
Kihara,I.L. Caldas, J. D. Szezech Jr, M. S. Baptista, E.E.M. Macau, C.
G. Antonopoulos, A.M. Batista, J. Kurths, Bistable Firing Pattern in
a Neural Network Model, Frontiers in Computational Neuroscience 13
(2019) 19.
[42] C. R. Noback, N. L. Strominger, R. J. Demarest, D. A. Ruggiero, The
Human Nervous System: Structure and Function (Sixth ed.). Totowa,
NJ: Humana Press, 2005.
[43] F. Gabbiani, C. Koch, Principles of spike train analysis, In C. Koch, I.
Segev, Methods in neural Modeling, Cambridge, MIT Press (1998) 313.
[44] B. Lindner, A brief introduction to some simple stochastic processes,
Stochastic Methods in Neuroscience, Oxford, Oxford University Press
(2009) 1.
[45] S. Grun, S. Rotter, Analysis of parallel spike trains, Springer Science &
Business Media (2010) 1.
[46] S. Wieland, D. Bernardi, T. Schwalger, B. Lindner, Slow fluctuations in
recurrent networks of spiking neurons, Physical Review E 92 (2015) 1.
[47] R.F.O. Pena, S. Vellmer, D. Bernardi, A.C. Roque, B. Lindner, Self-
consistent scheme for spike-train power spectra in heterogeneous sparse
networks, Frontiers in Computational Neuroscience 12 (2018) 1.
[48] D. Cox, P. Lewis, The statistical analysis of series of events, Netherlands,
Springer, 1966.
[49] N. Maier, V. Nimmrich, A. Draguhn, Cellular and network mecha-
nisms underlying spontaneous sharp wave-ripple complexes in mouse
hippocampal slices, Journal of Physiology 550 (2003) 873.
25
[50] P. Giannopoulos, C. Papatheodoropoulos, Effects of mu-opioid receptor
modulation on the hippocampal network activity of sharp wave and
ripples, British Journal of Pharmacology 168 (2013) 1146.
[51] M. Bazelot, M.T. Telenczuk, R. Miles, Single CA3 pyramidal cells trigger
sharp waves in vitro by exciting interneurones, Journal of Physiology 594
(2016) 2565.
[52] K. Mizuseki, G. Buzsaki, Preconfigured, skewed distribution of firing
rates in the hippocampus and entorhinal cortex, Cell Reports 4 (2013)
1010.
[53] A. Peyrache, N. Dehghani, E.N. Eskandar, J.R. Madsen, W.S. Anderson,
J.A. Donoghue, L.R. Hochberg, E. Halgren, S.S. Cash, A. Destexhe,
Spatiotemporal dynamics of neocortical excitation and inhibition during
human sleep, Proceedings of the National Academy of Sciences of the
United States of America 109 (2012) 1731.
[54] R. Brette, W. Gerstner, adaptive, exponential integrate-and-fire model
as an effective description of neural activity, Journal of Neurophysiology
94 (2005) 3637.
[55] G. Maimon, J.A. Assad, Beyond Poisson: increased spike-time regularity
across primate parietal cortex, Neuron 62 (2009) 426.
[56] Y. Mochizuki Y, et al., Similarity in neural Firing Regimes across Mam-
malian Species, Journal of Neuroscience 36 (2016) 5736.
[57] M. Lundqvist, A. Compte, A. Lansner, Bistable, irregular firing and
population oscillations in a modular attractor memory network, Plos
Computational Biology 6 (2010) 1.
[58] C.A. van Vreeswijk, H. Sompolinsky, Irregular activity in large networks
of neurons, Methods and Models in Neurophysics, eds C. Chow, B.
Gutkin, D. Hansel, C. Meunier, J. Dalibard, Amsterdam, Elsevier 80
(2004) 341.
[59] F. Mastrogiuseppe, S. Ostojic, Intrinsically-generated fluctuating activ-
ity in excitatory-inhibitory networks, PLoS Computational Biology 13
(2017) 1.
26
[60] T.P. Vogels, K. Rajan, L.F. Abbott, Neural network dynamics, Annual
Review of Neuroscience 28 (2005) 357.
[61] M. Rudolph, M. Pospischil, I. Timofeev, A. Destexhe, Inhibition de-
termines membrane potential dynamics and controls action potential
generation in awake and sleeping cat cortex, Journal of Neuroscience 27
(2007) 5280.
27
|
1905.00319 | 1 | 1905 | 2019-05-01T14:06:02 | Beta Power May Mediate the Effect of Gamma-TACS on Motor Performance | [
"q-bio.NC"
] | Transcranial alternating current stimulation (tACS) is becoming an important method in the field of motor rehabilitation because of its ability to non-invasively influence ongoing brain oscillations at arbitrary frequencies. However, substantial variations in its effect across individuals are reported, making tACS a currently unreliable treatment tool. One reason for this variability is the lack of knowledge about the exact way tACS entrains and interacts with ongoing brain oscillations. The present crossover stimulation study on 20 healthy subjects contributes to the understanding of cross-frequency effects of gamma (70 Hz) tACS over the contralateral motor cortex by providing empirical evidence which is consistent with a role of low- (12~-20 Hz) and high- (20-~30 Hz) beta power as a mediator of gamma-tACS on motor performance. | q-bio.NC | q-bio | Beta Power May Mediate the Effect of Gamma-TACS on
Motor Performance
Atalanti A. Mastakouri1, Bernhard Scholkopf1 and Moritz Grosse-Wentrup1,2
1Max Planck Institute for Intelligent Systems, Empirical Inference department, T ubingen, Germany
2Research Group Neuroinformatics, Faculty of Computer Science, University of Vienna, Austria
still not fully understood [22,23].
A short overview of the role of physiological β-
and γ-oscillations in movement is important for
the construction of our main argument. Activity
in the γ-band has been associated both with cued
and self-paced transient finger movements [18].
Furthermore, relatively large ballistic movements
of greater movement amplitude were associated
with increased motor cortex γ-power [18]. Moreover,
Gaets et al. gave evidence for a motor γ-band
network for response selection and maintenance
of planned behaviour [24]. These observations jus-
tify our selection of 70 Hz for stimulation of the
contralateral motor cortex in the present study.
On the other hand, β-oscillatory activity has
been found to be significantly elevated in patients
with motor disorders (tremors, slowed movements,
trouble initiating movements) such as Parkinson's
disease [21,25,26]. Furthermore, for healthy subjects,
it was reported that movements preceded by a
reduction in β-power exhibited significantly faster
reaction times than movements preceded by an
increase in β-power [26]. It has also been proposed
that β-activity represents the status quo [27], sug-
gesting that enhanced β-activity prevents change
from the current state [28,29].
Based on existing knowledge about the role of β-
oscillations in the inhibition of movement speed [21],
and about the effect of high stimulation frequencies
on the decrease of β-power [20], we hypothesize that
the modulation of the ongoing β-activity mediates
the effect of γ-tACS on the behavioural response to
the stimulation. This hypothesis has two empirically
testable implications: First, we expect arm speed to
be affected by stimulation. We examine the effect of
γ-tACS on movement in Section II-B. Seccond, we
expect β-power to be affected by the stimulation.
For that purpose, we examine if this is true and,
if so, in which brain areas a modulation of β-
9
1
0
2
y
a
M
1
]
.
C
N
o
i
b
-
q
[
1
v
9
1
3
0
0
.
5
0
9
1
:
v
i
X
r
a
Abstract -- Transcranial alternating current stimula-
tion (tACS) is becoming an important method in the
field of motor rehabilitation because of its ability to
non-invasively influence ongoing brain oscillations at
arbitrary frequencies. However, substantial variations
in its effect across individuals are reported, making
tACS a currently unreliable treatment tool. One reason
for this variability is the lack of knowledge about the
exact way tACS entrains and interacts with ongoing
brain oscillations. The present crossover stimulation
study on 20 healthy subjects contributes to the under-
standing of cross-frequency effects of gamma (70 Hz)
tACS over the contralateral motor cortex by providing
empirical evidence which is consistent with a role of
low- (12 -20 Hz) and high- (20- 30 Hz) beta power as a
mediator of gamma-tACS on motor performance.
I. INTRODUCTION
Transcranial alternating current
stimulation
(tACS) modulates neural activity and behaviour
through the creation of an electric field inside
the brain [1,2]. More specifically, tACS applies a
weak electrical alternating current on the scalp
[3] and changes the membrane potential of the
affected neurons [4]. TACS has been used broadly in
behavioural studies [5,6] as well as for the treatment
of neurological disorders [7,8], although its exact
neurophysiological effect on brain networks has not
yet been fully understood [9].
TACS studies targeting motor cortex have re-
ported considerable variability in stimulation re-
sponse across individual subjects, with large per-
centages of non-responders [5,6]. Although tACS
in the γ- (∼ 70 Hz) and β- (∼ 20 Hz) range has
been proposed to facilitate and inhibit movement,
respectively [10] -- [13], contradictory outcomes have
been reported regarding the significance of these
effects [14] -- [16]. Much light has been shed on the
role of physiological γ- [17,18] and β-oscillations
in movement [19] -- [21]. However, the manner that
tACS entrains these ongoing brain oscillations is
power can be observed (cf. Section II-C). We expect
that changes in β-power induced by γ-tACS are
only significant in the subgroup of subjects that
exhibit a behavioural response to the stimulation.
While these two empirical observations would be
consistent with a role of β-power as mediator of
the effect of γ-tACS on behavioural performance,
causal relations between brain oscillations can not
be substantiated by correlational evidence only. We
thus perform an additional causal analysis, based
on the Information Geometric Causal Inference
algorithm [30], applied here for the first time to
EEG data, examining the effect of β-power change
on motor performance (Section II-D). The results
of this test lend further support to a potentially
causal role of β-power in the response to γ-tACS.
We discuss the neurophysiological plausibility of
our findings in Section IV.
The study conformed to the Declaration of
Helsinki, and the experimental procedures involv-
ing human subjects described in this paper were
approved by the Ethics Committee of the Med-
ical Faculty of the Eberhard Karls University of
T ubingen.
II. METHODS
A. Experimental Setup
Subjects: Twenty healthy, right-handed subjects
participated in this study. One of the subjects (ID
10) did not participate in the second day of the
recordings and was excluded from the analysis. The
remaining 19 healthy participants (nine female, ten
male) were 28.36 ± 8.57 years old.
Stimulation parameters: We chose a crossover
design in which both real- and sham stimulation
are applied to each subject in a randomised order.
High-definition-tACS (HD-tACS) was used for the
stimulation (DC Stimulator Plus, Neuroconn). The
HD 4×1 setup was preferred over the common two-
electrode setup in order to increase the focality of
the stimulation on the preferred motor area [31].
The equalizer extension box was used to extend the
two ordinary square sponge electrodes into a 4×1
set of round rubber electrodes of 20 mm diameter,
with one anode on the region to be stimulated and
four cathodes in a square around it, each cathode
at 7.5 cm from the central anode [32]. The anode
was placed on channel C3 (primary motor cortex
-- M1) and the four cathodes on Cz, F3, T7, and
P3, following the instructions described in [32]. For
both, real- and sham blocks, a duration of 15 min
was chosen [33].
For the real stimulation, a sinusoidal signal mode
at 70 Hz with a peak-to-peak amplitude of 1 mA
was used. For the sham stimulation, a sinusoidal
signal at 85 Hz with a peak-to-peak amplitude of
1uA was selected.
Paradigm: Each participant attended two sessions,
separated by a one-day break. On the first session
(day 1), participants performed a visuomotor target-
reaching task with their right arm, consisting of
three blocks of 50 trials each, while their brain
activity was recorded with EEG. Each block was
separated by a 5-minutes resting-state period, dur-
ing which the participant was asked to relax and
focus on a white cross on the black screen in front
of them. On the second session (day 2), either real
HD-tACS or sham-stimulation was applied in a
randomised order (blinded to the subject) during
the second and third block, respectively. The order
of real-/sham stimulation in the second and third
block was randomised across subjects in order to
compensate for unknown factors such as learning
effects or tiredness over time. The second session
consisted of three blocks, with a 20-minute break
between the second and third block to avoid carry-
over effects between the two blocks. Each block
consisted of as many random reaching -- trials as the
subject could complete in 15 minutes.
The participants were seated on a chair in the
middle of four infrared motion tracking cameras
(PhaseSpace), facing a screen and wearing a spe-
cially designed glove with three LEDs for real-time
tracking of their arm position, which was depicted
in real-time as a 3D sphere, as shown in Fig. 1.
Experimental data: The experimental data include
motion tracking data of the subjects' arm position,
recorded with fs = 960 Hz, and EEG data from
high-density EEG (128 channels, fs = 500 Hz, Brain
Products GmbH).
B. Motor Response to γ-tACS
For the analysis in the present paper, we focused
on the data of the stimulation day (day 2). For
each subject and trial, we computed the mean
movement velocity. The trials of each block, in
which movement speed exceeded three standard
deviations, were excluded as outliers.
Division into responders and non-responders based on
movement velocity: Based on the movement response
of each subject to stimulation, we categorised sub-
jects into two groups. For each subject, we tested
the null hypothesis that the average movement
velocity over the stimulation block is the same as
Fig. 1: Paradigm: The white sphere represents the real-time position of the subject's right arm. The
yellow/green sphere represents the target, appearing at a random location in each trial. Participants
were instructed to reach for the target when its color changed from yellow to green. After the subject
successfully reached the target, a green sphere appeared at the starting position, indicating to return their
hand to the starting position to complete the trial.
the movement velocity during the sham block. For
each subject, we performed a permutation t-test.
To build the null-distribution, we concatenated the
velocities of the two blocks and permuted them
104 times, computing the mean velocity of each
of the two blocks after every permutation. We
calculated the p-value as the frequency at which
we found the absolute difference between mean
velocity during real- and sham stimulation not to be
larger than when drawing from the null-distribution
(two-sided test). Setting a threshold α = 0.05, we
categorised subjects into two groups: Responders if
p < α and the average movement velocity during
the stimulation block was greater than during the
the sham block, and non-responders otherwise, i.e.,
subjects who did not show a significant increase
or who show a decrease in movement speed in
response to stimulation.
Effect size: We quantified the effect size of γ-
tACS over contralateral M1 for each subject as
the difference between the average arm speed
during the stimulation and during the sham block,
normalized by the standard deviation during the
sham block.
C. Effect of γ-tACS on β-power
To attenuate non-cortical artifacts in the EEG data,
we concatenated the EEG signal of each subject's
resting states between blocks, high-pass filtered the
data with a Butterworth filter with cut-off frequency
at 3Hz, applied a common-average reference filter,
and then performed SOBI Independent Component
Analysis (ICA) followed by manual rejection of non-
cortical sources [34].
To examine the effect of γ-stimulation on β-
activity, and the relation of β-power with motor
performance, we calculated the log-bandpower of
the 116 z-scored channels for each subjects (after
having removed the channels used for stimulation)
for the low- (12 -- 20) Hz and high- (20 -- 30) Hz β-
range. For visualization purposes, we calculated
the group grand-averages of the difference between
β-log-bandpower after and before the stimulation
for the stimulation- as well as for the sham block.
To test within each group (responders and non-
responders) whether the changes observed in β-
power are statistically significant, we performed
a permutation paired t-test with 104 permutations
(two-sided): For each channel, we tested the null
hypothesis that the neurophysiological changes in
β-log-bandpower during stimulation come from the
same distribution (across subjects) as those during
sham. We performed FDR-correction for multiple
testing at significance level α = 0.05 [35].
D. Causal analysis of β-power and movement speed
To directly test for causal relationships between
neurophysiological changes in β-power and motor
performance, we applied a causal inference anal-
ysis on the channels of each group. More specif-
ically, we examined which channels express the
causal relationship ∆βBPstimulation → Effect size, where
∆βBPstimulation = βBPafter stimulation − βBPbefore stimulation. To do so,
we used the Information Geometric Causal Inference
(IGCI) algorithm proposed by [30], which is applied
here for the first time to study causal relationships
between brain oscillations and behaviour. The IGCI
inference algorithms is based on the assumption that
if X → Y , the distribution of X and the function f
that maps X to Y are independent, i.e., it assumes
that the mechanism and the data that it processes
are not co-adapted. Independence between the
function f and the distribution of X is computed by
the relative entropy distance D(., .), which is then
used to estimate CX→Y = D(pXEX ) − D(pY EY ).
Based on the sign of CX→Y , the IGCI algorithm
decides which causal direction is more likely. If
CX→Y > 0, then X is inferred as the cause of Y .
We note that this test does not take into account
the possibility of hidden confounders.
III. RESULTS
A. Motor response to γ-tACS
Out of the original population of 19 subjects, only
six subjects responded positively to γ-tACS over
contralateral M1. Figure 2 depicts the effect size
for each subject, with the color indicating whether
the subject was a responder or a non-responder.
Contralateral M1 γ-tACS has been proposed in
several studies as a stimulation setup that facilitates
movement. Here, however, the overall effect size
is quite small (0.2366) due to the existence of
responders and non-responders with effect sizes
0.9073 and −0.1547, respectively. Based on these
observations, we next examined the effect of γ-tACS
on β-powerfor each of the two groups individually.
B. Effect of γ-tACS on β-power
We hypothesized that subjects, who exhibit a
larger decrease in β-log-bandpower over the con-
tralateral motor cortex, are those that respond posi-
tively to the stimulation, i.e., with faster movements.
In Figure 3, we see that the group of subjects, that
responded positively to the γ-stimulation, show
a larger decrease of β-power, mostly in the high
β-range [20 30] Hz and spread out over the con-
tralateral motor cortex, compared to the group of
the subjects that did not respond to stimulation.
Moreover, for the sham condition in the group
of responders, we found little change of β-log-
bandpower over contralateral motor cortex. In
contrast, the group of non-responders exhibits a
bilateral increase of β-power.
Figure 4 depicts the channels that exhibit FDR-
corrected statistical significance of differences in
β-power, between the conditions of real- and sham
stimulation, for each group. For the responders
group, these channels are found to be located over
the contralateral motor cortex, FC1, C1 and CCP3h
for the high β-range and FC1 for the low β-range.
For the non-responders, in contrast, we found no
channel with a significant difference between the
two conditions.
C. Causal analysis of β-power and movement speed
The causal inference analysis, for the detection of
channels that satisfy the relationship ∆βBPstimulation →
Effect size, was applied independently for each of
the two groups as well as for the low- and high β-
range. In the low β-range, no channel was found to
satisfy the above relationship. For the high β-range,
the channels that did not satisfy the condition were
set to zero. The remaining channels are depicted
in Fig. 5, colour-coded according to the difference
between high β-log-bandpower after and before
stimulation. We observe that for the responders
group the left motor cortex exhibits the above causal
relationship. In contrast, the majority of the channels
for the group of non-responders do not satisfy the
causal relationship.
IV. DISCUSSION
Applying HD-tACS at 70 Hz over contralateral
motor cortex on 20 healthy subjects, we found a
significant increase of upper-limb movement speed
in 36% of the original population. Consistent with
the results in [5], we found a large number of non-
responders. Considering the fact that γ-stimulation
is believed to facilitate movement [5,10], as well as
that an increase of γ-oscillatory activity has been
associated with large ballistic movements [18], we
decided to investigate the underlying modulation
of the antikinetic β-oscillations [36] as a possible
cause of this variability. We hypothesized that if
γ-stimulation is affecting the ongoing β-oscillations,
then the subjects that exhibit a larger decrease
in β-power should be those that respond to
the stimulation. Our EEG findings support this
hypothesis.
Taken together, our findings support a potential
role of β-power as a mediator of γ-tACS on motor
performance. In particular, the results reported in
Section III-A establish that γ-tACS (S) has an effect
on movement performance (P ), as measured by
movement velocity. Because S is a randomised
treatment, we can infer the direction of this relation
as S → P . The results in Section III-B, on the
other hand, demonstrate an effect of γ-tACS on
β-power, i.e., a causal path S → ∆β. It then remains
to distinguish between the two causal models
S → ∆β → P (with potentially an additional
path S → P that does not pass through ∆β) and
∆β ← S → P , both of which are consistent with
our evidence to this point. The results of the
causal analysis in Section III-C indicate that in the
stimulation condition ∆β → P , which is consistent
with the former and not with the latter causal
model. We thus argue that our empirical results
are in favour of the causal model S → ∆β → P ,
Fig. 2: Individual effect sizes for stimulation as measured by changes in movement velocity. The p-value
(rounded up to two decimals) for each subject is shown on top of each bar (cf. II-B). Green bars: Subjects
that performed significantly better during stimulation (responders). Brown bars: Subjects who either did
not respond to the tACS, or who performed significantly worse compared to sham (non-responders).
Green line: Average effect size of responders. Red line: Average effect size of non-responders. Yellow line:
Overall effect size of the whole population.
Fig. 3: Difference between β-power after and before stimulation (left) and sham (right), in the low- (12 -- 20)
Hz and high- 20 -- 30 Hz β-range for the groups of responders and non-responders.
Fig. 4: Difference of β-power after and before stimulation for the real- and sham stimulation in the low
(12 -- 20 Hz) and high- (20 -- 30 Hz) β-range, for the group of responders. FDR-corrected channels, that do
not exhibit significant neurophysiological differences, are set to zero.
Fig. 5: Difference of β-power after and before stimulation, in the low- (left) and high (right) β-range, with
channels that do not satisfy the causal relationship ∆βBPstimulation → Effect size set to zero.
i.e., that β-power may mediate the effect of γ-tACS
on motor performance. We stress, however, that
causal inference methods as applied here can not
prove but only provide empirical results consistent
with causal relationships.
We further note, however, that the causal model
S → ∆β → P is neurophysiologically plausible.
In the context of neurophysiological procedures
underlying the effect of γ-stimulation on β-power
and on the observed motor behaviour, a possible
explanatory factor could be the modulation of γ-
aminobutyric acid (GABA) concentration. We sup-
port this claim with the following argument. First,
β-oscillations have been shown to be the summed
output of principal cells temporally aligned by
GABAergic interneuron rhythmicity [37]. Specifi-
cally, GABA levels have been found to strongly
correlate with β-power and to exhibit elevated
values in bradykinesia and in Parkinson's disease
[21]. Secondly, high-γ deep brain stimulation in
motor cortex has been reported to cause a sig-
nificant decrease in β-power [20], supporting our
finding of the inhibitory effect of γ-stimulation on
the ongoing β-oscillations. Combining these two
literature research conclusions, we argue that the
behavioural response to γ-tACS may be explained
by a decrease of β-power and hence of GABA levels,
modulated by the stimulation. We note that it is
also conceivable that whenever γ-tACS leads to the
inhibition of human movements, this may be caused
by an increase in GABAergic drive, which hinders
β-power to be decreased.
REFERENCES
[1] C. Herrmann, S. Rach, T. Neuling, and D. Str uber, "Tran-
scranial alternating current stimulation: a review of the
underlying mechanisms and modulation of cognitive pro-
cesses," Front Hum Neurosci, vol. 7, p. 279, 2013.
[2] S. Bestmann and V. Walsh, "Transcranial electrical
stimulation," Curr. Biol., vol. 27, no. 23, pp. R1258 -- R1262,
2017. [Online]. Available: http://www.sciencedirect.com/
science/article/pii/S0960982217314434
[3] N. MA and P. W, "Excitability changes induced in the
human motor cortex by weak transcranial direct current
stimulation," J. Physiol., vol. 527, pp. 1469 -- 7793, 2000.
[4] A. Antal and C. S. Herrmann., "Current and random noise
stimulation: Possible mechanisms," Neural Plast., 2016.
[5] V. L´opez-Alonso, B. Cheeran, D. R´ıo-Rodr´ıguez, and M. F.
del Olmo, "Inter-individual variability in response to non-
invasive brain stimulation paradigms," Brain Stimul, vol. 7,
no. 3, pp. 372 -- 380, 2014.
[6] W. Strube, T. Bunse, B. Malchow, and A. Hasan, "Efficacy
and interindividual variability in motor-cortex plasticity
following anodal tdcs and paired-associative stimulation,"
Neural Plast., no. 530423, 2015.
[7] R. Schulz, C. Gerloff, and F. C. Hummel, "Non-
invasive brain stimulation in neurological diseases,"
Neuropharmacology, vol. 64, pp. 579 -- 587, 2013.
[8] F. Fregni, D. K. Simon, A. Wu, and A. Pascual-Leone,
"Non-invasive brain stimulation for Parkinson's disease:
a systematic review and meta-analysis of the literature,"
J. Neurol. Neurosurg. Psychiatry, vol. 76, no. 12, pp. 1614 --
1623, 2005.
[9] J. Vosskuhl, D. Str uber, and C. S. Herrmann, "Non-invasive
brain stimulation: A paradigm shift in understanding brain
oscillations," Front Hum Neurosci, vol. 12, p. 211, 2018.
[10] R. A. Joundi, N. Jenkinson, J.-S. Brittain, T. Z. Aziz, and
P. Brown, "Driving oscillatory activity in the human cortex
enhances motor performance," Curr. Biol., vol. 22, no. 5,
pp. 403 -- 407, 2012.
[11] A. Pogosyan, L. Doyle Gaynor, A. Eusebio, and P. Brown,
"Boosting cortical activity at beta-band frequencies slows
movement in humans," Curr. Biol., vol. 19, no. 19, pp. 1637
-- 1641, 2009.
[12] V. Moliadze, A. Antal, and W. Paulus, "Boosting brain
excitability by transcranial high frequency stimulation in
the ripple range," J. Physiol., vol. 588, no. 24, pp. 4891 -- 4904,
2010.
[13] C. Wach, V. Krause, V. Moliadze, W. Paulus, A. Schnitzler,
and B. Pollok, "Effects of 10hz and 20hz transcranial
alternating current stimulation (tacs) on motor functions
and motor cortical excitability," Behav. Brain Res., vol. 241,
pp. 1 -- 6, 2013.
[14] M. Moisa, R. Polania, M. Grueschow, and C. C. Ruff,
"Brain network mechanisms underlying motor enhancement
by transcranial entrainment of gamma oscillations," J.
Neurosci., vol. 36, no. 47, pp. 12 053 -- 12 065, 2016.
[15] S. L. Gonzalez Andino, C. M. Michel, G. Thut, T. Landis,
and R. Grave de Peralta, "Prediction of response speed by
anticipatory high-frequency (gamma band) oscillations in
the human brain," Hum. Brain Mapp., vol. 24, no. 1, pp.
50 -- 58, 2005.
[16] A. Antal, K. Boros, C. Poreisz, L. Chaieb, D. Terney, and
W. Paulus, "Comparatively weak after-effects of transcranial
alternating current stimulation (tACS) on cortical excitability
in humans," Brain Stimul, vol. 1, no. 2, pp. 97 -- 105, 2008.
[17] M. Nowak, C. Zich, and C. Stagg, "Motor cortical gamma
oscillations: What have we learnt and where are we
headed?" Curr Behav Neurosci Rep, vol. 136, no. 5, 2018.
[18] S. D. Muthukumaraswamy, "Functional properties of
human primary motor cortex gamma oscillations," J.
Neurophysiol., vol. 104, no. 5, pp. 2873 -- 2885, 2010.
[19] S. Espenhahn, "The relationship between cortical beta
oscillations and motor learning," Doctoral Thesis, UCL,
2018.
[20] A. Gulberti, C. Moll, W. Hamel, C. Buhmann, J. Koeppen,
K. Boelmans, S. Zittel, C. Gerloff, M. Westphal, T. Schneider,
and A. Engel, "Predictive timing functions of cortical
beta oscillations are impaired in Parkinson's disease and
influenced by l-dopa and deep brain stimulation of the
subthalamic nucleus," NeuroImage: Clinical, vol. 9, pp. 436
-- 449, 2015.
[21] C. J. McAllister, K. C. Ronnqvist, I. M. Stanford, G. L.
Woodhall, P. L. Furlong, and S. D. Hall, "Oscillatory
beta activity mediates neuroplastic effects of motor cortex
stimulation in humans," J. Neurosci., vol. 33, no. 18, pp.
7919 -- 7927, 2013.
[22] N. Davis and M. Koningsbruggen, "Non-invasive brain
stimulation is not non-invasive," Front Syst Neurosci, vol. 7,
p. 76, 2013.
[23] R. F. Helfrich, C. S. Herrmann, A. K. Engel, and T. R.
Schneider, "Different coupling modes mediate cortical cross-
frequency interactions," NeuroImage, vol. 140, pp. 76 -- 82,
2016.
[24] W. Gaetz, C. Liu, H. Zhu, L. Bloy, and T. Roberts, "Evidence
for a motor gamma-band network governing response
interference," NeuroImage, vol. 74, pp. 245 -- 253, 2013.
[25] P. Brown, "Abnormal oscillatory synchronisation in the
motor system leads to impaired movement," Curr. Opin.
Neurobiol., vol. 17, no. 6, pp. 656 -- 664, 2007.
[26] P. Khanna and J. M. Carmena, "Beta band oscillations in
motor cortex reflect neural population signals that delay
movement onset," ELife, vol. 6, 2017.
[27] A. K. Engel and P. Fries, "Beta-band oscillations -- signalling
the status quo?" Curr. Opin. Neurobiol., vol. 20, no. 2, pp.
156 -- 165, 2010.
[28] A. Schnitzler and J. Gross, "Normal and pathological os-
cillatory communication in the brain," Nat. Rev. Neurosci.,
vol. 6, p. 285, 2005.
[29] N. J. Davis, S. P. Tomlinson, and H. M. Morgan, "The role
of beta-frequency neural oscillations in motor control," J.
Neurosci., vol. 32, no. 2, pp. 403 -- 404, 2012.
[30] P. Daniusis, D. Janzing, J. Mooij, J. Zscheischler, B. Steudel,
K. Zhang, and B. Scholkopf, "Inferring deterministic
causal relations," in Proceedings of the 26th Conference
on Uncertainty in Artificial Intelligence, 2010, pp. 143 -- 150.
[31] J. P. Dmochowski, A. Datta, M. Bikson, Y. Su, and L. C.
Parra, "Optimized multi-electrode stimulation increases
focality and intensity at target," J Neural Eng, vol. 8, p.
046011, 2011.
[32] M. F. Villamar, M. S. Volz, M. Bikson, A. Datta, A. F. DaSilva,
and F. Fregni, "Considerations in the use of 4x1 ring high-
definition transcranial direct current stimulation (hd-tdcs),"
JoVE, vol. 77, p. e50309, 2013.
[33] D. Str uber, S. Rach, T. Neuling, and C. S. Herrmann, "On
the possible role of stimulation duration for after-effects
of transcranial alternating current stimulation," Front Cell
Neurosci, vol. 9, p. 311, 2015.
[34] B. W. McMenamin, A. J. Shackman, J. S. Maxwell, D. R.
Bachhuber, A. M. Koppenhaver, L. L. Greischar, and R. J.
Davidson, "Validation of ica-based myogenic artifact cor-
rection for scalp and source-localized eeg," NeuroImage,
vol. 49, no. 3, pp. 2416 -- 2432, 2010.
[35] Y. Benjamini and Y. Hochberg, "Controlling the false dis-
covery rate: A practical and powerful approach to multiple
testing," J R Stat Soc Series B Stat Methodol, vol. 57, no. 1,
pp. 289 -- 300, 1995.
[36] L. Brinkman, A. Stolk, H. C. Dijkerman, F. P. de Lange,
and I. Toni, "Distinct roles for alpha- and beta-band oscil-
lations during mental simulation of goal-directed actions,"
J. Neurosci., vol. 34, no. 44, pp. 14 783 -- 14 792, 2014.
[37] N. Yamawaki, I. Stanford, S. Hall, and G. Woodhall,
"Pharmacologically induced and stimulus evoked rhythmic
neuronal oscillatory activity in the primary motor cortex in
vitro," Neuroscience, vol. 151, no. 2, pp. 386 -- 395, 2008.
|
1706.00085 | 1 | 1706 | 2017-05-31T20:58:20 | A Formal Approach to Modeling the Cost of Cognitive Control | [
"q-bio.NC"
] | This paper introduces a formal method to model the level of demand on control when executing cognitive processes. The cost of cognitive control is parsed into an intensity cost which encapsulates how much additional input information is required so as to get the specified response, and an interaction cost which encapsulates the level of interference between individual processes in a network. We develop a formal relationship between the probability of successful execution of desired processes and the control signals (additive control biases). This relationship is also used to specify optimal control policies to achieve a desired probability of activation for processes. We observe that there are boundary cases when finding such control policies which leads us to introduce the interaction cost. We show that the interaction cost is influenced by the relative strengths of individual processes, as well as the directionality of the underlying competition between processes. | q-bio.NC | q-bio | A Formal Approach to Modeling the Cost of Cognitive Control
Kayhan Ozcimder 1,2,∗, Biswadip Dey2,∗, Sebastian Musslick1,∗,
Giovanni Petri3, Nesreen K. Ahmed4, Theodore L. Willke4, Jonathan D. Cohen1
1Princeton Neuroscience Institute, Princeton University, Princeton, NJ 08540, USA.
2Department of Mechanical and Aerospace Engineering, Princeton University, Princeton, NJ 08544, USA.
3 ISI Foundation, Via Alassio 11/c, 10126 Torino, Italy.
4Intel Labs, Santa Clara, CA 95054, USA.
∗Equal Contribution, Corresponding Author: [email protected]
7
1
0
2
y
a
M
1
3
]
.
C
N
o
i
b
-
q
[
1
v
5
8
0
0
0
.
6
0
7
1
:
v
i
X
r
a
Abstract
This paper introduces a formal method to model the level of de-
mand on control when executing cognitive processes. The cost
of cognitive control is parsed into an intensity cost which en-
capsulates how much additional input information is required
so as to get the specified response, and an interaction cost
which encapsulates the level of interference between individ-
ual processes in a network. We develop a formal relationship
between the probability of successful execution of desired pro-
cesses and the control signals (additive control biases). This
relationship is also used to specify optimal control policies to
achieve a desired probability of activation for processes. We
observe that there are boundary cases when finding such con-
trol policies which leads us to introduce the interaction cost.
We show that the interaction cost is influenced by the relative
strengths of individual processes, as well as the directionality
of the underlying competition between processes.
Keywords: cognitive control; multi-tasking; intensity; iden-
tity
Introduction
A long standing focus in cognitive research has been to-
wards understanding the ability to execute tasks/processes1
that demand cognitive control. In this context, cognitive con-
trol is defined as the set of mechanisms required to pursue
a goal, especially when distraction or strong competing re-
sponses (interferences) must be overcome (Posner & Sny-
der, 1975; Shiffrin & Schneider, 1977; Cohen, Dunbar, &
McClelland, 1990). Earlier work (Posner & Snyder, 1975;
Shiffrin & Schneider, 1977; Cohen et al., 1990; Botvinick &
Cohen, 2014) has argued that the processes demanding con-
trol can be distinguished from automatic processes in terms
of the strength of the associations in the pathways underlying
processing: automatic processes are characterized by path-
ways with associations strong enough to resist interference
from competing processes, whereas controlled processes are
weaker, and therefore rely on input from control mechanisms
to support their execution against interference.
Another longstanding observation is that the allocation of
cognitive control is costly (often discussed in terms of "men-
tal effort" (Posner & Snyder, 1975; Botvinick & Braver,
2015; Shenhav et al., 2017)). This cost has been interpreted in
physical terms (such as metabolic demands (Muraven, Tice,
1A task/process/input-output mapping is defined as a unique
mapping from all possible vectors in the input subspace to corre-
sponding vectors in the output subspace, that is independent of the
mappings for all other combinations of input and output components
in the network.
& Baumeister, 1998)) or in terms of an opportunity cost re-
flecting the allocation of a limited resource (Kurzban, Duck-
worth, Kable, & Myers, 2013). Elsewhere (Feng, Schwem-
mer, Gershman, & Cohen, 2014; Musslick et al., 2016), we
have proposed that limitations in the capacity for control-
dependent processing reflect the purpose of control to di-
minish interference rather than any intrinsic limitation in the
mechanism responsible for control. This view suggests that
the architecture of the processing system as a whole con-
strains the opportunities for control-dependent processing, re-
sulting in opportunity costs associated with allocating control
to any particular task(s).
Here, we build on a closely related proposal, by Koechlin
and Summerfield (2007), to define the cost of control in terms
of internal representational requirements to insure that a given
stimulus (or a set of stimuli) produces the desired response
(or a set of responses), given the intrinsic architecture of the
system. Their work focused on a single task. Here, we ex-
tend this to consider an arbitrary number of tasks and thus
accommodate their possibility for, and costs of, multitasking
(i.e.parallel processing of task pathways). To do so, we follow
the framework proposed by Shenhav, Botvinick, and Cohen
(2013) that distinguishes two components of control signals:
intensity and identity. Specifically, Shenhav et al. (2013) de-
fined the intensity of a control signal as the strength of the sig-
nal needed to insure performance of a particular task, and the
identity as which control signal should be selected to achieve
a desired objective given environmental conditions. Here we
build on that distinction to define two corresponding com-
ponents of the cognitive control costs - a cost associated with
intensity, and a cost associated with interaction. Furthermore,
we define the interaction cost to capture the level of interfer-
ence between the processes in a network.
In this paper, we begin by introducing formal constructs
for intensity and interaction costs by using the graph theoretic
representation of a neural network and terms/notions adopted
from probability theory. We describe an intensity cost that
represents the control signals (as biases infused into a neural
network), above and beyond the specified strength of the sig-
nal (stimulus) itself. This is achieved by developing a formal
relationship between the probability of successful execution
of desired processes and the control signals. In turn, this de-
fines an optimization problem, which can then be solved to
find optimal control signals that achieve a specified activa-
tion for desired processes. However, we observe that there
2016), we consider a single-layered, feed-forward network
with N input and M output layer components to formalize the
notion of intensity cost in a cognitive control context (Fig. 1
shows a simple example of such a network). In this frame-
work, each component represents an input/stimulus or out-
put/response dimension (vector subspace), and the connec-
tion from an input to an output component constitutes the
processing pathway for a given task. This allows us to de-
fine an abstraction of the network as a directed bipartite graph
GB = (V ,E), wherein the set of vertices V can be partitioned
into two disjoint sets Vin and Vout, representing the input and
output layer components respectively. Moreover, a directed
edge (i, j) ∈ E ⊆ Vin×Vout represents a connection from the
vertex i in the input layer to vertex j in the output layer (i.e.,
a task). In this setting, we represent the processing pathway
by introducing a weight matrix W with elements wi j. As we
will see later, this abstraction plays an important role in for-
malizing the interaction cost of cognitive control.
In this setting, we assume that control signals bias the pro-
cessing of a stimulus towards a specified response at two
different levels, i.e. gi and b j, which we refer to as pre-
interaction and post-interaction control biases, respectively.
This complies with early computational models of cognitive
control in which control signals act as an increase in gain of
non-linear processing units (Cohen et al., 1990; Botvinick et
al., 2001) and allows us to treat such control biases as key
contributing factors towards the intensity cost for cognitive
control. It is worth noting here that for simplicity the only
sources of nonlinearity in this setting are the logistic2 acti-
vation functions which act upon the linearized output vec-
tor y = [ Y1, Y2, . . . , YM]. Without loss of generality, in what
follows we consider the individual features to be scalar, and
carry out a formal investigation on how these control biases
g = [g1, . . . ,gN] and b = [b1, . . . ,bM] influence the response
from this cognitive architecture. In our formulation, the cor-
responding magnitude, i.e. (cid:107)g(cid:107)2 +(cid:107)b(cid:107)2, can be treated as a
measure of control intensity applied to the system, and there-
fore the cost for cognitive control.
We begin our analysis by assuming the vector of fea-
tures s = [S1,S2, . . . ,SN] to be an N-dimensional multivariate
Gaussian random variable with mean µS and covariance ΣS.
(The assumption of Gaussianity is motivated by the technical
tractability). With this assumption, the vector [X1,X2, . . . ,XN]
becomes an N-dimensional multivariate Gaussian random
variable with a shifted mean and same covariance. Further-
more, as all the transformations (before the nonlinear logis-
tic activation function) are linear in nature, [Y1,Y2, . . . ,YN]
also remains a multivariate Gaussian random variable whose
mean and covariance are given by µY = W (µS + g) and ΣY =
WΣSW T , respectively. Similarly, the vector of linearized out-
puts [ Y1, Y2, . . . , YN] is also a multivariate Gaussian, with a
shifted mean and the same covariance.
2Although we are restricting ourselves to logistic functions with
unit steepness, in a more general setting one can use the steepness
as another design parameter.
Figure 1:
Illustration of a single-layered, feed-forward network
with 3-input and 3-output layer components, wherein the individual
features are scalar in nature.
are boundary cases in which this optimization problem can
not be solved. These cases reflect situations in which the si-
multaneous execution of the processes is not feasible due to
interference. Hence, interaction cost analysis motivates an
additional investigation towards finding a proper metric that
continuously measures the level of interference between pro-
cesses. To achieve this we introduce the definition of inter-
action cost associated with process mappings in a network
configuration. Specifically, it measures the level of interfer-
ence introduced by competing processes that interfere with
the tasks of interest. In their study, Koechlin and Summer-
field (2007) have already used information theoretic terms to
measure cognitive control. However, in order to apply these
metrics to neural networks while considering parallelism, we
augment some of these measures, and through simulations we
will demonstrate how interaction cost can be used to predict
interference in neural network architectures. Finally, we will
discuss general research directions revealed by the analysis
presented here.
Intensity: the cost of control
Cognitive control is defined as the underlying mechanism
that biases the processing of a task in order to maximize the
reward (Botvinick, Braver, Barch, Carter, & Cohen, 2001;
Botvinick, Cohen, & Carter, 2004; Bogacz, Brown, Moehlis,
Holmes, & Cohen, 2006; Botvinick, 2007). Here, we adapt
the notion of intensity cost from Shenhav et al. (2013) as a
function of the amount of control bias that cognitive con-
trol applies to the system. However, Shenhav et al. (2013)
described this function in qualitative rather than quantitative
terms. In this work, we provide an explicit characterization
of the cost of cognitive effort in terms of a set of physically
meaningful parameters, which allow the manipulation of the
response of a cognitive architecture.
Following earlier works (Feng et al., 2014; Musslick et al.,
Hence, each individual linearized output Yi is itself a Gaus-
sian random variable with
mean:
µ Y
i =
and variance: σ Y
ii =
N
∑
j=1
N
∑
j=1
w ji(µ j + g j) + bi
N
∑
k=1
w jiwkiσk j,
where µ j is the mean of stimulus Sk and σk j is the covariance
between stimuli Sk and S j. As a consequence, the correspond-
ing output (response) will have a logit-normal distribution,
and this leads us to our key result in this section.
As outlined by Shenhav et al. (2013), the response Oi
should overcome a specified threshold in order to execute
the corresponding process (task). Then, by letting αi ∈ (0,1)
represent this activation threshold associated with output Oi,
the corresponding probability of task execution (probability
of the output Oi surpassing the threshold αi) is expressed as
Here we have exploited the monotonicity of the logistic func-
tion to compute its inverse. Then the result follows from the
cumulative distribution function of Yi.
M
∑
i=1
N
∑
i=1
i ≤ Cg and
g2
We characterized the activation probability of a given net-
work in terms of the pre-interaction and post-interaction con-
trol biases. This is crucial because it provides new directions
to incorporate the cost of control into the design of a cognitive
network architecture. For example, the problem of allocating
a limited amount of cognitive control into different compo-
nents of the network to maximize the associated probability
of activation can be formulated as an optimization problem in
which the goal becomes minimizing f (αi,bi,w,µ,g,ΣS) over
i ≤ Cb,
b2
g and b subject to the constraints
where Cg and Cb define the maximum amount of control that
can be applied. Alternatively, in this setting, we can also ap-
proach the problem of minimizing the cost of control, while
still maintaining a desired value of probability of activation.
One can consider the joint distribution of the processes of
interest to incorporate the effects of interaction between tasks.
To be consistent with (Feng et al., 2014; Musslick et al.,
2016), it is reasonable to begin with a focus narrowed to the
situation where the choice of interaction weights and the prior
distribution of the stimuli render the interactions undesirable.
Then the effect of multitasking can measured by introducing
a suitable distance metric (for example, the Kullback-Leibler
divergence (Ortega & Braun, 2013)) between the joint distri-
butions of relevant processes and the product of correspond-
ing marginals, and one can attempt to minimize this distance
log
(cid:16) αi
(cid:115)
1−αi
(cid:17)− bi − N
(cid:123)(cid:122)
N
∑
k=1
N
∑
j=1
2
f (αi,bi,w,µ,g,ΣS)
w ji(µ j + g j)
∑
j=1
w jiwkiσk j
P[Oi ≥ αi] =
1
2
− 1
2
erf
(cid:124)
(cid:125)
.
at the expense of a limited amount of cognitive control. How-
ever, as one might expect, this optimization problem can have
an empty solution set under certain values of the interaction
weights and activation thresholds, meaning that certain net-
work configurations strictly prohibit successful multitasking
performance (Musslick et al., 2016). Before approaching this
computation in detail, it would be beneficial to investigate
how the interaction structure influences the solution space,
and that leads us to our next section wherein we introduce the
notion of interaction cost.
Interaction: the cost of mapping
In this section, we will introduce a detailed formalism of the
interaction cost associated with process mappings in a net-
work configuration to accommodate the possibility for multi-
tasking. In our earlier work (Musslick et al., 2016), we have
formalized three distinct types of interference (Fig. 2).
Convergent interference (Fig. 2a) occurs when two in-
puts/stimuli (e.g. S1 and S2) compete to determine a common
output (e.g. O1). We also consider divergent interference in
our analysis (Fig. 2b). Although this does not pose an im-
pediment to performance, i.e. it is possible to generate two
distinct outputs (e.g. O1 and O2) to the same input (e.g. S1),
it represents a restriction on the number of independent stim-
uli (and therefore the number of tasks) that the system can
process at once, and thus was treated formally as a type of
interference due to this dependency in our analysis of paral-
lel processing capability. Finally, we consider a third, indi-
rect interference that supervenes on the first two (shown in
Fig. 2c and Fig. 2d). In this case, the two tasks with strengths
w11 and w22 in question do not directly interfere with one
another. However, their simultaneous execution would nec-
essarily engage a third task with strength w21 (also possibly
a fourth task with strength w12) that would produce interfer-
ence in output O1 (and O2). While Musslick et al. (2016)
treated these three types of interference identically in terms
of their effect on the overall parallel processing capability of
a network, the proposed interaction cost will also distinguish
between these three types of interference.
In interaction cost analysis, we will assume that a stimulus
is of value 1 when it is active, and 0 otherwise. Moreover, to
increase tractability, we will consider linear activation at the
output level, which also implies that without loss of general-
ity the pre- and post-interaction biases can be assumed to be
zero. A more detailed version of the interaction cost analysis,
involving the strength of stimuli, as well as the nonlinear acti-
vation function, will be discussed in subsequent publications.
To introduce the interaction cost, we take an approach sim-
ilar to the one adopted by Koechlin and Summerfield (2007).
In their work, Koechlin and Summerfield (2007) proposed a
metric for selecting a single action among multiple alterna-
tives. Here, we will refine this metric to introduce the inter-
action cost for neural network architectures. Towards this ob-
jective, we first leverage the assumptions discussed earlier in
Equation 1 implies the P[a j = i] = 1 when only the rel-
evant stimulus Si associated with task i j is activated in the
network. Hence the interaction cost is computed as Ψ(a j =
i) = 0 which implies that there is no interaction cost. More-
over, when multiple processes are competing due to the ac-
tivation of multiple stimuli, P[a j = i] → 0 as the compe-
tition increases, and as a consequence the interaction cost
Ψ(a j = i) → ∞.
We further extend equation 1, to encapsulate the proba-
bility associated with parallel processing of task pathways
in the network. Thus, we introduce the joint probability
of distinct output components responding to a set of stim-
uli. For instance, let us consider the parallel processing of
tasks with strength w11 and w21 in Fig. 2a, and calculate
P[a1 = 1,a1 = 2]. This is the probability of output compo-
nent O1 responding to both S1 and S2, and by definition we
know that this probability is zero. For the case illustrated in
Fig. 2b, the joint probability P[a1 = 1,a2 = 1] = 1 since ac-
tivation of S1 will activate both processes with strengths w11
and w12, and there is no competition in outputs O1 and O2.
This result is parallel to the observation made by (Musslick
et al., 2016), who stated that divergent interference is not ac-
tually an interference but a dependency on the stimuli.
Now let us consider the case introduced in Fig. 2c which
can be thought of as the composition of the two cases pre-
sented in Fig. 2a-b. We compute the interaction cost of paral-
lel processing the tasks with strengths w11 and w22. This re-
quires simultaneous activation of S1 and S2, which indirectly
engages the task with strength w21, and initiates a competition
in the output O1. Thus, the interaction cost of parallelism be-
tween tasks represented by w11 and w22 is given by
Ψ1(a1 = 1,a2 = 2) = −log(P[a2 = 2]· P[a1 = 1a2 = 2])
(cid:18)
= −log
1·
w11
w11 + w21
(cid:19)
.
Here P[a2 = 2] = 1 since task with weight w22 is not compet-
ing with any other process in the output O2. The competition,
however, takes place in O1, and the interaction cost associ-
ated with w11 for this case has already been computed when
we discussed the case in Fig. 2a.
In a similar way, we can compute the interaction cost
of parallelism between tasks represented by w11 and w22 in
Fig. 2d, and we have
Ψ2(a1 = 1,a2 = 2) = −log(P[a2 = 2]· P[a1 = 1a2 = 2])
(cid:18) w11
= −log
w11 + w21
·
w22
w22 + w12
(cid:19)
.
In this case, simultaneous activation of S1 and S2 causes com-
petition in both outputs O1 and O2. Thus, by revealing further
insight about the strength and directionality of interference,
the interaction cost serves as an extension of the interference
definition presented by (Musslick et al., 2016). For instance,
for the same values of w11,w21,w22 ≥ 0 in both configura-
tions in Fig. 2c-d, and given w12 ≥ 0 for the configuration in
Figure 2: The illustration of convergent, divergent, asymmetric,
and symmetric interference.
the section, and abstract out the network configurations pre-
sented in Fig. 2 from the network shown in Fig. 1. We also
assume that the strength of a task i j from stimulus Si to output
O j is represented by its non-negative weight wi j ≥ 0.
Let us first consider the case shown in Fig. 2a. It is obvious
that the response in the output component O j is completely
determined by the stimulus if either S1 or S2 is activated in
the network (executing a single process). However, activating
both stimuli S1 and S2 simultaneously creates a conflict, since
the output can not respond to two distinct stimuli simultane-
ously (as the activations are linear, the network will always
have a response in the output level). In order to measure the
level of this competition between stimuli, we define a random
variable a1 associated with the output O1 such that a1 ∈ {1,2}
(Fig. 2a). This implies that a1 = 1 or a1 = 2 when the out-
put O1 is driven completely by S1 or S2, respectively. Since a
stronger task will have a higher probability of being selected
to generate the response, we consider the relative strengths of
the task pathways (with associated strengths w11 and w21) in
order to define the probability of the possible outcomes, i.e.
the probability of a1 = 1 and a1 = 2 when both S1 and S2
activated. Hence, we compute the probability as
P[a1 = 1] =
w11
w11 + w21
,
and, P[a1 = 2] =
w21
w11 + w21
.
Next, we extend this framework to consider networks
j ∈
with N stimuli and M outputs, wherein an output O j,
{1, . . . ,M} responds to a set of stimuli. Let us assume that
there are n ≤ N incoming edges to a particular output O j,
and each edge is originated from a distinct stimulus Si, i =
i1, . . . ,in, where ik ∈ {1, . . . ,N}. Then the probability of the
event that output O j is responding to stimulus Si is given by,
P[a j = i] =
wi j 1(Si)
wik j 1(Sik )
n
∑
k=1
,
(1)
where 1(Si) is the indicator function that represents the ac-
tivation of stimulus Si such that 1(Si) = 1 if stimulus Si is
active and 1(Si) = 0, otherwise. Then, by building upon the
ideas proposed by (Koechlin & Summerfield, 2007), we de-
fine the interaction cost as
Ψ(a j = i) = −log(P[a j = i]),
where the logarithm is with respect to base 2.
(2)
Fig. 2d, it is obvious that
Ψ1(a1 = 1,a2 = 2) ≤ Ψ2(a1 = 1,a2 = 2).
Neural Network Simulation
In order to investigate the effect of directionality for the third
indirect interference during parallel processing, we imple-
mented a synthetic neural network simulation3 identical to
our earlier work (Musslick et al., 2016). The neural net-
work used for this simulation maps stimulus input encoded
at a stimulus layer via a non-linear associative layer to non-
linear response layer. A separate task input layer encodes
the current task to be performed with respect to that stimulus
and projects to both the associative layer and response layer.
Units in the stimulus layer were grouped into six stimulus di-
mensions with three units per dimension. Similarly, units in
the response layer was grouped into six response dimensions
with three units per dimension. The network was trained on
12 tasks, where each task corresponds to a one-to-one map-
ping between a subset of three input features in a stimulus
layer to a subset of three response units in an output layer.
We then used the methods described in Musslick et al.
(2016) to extract a bipartite task graph from single represen-
tations encoded at the associative layer. The representations
associated with each task can be characterized by calculating,
for each unit in the associative and output layers, the mean of
its activity over all of the stimuli for a given task; this mean
pattern of activity can then be used as a representation of the
task. Correlating these patterns of activity across tasks yields
a task similarity matrix that can be examined separately for
the associative and output layers of the network. This can
then be used to assess the extent to which different tasks rely
on similar or different representation within each layer of the
network. Tasks that have similar representations over the as-
sociative layer can be inferred to rely on the same input di-
mension that is, they share an input component in the bipar-
tite graph representation of the network and tasks that are
similar at the output layer can be inferred to share an output
component. Accordingly, a bipartite graph can be constructed
by measuring the patterns of activity observed in the network
while it performs each individual task.
The extracted bipartite graph can be used to extract inter-
ference patterns between pairs of tasks (cf. Fig. 2). We use
this to extract all possible learned task-pairs involving no in-
terference case (Fig. 3a), and two distinct cases of interfer-
ence as shown in Fig. 3b and Fig. 3c from single-task repre-
sentations. Fig. 3 shows the activation patterns of the output
units for the simultaneous execution of two tasks, averaged
across the patterns of all task pairs for a given interference
structure. That is, no interference (Fig. 3a) leads to very ac-
curate response patterns (i.e. the current activation shown in
orange) is very close to the desired activation pattern shown
in grey). For the case in Fig. 3b, the response pattern of task
(S1 : O1) is primarily impaired due to the interference aris-
3Simulation details are omitted due to space constraints.
ing from task (S2 : O1). However, for the case of interfer-
ence illustrated in Fig. 3c, the response patterns for both tasks
(S1 : O1) and (S2 : O2) are impaired as observed by the activa-
tion patterns. These simulation results reflect the influence of
the directionality of interference between tasks as predicted
by the proposed interaction cost.
Figure 3: This figure illustrates the performance of a task-pair for a
given interference pattern. Each tasks maps a subset of three stimu-
lus input units onto three response units (see text). The orange color
in the bar plots indicates unit activation of response units relevant to
the depicted tasks, while gray indicates desired response pattern of
those units.
General Discussion and Conclusion
In this study, we have introduced two new measures to de-
termine costs associated with intensity and interaction for the
demand on control. First, we quantify the intensity cost as
a function of the amount of control bias that is supplemen-
tary to stimulus-specific processing in order to achieve a de-
sired response from the network. Doing so, we formalize
the probability of achieving a desired task given the stim-
ulus, weights and biases infused to the network. Since the
stimuli and weights are considered as network properties, the
intensity cost to achieve desired response is defined as the
amount (value) of control biases required to be injected to the
input and output components of the network. The detailed
analysis of intensity cost revealed an interesting optimization
problem to maximize the probability of surpassing a speci-
fied activation for a given budget of resources (i.e. an upper
bound on the control biases). The existence of a solution of
this optimization problem implicitly reveals whether the de-
sired objective is feasible. However, as it can be foreseen that
under certain circumstances the solution does not exist due to
interference between the involved processes. Such boundary
conditions motivated the second metric introduced in our pa-
per in which we formalize the interaction cost to measure the
level of interactions/interference between processes by means
of their type of connections and weights.
With the introduced characterization of intensity and inter-
action costs, it is possible to formally define whether a pro-
cess is considered a reflex, automatic or controlled. Con-
cretely, a process is considered a reflex if the underlying
weight guarantees a successful execution. In other words, a
reflex can be successfully executed without any intensity or
interaction costs. We assume that the execution of both con-
trolled and automatic processes carries with it an intensity
cost as some amount of control bias is needed to elicit a re-
sponse. However, unlike the former, controlled processes are
subject to interference and thus yield interaction costs large
than zero.
The metrics proposed here can also be used towards fur-
ther understanding cognitive effort as well as synthetic neu-
ral networks designed to achieve goal-driven tasks. By using
the intensity cost, which reveals the interrelationship between
control bias and probability of achieving a desired objective,
we will investigate the limitations of any given neural net-
work architecture by allocating a budget of control bias. The
intensity cost can also be used to investigate the feasibility of
achieving a desired objective defined by the set of processes
of interest in a network.
In the interaction cost analysis, we have assumed that there
exist a response in the output for any stimulus activation and
this may not be the case for a nonlinear activation in the out-
put components. Hence, one major research direction is the
detailed analysis for the classification of processes with non-
linear activation in output. Another possible direction for fu-
ture work is to further analyze the interaction cost in order
to capture the properties of the overall network (not only a
subset of tasks of interest). This will allow one to use the
interaction cost as an objective function for network train-
ing. Another possible direction is to explore the interrela-
tionship between intensity and interaction cost. In our work
(Musslick et al., 2016), we noticed a fundamental trade-off
between shared representations in a network and its parallel
processing capability (separated representations). Intuitively,
we envision this separation will decrease the interaction cost
while increasing the likelihood of successful execution for a
given budget of control bias.
Acknowledgements
We would like to thank Zahra Aminzare, Adam Charles,
Jonathan Pillow and Vaibhav Srivastava for their help in the
formalism of the problem.
References
Bogacz, R., Brown, E., Moehlis, J., Holmes, P., & Co-
hen, J. D.
(2006). The physics of optimal decision
making: A formal analysis of models of performance in
two-alternative forced-choice tasks. Psychological Review,
113(4), 700 -- 765.
Botvinick, M. M. (2007). Conflict monitoring and decision
making: Reconciling two perspectives on anterior cingu-
late function. Cognitive, Affective, & Behavioral Neuro-
science, 7(4), 356 -- 366.
Botvinick, M. M., & Braver, T. (2015). Motivation and Cog-
nitive Control: From Behavior to Neural Mechanism. An-
nual Review of Psychology, 66(1), 83 -- 113.
Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S.,
& Cohen, J. D. (2001). Conflict monitoring and cognitive
control. Psychological Review, 108(3), 624 -- 652.
Botvinick, M. M., & Cohen, J. D. (2014). The computational
and neural basis of cognitive control: Charted territory and
new frontiers. Cognitive Science, 38(6), 1249 -- 1285.
Botvinick, M. M., Cohen, J. D., & Carter, C. S. (2004). Con-
flict monitoring and anterior cingulate cortex: an update.
Trends in Cognitive Sciences, 8(12), 539 -- 546.
Cohen, J. D., Dunbar, K., & McClelland, J. L. (1990). On the
control of automatic processes: a parallel distributed pro-
cessing account of the stroop effect. Psychological Review,
97(3), 332 -- 361.
Feng, S. F., Schwemmer, M., Gershman, S. J., & Cohen, J. D.
(2014). Multitasking vs. Multiplexing: Toward a norma-
tive account of limitations in the simultaneous execution of
control-demanding behaviors. Cognitive, Affective, & Be-
havioral Neuroscience, 14(1), 129-146.
Koechlin, E., & Summerfield, C. (2007). An information the-
oretical approach to prefrontal executive function. Trends
in Cognitive Sciences, 11(6), 229 -- 235.
Kurzban, R., Duckworth, A., Kable, J. W., & Myers, J.
(2013). An opportunity cost model of subjective effort
and task performance. The Behavioral and Brain Sciences,
36(6), 661 -- 679.
Muraven, M., Tice, D. M., & Baumeister, R. F. (1998). Self-
control as a limited resource: Regulatory depletion pat-
terns. Journal of Personality and Social Psychology, 74(3),
774-789.
Musslick, S., Dey, B., Ozcimder, K., Patwary, M. M. A.,
Willke, T. L., & Cohen, J. D. (2016). Controlled vs. Au-
tomatic Processing: A Graph-Theoretic Approach to the
Analysis of Serial vs. Parallel Processing in Neural Net-
In Proceedings of the 38th Annual
work Architectures.
Conference of the Cognitive Science Society (pp. 1547 --
1552). Philadelphia, PA.
Ortega, P. A., & Braun, D. A. (2013). Thermodynamics as
a theory of decision-making with information-processing
Proceedings of Royal Society A, 469(2153),
costs.
20120683.
Posner, M. I., & Snyder, C. R. R.
(1975). Attention and
cognitive control. In R. L. Solso (Ed.), Information Pro-
cessing and Cognition: The Loyola Symposium (p. 55-85).
Lawrence Erlbaum.
Shenhav, A., Botvinick, M. M., & Cohen, J. D. (2013). The
expected value of control: an integrative theory of anterior
cingulate cortex function. Neuron, 79(2), 217 -- 240.
Shenhav, A., Musslick, S., Lieder, F., Kool, W., Griffiths,
T. L., Cohen, J. D., & Botvinick, M. M. (2017). Toward
a rational and mechanistic account of mental effort. (sub-
mitted to Annual Reviews of Neuroscience)
Shiffrin, R. M., & Schneider, W. (1977). Controlled and auto-
matic human information processing: II. Perceptual learn-
ing, automatic attending and a general theory. Psychologi-
cal Review, 84(2), 127-190.
|
1002.4835 | 2 | 1002 | 2010-11-06T11:59:27 | From modular to centralized organization of synchronization in functional areas of the cat cerebral cortex | [
"q-bio.NC",
"nlin.AO",
"physics.bio-ph"
] | Recent studies have pointed out the importance of transient synchronization between widely distributed neural assemblies to understand conscious perception. These neural assemblies form intricate networks of neurons and synapses whose detailed map for mammals is still unknown and far from our experimental capabilities. Only in a few cases, for example the C. elegans, we know the complete mapping of the neuronal tissue or its mesoscopic level of description provided by cortical areas. Here we study the process of transient and global synchronization using a simple model of phase-coupled oscillators assigned to cortical areas in the cerebral cat cortex. Our results highlight the impact of the topological connectivity in the developing of synchronization, revealing a transition in the synchronization organization that goes from a modular decentralized coherence to a centralized synchronized regime controlled by a few cortical areas forming a Rich-Club connectivity pattern. | q-bio.NC | q-bio | From modular to centralized organization of synchronization in
functional areas of the cat cerebral cortex
Jes´us G´omez-Gardenes1,2,∗, Gorka Zamora-L´opez3, Yamir Moreno1,4, Alex Arenas1,5
1 Instituto de Biocomputaci´on y F´ısica de Sistemas Complejos (BIFI), Universidad de
Zaragoza, 50009 Zaragoza, Spain
2 Departamento de Matem´atica Aplicada, Universidad Rey Juan Carlos (ESCET), 95123
M´ostoles (Madrid), Spain
3 Interdisciplinary Center for Dynamics of Complex Systems, University of Potsdam,
D-14476 Potsdam, Germany
4 Departamento de F´ısica Te´orica, Universidad de Zaragoza, 50009 Zaragoza, Spain
5 Departament d'Enginyeria Inform´atica y Matematiques, Universitat Rovira i Virgili,
43007 Tarragona, Spain
∗ E-mail: [email protected]
Abstract
Recent studies have pointed out the importance of transient synchronization between widely distributed
neural assemblies to understand conscious perception. These neural assemblies form intricate networks
of neurons and synapses whose detailed map for mammals is still unknown and far from our experimental
capabilities. Only in a few cases, for example the C. elegans, we know the complete mapping of the
neuronal tissue or its mesoscopic level of description provided by cortical areas. Here we study the
process of transient and global synchronization using a simple model of phase-coupled oscillators assigned
to cortical areas in the cerebral cat cortex. Our results highlight the impact of the topological connectivity
in the developing of synchronization, revealing a transition in the synchronization organization that goes
from a modular decentralized coherence to a centralized synchronized regime controlled by a few cortical
areas forming a Rich-Club connectivity pattern.
0
1
0
2
v
o
N
6
]
.
C
N
o
i
b
-
q
[
2
v
5
3
8
4
.
2
0
0
1
:
v
i
X
r
a
2
Introduction
Processing of information within the nervous system follows different strategies and time-scales. Particu-
lar attributes of the sensory stimuli are transduced into electrical signals of different characteristics, e.g.
regular or irregular spiking, the rate of firing, etc. Further aspects of the information are "encoded" by
specialization of neurons, e.g. the color and orientation of a visual stimulus will activate only a set of
neurons and leave others silent. For higher order processes such as feature binding and association, the
synchronization between neural assemblies plays a crucial role [1 -- 5]. For example, subliminal stimulation
which is not consciously perceived, triggers a similar cascade of activation in the sensory system but fails
to elicit a transient synchronization between distant cortical regions [6].
The neurons comprising the nervous system form a complex network of communications. To what
extent this intricate architecture supports the richness and complexity of the ongoing dynamical activity
in the brain is a fundamental question [7]. A detailed map of the neurons and their synapses in mammals
is still unknown and far from our experimental capabilities. Only in a few cases, for example the nematode
C. elegans, we know the complete mapping of the neuronal tissue. In the cases of macaque monkeys and
cats a mesoscopic level of description is known, composed of cortical areas and the axonal projections
between them. These areas are arranged into modules which closely follow functional subdivisions by
modality [8 -- 13]. Two cortical areas are more likely connected if both are involved in the processing of
the same modal information (visual, auditory, etc.) Beyond this modular organization, some cortical
areas are extensively connected (referred as hubs) with projections to areas in all modalities [14, 15].
For the corticocortical network of cats, these hubs are found to be densely interconnected forming a
hidden module [16], at the top of the cortical hierarchy which might be responsible for the integration
of multisensory information. A core of cortical areas has also been detected in estimates of human
corticocortical connectivity by Diffusion Spectrum Imaging [17]. Following the above discussion that
synchronization plays a major role in the processing of high level information, it would be important
to analyze the synchronization behavior of these networks in relation to their modular and hierarchical
organization. To this end, we simulate the corticocortical network of the cat by non-identical phase
oscillators and we follow the evolution of their synchronization from local to global.
The study of synchronization phenomena is a useful tool to analyze the substrate of complex networks.
The dynamical patterns under different parameters unveil features of the underlying microscopic and
3
mesoscopic organization [18].
In particular, recent studies highlight the impact that the topological
properties such as the degree heterogeneity, the small-world effect and the modular structure have on the
path followed from local to global synchronization [18 -- 21].
In this work we study the routes to synchronization in the corticocortical network of cats brain (see
Figure 1) by modelling each cortical area as a phase oscillator with an independent internal frequency.
This assumption considers that the ensemble of neurons contained within a cortical area behaves coher-
ently having a well defined phase average whose dynamics is described by the internal frequency [22]. The
coupling between areas is modelled using the Kuramoto nonlinear coupling and its relative strength can
be conveniently tuned to allow for the observation of synchronization at different scales of organization.
This seemingly crude approximation allows to obtain similar synchronization patterns as those observed
with more refined models based on neuronal ensembles placed at each cortical area [23 -- 25]. For instance,
using the Kuramoto model one obtains that highly connected areas promote synchronization of neural
activity just as revealed by the more stylized model used for the dentate gyrus [26].
Our results point out that complex structures of highly connected areas play a key role in the synchro-
nization transition. In contrast to the usual partition of the brain cortex into four sets of modally-related
areas, we find that this modular organization plays a secondary role in the emergence of synchroniza-
tion patterns. On the contrary, we unveil that a new module made up of highly connected areas (not
necessarily modally-related) drives the dynamical organization of the system. This new set is seen as
the Rich-Club of the corticocortical network. Surprisingly, the new partition of the network including
the Rich-Club as a module preserves the modular behavior of the system's dynamics at low intercortical
coupling strengths. This modular behavior transforms into a centralized one (driven by the Rich-Club)
at the onset of global synchronization highlighting the plasticity of the network to perform specialized
(modular) or integration (global) tasks depending on the coupling scale.
Results
As introduced above, we describe the dynamical behavior of the cortical network using the Kuramoto
model [27], where the time evolution of the phase of each cortical area, θi(t), is given by
θi = ωi + λ
N
Xj=1
Wij sin(θj − θi),
(1)
4
where ωi is the internal frequency associated to area i and Wij is the weighted inter-cortical coupling
matrix that takes a value 0 if area j does not interact with the dynamics of area i while Wij > 0 otherwise.
In this latter case Wij can take values 1, 2 or 3 depending on the axon density going from area j to area
i. Let us remark that the inter-cortical coupling matrix is not symmetric so that in general Wij 6= Wji.
Besides, the inter-cortical dynamical coupling is modeled as the sine of the phase differences between two
connected areas such that when θj > θi the average phase of area i accelerates while that of area j slows
down to approach each other. Finally, the parameter λ accounts for the strength of the inter-cortical
coupling.
In a system composed of all-to-all coupled oscillators, the Kuramoto model shows a transition from
incoherent dynamics to a synchronized regime as λ increases [28, 29]. However, when the system has a
nontrivial underlying structure this transition does not take place in an homogeneous manner. In complex
topologies, and for moderate coupling values, certain parts of the system become synchronized rather fast
whereas other regions still behave incoherently. Therefore, one can monitor the synchronization patterns
that appear as the coupling λ is increased and describe the path to synchronization accurately [18]
by reconstructing the synchronized subgraph composed of those nodes and links that share the largest
degree of synchronization (see Materials and Methods). The study of these synchronization clusters as the
coupling λ is increased allow to unveil the important set of nodes that drives these dynamical processes
in the system.
We will analyze different scales of organization: i) the macroscopic scale referring to global synchro-
nization of the network; ii) the microscopic scale of organization corresponding to the individual state
of the oscillators; and finally iii) the intermediate mesoscopic scale of dynamical organization between
the macroscopic and microscopic scales. Usually, it consists of groups of nodes classified according to a
certain additional information, for example that provided by the anatomico-functional modules. Every
scale of organization requires a particular set of statistical descriptors. This is especially important in
the mesoscopic scale where changing the groups, the characterization of the system is also changed.
Macroscopic analysis
We start by describing how global synchronization is attained as the inter-cortical coupling λ is increased.
Global synchronization is characterized by the usual Kuramoto order parameter, r, and the fraction of
links that are synchronized rlink [20,21] (see Materials and Methods). Both parameters take values in the
5
region [0, 1], being close to 0 when no dynamical coherence is observed and close to 1 when the system
approaches to full synchronization.
In Fig. 2 we show the evolution of r and rlink as a function of λ. The plot reveals a well defined
transition from incoherent to globally synchronized, the onset of synchronization occurring for coupling
strength 0.011 ≤ λ ≤ 0.021. When λ ≃ 0.2, the system reaches the fully synchronized state. In the
following, we will explore this transition in more detail and at lower scales of dynamical organization.
Mesoscopic analysis as described by the four anatomical modules
The measures r and rlink describe completely the dynamical state of the system if one assumes that all
the cortical areas behave identically. However, it is possible to extract more information about the local
dynamical properties of the system. In particular, for a given value of λ we can monitor the degree of
synchronization between two given areas i and j, rij ∈ [0, 1] (see Materials and Methods).
The studies of the transition to synchronization in modular architectures [19,21] show that synchrony
patterns appear first at internal modules, i.e. synchrony shows up among the nodes that belong to the
same module due to a larger local density within the module and similar pattern of inputs of the nodes.
As the coupling λ is increased, synchrony starts to affect the links connecting nodes of different clusters
and finally spreads to the entire system. Now, we analyze whether the four anatomico-functional modules
of the corticocortical network of the cat act also as dynamical clusters in the synchronization transition.
To this end, we have analyzed the evolution of the average synchronization within and between the
four anatomical modules taking into account solely the information about the dynamical coherence rij
between the network's areas. We define the average synchronization between module α and module β as:
rαβ =
1
Lαβ Xi∈α, j∈β
rij ,
(2)
where Lαβ is the number of possible pairs of areas from modules α and β, i.e., Lαβ = NαNβ/2 where
Nα and Nβ are the number of areas of modules α and β respectively. If α = β equation (2) denotes the
intramodule average synchronization.
The histograms in the left columns of Figure 3 and Figure 4 show the values of the set {rαβ} for several
values of λ corresponding to the region before (Fig. 3) and at (Fig. 4) the onset of synchronization. From
the histograms it is clear that the degree of synchronization grows with λ as it occurs for the global
6
parameters r and rlink in the macroscopic description. Besides, the histograms inform us about the
importance of the anatomical partition in the dynamical organization of the cat cortex. From Figure 3 it
becomes clear that the average dynamical correlation within areas of the same anatomical module is higher
than that between areas belonging to different modules. Moreover, before the onset of synchronization,
for λ = 0.007 to λ = 0.11 all the modules satisfy rαα ≥ rαβ for α 6= β. At the onset of synchronization (left
column of Figure 4) we observe that the initial intra-module synchronization is progressively compensated
by the increase of inter-module dynamical coherence. In particular, the influence of the Somatosensory-
Motor on the remaining modules is remarkably relevant during the onset of synchronization and, for
λ > 0.021, this module shows the largest degree of synchronization with the rest of the system.
Microscopic analysis: Unveiling the dynamical organization
The mesoscopic analysis based on the partition of the cortex into four modules has revealed a fingerprint
of a hierarchical organization of the synchronization based on the dominating role of the Somatosensory-
Motor module. Here we will analyze the microscopic correlation between all the areas of the cortical
network to unveil whether there is a group of nodes that lead the onset of synchronization in the system.
To this purpose we study the subgraphs formed by those pairs of areas sharing an average synchronization
value rij larger than a threshold T . Certainly when T = 1 the subgraph is the null (empty) graph and
for T = 0 the subgraph is the whole cortex.
In Figure 5 we show a ranking of the cortical areas at coupling strengths λ = 0.015, 0.017, 0.019 and
0.021 corresponding to the onset of synchronization. The rankings are made by labeling the area i with
the largest value of the threshold at which the area is incorporated into the synchronized subgraph as T
is tuned from 1 to 0. Additionally, the modular origin of the areas has been color coded to distinguish
the role of each module. From the rankings we find that there are three areas 36, 35 and Ig, from the
Fronto-Limbic system, that share the largest degree of synchrony. In all the cases, several jumps in the
threshold are observed that distinguishes those groups of cortical areas that are more synchronized than
the rest of the network. For instance, at λ = 0.017 we observe 14 areas spanning from the 36 area to the
2 (Somatosensory-Motor system) while for λ = 0.021 we find 16 areas ranging from the 36 to the area
4 (Somatosensory-Motor system). From these figures it is no clear that there is one module dominating
the synchronization. Quite on the contrary both Somatosensory-Motor and Fronto-Limbic systems are
well represented among the most synchronized areas.
7
A further analysis of the composition of these highly synchronized areas reveals that most of them
take part in a higher-order topological structure of the cortical network: a Rich-Club (see Materials
and Methods). The Rich-Club of a given network is made up of a set of nodes with high connectivity,
which at the same time, form a tightly interconnected community [30, 31]. Therefore, the Rich-Club of
a network can be described as a highly cohesive set of hubs, that form a dominant community in the
hierarchical organization. The Rich-Club of the cat cortex is composed of 11 cortical areas of different
modalities: 3 visual areas (20a, 7 and AES), 1 area from the Auditory system (EPp), 2 areas of the
Somatosensory-Motor system (6m and 5Al) and 5 fronto-limbic areas (Ia, Ig, CGp, 35 and 36). In Figure
6 we show again the ranking of areas for the cases λ = 0.015, 0.017, 0.019 and 0.021 but highlighting those
areas belonging to the Rich-Club in black. From the plots it is clear that most of the Rich-Club areas
are largely synchronized. In particular for λ = 0.019 and 0.021 the 8 out of the 10 largest synchronized
areas of the network belong to the Rich-Club, although originally they belong to the Fronto-Limbic,
Somatosensory-Motor and the Visual systems in the partition into four modules.
Mesoscopic analysis of synchronization including the Rich-Club
Looking at the composition of the Rich-Club we observe that it is mainly composed of fronto-limbic
areas. Taking into account that we previously observed how the Somatosensory-Motor system took the
leading role within the description with 4 modules of the synchronization transition, this dominance of the
Fronto-Limbic system in the Rich-Club may seem counterintuitive. To test the role of the Rich-Club in
the synchronization transition we define a new partition of the cortical network into 5 clusters composed
of the Rich-Club (as defined above) and the 4 original modules, but with the corresponding areas of the
Rich-Club removed from them.
At the mesoscopic scale, we investigate the self-correlation of the new five clusters and their cross
correlation according to Equation (2).
In the right columns of Figure 3 and Figure 4 we present the
histograms of the inter and intra correlations for different values of the coupling. In both figures the
role of the Rich-Club orchestrating the process towards synchrony while increasing the coupling strength
becomes clear. More importantly, the addition of the Rich-Club to the partition helps to elucidate
the patterns of synchrony: both the dynamical self-correlation of the four modally-related clusters and
their correlation with the Rich-Club remain large.
In particular, we observe in Figure 3 that, before
the onset of synchronization, these new five modules keep a large self-correlation during this stage. On
8
the other hand, at the onset of synchronization (Figure 4) the four modally-related clusters loose their
large self-correlation in the following sequence: The "Fronto-Limbic" cluster remains autocorrelated until
λ = 0.013, the "Visual" one until λ = 0.017, the "Auditory" system until 0.019 and the "Somatosensory-
Motor" cluster until λ = 0.021. For larger couplings, all the clusters switch from autocorrelation to be
synchronized with the Rich-Club, which acts as a physical mean-field of the system. Moreover, during the
whole synchronization path the Rich-Club is always the cluster with the largest self-correlation. Thus, the
distinction of Rich-Club in the partition preserves the autocorrelation of the four modal clusters before
the onset of synchronization while, at the same time, rules the path to complete dynamical coherence
during the onset of synchronization. This two-stage dynamics (modal cluster synchronization followed by
a sequential synchronization with the Rich-Club) supports the idea that the modular organization with
a centralized hierarchy described in [16] facilitates the segregation and integration of information in the
cortex.
Characterization of the transition from modular to centralized synchronization
The results so far indicate a plausible transition from modular to centralized organization in the cortex,
depending on the coupling strength. In particular, we have shown patterns of synchronization that change
the behavior while increasing the coupling λ. Now we propose a characterization of this change in terms
of statistical descriptors. To this end, we define two different measures: (i) the dynamical modularity
(DM) and (ii) the dynamical centralization (DC). The dynamical modularity compares the degree of
internal synchrony within the clusters with the average dynamical correlation across clusters. With this
aim we define the DM as the fraction of the average self-correlation of clusters and the average intercluster
cross-correlation. For a network composed of m clusters we have:
DM =
Pα rαα/m
Pα,β6=α rαβ/[m(m − 1)]
.
(3)
The DM will take values above 1 when the system contains true dynamical clusters while DM < 1
means that the entitity of the partition is not consistent with a clustered behavior. On the other hand,
the dynamical centralization of the network measures the relative difference in synchrony between the
maximum among the m clusters of rα = Pβ rαβ/m and the average degree of synchrony over clusters,
hrαi = Pα rα/m:
DC =
maxα{rα}− < rα >
< rα >
.
9
(4)
In the case of the DC we always obtain positive values. A large value of DC means that the system
displays a highly centralized dynamical behavior around a leading cluster while we will obtain DC values
approaching to 0 when the system behaves homogeneously, i.e. when there is no leading cluster that
centralizes the dynamics.
We have measured both DM and DC for the original partition into m = 4 modules and the new
partition with m = 5 incorporating the Rich-Club. In Figure 7 we show the evolution of the two quantities
as a function of the coupling parameter. For the case of the DM we confirm that the partition with the
Rich-Club keeps the modular behavior of the original partition along the whole synchronization path. The
DM is remarkably high for low values of the coupling λ pointing out that before the synchronization onset
the internal synchronization dominates over the cross-correlation between the clusters. Regarding the
DC we find that in both partitions DC increases with λ reaching a maximum around the synchronization
onset, signaling that at this point the synchronization is driven hierarchically and lead by one of the
modules. However, the partition that incorporates the Rich-Club shows the remarkably larger values of
DC along the whole path, specially around the synchronization onset. In particular, the dominant role
of the Rich-Club is clearly highlighted by the maximum of the DC at λ = 0.015, just at the onset of
synchronization. In order to verify that the Rich-Club is the cluster contributing to the term maxα{rα}
in the dynamical centrality of the network we plot in Figure 7 the evolution of the DC considering each
of the 5 modules as the the central cluster by substituting maxα{rα} by the corresponding value of rα.
From the plot it is clear that the Rich-Club is the central cluster orchestrating the dynamics of the system
at the onset of synchronization.
The coupled evolution of DM and DC corroborates the two-mode operation of the cortical network
when described with the Rich-Club and the remaining parts of the four original modules: At low values
of the coupling, the modular structure of the network dominates the synchronization dynamics, pointing
out the capacity to concentrate sensory stimuli within its corresponding module. When the coupling is
increased the dynamical organization is driven by a leading subset of nodes, organized in a topological
Rich-Club, that integrates information between different regions of the cortex.
10
Discussion
Previous simulations performed in the cat cortical network [23 -- 25] have dealt with its synchronization
properties.
In these works, the transition towards synchronization is studied by using ensembles of
neurons coupled through a small-world topology placed inside each cortical area whereas different neuronal
populations are dynamically coupled accordingly to the topology of the cat cortical network. By means
of this two-level dynamical model, numerical simulations allowed to find different clusters of synchrony
as the coupling between the cortical areas is increased. It was found that only for weak coupling these
clusters were closely related to the four modal clusters. In the light of these previous studies, and the
recent report of a novel modular and hierarchical organization of the corticocortical connectivity [16],
the issue regarding the relation between the mesoscopic structure of the cat cortex and its dynamical
organization remains open.
Here, we have investigated the evolution of synchronization in a network representing the actual con-
nectivity among cortical areas in the cat's brain. We have confirmed, that the role of the different areas
in the path towards synchrony is difficult to assess using the traditional partition into four groups of
modally-related areas. On the contrary, we have shown that a subset of areas, forming a topological
Rich-Club, orchestrates this process. The distinction of this subset permits the interpretation of a new
mesoscale formed by the four modules, excluding some nodes that form the Rich-Club, which are con-
sidered here as the fifth module. This proposed structure allows us to reveal a transition in the path
to synchronization as a function of the coupling strength, that seems to indicate a two-mode operation
strategy. For low values of the coupling, a state of weak internal coherence within the five modules
governs the coordination dynamics of the network. As the coupling strength is increased, the Rich-Club
becomes the responsible of centralizing the network dynamics and leads the transition towards global
synchronization.
Finally, the composition of the Rich-Club allows to make some additional biologically relevant ob-
servations. First, the Rich-Club comprises of cortical areas of the four different modalities, supporting
the hypothesis of distributed coordination dynamics at the highest levels of cortical processing such as
integration of multisensory information [32]. Second, the Rich-Club comprises of most of the frontal
areas in the Fronto-Limbic module. Moreover, the areas of the Rich-Club collected from the original
Somatosensory-Motor system contain the so-called supplementary motor area (SMA). The SMA is a
11
controversial region of the motor cortex, since in contrast with the rest of somatosensory-motor areas
it is in charge of the initiation of planned or programmed movements [33]. Furthermore, the area AES
of the Rich-Club, originally assigned to the Visual module, is believed to integrate all visual and even
auditory signals for their multimodal processing and transference as coherent communication signals [34].
Summing up, the Rich-Club is basically made up of areas involved in higher cognitive tasks devoted to
planning and integration. The prominent role of the aforementioned regions in the cortex activity is
unveiled from our network perspective in terms of a Rich-Club leading the path to synchronization. Our
proposal, after this observation, is to investigate the evolution of synchronization in the cat cortex by
tracking the transient of five modules corresponding to the anatomico-functional areas (S-M, F-L, Aud,
Vis) and the Rich-Club as a separate (but interrelated) functional entities.
Materials and Methods
Cortico-cortical network of cats' brain
After an extensive collation of literature reporting anatomical tract-tracing experiments, Scannell and
Young [8,9] published a dataset containing the corticocortical and cortico-thalamical projections between
regions of one brain hemisphere in cats. The connections were weighted according to the axonal density
of the projections. Connections originally reported as weak or sparse were classified with 1 and, the
connections originally reported as strong or dense with 3. The connections reported as intermediate
strength, as well as those connections for which no strength information was available, were classified
with 2, see Figure 1(b). Here we make use of a version of the network [10] consisting of N = 53 cortical
areas interconnected by L = 826 directed corticocortical projections.
Rich-Club areas
A key factor of the hierarchical organization of the corticocortical network of the cat is that the hub
areas (those with the largest number of projections) are very densely connected between them [16]. The
Rich-Club phenomenon [30, 31] is characterised by the growth of the internal density of links between all
nodes with degree larger than a given k′, referred as k-density, φ(k′):
φ(k′) =
Lk′
Nk′ (Nk′ − 1)
,
12
(5)
where Nk′ is the number of nodes with k(v) ≥ k′ and Lk′ is the number of links between them. As φ(k)
is an increasing function of k, a conclusive interpretation requires the comparison with random surrogate
networks with the same degree distribution. The question is then whether φ(k) of the real network grows
faster or slower with k than the expected k-density of the surrogate networks. If φ(k) grows slower, it
means that the hubs are more independent of each other than expected.
In our case, the network is directed but the input degree kin(v) and the output degree kout(v) of the
areas are highly correlated. Hence, we compute the k-density of the corticocortical network of the cat,
φcat(k), considering the degree of every area as: k(v) = 1
2 (kin(v) + kout(v)). The result is presented in
Figure 8 together with the ensemble average φ1n(k) of 100 surrogate networks. At low degrees φcat(k)
follows very close the expectation, but for degrees k > 13, φcat(k) starts to grow faster. The largest
difference occurs at k = 22, comprising of a set of eleven cortical hubs of the four modalities: visual areas
20a, 7 and AES; auditory area EPp, somatosensory-motor areas 6m and 5Al; and fronto-limbic areas Ia,
Ig, CGp, CGa, 35 and 36.
Numerical simulation details
We integrate the Kuramoto equations, see equation (1), using a fourth order Runge-Kutta method with
time step δt = 10−2. The system is set up by randomly assigning the initial conditions {θi(0)} and the
internal frequencies {ωi} randomly in the intervals [−π, π] and [−1/2, 1/2] respectively. The integration
of the Kuramoto is performed for a total time T = 700. After a transient time of τ = 300 we start the
computation of the different dynamical measures such as the order parameters r and rlink.
13
Synchronization order parameters
The dynamical coherence of the population of N oscillators (areas) is measured by means of two different
order parameters r and rlink. The first one is obtained from a complex number z(t) defined as follows:
z(t) = r(t) exp [iφ(t)] =
N
Xj=1
exp [iθj(t)] .
(6)
The modulus of z(t), r(t), measures the phase coherence of the population while φ(t) is the average phase
of the population of oscillators. Averaging over time the value of r(t) we obtain the order parameter
r = hr(t)i.
The second order parameter, rlink, is measured looking at the local synchronization patterns, allowing
for the exploration of how global synchronization is attained. We define rlink by measuring the degree of
synchrony between two connected areas i and j:
Cij = lim
∆t→∞(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
Aij
∆t Z τ +∆t
τ
ei[θi(t)−θj (t)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,
(7)
where Aij is the adjacency matrix of the network, being Aij = 1 when Wij > 0 and Aij = 0 otherwise.
Each of the values {Cij} are bounded in the interval [0, 1], being Cij = 1 when the connected areas i
and j are fully synchronized and rij = 0 when these areas are dynamically uncorrelated. Note that for
a correct computation of Cij the averaging time ∆t should be taken large enough (in our computations
∆t = 400) in order to obtain good measures of the degree of coherence between each pair of areas. Since
Cij = 0 for the areas that are not physically connected we construct the N × N matrix C and define the
global order parameter rlink as follows:
rlink =
1
L Xi,j
Cij .
(8)
Therefore, the parameter rlink measures the fraction of all possible links that are synchronized in the
network.
In the more general case in which all possible pairs of areas are taken into account to compute the
average synchronization between cortical areas, Eq.(7) and Eq.(8) can be rewritten as:
C ∗
ij = lim
∆t→∞(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
1
∆t Z τ +∆t
τ
ei[θi(t)−θj(t)](cid:12)(cid:12)(cid:12)(cid:12)(cid:12)
,
and
r∗
link =
2
N (N − 1) Xi,j
C ∗
ij ,
14
(9)
(10)
respectively. Note that C ∗
ij and r∗
link account for the degree of synchronization between areas i and j
regardless of whether or not they are connected.
Defining the average synchronization between areas
To label two areas i and j as synchronized or not one has to analyze the matrix C ∗ and construct a
filtered matrix F whose elements are either Fij = 1 if i and j are considered as synchronized or Fij = 0
otherwise. From the computation of r∗
link, equation (10), one knows the fraction of all possible pairs
of areas that are synchronized. Therefore, one would expect that N (N − 1) · r∗
link/2 elements of the
matrix F have Fij = 1, while the remaining elements are Fij = 0. The former elements correspond to
the N (N − 1) · r∗
link/2 pairs with the largest values of C ∗
ij.
In order to measure the average degree of synchronization between pairs of areas one have to average
over different n realizations using different initial conditions {θi(0)} and different internal frequencies
{ωi} (typically we have used n = 5 · 103 different realizations for each value of λ studied). To this purpose
we average the set of filtered matrices {F l} (l = 1,...,n) of the different realizations to obtain the average
degree of synchronization between areas:
rij =
1
n
n
Xl=1
F l
ij .
(11)
In this way the value for rij ∈ [0, 1] accounts for the probability that areas i and j are considered as
synchronized.
Acknowledgments
The authors are grateful to Jes´us G´omez-Tol´on for useful discussions and suggestions that helped to
15
improve the manuscript.
References
1. Engel AK, Singer W (2001) Temporal binding and the neural correlates of sensory awareness.
Trends Cogn. Sci. 5: 16 -- 25.
2. Engel AK, Fries P, Singer W (2001) Rapid feature selective neuronal synchronization through
correlated latency shifting. Nat. Rev. Neurosc. 2: 704 -- 716.
3. Fahle M (1993) Figure -- Ground Discrimination from Temporal Information. Proc. R. Soc. Lond. B
254: 199 -- 203.
4. Singer W, Gray CM (1995) Visual feature integration and the temporal correlation hypothesis.
Ann. Rev. Neurosci. 18: 555 -- 586.
5. Ulhaas PJ, Pipa G, Lima B, Melloni L, Neuenschwander S et al. (2009) Neural synchrony in cortical
networks: history, concept and current status. Frontiers Int. Neurosc. 3, 17.
6. Melloni L, Molina C, Pena M, Torres D, Singer W, et al. (2007) Synchronization of neural activity
across cortical areas correlates with conscious perception. J. Neurosc 27: 2858 -- 2865.
7. Boccaletti S., Latora V. and Moreno Y.(eds.) (2009), Handbook on Biological Networks, Singapore:
World Scientific.
8. Scannell JW, Blakemore CW, Young MP (1995) Analysis of connectivity in the cat cerebral cortex.
J. Neurosc. 15: 1463 -- 1483.
9. Scannell JW, Burns GAPC, Hilgetag CC, O'Neill MA, Young MP (1999) The connectional orga-
nization of the cortico-thalamic system of the cat. Cer. Cortex 9: 277 -- 299.
10. Hilgetag CC, Burns GAPC, O'Neill MA, Scannell JW, Young MP (2000) Anatomical connectivity
defines the organization of clusters of cortical areas in the macaque monkey and the cat. Phil.
Trans. R. Soc. London B 355: 91 -- 110.
16
11. Hilgetag CC, O'Neill MA, Young MP (2000) Hierarchical organization of macaque and cat cortical
sensory systems explored with a novel network processor. Phil. Trans. R. Soc. London B 355:
71 -- 89.
12. Sporns O, Chialvo DR, Kaiser M, Hilgetag CC (2004) Organization, development and function of
complex brain networks. Trends Cogn. Sci. 8: 418 -- 425.
13. Hilgetag CC, Kaiser M (2004) Clustered organization of cortical connectivity. Neuroinformatics 2:
353 -- 360.
14. Zamora-L´opez G, Zhou CS, Kurths J (2009) Graphs analysis of cortical networks reveals complex
anatomical communication substrate. Chaos 19: 015117.
15. Sporns O, Honey CJ, Kotter R (2007) Identification and classification of hubs in brain networks.
PLoS ONE 10: e1049.
16. Zamora-L´opez G., Zhou CS, Kurths J (2010) Cortical hubs form a module for multisensory inte-
gration on top of the hierarchy of cortical networks. Frontiers Neuroinf. in press.
17. Hagmann P, Cammoun L, Gigandet X, Meuli R, Honey CJ et al. (2008) Mapping the Structural
Core of Human Cerebral Cortex. PLoS Biol 6(7): e159.
18. Arenas A, D´ıaz-Guilera A, Kurths J, Moreno Y, Zhou CS (2008) Synchronization in complex
networks. Phys. Rep. 469: 93 -- 153.
19. Arenas A, D´ıaz-Guilera A, P´erez-Vicente CJ (2006) Synchronization reveals topological scales in
complex networks. Phys. Rev. Lett. 96: 114102.
20. G´omez-Gardenes J, Moreno Y,Arenas A (2007) Paths to synchronization in complex networks.
Phys. Rev. Lett.98: 034101.
21. G´omez-Gardenes J, Moreno Y, Arenas A (2007) Synchronizability determined by coupling strengths
in complex networks. Phys. Rev. E 75: 066106.
22. Rulkov NF (2001) Regularization of synchronized chaotic burst. Phys. Rev. Lett. 86: 183.
23. Zemanova L, Zhou CS, Kurths J (2006) Structural and functional clusters of complex brain net-
works. Physica D 224: 202 -- 212.
17
24. Zhou CS, Zemanova L, Zamora G, Hilgetag CC, Kurths J (2006) Hierarchical organization unveiled
by functional connectivity in complex brain networks. Phys. Rev. Lett. 97: 238103.
25. Zhou CS, Zemanova L, Zamora G, Hilgetag CC, Kurths J (2007) Structure-function relationship
in complex brain networks expressed by hierarchical synchronization. New J. Phys. 9: 178.
26. Morgan RJ, Soltesz I (2008), Nonrandom connectivity of the epileptic dentate gyrus predicts a
major role for neuronal hubs in seizures. Proc. Natl. Acad. Sci. (USA) 105: 6179 -- 6184.
27. Kuramoto Y (1984) Cooperative Dynamics of Oscillator Community. Prog. Theor. Phys. 79: 223 --
240.
28. Strogatz SH (2000) From Kuramoto to Crawford: exploring the onset of synchronization in popu-
lations of coupled oscillators. Physica D 143: 1 -- 20.
29. Acebron JA, Bonilla LL, Perez Vicente CJ, Ritort F, Spigler R (2005) The Kuramoto model: A
simple paradigm for synchronization phenomena. Rev. Mod. Phys. 77: 137 -- 185.
30. Zhou S and Modrag´on RJ (2004) The rich-club phenomenon in the internet topology. IEEE Comm.
Lett. 8(3): 80 -- 182.
31. Colizza V, Flammini A, Serrano MA, Vespignani A (2006) Detecting Rich-Club ordering in complex
networks. Nature Phys. 2: 110-115.
32. Bressler SL and Kelso JAS (2001) Cortical coordination dynamics and cognition. Trends Cogn.
Sci. 5, 26 -- 36.
33. Nachev P, Kennard Ch, Husain M (2008) Functional role of the supplementary and pre-
supplementary motor areas. Nature Rev. Neuroscience 9: 856 -- 869.
34. Stein BE, Stanford TR (2008) Multisensory integration: current issues from the perspective of the
single neuron. Nature Rev. Neuroscience 9: 255 -- 266.
Figure Legends
18
Figure 1. The brain cortical network of the cat. On the left we show the topology of the nodes
(areas) and links (axon interconnections) between them. On the right the weighted adjacency matrix is
shown. The weight of the links denote the axon density between two connected areas. Besides the
matrix shows the partition of the network into four main modules of modally-related areas: Visual,
Auditory, Somatosensory-Motor and Fronto-Limbic.
Figure 2. Synchronization diagrams. The figure shows the evolution of the Kuramoto order
parameter r and the fraction of synchronized links rlink as the coupling strenght is increased. The
transition from asynchronous dynamics to global dynamical coherence as λ grows is clear from the two
curves. The region in blue corresponds to the onset of synchronization.
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
λ=0.007
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
λ=0.009
Vis.
A
u
d.
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
S
.-
M
.
F
.-
L.
λ=0.011
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
19
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
λ=0.007
Vis.
A
u
d.
S
.-
M
.
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
Vis.
A
u
d.
S
.-
M
.
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
R
.-
C
.
λ=0.009
F
.-
L.
R
.-
C
.
λ=0.011
F
.-
L.
R
.-
C
.
Figure 3. Dynamical correlation within the 4 modal clusters (left column) and the new 4
modally-related clusters and the Rich-Club (right column) before the onset of
synchronization. The bars of the histograms show the values of the dynamical correlation rαβ (see
Equation (2)) between the 4 original modules (left) and the new 4 clusters and the Rich-Club (right).
From top to bottom we show the cases for λ = 0.007, 0.009 and 0.011 that correspond to the region
before the onset of synchronization.
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
λ=0.013
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
λ=0.015
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
λ=0.017
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
Visual
Auditory
Somatosensory-Motor
Fronto-limbic
λ=0.019
Vis.
A
u
d.
S
.-
M
.
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
λ=0.021
F
.-
L.
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
20
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
λ=0.013
Vis.
A
u
d.
S
.-
M
.
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
Vis.
A
u
d.
S
.-
M
.
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
Vis.
A
u
d.
S
.-
M
.
Visual
Auditory
Somato-motor
Fronto-limbic
Rich-club
Vis.
A
u
d.
S
.-
M
.
λ=0.021
Vis.
A
u
d.
S
.-
M
.
F
.-
L.
R
.-
C
.
λ=0.015
F
.-
L.
R
.-
C
.
λ=0.017
F
.-
L.
R
.-
C
.
λ=0.019
F
.-
L.
F
.-
L.
R
.-
C
.
R
.-
C
.
Figure 4. Dynamical correlation within the 4 modal clusters (left column) and the 5
dynamical clusters (right column) at the onset of synchronization. The bars of the histograms
show the values of the dynamical correlation rαβ (see Equation (2)) between the 4 modules (left) and
c
o
l
o
u
r
o
f
e
a
c
h
b
a
r
a
c
c
o
u
n
t
s
f
o
r
t
h
e
c
o
r
r
e
s
p
o
n
d
n
g
m
o
d
u
l
e
i
o
f
t
h
e
c
o
r
t
i
c
a
l
a
r
e
a
.
v
a
l
u
e
o
f
t
h
e
t
h
r
e
s
h
o
l
d
,
T
i
,
a
t
w
h
i
c
h
t
h
e
a
r
e
a
i
s
i
n
c
o
r
p
o
r
a
t
e
d
i
n
t
h
e
s
y
n
c
h
r
o
n
i
z
e
d
s
u
b
g
r
a
p
h
.
B
e
s
i
d
e
s
,
t
h
e
F
i
g
u
r
e
5
.
S
y
n
c
h
r
o
n
y
R
a
n
k
o
f
a
r
e
a
s
:
U
n
v
e
i
l
i
n
g
a
n
a
t
o
m
i
c
a
l
s
t
r
u
c
t
u
r
e
o
f
t
h
e
l
a
r
g
e
s
t
f
o
r
λ
=
0
.
0
1
5
(
a
)
,
0
.
0
1
7
(
b
)
,
0
.
0
1
9
(
c
)
a
n
d
0
.
0
2
1
(
d
)
.
T
h
e
h
e
i
g
h
t
o
f
t
h
e
b
a
r
s
a
c
c
o
u
n
t
o
f
t
h
e
m
a
x
i
m
u
m
s
y
n
c
h
r
o
n
i
z
e
d
a
r
e
a
s
.
I
n
t
h
e
s
e
p
l
o
t
s
w
e
s
h
o
w
t
h
e
r
a
n
k
o
f
a
r
e
a
s
f
r
o
m
t
h
e
m
o
s
t
t
o
t
h
e
l
e
s
s
s
y
n
c
h
r
o
n
i
z
e
d
36
35
Ig
AES
CGp
6m
4g
SII
Ia
7
3a
3b
2
6l
1
4
20a
5Bm
19
SIV
PLLS
21a
EPp
5Al
PMLS
SSAo
18
20b
AMLS
5Am
5Bl
PFCL
SSSAi
CGa
PS
17
Enr
PFCMil
21b
PFCMd
RS
Tem
ALLS
DLS
AII
Sb
P
pSb
VLS
AI
AAF
Hipp
VP
36
Ig
35
AES
CGp
6m
7
4g
SII
Ia
3a
3b
2
6l
1
4
SIV
20a
19
5Bm
PLLS
EPp
5Al
21a
20b
PMLS
SSAo
AMLS
5Bl
5Am
PS
18
SSSAi
PFCL
CGa
17
Enr
21b
PFCMil
PFCMd
RS
ALLS
Tem
Sb
DLS
AII
P
pSb
VLS
AI
AAF
VP
Hipp
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
t
o
r
F
r
o
n
t
i
o
-
L
m
b
c
i
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
c
)
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
o
r
t
t
F
r
o
n
o
-
L
m
b
c
i
i
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
d
)
36
35
Ig
CGp
6m
SII
3a
4g
3b
7
Ia
1
2
AES
6l
4
20a
19
5Bm
21a
18
PMLS
PLLS
5Al
SIV
Enr
17
5Am
CGa
EPp
PFCL
5Bl
SSSAi
SSAo
AMLS
20b
PS
PFCMd
PFCMil
21b
Tem
RS
DLS
VLS
Sb
ALLS
pSb
P
AII
AAF
AI
Hipp
VP
36
35
Ig
CGp
6m
4g
SII
3a
AES
Ia
3b
7
1
2
6l
4
20a
19
5Bm
PLLS
SIV
21a
5Al
PMLS
18
EPp
5Am
CGa
5Bl
SSAo
PFCL
17
20b
SSSAi
Enr
AMLS
PS
PFCMil
PFCMd
21b
RS
Tem
DLS
ALLS
AII
Sb
pSb
P
VLS
AAF
AI
Hipp
VP
0
.
3
0
.
3
5
0
.
4
0
.
4
5
0
.
5
0
.
5
5
0
.
6
0
.
6
5
0
.
7
0
.
7
5
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
t
o
r
F
r
o
n
t
i
o
-
L
m
b
c
i
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
a
)
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
o
r
t
t
F
r
o
n
o
-
L
m
b
c
i
i
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
b
)
2
1
c
o
m
p
o
s
i
t
i
o
n
o
f
t
h
e
l
a
r
g
e
s
t
s
y
n
c
h
r
o
n
i
z
e
d
a
r
e
a
s
.
λ
=
0
.
0
1
5
(
a
)
,
0
.
0
1
7
(
b
)
,
0
.
0
1
9
(
c
)
a
n
d
0
.
0
2
1
(
d
)
)
.
W
e
h
a
v
e
r
e
c
o
l
o
r
e
d
t
h
e
b
a
r
s
o
f
t
h
o
s
e
a
r
e
a
s
c
o
r
r
e
s
p
o
n
d
n
g
i
t
o
t
h
e
R
i
c
h
-
C
u
b
l
t
o
h
i
g
h
l
i
g
h
t
t
h
e
i
d
o
m
n
a
n
t
r
o
l
e
o
f
t
h
i
s
t
o
p
o
l
o
g
i
c
a
l
s
t
r
u
c
t
u
r
e
i
n
t
h
e
F
i
g
u
r
e
6
.
S
y
n
c
h
r
o
n
y
R
a
n
k
o
f
a
r
e
a
s
:
S
t
r
u
c
t
u
r
e
o
f
t
h
e
l
a
r
g
e
s
t
s
y
n
c
h
r
o
n
i
z
e
d
a
r
e
a
s
a
s
d
e
s
c
r
i
b
e
d
b
y
t
h
e
R
i
c
h
-
C
u
b
l
.
T
h
e
t
w
o
p
l
o
t
s
s
h
o
w
t
h
e
s
a
m
e
s
y
n
c
h
r
o
n
y
r
a
n
k
s
a
s
i
n
F
i
g
u
r
e
5
(
a
g
a
i
n
36
35
Ig
AES
CGp
6m
4g
SII
Ia
7
3a
3b
2
6l
1
4
20a
5Bm
19
SIV
PLLS
21a
EPp
5Al
PMLS
SSAo
18
20b
AMLS
5Am
5Bl
PFCL
SSSAi
CGa
PS
17
Enr
PFCMil
21b
PFCMd
RS
Tem
ALLS
DLS
AII
Sb
P
pSb
VLS
AI
AAF
Hipp
VP
36
Ig
35
AES
CGp
6m
7
4g
SII
Ia
3a
3b
2
6l
1
4
SIV
20a
19
5Bm
PLLS
EPp
5Al
21a
20b
PMLS
SSAo
AMLS
5Bl
5Am
PS
18
SSSAi
PFCL
CGa
17
Enr
21b
PFCMil
PFCMd
RS
ALLS
Tem
Sb
DLS
AII
P
pSb
VLS
AI
AAF
VP
Hipp
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
t
o
r
F
r
o
n
i
R
c
h
-
C
u
b
l
t
i
o
-
L
m
b
c
i
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
c
)
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
1
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
o
r
t
t
F
r
o
n
o
-
L
m
b
c
i
i
i
R
c
h
-
C
u
b
l
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
d
)
36
35
Ig
CGp
6m
SII
3a
4g
3b
7
Ia
1
2
AES
6l
4
20a
19
5Bm
21a
18
PMLS
PLLS
5Al
SIV
Enr
17
5Am
CGa
EPp
PFCL
5Bl
SSSAi
SSAo
AMLS
20b
PS
PFCMd
PFCMil
21b
Tem
RS
DLS
VLS
Sb
ALLS
pSb
P
AII
AAF
AI
Hipp
VP
36
35
Ig
CGp
6m
4g
SII
3a
AES
Ia
3b
7
1
2
6l
4
20a
19
5Bm
PLLS
SIV
21a
5Al
PMLS
18
EPp
5Am
CGa
5Bl
SSAo
PFCL
17
20b
SSSAi
Enr
AMLS
PS
PFCMil
PFCMd
21b
RS
Tem
DLS
ALLS
AII
Sb
pSb
P
VLS
AAF
AI
Hipp
VP
0
.
3
0
.
3
5
0
.
4
0
.
4
5
0
.
5
0
.
5
5
0
.
6
0
.
6
5
0
.
7
0
.
7
5
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
t
o
r
F
r
o
n
i
R
c
h
-
C
u
b
l
t
i
o
-
L
m
b
c
i
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
a
)
0
.
3
0
.
4
0
.
5
0
.
6
0
.
7
0
.
8
0
.
9
S
o
m
a
t
o
s
e
n
s
o
r
y
-
M
o
o
r
t
t
F
r
o
n
o
-
L
m
b
c
i
i
i
R
c
h
-
C
u
b
l
A
u
d
i
t
o
r
y
i
V
s
u
a
l
(
b
)
2
2
23
Anatomical Clusters
Dynamical Clusters
2.2
2
1.8
M
D
1.6
1.4
1.2
1
0.005
0.01
0.015
0.025
0.03
0.035
0.02
λ
C
D
0.5
0.45
0.4
0.35
0.3
0.25
0.2
0.15
0.1
0.05
0
0.005
0.4
0.2
0
-0.2
0.01
0.015
0.02
λ
0.025
0.03
0.035
-0.4
0.005 0.01 0.015 0.02 0.025 0.03 0.035
λ
Figure 7. Transition from Modular to Centralized organization of synchronization. The two
upper plots show the evolution of the dynamical modularity DM and the dynamical centralization DC
as a function of λ. Both panels show the evolution of the above properties for the network described by
means of both the original 4 modal clusters and the 5 new modules including the Rich-Club. The panel
in the bottom shows the evolution of the DC of the 5 new modules obtained by replacing in Equation
(4) the term maxα{rα} by each rα corresponding to the new 5 modules. From the curves it is clear that
the dynamics is centralized around the Rich-Club.
24
25
30
1
)
k
(
φ
0.8
0.6
0.4
0.2
0
5
Cat
Rewired
15
20
10
Degree
Figure 8. K-density of the corticocortical directed network of the cat φ(k), compared to the
expectation out of the surrogate ensemble. The largest difference occurs at k = 22 (vertically dashed
line) giving rise to a Rich-Club composed of eleven cortical areas.
|
1801.01362 | 1 | 1801 | 2018-01-04T14:17:15 | A simple model for low variability in neural spike trains | [
"q-bio.NC",
"cond-mat.dis-nn"
] | Neural noise sets a limit to information transmission in sensory systems. In several areas, the spiking response (to a repeated stimulus) has shown a higher degree of regularity than predicted by a Poisson process. However, a simple model to explain this low variability is still lacking. Here we introduce a new model, with a correction to Poisson statistics, which can accurately predict the regularity of neural spike trains in response to a repeated stimulus. The model has only two parameters, but can reproduce the observed variability in retinal recordings in various conditions. We show analytically why this approximation can work. In a model of the spike emitting process where a refractory period is assumed, we derive that our simple correction can well approximate the spike train statistics over a broad range of firing rates. Our model can be easily plugged to stimulus processing models, like Linear-nonlinear model or its generalizations, to replace the Poisson spike train hypothesis that is commonly assumed. It estimates the amount of information transmitted much more accurately than Poisson models in retinal recordings. Thanks to its simplicity this model has the potential to explain low variability in other areas. | q-bio.NC | q-bio | A simple model for low variability in neural spike trains
Ulisse Ferrari,1, ∗ St´ephane Deny,2 Olivier Marre,1, † and Thierry Mora3, †
1Sorbonne Universit´e, INSERM, CNRS, Institut de la Vision, 17 rue Moreau, 75012 Paris, France.
2Neural Dynamics and Computation Lab, Stanford University, California, U.S.A
3Sorbonne Universit´e, Universit´e Paris-Diderot and ´Ecole normale sup´erieure (PSL), 24, rue Lhomond, 75005 Paris, France
Neural noise sets a limit to information transmission in sensory systems. In several areas, the spiking response
(to a repeated stimulus) has shown a higher degree of regularity than predicted by a Poisson process. However,
a simple model to explain this low variability is still lacking. Here we introduce a new model, with a correction
to Poisson statistics, which can accurately predict the regularity of neural spike trains in response to a repeated
stimulus. The model has only two parameters, but can reproduce the observed variability in retinal recordings
in various conditions. We show analytically why this approximation can work. In a model of the spike emitting
process where a refractory period is assumed, we derive that our simple correction can well approximate the
spike train statistics over a broad range of firing rates. Our model can be easily plugged to stimulus processing
models, like Linear-nonlinear model or its generalizations, to replace the Poisson spike train hypothesis that is
commonly assumed. It estimates the amount of information transmitted much more accurately than Poisson
models in retinal recordings. Thanks to its simplicity this model has the potential to explain low variability in
other areas.
8
1
0
2
n
a
J
4
]
.
C
N
o
i
b
-
q
[
1
v
2
6
3
1
0
.
1
0
8
1
:
v
i
X
r
a
∗ Correspondence should be sent to [email protected].
† These authors contributed equally.
I.
INTRODUCTION
2
Neural variability imposes constraints on the way neurons transmit and process information [1] and has been extensively
studied in the mammalian visual system [2–5] and beyond [6–8]. The Fano Factor, i.e. the ratio of the spike count variance to
the mean, is often used to characterize such variability. Many studies have reported various values for Fano Factor depending
on brain area [5] or experimental condition [8]. A Fano Factor lower than one suggests that neurons spike regularly, a condition
sometimes termed as "under-dispersion" [5, 7, 9–11]. A Fano Factor above one is termed over-dispersion [6, 8].
A common model for neural spike count is to assume Poisson statistics, where the variance in the number of spikes emitted
is equal to its mean and the Fano Factor is thus one. Like the classical "LNP" (Linear Nonlinear Poisson) [12], many models
describing how stimuli are processed by individual neurons rely on this assumption. New models are therefore needed to account
for the deviations from Poisson statistics observed in the data. Several models have been proposed to account for over-dispersion
in the spike count distribution [13–15]. For under-dispersed spike count distribution, a few models have been proposed, but they
come with specific constraints. One possibility is to use a spike history filter that allows past spikes to inhibit the probability of
emitting a new spike at present [16]. However, this approach requires defining the probability of spiking over very small time
bins (ex: 1ms), and consequently needs several parameters or strong regularization to describe the spike history filter. Other
models have been proposed, but some have many parameters (for example, ∼ 7 for each cell in [17]), and may therefore require
very large datasets to learn the model, or make specific assumptions that may not always be verified in the data [18, 19]. Overall,
it is unclear if a general yet simple model can account for the spike count distribution found in the data [15].
Here we present a simple model that can account for under-dispersion in the spike count variability, with only two parameters.
Our starting assumption is that the deviation from Poisson statistics comes from the refractory period, i.e. the fact that there
is a minimal time interval between two spikes. We start from the analytic form of the spike count probability taken by a
spike generation process composed of an absolute refractory period followed by a Poisson process, and show that it can be
approximated by a model with only one parameter to fit to the data. We further simplify the model to make it amenable to
log-likelihood maximization. We then relax our assumption by allowing for a relative refractory period [20, 21] and we derive
a more flexible model with two parameters that can accurately predict neural variability. The simple form of this model makes
it possible to plug it on classical stimulus processing models (e.g. Linear-Nonlinear (LN) model [16] or more complex cascade
models [22]), that can be fitted with log-likelihood maximization. We test our model on retinal data and find that it outperforms
the Poisson model, but also other statistical models proposed in the literature to account for under-dispersion. Either our model
performs better at describing the data, or other models need more parameters for an equally accurate description. When combined
with classical stimulus processing models, our model is able to predict the variance of the spike count over time, as well as the
amount of information conveyed by single neurons, much better than classical models relying on Poisson processes. We thus
propose a simple model for neural variability, with only two parameters, that can be used in combination with any stimulus
processing model to account for sub-Poisson neural variability.
3
II. SUB-POISSON BEHAVIOR OF RETINAL GANGLION CELLS
FIG. 1. Sub-Poisson behavior of RGCs. A) Top: Example raster plot for one OFF cell. Bottom: Firing rate behavior for the same cell.
Colored (resp. black) area delimit the empirical (resp. Poisson) noise, estimated as mean +/- one standard deviation. B) As panel A, but for an
example ON cell. C) OFF cells show sub-Poisson behavior. Top: histogram of the observed means of the spike count across stimulus repetition
(time-bin of 16.7ms). Time-bins with zero mean have been excluded. Bottom: Mean of spike count across stimulus repetitions plotted against
its variance. Each point correspond to one cell in one time bin. Multiple points are superimposed. Green full line: average of points with
similar mean. Black line: prediction from a Poisson distributed spike count (mean equals to variance). D) ON cells show sub-Poisson behavior.
As in C but for ON cells. Note the increase of activity with respect to OFF cells.
4
We used previously published data [22] where we recorded ganglion cells from the rat retina using multi-electrode arrays
[23, 24]. Two types of ganglion cells, ON and OFF, with receptive fields tiling the visual space, were isolated in these experiments
(n=25 and n=19 cells, respectively). Cells were then stimulated with the video of two parallel horizontal bars performing a
Brownian motion along the vertical direction [22]. Some stimuli sequences were repeated and triggered reliable responses from
ganglion cells (Fig. 1A&B top panels).
We then binned the RGC responses with 16.7ms time windows ∆T locked to the frame update of the video stimulation
(60Hz). For each cell i in each time bin t and for each stimulus repetition rep we associated an integer spike count nrep
i
(t)
equal to the number of emitted spikes in that time-window. In order to analyze the cell reliability as a function of the firing rate,
for each cell and time bins we computed the mean (λi(t)) and variance (Vi(t)) of this spike count across repetitions of the same
stimulus. In Fig. 1A&B (bottom panels) we show the firing rate of two example ON and OFF cells, together with its fluctuation
across the repetitions (colored areas). For comparison we also show the fluctuations that can be expected from a Poisson process
; that is, with a variance equal to its mean. During transients of high activity, cells are more reliable than what can be expected.
Figs. 1C&D show that both ON and OFF cells have a sub-Poisson variability. Variances of spike counts are much lower than
their means and their ratio (the Fano factor, equal to 1 for Poisson distributions) is not only smaller than one, but decreases with
the mean. Note that here and for the rest of the paper, we pool together all the cells of the same type together when estimating
the parameters of the spike generators. We found that spike statistics were remarkably homogeneous across cells of the same
type, and this gives us more data to fit different models. However, the same process could have been applied on a single cell,
with enough repetitions of the same stimulus.
These results are consistent with previous findings [5, 20, 21]: ganglion cells emit spikes more reliably than a Poisson process.
Consequently, a model for predicting RGC response that accounts for noise with a Poisson generator, for example the LNP model
[12], largely under-estimates this spiking precision. We then search for a simple model to account for this relation between signal
and noise.
III. SPIKE COUNT STATISTICS FOR A REFRACTORY NEURON
Absolute or relative refractory periods are known to impose some regularity on the sequence of emitted spikes [5, 20] and thus
to decrease the neural variability. We aim at looking how refractoriness impacts the spike count distribution of the spike train.
We consider a model of refractory neuron where the instantaneous rate is inhibited by an absolute refractory period of duration
τ. Such a neuron has the following Inter-Spike interval (ISI) distribution:
ρΘ(t) = Θ(t − τ )re−r (t−τ ) .
5
(1)
where Θ(u) the Heaviside unit step (Θ(u) = 1 for u > 0 and zero otherwise) and r is the firing rate in absence of refractoriness.
From the ISI distribution (1) we can estimate the probability distribution of the spike count n, i.e. the probability distribution of
the number of spike emitted in a time bin ∆t. For the particular case of τ = 0, then λ ≡ (cid:104)n(cid:105) = r∆t and n follows a Poisson
distribution:
(cid:0)n(cid:12)(cid:12)λ(cid:1) =
PPois
e−λ.
λn
n!
(2)
For τ > 0, n follows a sub-Poisson distribution PΘ( n r, ∆t, τ ), which admits an analytic (albeit complex) expression [25] (see
Eq. (13) in the Methods and the derivation in the Supplementary Information).
FIG. 2. A simple neuron model with an absolute refractory period accounts for the observed spike count statistics. A) Relation between
mean and variance predicted by the model (pink) and measured for the OFF population (green, same data as Fig. 1). B) Same as A, but for the
ON population.
To test if this expression accounts well for the data, we need to adjust the value of the refractory period τ for both cell type.
Because the complex expression of PΘ( n r, ∆t, τ ) makes it hard to apply likelihood maximization, we perform the inference
of τ by minimizing the Mean-Square Error (MSE) of the mean-variance relation, summed over all the cells and time bins of
the training set (see Methods). We estimate τ for the two populations separately, and found τ = 8.8ms for OFF cells and
AB00.5100.20.40.60.811.2OFF-cells00.511.5200.511.52ON-cells τ = 3.1ms for ON cells. In Fig. 2 we compare the model prediction, with the mean-variance relation measured on the testing
set. The good agreement of the predictions suggests that this simple model of refractory neurons accounts well for the neural
variability, even when the firing rate is large. A simple model that takes into account the deviation from Poisson statistics due to
the refractory period is thus able to predict the mean variance relation observed in the data.
6
IV. SIMPLE MODELS FOR THE SPIKE COUNT STATISTICS
We have shown how a rather simple model of refractory neurons accurately reproduces the mean-variance relation of the
recorded cells. However, the model has a complicated analytic form, which is not easily amenable to a likelihood-maximization
approach, and it cannot be considered as an alternative to models based on Poisson generators such as the LNP model. To
overcome this limitation, we propose a further simplification of the refractory neuron that is much more efficient and tractable.
For τ = 0, the refractory distribution PΘ( n r, ∆t, τ ) reduces to the Poisson distribution (Eq. 2). The small values of the
refractory periods relative to the bin size suggests that an expansion at small τ could capture most of the model's behavior. We
expand PΘ( n r, ∆t, τ ) around the Poisson distribution to the second order in the small parameter f = τ /∆T (see supp. B for
details):
(cid:26)
Θ ( n λ, f ) = exp
P (2nd)
θλn − (f − f 2)n2 − f 2
2
n3 − log n! − log Zλ
(cid:27)
θλ
such that
(cid:104)n(cid:105)P (2nd)
Θ ( n λ,f ) = λ .
(3)
(4)
The"Second-Order" model (3) has only one free parameter, f or equivalently τ. θλ is a function of λ and has to be estimated
numerically in order to reproduce the mean spike count λ once the refractory period τ has been fixed. For instance, for τ = 0,
θλ = log λ and the Second-Order reduces to the Poisson distribution. Importantly, the coefficients of the n2 and n3 terms do not
depend on the firing rate r, insofar as the refractory period does not either. The exponential form of the model makes it easy to
calculate its derivatives and to use with maximum-likelihood methods.
We infer τ through a log-likelihood maximization (see Methods) using Eq. 3, for OFF and ON populations separately. From
these values we can estimate the refractory period for the Second-Order model: τ = 10.8ms for OFF cells and τ = 3.0ms
for ON cells. In Fig. 3 we compare the model predictions for spike count variance V with the empirical values. Unlike in
the previous section, model parameters were not optimized to best fit this curve. Yet the Second-Order model shows a high
performance for the OFF population, even at large firing rates. Despite the approximation, the model is therefore still able to
describe accurately the mean variance relation, and can now be fitted using maximum likelihood approaches. For ON cells, the
7
FIG. 3. Second-Order and Effective models account well for observed spike count statistics. A) The empirical behavior for the OFF
population estimated from the repeated two bars stimulus (green, same data as Fig. 1) is compared with the prediction of Second-Order
(purple) and Effective (red) models. Lines superimpose as the two models take a very similar form for the OFF population. For comparison,
Poisson prediction is shown in black (equality line) Inset: the function θλ of the two models (see text), is compared with θλ = log λ for the
Poisson model (black line) B) Same as A, but for the ON population. In all cases, models are learned on a separate training set.
model outperforms the Poisson model but the mean variance relation is not perfectly predicted. We will then explore why this
could be the case and improve our model to get a more flexible one, that can be suited for a broader range of experimental cases.
V. A SIMPLE EFFECTIVE MODEL TO DESCRIBE SPIKE TRAINS STATISTICS
The previous results hold for a very specific assumption about how refractoriness constrains firing, with just a single fitting
parameter – the refractory period. We thus wondered whether more general rules of refractoriness could give rise to a broader
class of sub-Poisson spiking distributions, allowing for a better agreement with the data.
We considered a general model in which the instantaneous spike rate is now inhibited by a time-dependent factor, α(r, u)r,
where r is the spike rate in absence of refractoriness, and u is the time following the previous spike, with α(r, 0) = 0 and
α(r, u ≥ τ ) = 1. Calculating the distribution of number spikes under this assumption is intractable analytically, but a Second-
Order expansion such as the one performed in the previous section can still be performed, and yields the following expression at
Second-Order (see Supp B):
AB00.5100.20.40.60.811.2OFF-cells0.511.5-50500.511.5200.511.52ON-cells0.511.5-505 ( n λ, r, f, ∆t ) = exp(cid:8)θλn − γαn2 − δαn3 − log n! − log Zλ
P (2nd)
α
(cid:19)
− g
(cid:19)
(cid:18) f 2
2
− g
(cid:18) f 2
2
γα = f − f 2 + (2 + r∆t)
δα = f 2 − g =
f 2
2
+
θλ
such that
(cid:104)n(cid:105)P (2nd)
α ( n λ,r,f,∆t ) = λ .
(cid:9)
8
(5)
(6)
where f ≡ (cid:82) ∞
0 du (1 − α(r, u))/∆t and g ≡ (cid:82) ∞
0 du u (1 − α(r, u))/∆t2. The special case of a pure refractory period,
α(r, u) = 0 for 0 ≤ u ≤ τ, gives back Eq. 3 and the previous definition of f. In this case, f 2/2 − g vanishes. This quantity can
thus be considered as an estimate of the deviation from the absolute refractoriness. Again, θλ should be adjusted to match the
average number spikes in each cell and time bin, λi(t).
This analytic development shows that the coefficients of n2 and n3, γα and δα, can have very different forms depending on
the exact form of the refractoriness. We thus decided to relax the assumption of a strict dependence between γα and δα. We
tested if a model with γα and δα that do not depend on r shows a good agreement with the data. In the following we thus treat
the coefficients of n2 and n3 in Eq. (5 as two parameters that are independent of each other, and also independent of the firing
rate.
The resulting Effective model is:
PEff( n λ, γ, δ ) = exp(cid:8)θλn − γn2 − δn3 − log n! − log Zλ
(cid:9)
θλ
such that
(cid:104)n(cid:105)PEff( n λ,γ,δ ) = λ .
(7)
(8)
where as before θλ is not a free parameter, but a uniquely defined function of λ with parameters γ and δ. The probability
distribution (7) belongs to the class of weighted Poisson distribution and its mathematical properties have been already studied
elsewhere [26].
We infer γ and δ through a log-likelihood maximization on the OFF and ON cell separately. For the OFF population we obtain
similar values than the Second-Order model (fig. 3A). By contrast, for the ON population the Effective model takes advantage
of the additional free parameter and uses it to improve its performance. We obtained the values γ∗
while the equivalent parameter values in the Second-Order model are f∗
ON = 0.15 and f∗2
ON = −0.52 and δ∗
ON/2 = 0.02.
ON − f∗2
ON = 0.15,
This Effective model is therefore a simple model able to describe accurately sub-Poisson neural variability in several cases,
with only two parameters.
VI. BENCHMARK OF PROPOSED MODELS
9
The Effective model outperforms a Poisson model at predicting the empirical spike count variance. In this section we compare
its performance with two other spike count models proposed in the literature. The Generalize Count (Gen.Count) model [17]
can be seen as a generalization of our Effective model and thus offers a larger flexibility to model the spike count statistics.
This however comes at the price of introducing more parameters to fit, and could potentially lead to overfitting. The Conwey-
Maxwell-Poisson (COMP) model [18] has been proposed to account for both under- and over-dispersed mean-variance relation
[19]. COMP is a one-parameter extension of Poisson, which differs from our Effective model, but it is still a particular case of
Gen.Count (see Methods for details). For the sake of completeness we also compare the Refractory model, PΘ of Eq. (13) and
its second order expansion, the Second-Order model, P 2nd
Θ of Eq.(3).
FIG. 4. Second-Order (eq. 3) and Effective (eq. 7) model performance compared with known spike count models: COMP (eq. 26) and
Generalized Count (eq. 25) Log-likelihood improvement over the Poisson for several models, all learned on a separate training set, for OFF
cells (A) and ON cells (B). To estimate log-likelihoods only time-bins with λi(t) > 0.3 are considered (see text).
We compare the performance of different models as improvement over the Poisson log-likelihood (fig. 4). In order to focus
on transient with high firing rates, Log-likelihood are estimated on the time-bins of the testing set with λi(t) > 0.3. For the OFF
population (fig. 4A), all models outperform Poisson and have similar performance. This is probably because OFF cells show
rather small firing rate and emit rarely more than two spikes in the same time-bin. The addition of one parameter (with respect
to Poisson) is thus enough for accurately modeling the spike count statistics. However, for the ON population (fig. 4B), while
all the considered models outperform Poisson, Effective and Gen.Count show the largest improvement. Remarkably, the one-
parameter Refractory and Second-Order models show very high performance as well, despite the first being learned by fitting
the mean-variance relation of the spike count rather than by maximizing the likelihood. Also in this case, the larger flexibility of
the Gen.Count model does not bring an improvement.
10
FIG. 5. Performance of different models on response to checkerboard stimulation Same as fig.4 but the testing set was the response to a
repeated checkerboard stimulation
To test if these models can generalize to other stimulus statistics, after learning the parameters of the model on the responses
to the moving bars stimuli, we test them on the responses of the same cells to a repeated sequence of checkerboard stimulus.
We find that neurons had also sub-Poisson behavior in response to this stimulus, with spike count variances smaller than the
corresponding means, very similar to fig. 1C&D. Fig. 5 shows the log-likelihood improvement over the Poisson model. All
models performed better than Poisson, except for the COMP model that for the ON cells does not seem to generalize to other
stimuli. In particular, for OFF cell we found that the Refractory model has the best performance, just a bit larger than the
Second-Order and the Effective model (both have very similar performances, see previous section). For the ON cells Effective
and Gen.Count models show the best performance. Also in this case, the larger flexibility of Gen.Count does not bring a
performance improvement. The Effective model is thus a simple model to describe deviation from Poisson statistics with only
two parameters. It works as well as the most general model, Gen.Count, to describe the sub-Poisson variability, but with less
parameters. It also performs better than other models with a low number of parameters.
VII.
IMPACT OF NOISE DISTRIBUTION ON INFORMATION TRANSFER
11
The ability of neurons to transmit information is limited by their variability [1]. If neurons are less reliable, or equivalently
have larger variance, their capacity to transmit information should be significantly decreased. To properly estimate the amount
of information transmitted by a neuron, a model should reproduce such variability. To test for this, we quantify the amount of
stimulus information encoded by emitted spikes as the mutual information M I between the spike count n and its mean λ:
M I(n, λ) = H[ P (n) ] −(cid:10) H[ P (nλ) ](cid:11)
λ
(9)
where H is the entropy function and P (n) is the distribution of the spike count without conditioning on the mean over repetitions.
(cid:104). . .(cid:105)λ is the average over the observed mean spike count and P (nλ) is the distribution of n at fixed λ, either empirically
estimated from the repeated data, or estimated with the model. To avoid under-sampling issues typical of time bin with low
activity and due to the finite number of stimulus repetitions, we restrict the averages to all cells and time-bins with λi(t) > 0.1.
FIG. 6. Second-Order and Effective model predicts empirical information transmission Empirical estimation of the mutual information
is compared with prediction for Poisson (black), Second-Order (pink) and Effective (red) models. using Poisson as noise distribution for the
spike count leads to a strong under-estimation, biasing the prediction of information transfer of stimulus processing models. Error-bars are
standard deviation of the mean over λ, see Eq. (9 ).
In Fig. 6, for both OFF and ON population, Poisson largely under-estimates mutual information, whereas the Effective model
predicts well the value of mutual information. Our model can thus be used to correctly estimate the mutual information thanks
to its accurate prediction of the mean-variance relation.
VIII.
IMPROVING STIMULUS PROCESSING MODELS
12
The Effective model, Eq. (7) describes efficiently the relation between mean and variance and predicts well the amount of
information transmitted by a neuron. Here we show how it can easily be plugged to any stimulus processing model like the
Linear-Nonlinear (LN) model, instead of a Poisson process. To estimate the parameters of LN models, a classical approach is
to assume a Poisson process for spike generation, and then to maximize the likelihood of the spiking data with respect to the
parameters of the model. The major advantage of this method is that the gradient of the likelihood has a very simple form that
allows for iterative log-likelihood maximization. Here we show that using the Effective model also leads to a tractable form
of the log-likelihood gradient, which can similarly be used for iterative optimization, but with the added advantage that the
mean-variance is accurately reproduced.
In general, a stimulus encoding model is defined by a series of computations - parametrized by parameters ψ - that takes the
past stimulus St as input and provides a prediction λψ(St) for the spike count mean as a function of time t. Only at the last stage
a stochastic spike counter P is introduced to predict the number of spikes n(t) emitted in the time-bin t:
St → λψ(St) → n(t) ∼ P(cid:0) n(t)(cid:12)(cid:12) λψ(St)(cid:1)
(10)
where for example, P is a Poisson distribution. The classical example for this is the Linear-Nonlinear-Poisson (LNP) model
[12], but many generalizations have been proposed, especially for the retina [22, 27, 28].
One of the major advantage of using a Poisson spike counter PPois is that it allows for a straightforward optimization of the
model parameters ψ [12, 27]. Thanks to the explicit expression of the log-likelihood (cid:96)(ψ) (see Methods), the log-likelihood
gradient ∇ψ(cid:96)(ψ) ≡ d(cid:96)(ψ)/dψ takes a very simple form:
(cid:88)
∇ψ(cid:96)(ψ) =
1
T
d log PPois
t
dλ
(cid:0) n(t)(cid:12)(cid:12) λ(cid:1)
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)λ=λψ(St)
∇ψ
λψ(St) =
(cid:32)
(cid:88)
t
1
T
n(t) − λψ(St)
λψ(St)
(cid:33)
∇ψ
λψ(St) ,
(11)
Once ∇ψ
λψ(St) is evaluated, (11) allows for iterative log-likelihood maximization.
In the previous section we showed that Effective model outperform Poisson as spike counter distribution.
Its particular
structure, Eq. (7), allows an easy estimation of the log-likelihood gradient (see Methods for the calculation details):
(cid:32)
(cid:88)
t
∇ψ(cid:96)(ψ) =
1
T
n(t) − λψ(St)
V Eff
λψ(St)
(cid:33)
∇ψ
λψ(St)
(12)
With respect to the Poisson case, the inference of the parameters ψ requires only the variance V Eff
λ
as a function of λ, which
depends only on the noise model and can be easily estimated from it before running the inference (see Methods for more details).
13
FIG. 7. Effective models improve performance of stimulus processing models A) PSTH for an example OFF cell (Green) is compared
with predictions by the LN2 models equipped with Poisson (black) or Effective (red) spike counter. B) As in panel A, but for the spike count
variance. Poisson model largely overestimates the empirical variance during transients of high activity. C) Fano Factor as function of time:
Effective model (red) accounts for the empirical behavior (green). Poisson prediction (black constant line) does not. The empirical trace is
not defined when the mean spike count is zero. D) Improvement with respect to Poisson in the mean square error (MSE) between empirical
and model variance, plotted as a function of the MSE between empirical and model mean spike count. Each point represents a cell of the OFF
population. Circled point refers to the example cell in panels A-C. E) Same as D, but for the Effective model applied on the ON population.
As in [22] and for both OFF and ON populations, we infer a Linear-Nonlinear-Linear-Nonlinear-Poisson (LN2-Pois) model
to predict the average cell response to the two bar stimulation. This is a stimulus processing model composed by a cascade of
two layers of linear-filtering&Nonlinear-transformation followed by a Poisson spike counter. Similarly, we infer a two-layer
cascade Effective model (LN2Eff) to predict the average cell response of OFF and ON cells respectively. LN2-Eff differs from
LN2-Pois in the noise generator, either Effective, see Eq. (7), either Poisson. The two models show very similar prediction for
the mean spike activity (fig. 7) while LN2-Pois largely over-estimate the spike count variance (fig. 7B). LN2-Eff predicts also
ABCDE00.51OFF-cellEffectivePoisson00.51910110.20.40.60.811.200.050.10.150.200.050.10.150.20.25OFF-cells00.10.20.300.050.10.150.20.25ON-cells well the Fano Factor estimated over time(fig. 7C). For most of the OFF cells, replacing Poisson by the Effective model leads to
a significant performance improvement when trying to predict the variance (fig. 7D). For OFF cells, the Second-Order model,
Eq. 3, performs as well as the Effective model. The Effective model also predicts well variance for ON cells (fig. 7E). Our
Effective model can therefore be plugged in encoding models to describe accurately the variance of the spike count over time.
14
DISCUSSION
We have shown that a simple model taking into account refractoriness in the spiking process explains most of the deviation
from Poisson that we observe in the spike count statistics. The model has only two parameters, can easily be plugged into any
encoding model to be fitted to sensory recordings. It allows for an accurate estimation of the relation between the mean spike
count and its variance, but also of the amount of information carried by individual neurons. The form of this model is inspired
by the regularity imposed on spike trains by the refractory period. However, it can potentially work for data with other sources
of regularity. In the retina, this model works for two different types of cells. Thanks to its simplicity and generality, this model
could potentially be used to account for mean-variance relation in neurons recorded in other sensory areas [5, 7, 9–11], even if
their neural variability is not solely determined by the refractory period in the spike generation of the neuron. Previous works
[1, 5] suggested that the refractory period present in the phenomenological model may reflect refractoriness at any stage in the
circuit and may not directly correspond to the refractoriness of the recorded cell.
We have compared our model with others already proposed in the literature: the Conwey-Maxwell-Poisson (COMP) [19] and
the Generalized Count (Gen.Count) [17]. We found the COMP model was quite inefficient at fitting our data. The Gen.Count
model can be considered as encompassing a much larger class of possible models, and includes ours as a special case. However,
this comes at the cost of having many more parameters to fit, which can lead to overfitting in some cases, as we have shown
in the results. Moreover, in our data, these additional parameters did not allow improving the performance of the model. The
model we propose has only two parameters, which makes it easy to fit and usable in many cases, even when the amount of data
is limited. It will be interesting to compare our effective models to the Gen.Count model in other sensory structures.
The relation between mean and variance of the spike count strongly depends on the bin size chosen to bin the cell response.
For very long windows, we cannot assume constant firing rate, and we observe a superimposition of many different rates. In
this case, and even for the retina, cells show much larger Fano Factor [5], that can also exceed unity. In other systems where
some of the sources of variability are not controlled, a similar heterogeneity of firing rates can be observed also because the
uncontrolled source will impose a different firing rate at each trial. Several solutions have been proposed to model this over
dispersion [13–15]. For very small time bins the number of emitted spikes rarely exceeds unity, so that the spike count becomes
15
a binary variable with fixed mean-variance relation V = λ(1 − λ). However, neuron refractoriness plays an important role
and the activity in consecutive small time-bins are strongly correlated, for example cells never spike in consecutive time-bins:
ni(t) = 0 if ni(t − 1) = 1. In this case a common solution is to introduce a spike history filter [16] that models the spike
probability with a dependence on the past activity. We have chosen a bin size between these two extremes, such that the spike
count is not a binary variable, but the firing rate stays roughly constant within a single time bin and the dependence between
consecutive time bins is relatively weak. The advantage of this choice of bin size is that the number of parameters needed to
describe the spike regularity is small: only two parameters. The bin size chosen also corresponds to the timescale of the retinal
code [21].
METHODS
Equilibrium Poisson process with absolute refractory period A Poisson process with firing rate Θ(u − τ )r, where Θ is
the Heaviside step, function of the time from previous spike, has inter-spike interval distribution given by Eq. (1). For such
(cid:0)n(cid:12)(cid:12)ν, f(cid:1), where f = τ /∆t and ν = r∆t, of
the number a spikes emitted in a time window ∆t, whose starting point is chosen as random [25, 29] (see SI for a derivation):
process it is possible to compute the probability distribution PΘ
(cid:0)n(cid:12)(cid:12)ν, f(cid:1) =
PΘ
1
1 + νf
Φ(n) + Θ(nMax − 2 − n)
(cid:0)n(cid:12)(cid:12)r, τ, ∆t(cid:1) = PΘ
n(cid:88)
(n + 1 − j)
νj(1 − (n + 1)f )je−ν(1−(n+1)f )
−2Θ(nMax − n − 1)
(n − j)
j=0
νj(1 − nf )je−ν(1−nf )
j!
+ Θ(nMax − n)
(n − 1 − j)
νj(1 − (n − 1)f )je−ν(1−(n−1)f )
j!
n−1(cid:88)
j=0
n−2(cid:88)
j=0
Φ(n) ≡
0
nMax(1 + νf ) − ν
ν − (nMax − 1)(1 + νf )
,
,
,
n ≤ nMax − 2
n = nMax − 1
n = nMax
j!
(13)
(14)
where nMax is the smallest integer larger than ∆t/τ and we used the convention Θ(0) = 1. The distribution (13) has expected
value given by:
EΘ(nν, f ) =
ν
1 + νf
(15)
and exact variance [25]:
VΘ(nν, f ) =
2(cid:80)nMax−1
n=0
(cid:104)
ν(1 − nf ) − n +(cid:80)n−1
j=0 (n − j) νj (1−nf )j e−ν(1−nf )
1 + νf
j!
(cid:105) − ν − ν2
1+νf
16
(16)
Inference of the absolute refractory period model To infer the value of f for the absolute refractory model, see Eq. (13) we
perform a mean square error (MSE) minimization of the mean-variance relation:
M SE ≡(cid:88)
(cid:2) Vi(t) − VarΘ(nνi(t), f )(cid:3)2
νi(t) =
i,t
λi(t)
1 − λi(t)f
(17)
(18)
where λi(t) and Vi(t) are the empirical mean and variance of n for the cell i in the time-bin t of the training set. We used
the exact expression (16) for VarΘ(nνi(t), f ) although an approximated results can be obtained by the simpler asymptotic
expression [25]:
VΘ(nν, f ) ≈
ν
(1 + νf )3
(19)
Inference of the Second-Order refractory model To infer the value of f for the Second-Order refractory model, see Eq. (3)
Θ (nλ, f ), the
we perform log-likelihood maximization with a steepest descent algorithm. Thanks to exponential form of P 2nd
derivative of the log-likelihood ((cid:96)) with respect to f takes a simple form:
(cid:96) ≡(cid:88)
(cid:88)
(2f − 1)(cid:0)(cid:104)ni(t)2(cid:105)data − (cid:104)ni(t)2(cid:105)modeli(t)
Θ (nλi(t), f )
log P 2nd
=
d (cid:96)
d f
i,t
i,t
(cid:1) − f(cid:0)(cid:104)ni(t)3(cid:105)data − (cid:104)ni(t)3(cid:105)modeli(t)
(cid:1) ,
(20)
(21)
where (cid:104)·(cid:105)modeli(t) means average with the model distribution P 2nd
parameter using the log-likelihood gradient and we adjust the function θλ accordingly.
Θ (nλ = λi(t), f ). At each iteration, we update the value of the
Inference of the Effective model To infer the value of γ and δ for the Effective model, see Eq. (7) we perform log-likelihood
maximization with Newton method, where at each iteration we update the values of the parameters using the log-likelihood
gradient:
(cid:88)
(cid:88)
i,t
i,t
d
d γ
d
d δ
log PEff(nλi(t), f ) = (cid:104)ni(t)2(cid:105)data − (cid:104)ni(t)2(cid:105)modeli(t)
log PEff(nλi(t), f ) = (cid:104)ni(t)3(cid:105)data − (cid:104)ni(t)3(cid:105)modeli(t)
and the Hessian:
H = −
(cid:104)ni(t)2(cid:105)modeli(t) − λi(t)2
(cid:104)ni(t)3(cid:105)modeli(t) − λi(t)(cid:104)ni(t)2(cid:105)modeli(t)
(cid:104)ni(t)3(cid:105)modeli(t) − λi(t)(cid:104)ni(t)2(cid:105)modeli(t)
(cid:104)ni(t)4(cid:105)modeli(t) − (cid:104)ni(t)2(cid:105)2
modeli(t)
17
(22)
(23)
(24)
where now (cid:104)·(cid:105)modeli(t) means average with the model distribution PEff(nλ = λi(t), γ, δ). After updating the parameter, we
adjust the function θλ and iterate until convergence.
Generalized Counter model In Ref [17], the authors define a Generalized Counter (Gen.Count) distribution, that in our
notation and framework reads:
PGen.Count( n λ ) = exp{θλ[G] n + G[n] − log n! − log Zλ[G]}
(25)
where G[n] is a generic real function defined on the non-negative integers n ∈ [0,∞]. To better characterize its λ dependence,
we have rewritten Gen.Count introducing the proper θλ[G] function.
For G[n] = −γn2 − δn3 the distribution (25) reduces to Effective model, see Eq. (7). The Gen.Count model is thus a
generalization of our model, and as such is potentially more flexible in modeling the spike count statistics. This however comes
at the price of introducing more parameters to fit. Specifically, in practical application one need to define G[n] up to an nMax,
that is nMax + 1 free parameters [17], but the model is invariant for G[n] → G[n] + const. and G[n] → G[n] + cn. Gen.Count
has thus nMax − 1 free parameters: if nMax = 1, 2, 3 then Gen.Count is equivalent to, respectively Poisson, Effective with
δ = 0 and Effective models. Otherwise is offers a potentially interesting way to generalize our cubic model at the price of
inferring the parameters G[n] for large values of n Thanks to its exponential form, we easily infer the Gen.Count parameters
with log-likelihood maximization and we have set nMax = 4 and 5 for, respectively OF F and ON populations.
Conwey-Maxwell-Poisson model In Ref. [19] the Conwey-Maxwell-Poisson (COMP) model [18] has been proposed to
account for both under- and over-dispersed mean-variance relation. In our notation, the COMP model reads:
PCOMP( n λ ) = exp{θλ[η] n − η log n! − log Zλ[η]}
(26)
18
For η = 1 the COMP reduces to the Poisson model, whereas for G[n] = −(η − 1) log n! the Generalized Count reproduces the
COMP. Also for the COMP, the exponential form allows for log-likelihood maximization.
Equipping stimulus processing model with Second-Order or Effective noise term In this appendix we detail the calcula-
tion for computing the log-likelihood gradient of Eq. (12). For a general stimulus processing model, see Eq. (10), equipped with
noise term P the log-likelihood reads:
(cid:88)
t
1
T
log P(cid:0) n(t)(cid:12)(cid:12) λψ(St)(cid:1)
(cid:96)(ψ) =
where the summation runs over the duration of the training set. Consequently the log-likelihood gradient ∇ψ(cid:96)(ψ):
d log P(cid:0) n(t)(cid:12)(cid:12) λ(cid:1)
(cid:88)
t
dλ
(cid:12)(cid:12)(cid:12)(cid:12)(cid:12)λ=λψ(St)
∇ψ
λψ(St) .
∇ψ(cid:96)(ψ) =
1
T
(27)
(28)
(29)
(30)
(31)
To estimate ∇ψ(cid:96)(ψ) we thus need to compute the derivative of the log-probability with respect to the mean spike count. If P is
the Poisson distribution PPois, then:
d
dλ
log PPois
If P is instead the Effective model:
(cid:0) n(t)(cid:12)(cid:12) λ(cid:1) =
(cid:16)
d
dλ
n(t) log λ − λ − log n(t)!
(cid:17)
n(t) − λ
λ
.
=
(cid:0) n(t)(cid:12)(cid:12) λ(cid:1) =
(cid:16)
d
dθλ
d
dλ
log PEff
(cid:0) n(t)(cid:12)(cid:12) λ(cid:1)(cid:17) d
(cid:19)−1
(cid:18) d
=
λ
dθλ
θλ
dλ
n(t) − λ
V Eff
λ
.
log PEff
= (n(t) − λ)
The very same expression, with replaced V Eff
λ by V 2nd
λ , holds for the Second-Order model. Note that (29) has the same form
of (31) because for Poisson V Pois
λ
(t) = λ. Eq. (12) is in fact a general result for spike counter P belonging to the exponential
family. Note that V Eff
λ
and V 2nd
λ
are properties of the noise distribution (respectively PEff and P 2nd
Θ ). These functions can be
numerically estimated before running the inference of the stimulus processing model, for example, by computing their values
for several values of λ and then interpolating with a cubic spline.
Acknowledgments
19
We like to thank R. Brette, M. Chalk, C. Gardella, B. Telenczuk and M. di Volo for useful discussion. This work was supported
by ANR TRAJECTORY, the French State program Investissements d'Avenir managed by the Agence Nationale de la Recherche
[LIFESENSES: ANR-10-LABX-65], a EC grant from the Human Brain Project (FP7-604102)), NIH grant U01NS090501 and
AVIESAN-UNADEV grant to OM
Appendix A: Equilibrium Poisson process with absolute refractory period
Here we provide a sketch of the derivation to obtain the complete expression (13). We are interested in computing the
(cid:0)n(cid:12)(cid:12)r, τ, ∆t(cid:1) of the number a spikes emitted in a time window ∆t, whose starting point is chosen as
probability distribution PΘ
random, when the inter-spike interval distribution is:
ρΘ(t) = Θ(t − τ )re−r (t−τ ) .
PΘ can be expressed as difference between its cumulative distribution:
∞(cid:88)
(cid:0)k(cid:12)(cid:12)∆t, r, τ(cid:1)
C(n∆t) ≡
(cid:0)n(cid:12)(cid:12)∆t, r, τ(cid:1) = C(n∆t) − C(n + 1∆t) .
PΘ
k=n
PΘ
Because C(n∆t) is the probability of having at least n spikes in the time-bin ∆t it can be computed as:
(cid:90) ∆t
0
C(n∆t) =
dt [ρE
Θ (cid:63) ρ(cid:63)(n−1)
Θ
](t)
(A1)
(A2)
(A3)
(A4)
where (cid:63) is the convolution symbol and (·)(cid:63)n means n-times self convolution. ρE
Θ is the distribution of the first spike when the
beginning of the time-bin is chosen at random (equilibrium process) and its distribution can be computed as [29]
(cid:82) ∞
(cid:82) ∞
t dt(cid:48) ρΘ(t(cid:48))
0 dt(cid:48) t(cid:48) ρΘ(t(cid:48))
Θ(t)Θ(τ − t) + Θ(t − τ )e−r(t−τ )
e−r max(t−τ,0)
1 + rτ
= r
1 + rτ
ρE
Θ(t) =
= r
(cid:0)rI[0,τ ](t) + ρΘ(t)(cid:1) ,
=
1
1 + rτ
(A5)
where I[a,b](t) = 1 if t ∈ [a, b] and zero otherwise. Thanks to the explicit decomposition of ρE
Θ we have:
Θ (cid:63) ρ(cid:63)(n−1)
ρE
Θ
=
1
1 + rτ
rI[0,τ ] (cid:63) ρ(cid:63)(n−1)
Θ
+ ρ(cid:63)n
Θ
.
(cid:17)
In order to estimate the ρ(cid:63)n
Θ we introduce the Laplace transform, which for a generic function h(t) reads:
(cid:16)
(cid:90) ∞
L[h(t)](s) =
dt e−sth(t) ,
and use it to get read of the multiple convolutions:
0
Θ = L−1 [ L[ ρ(cid:63)n
ρ(cid:63)n
Θ ] ] = L−1 [ L[ ρΘ ]n ] = L−1
(cid:20) (cid:18) r
r + s
e−sτ
(cid:19)n (cid:21)
rn(t − nτ )(n−1)e−r(t−nτ )
(n − 1)!
=
Θ(t − nτ ) ≡ γ[n, r](t − nτ )
where we have introduced the Gamma distribution γ[n, r](t) ≡ Θ(t) rnt(n−1)e−r/(n − 1)!. From (A4) it follows:
C(n, ∆t) =
1
1 + rτ
dt(cid:48) I[0,τ ](t − t(cid:48))γ[n − 1, r](t(cid:48) − (n − 1)τ )
(cid:90)
(cid:32) (cid:90) ∆t
(cid:90) ∆t
dt
0
+
0
(cid:33)
dt γ[n, r](t − nτ )
.
20
(A6)
(A7)
(A8)
(A9)
C(n∆t) can be computed by integrating several times the gamma distribution using the following relation:
(cid:90) a
0
1
(n − 1)!
(cid:32)
(cid:33)
n−1(cid:88)
k=0
ak
k!
dt t(n−1)e−tΘ(t) =
1 − e−a
Θ(a) .
(A10)
After some algebra [25, 29] this calculation provides Eq. (13).
Appendix B: Small f expansion of the refractory neuron model
Here we derive the result (5), for which (3) is a particular case. We considered a general model in which the instantaneous spike
rate is modulated by a time-dependent factor, α(u)r, where r is the spike rate in absence of refractoriness, with α(u ≥ τ ) = 1,
and u is the time following the previous spike. In this case the inter-spike interval distribution is:
(cid:27)
(cid:26)
−ν
(cid:90) t
0
ρα(t) = να(t) exp
α(t)
Θ(t) ,
(B1)
21
and our goal is to expand, for small τ, Pα(n∆t, r, α), the probability of having n spikes within the time-bin ∆t. At first we
introduce two useful quantities:
(cid:90) ∞
(cid:90) ∞
0
0
dt(cid:0)1 − α(t)(cid:1)
dt t(cid:0)1 − α(t)(cid:1)
f ≡ 1
∆t
g ≡ 1
∆t
(B2)
(B3)
Such that, if α(t) = 1 for all t > 0, then f = g = 0 and we expect to recover the Poisson case, and if α(t) = 0 for all t < τ,
then f = τ /∆t and g = τ 2/∆t2/2 and we expect to recover the absolute refractory case, i.e. Eq. (3).
Much like the absolute refractory case, we consider the cumulative distribution of Pα(nr, α, ∆t, ):
(cid:90) ∆t
0
Cα(n∆t) =
α (cid:63) ρ(cid:63)(n−1)
[ρE
α
](t) ,
(B4)
where, as before, ρE
α is the distribution of the first spike for an equilibrium process and ρ(cid:63)(n−1)
α
is ρα self-convoluted n − 1
times. To perform the expansion in small τ, we decompose ρE
α and ρα around an exponential distribution:
α (t) =(cid:0)1 + ν2g(cid:1) ρ(t) + δρE
α (t)
ρα(t) = er(1−A)τ ρ(t) + δρα(t) .
ρE
(B5)
(B6)
In the following we will first perform the computation with ρ(t) instead of ρE
α , leaving for the end the corrections due to the
factor 1 + ν2g and term δρE
α . This is equivalent to consider a shifted process [29] where the time-bin start after the end of the
last refractory period, instead of the equilibrium process we are considering here. As for ρ(cid:63)n
Θ (t), see Eq. (A8), we can use the
Laplace transform to compute ρ(cid:63)n(t) = γ[n, r](t). This allows us to perform the following expansion:
[ρ (cid:63) ρ(cid:63)(n−1)
α
](t) = ρ (cid:63)
≈ e(n−1)νf γ[n, r](t)
i
i=0
(cid:3)(t)
(cid:18)n − 1
(cid:19)
n−1(cid:88)
e(n−i−1)νf(cid:2) ρ(cid:63)(n−1−i) (cid:63) δρ(cid:63)i
+(n − 1)e(n−2)νf(cid:2) γ[n − 1, r] (cid:63) δρα
(cid:3)(t)
(cid:19)
(cid:18)n − 1
(cid:3)(t) .
e(n−3)νf(cid:2) γ[n − 2, r] (cid:63) δρ(cid:63)2
+
α
α
2
(B7)
(B8)
To perform the integration of (B8) we use the following approximation:
(cid:90) ∞
(cid:90) ∞
(cid:90) ∞
0
0
0
dt δρα(t) ≈ −νf − 1
2
ν2f 2 + O(τ 3)
dt δρ(cid:63)2
α (t) ≈ ν2f 2 + O(τ 3)
dt t δρα(t) ≈ −νg∆t + O(τ 3) ,
which after some algebra we obtain for P S
α (nν), the distribution for the shifted process:
(cid:16)
(cid:17)(cid:3)(t) ∝ exp(cid:8)n log ν − log n! + c1n + c2n2 + c3n3(cid:9)
ρ(cid:63)(n−1)
− ρ(cid:63)n
α
(cid:104)
(cid:90) ∆t
0
dt
ρ (cid:63)
α
α (nν) =
P S
with
(cid:19)
(cid:18)
f − f 2
2
c1 = (1 + ν)
c2 = −f + f 2 − (2 + ν)
c3 = −f 2 + g
(cid:19)
(cid:18) f 2
(cid:18) f 2
−
2
− g
(cid:19)
2
− g
We need now to account for the full ρE
α (t). Because
(cid:90) ∞
0
dt δρE
Θ(t) ≈ −ν2g
is of the order τ 2 we can estimate Pα(n∆t, r) = Pα(nν) as:
Pα(nν) =(cid:0)1 + ν2g(cid:1) P S
α (nν) + PPois(n − 1ν)
∝ exp(cid:8)n log ν − log n! + (c1 − νg) n + c2n2 + c3n3(cid:9) .
dt δρE
0 (t)
0
(cid:90) ∞
22
(B9)
(B10)
(B11)
(B12)
(B13)
(B14)
(B15)
(B16)
(B17)
which is equivalent to (5).
[1] JA Movshon. Reliability of neuronal responses. Neuron, 27(3):412, 2000.
[2] HB Barlow and WR Levick. Three factors limiting the reliable detection of light by retinal ganglion cells of the cat. The Journal of
Physiology, 200(1):1–24, 1969.
23
[3] P Heggelund and K Albus. Response variability and orientation discrimination of single cells in striate cortex of cat. Experimental Brain
Research, 32(2):197–211, 1978.
[4] DJ Tolhurst, JA Movshon, and ID Thompson. The dependence of response amplitude and variance of cat visual cortical neurones on
stimulus contrast. Experimental brain research, 41(3-4):414–419, 1981.
[5] Prakash Kara, Pamela Reinagel, and R Clay Reid. Low response variability in simultaneously recorded retinal, thalamic, and cortical
neurons. Neuron, 27(3):635–646, 2000.
[6] Roland Baddeley, Larry F Abbott, Michael CA Booth, Frank Sengpiel, Tobe Freeman, Edward A Wakeman, and Edmund T Rolls.
Responses of neurons in primary and inferior temporal visual cortices to natural scenes. Proceedings of the Royal Society of London B:
Biological Sciences, 264(1389):1775–1783, 1997.
[7] Michael R DeWeese and Anthony M Zador. Binary coding in auditory cortex. In Advances in neural information processing systems,
pages 117–124, 2003.
[8] Mark M Churchland, M Yu Byron, John P Cunningham, Leo P Sugrue, Marlene R Cohen, Greg S Corrado, William T Newsome, An-
drew M Clark, Paymon Hosseini, Benjamin B Scott, et al. Stimulus onset quenches neural variability: a widespread cortical phenomenon.
Nature neuroscience, 13(3):369–378, 2010.
[9] Moshe Gur, Alexander Beylin, and D Max Snodderly. Response variability of neurons in primary visual cortex (v1) of alert monkeys.
Journal of Neuroscience, 17(8):2914–2920, 1997.
[10] Crista L Barberini, Gregory D Horwitz, and William T Newsome. A comparison of spiking statistics in motion sensing neurons of flies
and monkeys. Computational, Neural and Ecological Constraints of Visual Motion Processing, pages 307–320, 2001.
[11] Gaby Maimon and John A Assad. Beyond poisson: increased spike-time regularity across primate parietal cortex. Neuron, 62(3):426–
440, 2009.
[12] E.J. Chichilnisky. A simple white noise analysis of neuronal light responses. Network: Computation in Neural Systems, 12(2):199–213,
2001.
[13] Robbe LT Goris, J Anthony Movshon, and Eero P Simoncelli. Partitioning neuronal variability. Nature neuroscience, 17(6):858–865,
2014.
[14] James Scott and Jonathan W Pillow. Fully bayesian inference for neural models with negative-binomial spiking. pages 1898–1906, 2012.
[15] Adam S Charles, Mijung Park, J Patrick Weller, Gregory D Horwitz, and Jonathan W Pillow. Dethroning the fano factor: a flexible,
model-based approach to partitioning neural variability. bioRxiv, page 165670, 2017.
[16] J.W. Pillow, J. Shlens, L. Paninski, A. Sher, A.M. Litke, E. J. Chichilnisky, and E.P. Simoncelli. Spatio-temporal correlations and visual
signalling in a complete neuronal population. Nature , 454:995–999, 2008.
[17] Yuanjun Gao, Lars Busing, Krishna V Shenoy, and John P Cunningham. High-dimensional neural spike train analysis with generalized
count linear dynamical systems. pages 2044–2052, 2015.
[18] Kimberly F Sellers, Sharad Borle, and Galit Shmueli. The com-poisson model for count data: a survey of methods and applications.
24
Applied Stochastic Models in Business and Industry, 28(2):104–116, 2012.
[19] Ian H Stevenson. Flexible models for spike count data with both over-and under-dispersion. Journal of computational neuroscience,
41(1):29–43, 2016.
[20] Michael J Berry II and Markus Meister. Refractoriness and neural precision. In Advances in Neural Information Processing Systems,
pages 110–116, 1998.
[21] Michael J Berry and Markus Meister. Refractoriness and neural precision. Journal of Neuroscience, 18(6):2200–2211, 1998.
[22] Stephane Deny, Ulisse Ferrari, Emilie Mace, Pierre Yger, Romain Caplette, Serge Picaud, Gasper Tkacik, and Olivier Marre. Multiplexed
computations in retinal ganglion cells of a single type. Nature communications, 8(1):1964, 2017.
[23] O. Marre, D. Amodei, N. Deshmukh, K. Sadeghi, F. Soo, T. Holy, and M.J. Berry. Recording of a large and complete population in the
retina. Journal of Neuroscience , 32(43):1485973, 2012.
[24] P. Yger, G. L. B. Spampinato, E. Esposito, B. Lefebvre, S. Deny, C. Gardella, M. Stimberg, F. Jetter, G. Zeck, S. Picaud, J. Duebel, and
O. Marre. Fast and accurate spike sorting in vitro and in vivo for up to thousands of electrodes. bioRxiv, 2016.
[25] Jorg W Muller. Some formulae for a dead-time-distorted poisson process: To andr´e allisy on the completion of his first half century.
Nuclear Instruments and Methods, 117(2):401–404, 1974.
[26] Joan del Castillo and Marta P´erez-Casany. Overdispersed and underdispersed poisson generalizations. Journal of Statistical Planning
and Inference, 134(2):486–500, 2005.
[27] James M McFarland, Yuwei Cui, and Daniel A Butts. Inferring nonlinear neuronal computation based on physiologically plausible inputs.
PLoS Comput Biol, 9(7):e1003143, 2013.
[28] Lane McIntosh, Niru Maheswaranathan, Aran Nayebi, Surya Ganguli, and Stephen Baccus. Deep learning models of the retinal response
to natural scenes. In Advances in Neural Information Processing Systems, pages 1361–1369, 2016.
[29] Jorg W Muller. Dead-time problems. Nuclear Instruments and Methods, 112(1-2):47–57, 1973.
|
1602.05220 | 1 | 1602 | 2016-02-16T22:09:07 | BioSpaun: A large-scale behaving brain model with complex neurons | [
"q-bio.NC",
"cs.AI"
] | We describe a large-scale functional brain model that includes detailed, conductance-based, compartmental models of individual neurons. We call the model BioSpaun, to indicate the increased biological plausibility of these neurons, and because it is a direct extension of the Spaun model \cite{Eliasmith2012b}. We demonstrate that including these detailed compartmental models does not adversely affect performance across a variety of tasks, including digit recognition, serial working memory, and counting. We then explore the effects of applying TTX, a sodium channel blocking drug, to the model. We characterize the behavioral changes that result from this molecular level intervention. We believe this is the first demonstration of a large-scale brain model that clearly links low-level molecular interventions and high-level behavior. | q-bio.NC | q-bio |
BioSpaun: A large-scale behaving brain model with
complex neurons
Chris Eliasmith, Jan Gosmann, Xuan Choo
Centre for Theoretical Neuroscience, University of Waterloo, Waterloo, Ontario, Canada
Abstract
We describe a large-scale functional brain model that includes detailed, conductance-
based, compartmental models of individual neurons. We call the model
BioSpaun, to indicate the increased biological plausibility of these neurons,
and because it is a direct extension of the Spaun model [1]. We demonstrate
that including these detailed compartmental models does not adversely affect
performance across a variety of tasks, including digit recognition, serial work-
ing memory, and counting. We then explore the effects of applying TTX, a
sodium channel blocking drug, to the model. We characterize the behav-
ioral changes that result from this molecular level intervention. We believe
this is the first demonstration of a large-scale brain model that clearly links
low-level molecular interventions and high-level behavior.
Keywords: Spaun, Neural Engineering Framework, Semantic Pointer
Architecture, conductance neurons, biological cognition
1. Introduction
Recently, several large-scale brain models have been described. These
include a biophysically detailed model from Markram's group in the Human
Brain Project (HBP) [2], which includes about 31,000 compartmental neu-
rons and 37 million synapses, modelled with many equations per cell. This
model is large-scale because of the amount of computation required to sim-
ulate its behavior at this level of biological detail. Another model reported
earlier by the Synapse project has simulated 500 billion neurons – more than
5x the number in the human brain – although each neuron is much simpler
than those in the HBP model, and the connectivity is far more limited [3, 4].
Preprint submitted to ArXiv
February 18, 2016
We have also previously proposed a large-scale model that includes 2.5 mil-
lion neurons, 8 billion connections, and, unlike these other large-scale models,
exhibits a wide variety of cognitive behavior [1]. However, our model uses
simple leaky integrate-and-fire neurons, and for this reason Markram has
claimed "It's not a brain model" [5].
In this paper we incorporate detailed compartmental models of the type
used in the recent HBP model into different cortical areas of our large-scale,
behaving brain model. We refer to this augmented model as "BioSpaun". We
show that the behavior of the original Spaun model is not adversely affected
by changing the neuron model. We further show that the additional com-
plexity can be used to test hypotheses not possible with the original model.
Specifically, we demonstrate that BioSpaun can be used to simulate the ef-
fects of adding the drug tetrodotoxin (TTX) to these areas of cortex. We
perform this manipulation to both visual cortex and frontal cortex, demon-
strating performance declines related to the dosage of drug applied, both
within and across different tasks. While much remains to be done to verify
the accuracy of these simulations in vivo, we believe this is the first demon-
stration of a large-scale behaving neural model that includes a high degree
of biophysical detail. Integrating these two aspects of brain modeling pro-
vides a new method for testing low-level molecular and other physiological
interventions on high-level behavior.
2. Methods
2.1. Modeling approach
The Neural Engineering Framework (NEF) identifies three quantitatively
specified principles that can be used to implement nonlinear dynamical sys-
tems in a spiking neural substrate [6]. These methods have been used to
propose novel models of a wide variety of neural systems including parts of
the rodent navigation system [7], tactile working memory in monkeys [8], and
simple decision making in humans [9] and rats [10]. These methods have also
been used to better understand more general issues about neural function,
such as how the variability of neural spike trains and the timing of individ-
ual spikes relates to information that can be extracted from spike patterns
[11], and how mixed weight neuron models can be transformed into models
respecting Dales Principle, while preserving function [12].
Conceptually, the NEF can be thought of as a neural compiler which
allows a researcher to specify a computation as a general nonlinear dynamical
2
system in some state space, which is then implemented in a spiking neural
substrate using an efficient optimization method. There are several sources
for detailed descriptions of these methods [6, 13, 14], so we do not describe
them here. Centrally, the NEF answers questions about how neural systems
might compute, but it does not address the issue of what, specifically, is
computed by biological brains.
In our more recent work, we address this second question by proposing a
general neural architecture that includes specific functional hypotheses. We
call this proposal the Semantic Pointer Architecture (SPA; [15]). The SPA
identifies a generic means of characterizing neural representation, semantic
pointers, that are used to capture central features of perceptual, motor, and
cognitive representation. The SPA uses semantic pointers to address percep-
tual categorization, motor control, working and long-term memory, as well as
conceptual binding and structure representation. As well, the SPA includes a
characterization of cognitive control that relies on basal ganglia and thalamic
interactions with cortex for understanding action selection.
We have developed several models using the SPA that address a variety
of cognitive abilities. For instance, we have demonstrated the encoding and
decoding of the 114,000 concepts and their relations in the WordNet database
[16]. We have shown an SPA model that matches human performance on the
full Ravens Progressive Matrices (RPM) intelligence test, and demonstrated
its ability to capture aging effects through biological manipulation [17]. We
have described how the SPA naturally unifies the three most prevalent the-
ories of conceptual representation[18]. We have also shown simple models
of language parsing [19], instruction following [20], and the n-back task [21].
Together, we believe this body of works demonstrates a uniquely scalable
and biologically plausible approach to understanding cognitive function.
2.2. The Semantic Pointer Architecture Unified Network (Spaun)
To demonstrate the SPA in detail, we proposed a mechanistic, functional
model of the brain that uses 2.5 million spiking neurons, has about 8 bil-
lion synaptic connections, and performs 8 different tasks [1, 15]. We refer to
this large-scale neural model as the Semantic Pointer Architecture Unified
Network (Spaun). Spaun consists of leaky integrate-and-fire (LIF) model
neurons. The physiological and tuning properties of the cells are statisti-
cally matched to the various anatomical areas included in the model. There
are about 20 anatomical areas accounted for (see Figure 1). Four types of
neurotransmitters are included in the model (GABA, AMPA, NMDA, and
3
Dopamine), and their known time constants and synaptic effects are simu-
lated.
What makes Spaun unique among large-scale brain models is its func-
tional abilities. Spaun receives input from the environment through its sin-
gle eye, which is shown images of handwritten or typed digits and letters,
and it manipulates the environment by moving a physically modelled arm,
which has mass, length, inertia, and so on. Spaun uses these natural in-
terfaces, in combination with internal cognitive processes, to perceive visual
input, remember information, reason using that information, and generate
motor output (writing out numbers or letters).
It uses these abilities to
perform eight different tasks, ranging from perceptual-motor tasks (recre-
ating the appearance of a perceived digit) to reinforcement learning (in a
gambling task) to language-like inductive reasoning (completing abstract
patterns in observed sequences of digits). These tasks can be performed
in any order, they are all executed by the same model, and there are no
changes to the model between tasks. To see the model perform the tasks,
see http://nengo.ca/build-a-brain/spaunvideos.
We have compared the performance of Spaun to human and animal data
at several levels of detail [1]. Along many metrics the two align; for exam-
ple, the model and the brain share (1) dynamics of firing rate changes in
striatum during the gambling task, (2) error rates as a function of position
when reporting digits in a memorized list, (3) coefficient of variation of inter-
spike intervals, (4) reaction time mean and variance as a function of sequence
length in a counting task, (5) accuracy rates of recognizing unfamiliar hand-
written digits, and (6) success rates when solving induction tasks similar to
those found on the Raven's Progressive Matrices (a standard test of human
intelligence), among other measures. It is these comparisons, demonstrat-
ing the range and quality of matches between the model and real neural
systems, that makes it plausible to suggest that Spaun is capturing some
central aspects of neural organization and function.
The Spaun model is simulated in the neural simulation package Nengo
[22], which natively implements the NEF and SPA methods.
2.3. The compartmental model
To replace the simple LIF neurons in our model, we have chosen a com-
partmental conductance model of cortical neurons that is similar in complex-
ity to those used in the recent HBP model [2]. In the HBP model, neurons
had up to 13 ion channel types and 4 compartments. In this work, we use
4
Figure 1: Functional and anatomical architecture of Spaun.
(a) Neuroanatomical
architecture of Spaun, with matching colors and line styles indicating correspond-
ing components in the functional architecture in b. Abbreviations: V1/V2/V4 (pri-
mary/secondary/extrastriate visual cortex), AIT/IT (anterior/inferotemporal cortex),
DLPFC/VLPFC/OFC (dorso-lateral/ventro-lateral/orbito- frontal cortex), PPC (poste-
rior parietal cortex), M1 (primary motor cortex), SMA (supplementary motor area), PM
(premotor cortex), v/Str (ventral/striatum), STN (subthalamic nucleus), GPe/i (globus
pallidus externus/internus), SNc/r (substantia nigra pars compacta/reticulata), VTA
(ventral tegmental area). (b) Functional architecture of Spaun. The working memory, vi-
sual input, and motor output components represent hierarchies that compress/decompress
neural representations between different representational spaces. The action selection com-
ponent chooses which action to execute given the current state of the rest of the system.
The five internal subsystems, from left to right, are used to 1) map visual inputs to
conceptual representations, 2) induce relationships between representations, 3) associate
input with reward, 4) map conceptual representations to motor actions, and 5) map motor
actions to specific patterns of movement. Reproduced from [1].
5
the pyramidal cell model developed and described in detail in [23]. This
model is reduced from a model with several hundred compartments, based
on the neural reconstructions in [24]. The reduced model that we are using
has 20 compartments across 4 functional areas (soma, basal dendrite, api-
cal dendrite, and apical dendritic tuft), 27 parameters, and 9 different ion
channels. After automatically choosing parameter values based on an opti-
mization method, the reduced model very closely replicated the behavior of
the complex model (see, e.g., Figure 2). The model is available for download
at http://senselab.med.yale.edu/ModelDB, and runs in the NEURON 7.1
simulator.
To replace neurons in the Spaun model with this compartmental model,
we developed a Python interface, called nengo detailed neurons, between
Nengo and NEURON. Consequently, all of BioSpaun except for the compart-
mental neurons were run in Nengo, while the Python interface to NEURON
was used to communicate between the two simulators. The Python interface
is available at https://github.com/nengo/nengo_detailed_neurons.
The NEF is defined such that the response properties of the specific neu-
ron models being employed can be taken into account during the optimiza-
tion process for determining connection weights. Typically, point neurons are
used so all connections are to the same compartment. In the case of employ-
ing compartmental models, with different compartments for different areas
of the cell, the weights of excitatory synapses were uniformly distributed
along the compartment modeling the apical dendrite. The synaptic weights
were linearly scaled with the distance from the soma (up to a factor of 2
if at the end of the dendritic compartment). This method approximately
preserved the somatic effects of arriving spikes, resulting in neuronal activity
that can be accounted for using the standard NEF methods. We intend to
replace this simple heuristic with more precise methods in the future. The
inhibitory synapses directly targeted the somatic compartment.
We exploited the inclusion of this more sophisticated neuron model by
testing the effects of TTX on high-level function. TTX is known to block
voltage-gated sodium channels in neurons. Consequently, to simulate the
effects of TTX on the compartmental neurons, different percentages of the
sodium channels were blocked to simulate different concentrations of the
drug.
6
Figure 2: The reduced model compared to the complex model in terms of passive response.
a) The morphology of a neuron from [24] used to create the reduced model. b) An
illustration of the functional sections, defined over 20 compartments, used to optimize
the reduced model. c) Voltage induced at different locations in the models as a function
of a -1nA current injected in the soma. d) The somatic impedance as a function of low-
amplitude oscillatory somatic input current for both models. (e) The soma phase-shift
between the oscillatory input current and membrane potential oscillation for both models.
Reproduced from [23].
7
3. Results
3.1. Comparison of LIF and compartmental information processing
To ensure that the NEF method was successfully applied to populations
of compartmental neurons, we constructed a circuit consisting of 200 in-
put neurons in population A and 50 output neurons in population B. The
function computed by the circuit is identity, implementing a communication
channel. We injected one period of a sine wave into the input population
and reconstructed the input from both LIF and compartmental neurons (see
Figure 3).
Both circuits are able to represent the input well, although the compart-
mental neurons are slightly noisier. Specifically, the RMS error in the LIF
circuit was 0.1 whereas the RMS error in the compartmental circuit was 0.21.
To demonstrate the effects of a high dose TTX application, we blocked 60%
of the sodium channels in the conductance neurons in this circuit (Figure
4. This manipulation significantly negatively impacts the response of the
neurons and increases the RMS error.
3.2. Effects of TTX application on visual processing
In the original Spaun model, the visual system consists of a ventral stream
model that includes V1, V2, V4 and IT. In the digit recognition task, Spaun
classifies human handwritten digits from the MNIST database based on the
neural representation in area IT. To initially examine the effects of using the
compartmental neurons in Spaun, we replaced the LIF neurons in IT with
compartmental neurons. This is a much more challenging task than a simple
communication channel because classification is a highly nonlinear function.
A video of normal performance on the digit classification task can be found
at https://youtu.be/QDSXhuPGHSs.
The original classification accuracy of Spaun (94%) was preserved with
the introduction of the compartmental models (see Figure 5, leftmost value).
We examined the effects of changing the dosage of TTX applied to these
model neurons on classification accuracy. The results are shown in Figure 5.
As can be seen, classification is robust to large numbers of blocked sodium
channels, with little decrease in accuracy at 50% channel blockage. However,
by 72% blockage, the model's accuracy is at chance levels (i.e. 10%).
3.3. Effects of TTX applied to frontal cortex across tasks
Application of TTX to visual cortex has a similar consequence across all
of the tasks performed by Spaun: the input is made noisier and less certain.
8
Figure 3: Simple information processing through a communication channel in LIF and
compartmental neurons. The first panel is the state variable input, which is encoded
into spikes by population A. The second panel shows the estimate of the state variable
from decoding spikes from population A. The third panel shows the estimate of the state
variable from decoding the spikes from population B. a) The channel implemented with
LIF neurons in both populations using standard NEF methods. b) The same channel
implemented with compartmental neurons in the output population.
9
Figure 4: TTX application to a communication channel. The same circuit as shown in
Figure 3b with 60% of the sodium channels blocked.
Figure 5: Handwritten digit recognition accuracy as a function of TTX dosage in
BioSpaun. Error bars are 95% confidence intervals.
10
This is unsurprising given the kind of effects demonstrated in the previous
section. However, those effects are demonstrated in a largely feedforward
network. As well, the effects are only explored in a single task. Here we
apply the manipulation to a recurrent network and examine the effects the
same TTX manipulation across different tasks.
To examine heterogeneous effects of the same manipulation in a recurrent
network, we introduced conductance neurons into part of the frontal cortex
of Spaun. Specifically, the part of frontal cortex that we changed is mapped
to OFC and acts as a memory that is responsible for keeping track of the task
currently being performed. As before, this introduction made no observable
functional difference across tasks (see Figures 6 and 7).
We subsequently introduced TTX to block about 20% of the sodium
channels in this frontal area. We examined the performance on two cognitive
tasks, the counting task and the list memory task. The counting tasks con-
sists of showing BioSpaun a starting digit (e.g. 3) and a count digit (e.g. 4),
and then asking for the result of counting by the count digit from the starting
digit (e.g. 7). Spaun replicated human reaction time data on this task. An
example run of the performance of BioSpaun on the counting task is shown
in this video https://youtu.be/FoOGqzG8_WU. As shown, BioSpaun is able
to encode the starting and count digits, and begins the internal counting pro-
cess, but that process is interrupted (i.e. the task state is forgotten) before
the counting completes. Consequently, BioSpaun produces no result. Figure
6 demonstrates performance across various numbers of counts. These results
demonstrate that the performance significantly worsens with the application
of TTX, both in terms of reaction time, and task completion. Notably, there
was a timeout of 2s to complete the task after the display of the last digit.
Consequently, the worst possible performance (i.e., not responding to the
task) was most common for 5 counts with TTX.
The second task we performed the same manipulation in is the list mem-
ory task. In the list memory task, BioSpaun is shown a list of digits (with
varying numbers of digits), which it must reproduce after a delay. The orig-
inal Spaun model captures primacy and recency effects in such a task. Per-
formance of BioSpaun under the influence of TTX is shown in this video
https://youtu.be/kpwoBccdmd8. In this case, BioSpaun is able to encode
the list correctly, and begins to write out digits but is interrupted before
completing the list. Performance across lists of various lengths is shown in
Figure 7.
Although the behavior across these tasks is in someways quite different,
11
Figure 6: Effect of TTX application on the counting task time to completion. The control
task reproduces the results of [1] where it was shown to statistically match human perfor-
mance times, while using detailed neurons. The effetcs of TTX application is shown by
the green line, which shows significantly worse timing, with many time outs. Error bars
are 95% confidence intervals.
as the model always produces some output in the list task, while often not
at all in the counting task, it is clear how these different behaviors are the
result of the same TTX-induced failure. That is, TTX consistently induces
a failure to retain the task state long enough to complete the given task.
While preliminary, this is a useful demonstration of how a single brain model
can explain the behaviorally different effects of a single underlying molecular
manipulation.
4. Conclusion
Many reasons have been offered as to why large-scale models are im-
portant to build. These include the ability to understand mysterious brain
disorders, from autism to addiction [25], to develop and test new kinds of
medical interventions, be they drugs or stimulation [26], and to provide a
way to organize and unify the massive amounts of data generated by the
neurosciences [27]. However, no question about brains seems to loom larger
than: How do brains control behavior? The vast majority of sophisticated
behavior is the result of the interactions between many brain areas, recruiting
12
Figure 7: Effect of TTX application on the list memory task. Top: Performance with no
TTX application reproduces the results reported in [1], which are statistically indistin-
guishable from human performance, while using detailed neurons. Bottom: The effects
of the same TTX application as in Figure 6.
In this case the recency effect is largely
destroyed, and after two items performance is at chance. Error bars are 95% confidence
intervals.
13
many millions of cells, each of which exhibits complex nonlinear dynamics.
Without constructing models that explore these complex interactions, and
how they relate to those dynamics, we are unlikely to be able to understand
how to help a distressed brain, or explain how biological mechanisms give
rise to cognitive behavior. In short, without large-scale models, we cannot
test large-scale hypotheses.
Until the Spaun model, past work on large-scale models has been surpris-
ingly silent on the connection between complex neural activity and observable
behavior. Even with the introduction of Spaun, the link between low-level
biophysical properties and cognitive behavior was only mildly elucidated be-
cause of the simplicity of the single cell models employed. With BioSpaun,
we have provided a preliminary but suggestive method for simulating brain
models that include extensive biophysical detail while not sacrificing a clear
connection to interesting behavior. We have only exploited this connection to
examine potential consequences of employing a drug whose effects are reason-
ably approximated through a simply mechanism (i.e.
local sodium channel
blockage). Nevertheless, we believe that this kind of model opens the pos-
sibility of examining how a wide variety of low-level biological interventions
can influence systemic behavior.
5. References
References
[1] C. Eliasmith, T. C. Stewart, X. Choo, T. Bekolay, T. DeWolf, Y. Tang,
D. Rasmussen, A large-scale model of the functioning brain, Science
338 (2012) 1202–1205.
[2] H. Markram, E. Muller, S. Ramaswamy, M. Reimann, M. Abdellah,
C. Sanchez, A. Ailamaki, L. Alonso-Nanclares, N. Antille, S. Arsever,
G. Kahou, T. Berger, A. Bilgili, N. Buncic, A. Chalimourda, G. Chin-
demi, J.-D. Courcol, F. Delalondre, V. Delattre, S. Druckmann, R. Du-
musc, J. Dynes, S. Eilemann, E. Gal, M. Gevaert, J.-P. Ghobril, A. Gi-
don, J. Graham, A. Gupta, V. Haenel, E. Hay, T. Heinis, J. Her-
nando, M. Hines, L. Kanari, D. Keller, J. Kenyon, G. Khazen, Y. Kim,
J. King, Z. Kisvarday, P. Kumbhar, S. Lasserre, J.-V. LeB´e, B. Mag-
alhaes, A. Merch´an-P´erez, J. Meystre, B. Morrice, J. Muller, A. Munoz-
C´espedes, S. Muralidhar, K. Muthurasa, D. Nachbaur, T. Newton,
14
M. Nolte, A. Ovcharenko, J. Palacios, L. Pastor, R. Perin, R. Ran-
jan, I. Riachi, J.-R. Rodr´ıguez, J. Riquelme, C. Rossert, K. Sfyrakis,
Y. Shi, J. Shillcock, G. Silberberg, R. Silva, F. Tauheed, M. Telefont,
M. Toledo-Rodriguez, T. Trankler, W. VanGeit, J. D´ıaz, R. Walker,
Y. Wang, S. Zaninetta, J. DeFelipe, S. Hill, I. Segev, F. Schurmann,
Reconstruction and Simulation of Neocortical Microcircuitry, Cell 163
(2015) 456–492.
[3] T. Wong, R. Preissl, P. Datta, F. Myron, R. Singh, S. Esser, E. Mc-
IBM
Quinn, R. Appuswamy, W. Risk, H. Simon, D. Modha, 1014,
Research Report: RJ10502(ALM1211-004) (2012).
[4] P. A. Merolla, J. V. Arthur, R. Alvarez-Icaza, A. S. Cassidy, J. Sawada,
F. Akopyan, B. L. Jackson, N. Imam, C. Guo, Y. Nakamura, B. Brezzo,
I. Vo, S. K. Esser, R. Appuswamy, B. Taba, A. Amir, M. D. Flickner,
W. P. Risk, R. Manohar, D. S. Modha, Artificial brains. A million
spiking-neuron integrated circuit with a scalable communication network
and interface., Science (New York, N.Y.) 345 (2014) 668–73.
[5] L. Sanders, Mind & brain: Model brain mimics human quirks: Com-
puter simulation turns decisions into plans for action, Science News 183
(2013) 13–13.
[6] C. Eliasmith, C. H. Anderson, Neural engineering: Computation, rep-
resentation and dynamics in neurobiological systems, MIT Press, Cam-
bridge, MA, 2003.
[7] J. Conklin, C. Eliasmith, An attractor network model of path integration
in the rat, Journal of Computational Neuroscience 18 (2005) 183–203.
[8] R. Singh, C. Eliasmith, Higher-dimensional neurons explain the tuning
and dynamics of working memory cells, Journal of Neuroscience 26
(2006) 3667–3678.
[9] A. Litt, C. Eliasmith, P. Thagard, Neural affective decision theory:
Choices, brains, and emotions, Cognitive Systems Research 9 (2008)
252–273.
[10] T. Bekolay, M. Laubach, C. Eliasmith, A spiking neural integrator
model of the adaptive control of action by the medial prefrontal cortex,
The Journal of Neuroscience 34 (2014) 1892–1902.
15
[11] B. P. Tripp, C. Eliasmith, Neural populations can induce reliable post-
synaptic currents without observable spike rate changes or precise spike
timing., Cerebral cortex (New York, N.Y. : 1991) 17 (2007) 1830–40.
[12] C. Parisien, C. H. Anderson, C. Eliasmith, Solving the problem of
negative synaptic weights in cortical models, Neural Computation 20
(2008) 1473–1494.
[13] C. Eliasmith, A unified approach to building and controlling spiking
attractor networks, Neural computation 17 (2005) 1276–1314.
[14] T. Stewart, C. Eliasmith, Large-Scale Synthesis of Functional Spiking
Neural Circuits, Proceedings of the IEEE 102 (2014) 881–898.
[15] C. Eliasmith, How to build a brain: A neural architecture for biological
cognition, Oxford University Press, New York, NY, 2013.
[16] E. Crawford, M. Gingerich, C. Eliasmith, Biologically plausible, human-
scale knowledge representation, Cognitive Science (2015).
[17] D. Rasmussen, C. Eliasmith, A spiking neural model applied to the
Intelligence 42
study of human performance and cognitive decline,
(2014) 53–82.
[18] P. Blouw, E. Solodkin, P. Thagard, C. Eliasmith, Concepts as semantic
pointers: A framework and computational model, Cognitive Science
(2015) 1–35.
[19] T. Stewart, F.-X. Choo, C. Eliasmith, Sentence processing in spik-
in: P. Bello,
ing neurons: A biologically plausible left-corner parser,
M. Guarini (Eds.), 36th Cognitive Science Conference, Quebec City.
[20] X. Choo, C. Eliasmith, General Instruction Following in a Large-Scale
Biologically Plausible Brain Model, in: 35th Annual Conference of the
Cognitive Science Society, Cognitive Science Society, pp. 322–327.
[21] J. Gosmann, C. Eliasmith, A spiking neural model of the n-back task,
in: 37th Annual Meeting of the Cognitive Science Society, pp. 812–817.
[22] T. Bekolay, J. Bergstra, E. Hunsberger, T. Dewolf, T. C. Stewart,
D. Rasmussen, X. Choo, A. R. Voelker, C. Eliasmith, Nengo: a Python
16
tool for building large-scale functional brain models., Frontiers in neu-
roinformatics 7 (2014) 48.
[23] A. Bahl, M. B. Stemmler, A. V. M. Herz, A. Roth, Automated optimiza-
tion of a reduced layer 5 pyramidal cell model based on experimental
data., Journal of neuroscience methods 210 (2012) 22–34.
[24] G. Stuart, N. Spruston, Determinants of voltage attenuation in neocor-
tical pyramidal neuron dendrites, J Neurosci 18 (1998) 3501–10.
[25] M. A. Just, T. A. Keller, V. L. Malave, R. K. Kana, S. Varma, Autism
as a neural systems disorder: a theory of frontal-posterior underconnec-
tivity, Neuroscience & Biobehavioral Reviews 36 (2012) 1292–1313.
[26] J. A. Beeler, M. J. Frank, J. McDaid, E. Alexander, S. Turkson, M. Sol
Bernandez, D. S. McGehee, X. Zhuang, A role for dopamine-mediated
learning in the pathophysiology and treatment of Parkinson's disease,
Cell reports (2012).
[27] J. Tyrcha, Y. Roudi, M. Marsili, J. Hertz, The effect of nonstationarity
on models inferred from neural data, Journal of Statistical Mechanics:
Theory and Experiment 2013 (2013) P03005.
17
|
1804.00404 | 3 | 1804 | 2019-02-16T06:33:45 | Three-dimensional alteration of neurites in schizophrenia | [
"q-bio.NC",
"physics.bio-ph"
] | This paper reports nano-CT analysis of brain tissues of schizophrenia and control cases. The analysis revealed that: (1) neuronal structures vary between individuals, (2) the mean curvature of distal neurites of the schizophrenia cases was 1.5 times higher than that of the controls, and (3) dendritic spines were categorized into two geometrically distinctive groups, though no structural differences were observed between the disease and control cases. The differences in the neurite curvature result in differences in the spatial trajectory and hence alter neuronal circuits. We suggest that the structural alteration of neurons in the schizophrenia cases should reflect psychiatric symptoms of schizophrenia. | q-bio.NC | q-bio | Translational Psychiatry 9, 85 (2019).
https://doi.org/10.1038/s41398-019-0427-4
Three-dimensional alteration of neurites in schizophrenia
Ryuta Mizutani1*, Rino Saiga1, Akihisa Takeuchi2, Kentaro Uesugi2, Yasuko Terada2,
Yoshio Suzuki3, Vincent De Andrade4, Francesco De Carlo4, Susumu Takekoshi5, Chie
Inomoto5, Naoya Nakamura5, Itaru Kushima6, Shuji Iritani6, Norio Ozaki6, Soichiro Ide7,
Kazutaka Ikeda7, Kenichi Oshima7, Masanari Itokawa7, and Makoto Arai7
1Department of Applied Biochemistry, Tokai University, Hiratsuka, Kanagawa
259-1292, Japan.
2Japan Synchrotron Radiation Research Institute (JASRI/SPring-8), Sayo, Hyogo
679-5198, Japan.
3Graduate School of Frontier Sciences, University of Tokyo, Kashiwa, Chiba 277-8561,
Japan
4Advanced Photon Source, Argonne National Laboratory, Lemont, IL 60439, USA.
5Tokai University School of Medicine, Isehara, Kanagawa 259-1193, Japan.
6Graduate School of Medicine, Nagoya University, Nagoya, Aichi 466-8550, Japan.
7Tokyo Metropolitan Institute of Medical Science, Setagaya, Tokyo 156-8506, Japan.
*Correspondence should be addressed to:
Department of Applied Biochemistry, Tokai University, Hiratsuka, Kanagawa 259-1292,
Japan
Phone: +81-463-58-1211; Fax: +81-463-50-2426
E-mail: [email protected]
1
Abstract
Psychiatric symptoms of schizophrenia suggest alteration of cerebral neurons. However,
the physical basis of the schizophrenia symptoms has not been delineated at the cellular
level. Here we report nanometer-scale three-dimensional analysis of brain tissues of
schizophrenia and control cases. Structures of cerebral tissues of the anterior cingulate
cortex were visualized with synchrotron radiation nanotomography. Tissue constituents
visualized in the three-dimensional images were traced to build Cartesian coordinate
models of tissue constituents, such as neurons and blood vessels. The obtained Cartesian
coordinates were used for calculating curvature and torsion of neurites in order to
analyze their geometry. Results of the geometric analyses indicated that the curvature of
neurites is significantly different between schizophrenia and control cases. The mean
curvature of distal neurites of the schizophrenia cases was approximately 1.5 times
higher than that of the controls. The schizophrenia case with the highest neurite
curvature carried a frame shift mutation in the GLO1 gene, suggesting that oxidative
stress due to the GLO1 mutation caused the structural alteration of the neurites. The
differences in the neurite curvature result in differences in the spatial trajectory and
hence alter neuronal circuits. It has been shown that the anterior cingulate cortex
analyzed in this study has emotional and cognitive functions. We suggest that the
structural alteration of neurons in the schizophrenia cases should reflect psychiatric
symptoms of schizophrenia.
Introduction
Schizophrenia is a chronic mental disorder that affects approximately 1% of the
population.1 Clinical symptoms of schizophrenia include hallucinations, delusions,
emotional disorders, and cognitive dysfunction. The development of these symptoms
suggests alterations in the connectivity between cerebral neurons. It has been reported
that dendritic spines of neurons are significantly decreased in the external pyramidal layer
of the cerebral cortex of schizophrenic brains.2-4 Since dendritic spines form the majority
of excitatory synapses, the loss of spines can directly impair neuronal connectivity. The
reduced neuropil hypothesis5 posits that reductions in neuron size and arborization are the
explanation for the reduced brain volume observed in schizophrenia.6-9 The reductions in
neuron size and arborization can perturb the neuronal structures, resulting in changes to
the neuronal circuits. However, studies of brain tissues of schizophrenia patients have
mainly been performed using two-dimensional images of tissue sections, whereas the
neurons themselves are three-dimensional in nature.
2
It has been reported that three-dimensional structures of brain tissues can be analyzed
with electron microscopy by reconstructing them from serially sectioned images.10-12
Since soft tissues are deformed by sectioning, the deformations are artificially corrected
in the three-dimensional reconstruction.13,14 Therefore, three-dimensional image
reconstructed from serial sections does not exactly reproduce the three-dimensional
structure of the tissue. Another method to visualize the three-dimensional structure of
biological tissue is confocal light microscopy. However, light microscopy cannot
visualize structures behind opaque objects. Its resolution is three-dimensionally
anisotropic and depends on the direction of the optical axis.15 Although neuronal
structures have been deposited in the NeuroMorpho.Org database16 including those of
human cerebral cortex,17 three-dimensional coordinates estimated from light microscopy
images show irregular displacements especially along the optical axis and are not
applicable to geometric analyses. Thus, resolution anisotropy, tissue opacity, and
sectioning deformation can degrade the three-dimensional features that may be relevant
to schizophrenia.
In this study, we analyzed brain tissue structures of schizophrenia patients and
control cases with synchrotron radiation nanotomography.18,19 X-ray microtomography
and nanotomography20-25 can visualize three-dimensional opaque objects with nearly
isotropic resolution.26 Its reconstruction process does not involve any deformation
correction, and hence, the obtained image reproduces the actual three-dimensional
structures. Although the low x-ray contrast of the brain tissue itself initially limited
visualizations to those of large-scale structures,27 the use of high-Z element staining has
since allowed neuronal networks in brain tissue to be observed.28,29 Tissue constituents
visualized in this study were traced to build three-dimensional Cartesian coordinate
models of tissue structures, such as neurons and blood vessels. The three-dimensional
tissue structures were reproduced as Cartesian coordinates through this process. The
resultant neuronal coordinates can be used for analyzing the geometry of neurons in
schizophrenia and control cases.
Cerebral tissues analyzed in this study are those of the anterior cingulate cortex
(Brodmann area 24). It has been shown that this brain area has emotional and cognitive
function.30,31 It has also been reported that the anterior cingulate cortex is related to
schizophrenia,32,33 attention-deficit/hyperactivity disorder,34 and obsessive-compulsive
disorder.35 Therefore, the structural character of neurons of the anterior cingulate cortex
can affect the mental capabilities or the psychiatric symptoms. In this study, we
3
examined differences of neuronal structures between schizophrenia and control cases
and also between control individuals.
Materials and Methods
Cerebral tissue samples
All post-mortem human cerebral tissues were collected with informed consent
from the legal next of kin using protocols approved by the clinical study reviewing
board of Tokai University School of Medicine (application no. 07R-018) and the ethics
committee of Tokyo Metropolitan Institute of Medical Science (approval no. 17-18).
This study was conducted under the approval of the ethics committee for the human
subject study of Tokai University (approval nos. 11060, 11114, 12114, 13105, 14128,
15129, 16157, and 18012). Schizophrenia patients S1 -- S4 (Supplementary Table S1)
were diagnosed according to the DSM-IV codes with the consensus of at least two
experienced psychiatrists. Control patients (Supplementary Table S1) had been
hospitalized because of a traffic injury (N1) or non-psychiatric lethal diseases (N2 -- N4)
and were not psychiatrically evaluated. Since the cause of death of the N1 case was
damage to the heart, histological changes in brain tissue specific to the injury can be
excluded. The number of cases was determined in consideration of the available
beamtime at the synchrotron radiation facilities. The control cases were selected so as to
match the gender and age of the schizophrenia cases. Cases in which hemorrhage,
infarction, or neoplasm were observed in the histological assessment of cerebral tissues
were excluded. No previous records of schizophrenia were found for the control cases.
The cerebral tissues of the anterior cingulate cortex (Brodmann area 24) were collected
from the left hemispheres of the biopsied brains and subjected to Golgi impregnation, as
described previously,29 in order to visualize neurons in x-ray images.
The Golgi-stained tissues were first soaked in neat ethanol, then in n-butyl-glycidyl
ether, and finally in Petropoxy 154 (Burnham Petrographics, USA) epoxy resin, as
described previously.36 The resin-soaked tissues were cut into rod shapes with
approximate widths of 0.5 mm and lengths of 3 -- 5 mm under a stereomicroscope and
then transferred to borosilicate glass capillaries (W. Müller, Germany) filled with resin.
The capillary diameter was approximately 0.8 mm. The capillaries were incubated at
90°C for 70 -- 90 hours for curing the resin.
Nanotomography
The distal end of the capillary sample was sleeved with a brass tube using epoxy
glue and secured with a setscrew to a brass or invar adapter specially designed for
4
nanotomography. The mounted samples were placed in the experiment hutch as soon as
possible in order to equilibrate their temperature with that of the sample stage.
Nanotomography experiments using Fresnel zone plate optics were performed at the
BL37XU18 and BL47XU37 beamlines of the SPring-8 synchrotron radiation facility and
at the 32-ID beamline19 of the Advanced Photon Source (APS) of the Argonne National
Laboratory. Since the nanotomography experiments were performed in the manner of
local computed tomography (CT), only the region of interest within the viewing field
(64 -- 122 μm diameter; Supplementary Table S2) was visualized.
In the experiments at the SPring-8 beamlines, the tissue samples were mounted on
a slide-guide rotation stage specially built for nanotomography (SPU-1A, Kohzu
Precision, Japan). Transmission images were recorded with a CMOS-based imaging
detector (ORCA-Flash4.0, Hamamatsu Photonics, Japan) using monochromatic
radiation at 8 keV. Examples of raw images and their reconstructed slices are shown in
Supplementary Figure S1. Octagonal sector condenser zone plates37 were used as beam
condensers. The photon flux at the sample position of the BL37XU optics of the 2017.4
setup was estimated to be 1.0 × 1014 photons/mm2/s using Al2O3:C dosimeters
(Nagase-Landauer, Japan). Since the specific gravity of Petropoxy 154 is 1.18 and its
attenuation coefficient was estimated to be 6.1 cm-1 at 8 keV, this x-ray flux
corresponds to an absorbed radiation dose by the resin of 7.0 × 104 Gy/s. In the
nanotomography experiments at the APS beamline, the tissue samples were mounted on
an air-bearing rotation stage (UPR-160AIR, PI miCos, Germany) during the 2016.6
beamtime, or on a motorized model 4R Block-Head air bearing spindle (Professional
Instruments Company, USA) during the 2017.10 beamtime. Zernike phase-contrast
images were recorded with a CMOS-based imaging detector (GS3-U3-51S5M-C, FLIR,
USA) using monochromatic radiation at 8 keV. A polygonal beam shaping condenser
and compound refractive lenses were used as beam condensers.19 Spatial resolutions
were estimated using three-dimensional square-wave patterns38 or from the Fourier
domain plot.39 The experimental conditions are summarized in Supplementary Table S2.
The data collection procedure is described in the Supplementary Materials and Methods.
While absorption contrast was sufficient for visualization of the neurons in the
SPring-8 BL37XU experiments, the Zernike phase-contrast method was used in order to
enhance the sample contrast in the experiments at the BL47XU beamline of SPring-8
and at the 32-ID beamline of APS. Although a number of contrast-enhancing methods
have been reported,20,27,40 methodological prerequisites, such as the strict requirement of
x-ray coherence in the holography, pose limitations in real applications. In this study,
approximately 4 × 105 images of human brain tissues were collected in order to examine
5
the statistical significance of their structural differences. Hence, practical aspects
including the time efficiency and the methodological tolerance against the beam
fluctuation should be taken into account. For these reasons, we used the Zernike method
to enhance the sample contrast.
Tomographic reconstruction
Tomographic slices perpendicular to the sample rotation axis were reconstructed
with the convolution-back-projection method using the RecView software.36 The
reconstruction calculation was performed by RS. Image pixels of APS datasets were
averaged by 2 × 2 binning prior to the tomographic reconstruction, since the original
pixel size of 26 nm was approximately half the pixel sizes of the SPring-8 experiments
(40.2, 48.3 and 59.6 nm, respectively; Supplementary Table S2) and the binned pixel
size of 52 nm was sufficiently fine compared to the spatial resolution of the APS
datasets (300 nm). Since the reconstruction of the SPring-8 datasets was performed
without binning, the reconstructed voxel size was the same as the original pixel size.
Multiple image sets taken by shifting the sample were aligned and stacked to obtain the
entire three-dimensional image. The entire reconstructed volume of each dataset was
subjected to further analysis.
Data blinding
Nanotomography datasets were analyzed with the role allotment of data
management to RS and data analysis to RM. The data manager reconstructed
tomographic slices of all datasets and coded the dataset name. The data manager
managed the information regarding the case assignment of each dataset but had no
access to the analysis results except for an indicator of the analysis amount. The data
analyst built Cartesian coordinate models from the datasets but had no access to the case
information. The whole model building procedure was performed by RM in order to
keep the modeling quality constant. The detailed protocol is described in the
Supplementary Materials and Methods.
Cartesian coordinate models and geometric analysis
Cartesian coordinate models were built in four steps: (1) manual assignment of
large structural constituents such as somata and blood capillary vessels, (2) automatic
tracing to build a computer generated model and its subsequent refinement, (3),
examination of the entire three-dimensional image and manual intervention to modify
the working model, and (4) final structural refinement. The geometric analysis was
6
performed in three steps: (1) cell typing, (2) structure annotation, and (3) geometric
parameter calculation. Detailed protocols are described in Supplementary Materials and
Methods.
Code availability
The RecView software36 that was used for the tomographic reconstruction is
available from https://mizutanilab.github.io under the BSD 2-Clause License. The
model building and geometric analysis procedures were implemented in the MCTrace
software41 available from https://mizutanilab.github.io under the BSD 2-Clause License.
Results
Three-dimensional tissue structure
The cerebral cortex tissues analyzed in this study were taken from the left
hemispheres of autopsied brains of four schizophrenia cases S1 -- S4 (age: mean 65 ±
standard deviation 6 yr; 2 males and 2 females) and four age/gender-matched control
cases N1 -- N4 (64 ± 5 yr; 2 males and 2 females). The N1 tissue may have suffered from a
prolonged post-mortem time (85 h; Supplementary Table S1) compared with the other
cases. Three-dimensional structures of these cerebral tissues were visualized with
nanotomography18,19 (Fig. 1a and Supplementary Figure S2) at spatial resolutions of
180 -- 300 nm (Supplementary Table S2). Tissue constituents, such as neurites and blood
vessels visualized in the 55 obtained datasets, were traced to build three-dimensional
Cartesian coordinate models of the tissues (Fig. 1b -- d and Supplementary Figs. S3 -- S4).
An example of a spiny dendrite superposed on a three-dimensional map is shown in Fig.
1e. Representative images and models of all datasets are shown in Supplementary Figs.
S2 -- S4. The statistics of the datasets and models are summarized in Supplementary Table
S3. Coordinate files are available from https://mizutanilab.github.io
(RRID:SCR_016529).
7
Figure 1. Three-dimensional visualization of cerebral cortex neurons and their models
represented with Cartesian coordinates. The pial surface is toward the top. The
three-dimensional image was rendered with VGStudio (Volume Graphics, Germany).
8
The models were drawn using MCTrace.41 Structural constituents of the model are
color-coded. Nodes composing each constituent are indicated with circles. Dots indicate
soma nodes. (a) Rendering of dataset N2C of the control N2 tissue. Voxel values of
500 -- 1600 were rendered with the scatter HQ algorithm. Scale bar: 20 μm. (b) Initial
model. Structures of somata and thick neurites were built manually in order to mask
them in the subsequent automatic model generation. (c) Automatically generated model
of the tissue structure. Neurites were searched by calculating the gradient vector flow61
throughout the image. The neurites found in the search were then traced using a
three-dimensional Sobel filter.62 (d) The computer-generated model was manually
examined and edited according to the method used in protein crystallography. The
obtained working model was refined with conjugate gradient minimization. The
geometric parameters were calculated from the three-dimensional Cartesian coordinates
of the refined model. (e) Stereo drawing of a spiny dendrite indicated with the arrow
head in panel d. The structure is superposed on a three-dimensional map of the observed
image. The map drawn in gray is contoured at 2.5 times the standard deviation (2.5 σ)
of the image intensity with a grid size of 96.6 nm.
Neurites in three-dimensional tissue can be regarded as three-dimensional curves. A
three-dimensional curve can be represented with two parameters: curvature and torsion.
Curvature corresponds to the reciprocal of the radius of the curve. Torsion represents the
deviation of the curve from a plane. The neurites were divided into segments at each
ramification point. The geometries of these neurite segments were analyzed by evaluating
their curvature and torsion (Table 1). Spine structures were analyzed using the parameters
of length, minimum radius, and maximum radius in addition to the curvature and torsion
(Supplementary Table S4), since dendritic spines have been classified into several
categories in terms of their neck width and length.42,43 Spine density was defined as the
number of spines per total length of spiny dendrites (Supplementary Tables S1 and S3).
Neurite structures
The curvatures and torsions of the neurite segments are summarized in Table 1. A
total of 2737 neurite segments from the schizophrenia cases and 2254 segments from the
control cases were analyzed. Figures 2a and 2b show the distributions of the curvature
and torsion of the segments. The curvature distribution of the schizophrenia cases
9
exhibited long upper tails (Fig. 2a), showing a 45% increase on average. This resulted in a
larger standard deviation in the curvature of the neurite segments of the schizophrenia
cases (Table 1; 0.28 -- 0.36 μm-1) in comparison with that of the control cases (0.21 -- 0.23
μm-1). The curvature median exhibited a significant difference even between the four
schizophrenia cases (p < 2.2 × 10-16 with the Kruskal-Wallis test) and between the four
control cases (p < 2.2 × 10-16). In contrast, the torsion showed no apparent difference
between the schizophrenia and control cases (Fig. 2b). There was no significant
difference in torsion median between all the cases (p = 0.44 with the Kruskal-Wallis test).
The torsion distribution of every case has a peak at the origin, indicating that the neurites
have no chiral bias (Fig. 2b).
Table 1. Geometric parameters of neurites.
Case
Curvature (μm-1)
Torsion (μm-1)
Total
0.46 (0.28) / 523
0.47 (0.32) / 754
0.60 (0.34) / 435
0.71 (0.36) / 880
0.33 (0.22) / 415
0.44 (0.21) / 731
0.37 (0.21) / 491
0.41 (0.23) / 432
Orphan neurite
0.58 (0.30) / 288
0.59 (0.34) / 450
0.78 (0.32) / 238
0.79 (0.36) / 700
0.38 (0.21) / 154
0.49 (0.21) / 422
0.46 (0.21) / 289
0.48 (0.24) / 252
S1
S2
S3
S4
N1
N2
N3
N4
Values represent mean (sample standard deviation) / number of observations. S1 -- S4 are
the schizophrenia cases, and N1 -- N4 are the control cases. The orphan neurite column
represents statistics of neurites whose somata were outside of the viewing field.
-0.03 (0.35) / 513
-0.02 (0.37) / 742
0.01 (0.41) / 426
0.00 (0.33) / 873
0.02 (0.44) / 389
0.01 (0.32) / 721
0.01 (0.27) / 484
-0.03 (0.34) / 426
10
Figure 2. (a) Distribution of neurite curvature. Quartiles are indicated with bars. (b)
Distribution of neurite torsion. (c) Distribution of curvature of orphan neurites without
soma in the viewing field. (d) Distribution of curvature of neurites whose somata were
visualized within the image. (e) Relative frequency of neurite in each 0.1 μm-1 bin of
curvature. The schizophrenia S1 case is plotted in red, S2 in orange, S3 in cyan, and S4
in purple. (f) Neurite curvature of the control N1 case is plotted in red, N2 in orange, N3
in cyan, and N4 in purple.
11
The neurites whose somata were visualized in the image are proximal ones within a
viewing field width of 64 -- 122 μm (Supplementary Table S2). Besides those laterally
proximal structures, neurites without soma were also visualized. These orphan neurites
should be distal parts of neuronal arbors whose somata were out of the viewing field.
These two categories of neurites are separately plotted in Figs. 2c and 2d. The orphan
neurites of the schizophrenia cases showed a wide curvature distribution, a 51% increase
on average compared with the controls (Table 1). The mean curvatures of the orphan
neurites were significantly different between the schizophrenia and control cases
(Welch's t-test, p = 0.020, 4 schizophrenia and 4 controls). These results indicate that
distal neurites have high curvature in the case of schizophrenia.
The profiles shown in Figures 2a -- d vary between cases, even within the control
cases, indicating that neuronal structures vary between individuals. The curvature
median showed significant differences between the four control cases and also between
the four schizophrenia cases, as described above. Figure 2e -- f plots the relative frequency
of the neurite curvatures. The frequency profiles were not identical between cases (Fig. 2e
and 2f). The structures of the S2, S3, N1, and N3 cases were analyzed using multiple
samples. The relative frequencies of the curvature in these multiple samples are
separately plotted in Supplementary Figure S5. Multiple samples from the same
individual have similar profiles, except that two samples of the N1 case show differences
probably due to its long post-mortem time. These results suggest that the neuronal
structures have features characteristic of each individual. A difference in neurite
curvature results in a difference in the spatial trajectory and hence alters the neuronal
circuits. The tissues analyzed in this study were taken from the anterior cingulate cortex.
It has been shown that the anterior cingulate cortex has emotional and cognitive
functions.30,31 Therefore, the structural differences of the neurons of the anterior
cingulate cortex should represent mental individuality or the psychiatric states of
individuals.
Figures 3a and 3b, along with Supplementary Figure S6, show scatter plots of the
mean curvature of the neurite trajectory and the mean radius of the neurite. The plots
indicate reciprocal relationships between the trajectory curvature and the neurite radius.
The schizophrenia cases show a wide distribution (Fig. 3a) for both spiny dendrites and
smooth neurites, while the control cases show a narrow distribution (Fig. 3b). These plots
indicate that the high-curvature neurites of the schizophrenia cases have a short radius.
Figure 3c shows representative structures of the neurites of the schizophrenia and
control cases. The neurites of the schizophrenia cases exhibit frequent changes in
12
direction (Fig. 3c), resulting in tortuous structures. The S4 neurite is thinner than the other
neurites. In contrast, the neurites of the control cases show gradual and broad curves. The
tissue structures of the S4 case and its age/gender-matched control N4 case are shown in
Fig. 3d and 3e, respectively (also in Supplementary Videos S1 and S2). The structure of
the S4 case is frizzy, whereas that of the N4 case is mostly straight. The S4 patient
showed severe schizophrenic symptoms and bore a frame shift mutation in the GLO1
gene.44,45 It has been shown that the GLO1 mutation can cause oxidative stress.46
Therefore, the structural alteration of the neurites of the S4 case is ascribable to oxidative
stress due to the GLO1 mutation.
13
Figure 3. Neurites in schizophrenia and control cases. (a) Scatter plot of curvature and
radius of neurites in the schizophrenia S2 case. (b) Scatter plot of control N2 case. The
horizontal axis represents the mean radius of the neurite as a fiber. Thin neurites are on
the left, and thick ones are on the right. The vertical axis represents the mean curvature
of the neurite trajectory. Spiny dendrites are indicated with red dots and smooth neurites
with blue. Neurites of which mean radii are larger than 3 μm are omitted. (c) Neurite
14
segments showing median curvature in the top quartile of each case. Mean curvature of
each neurite is shown in parenthesis. The neurite of N1 is a branch on an apical dendrite
of a pyramidal neuron. Others are orphan neurites of which somata are not visualized
within the image. (d) Schizophrenia S4A structure. (e) Control N4A structure. Panels d
and e were produced by placing the soma node of the largest pyramidal neuron at the
figure center. The pial surface is toward the top. Structures are color-coded. Scale bars:
5 μm.
Dendritic spine structures
The geometric analyses of the dendritic spines are summarized in Supplementary
Table S4. A total of 15116 spines of schizophrenia cases and 12885 spines of control
cases were analyzed. It has been reported that dendritic spines can be categorized into
several groups, such as mushroom spine and stubby spine.42 Thin-necked and long spines
can be characterized with a small minimum radius and long length, while stubby spines
can be characterized with a large minimum radius and short length. Scatter plots of the
spine length and minimum radius are shown in Fig. 4 and Supplementary Fig. S7. The
plots of the schizophrenia and control cases have similar profiles composed of a
wedge-shape domain and a triangle domain (Fig 4a, 4c, Supplementary Fig. S7). This
indicates that dendritic spines can be categorized into two groups. Spines in the wedge
domain exhibit short lengths and hence correspond to stubby spines. Their radius exhibits
a nearly linear correlation to the length. Spines of the triangle domain exhibit long lengths
and comparably short minimum radii and hence correspond to necked mushroom spines.
The triangle distribution indicates that a longer spine has a thinner neck (Fig. 4a and 4c).
The presence of a neck affects the ratio between maximum and minimum radii. However,
the scatter plots of the radius ratio and spine length (Fig. 4b, 4d, Supplementary Fig. S7)
show no distinctive domains. This indicates that the neck morphology is continuous.43
The schizophrenia and control cases show similar profiles in all the plots, suggesting that
structures of dendritic spines in the schizophrenia cases analyzed in this study were the
same as those of the controls.
15
Figure 4. (a) Scatter plot of spine parameters in the schizophrenia S4 case. Minimum
node radius is plotted against length. Three outliers (length/radius = 7.51/0.13, 6.24/0.10,
and 6.23/0.10) were omitted. (b) Radius ratio between maximum / minimum radii of S4
is plotted against length. Four outliers (length/ratio = 7.51/2.77, 6.24/2.90, 6.23/3.70,
and 3.55/7.43) were omitted. (c) Minimum radius of N2. An outlier (length/radius =
7.88/0.14) was omitted. (d) Radius ratio of N2. An outlier (length/ratio = 7.88/3.14) was
omitted. (e) Distribution of spine curvature. Quartiles are indicated with bars. (f)
Distribution of spine torsion.
16
It has been reported that spine density in the external pyramidal layer (layer III) of the
frontal cortex2 and the temporal cortex3 is significantly lower in schizophrenia cases than
in control cases. It has also been reported that the difference in spine density is not
significant in the internal pyramidal layer of the frontal cortex47 or in the external
pyramidal layer of the occipital cortex.2 In this study, we analyzed spine structures mainly
in the internal pyramidal layer of the anterior cingulate cortex. The obtained spine density
per dendritic length is summarized in Supplementary Table S1. The spine density was
comparable to those observed in the Golgi-stained frontal cortex.48 No significant
difference in mean spine density was found between the four schizophrenia and four
control cases (Welch's t-test, p = 0.56), though the sample size of this study is small. The
spine density of the N1 case was lower than in the other cases, presumably due to the long
post-mortem time. Except for this case, the curvatures and torsions of the spines showed
similar distribution profiles (Fig. 4e and 4f), suggesting that the schizophrenia and control
cases analyzed in this study share common spine structures.
Discussion
In the S4 schizophrenia case, the high curvature and short radius of the neurites
should stem from the GLO1 frameshift mutation.45 Although similar structural alterations
were also observed in the S1, S2, and S3 schizophrenia cases, their causes are not clear at
present. Some adverse effects similar to oxidative stress of the GLO1 mutation should
have degraded the neuronal structures. Tortuous neurites have been observed in the
cerebral cortex of schizophrenic brain.49 It has been reported that the
schizophrenia-susceptible DISC1 protein50 interacts with a number of factors associated
with neuronal functions.51 Apical dendrites of dentate gyrus neurons showed
morphological alterations in mice carrying a DISC1 mutation.52 A significant decrease in
dendritic diameter was reported for a Shn2 knockout mouse with schizophrenia-like
symptoms.53 The disruption of any susceptible genes related to the neuronal structure or
environmental risk factors that affect brain development can alter geometry of neurites.
N-methyl-D-aspartate (NMDA) receptor antagonists including phencyclidine cause
psychiatric symptoms similar to those of schizophrenia.54 A corkscrew deformity of
dendrites was reported in an animal model of the NMDA receptor hypofunction.55 These
reports suggest that the high-curvature neurites observed in this study are related to
schizophrenia symptoms. However, we cannot exclude the possibility that antipsychotics
affected the neuronal structures. Such drug effects can be elucidated by using
nanotomography to analyze brain tissues of drug-treated animals.
17
The neurite of the schizophrenia cases showed higher curvature and shorter radius
compared with the controls. The radius of the neurite affects its conductivity, resulting in
altered connections between neurons. This neurite thinning should have a relationship to
the tissue volume reduction observed in schizophrenia.6-8 It has been suggested that a
neurodevelopmental defect in the neuropil can explain the loss of cortical volume
without loss of neurons.56 The curvature of a neurite determines its spatial trajectory. A
curve with a higher curvature reaches more positions in its three-dimensional vicinity, but
needs to be longer to reach a distal position compared with a straight line. This can alter
the connectivity of the neuronal circuit. It would be difficult to relieve or restore these
nanometer-scale structural alterations of the tissue. Therefore, the deteriorative outcome
of structurally altered neurons should be prevented in advance of their incorporation into
the neuronal circuit. The results obtained in this study hence support the consensus that
early diagnosis and treatment is important for a better prognosis of schizophrenia.
A disadvantage of x-ray visualization of brain tissue is that neurons show little
contrast in x-ray images, since they are composed of light elements. In this study, cerebral
tissues were stained with Golgi impregnation in order to label their neurons with silver.
Therefore, the obtained results are tempered by the limitations of the staining method.
Since only a small number of neurons are stochastically visualized in Golgi impregnation,
the labelled neurons are limited representatives of the neuronal population.57 The viewing
field width and number of cases are also limitations. Although hundreds of neurites and
thousands of spine structures were analyzed for each case, the results reported in this
paper are of millimeter-sized tissues of the anterior cingulate cortex of four schizophrenia
and four control cases. Another limitation of our case-control study is that the controls
were not psychiatrically evaluated. There is a possibility that the control cases had latent
mental diseases, although no geometric hallmark of schizophrenia was observed in them.
Biological individuality is encoded in the genome. At the macroscopic level,
individuals are identified from the body structure, such as face, fingerprints, or overall
brain connectivity.58 However, at the cellular level, little evidence of mental personality
has been delineated on the basis of neuronal structures, though histological studies
suggested differences of cellular structures between individuals.59,60 The results reported
in this paper reveal differences in the tissue structure of the anterior cingulate cortex
between schizophrenia and control cases, and also between control individuals. This
suggests that geometric profiles of brain tissue are different between individuals.
Structural differences of cerebral neurons can result in differences in brain circuits, and
hence affect the mental individuality.
18
Human mental activities are performed through coordination of many diverse areas of
the brain and cannot be explained from one study of a single area. The temporal lobe has
been reported to show a tissue volume reduction in schizophrenia9, and hence, it should
be analyzed with the same strategy. The results of this study also suggest that humans
have nanometer- to micrometer-scale structural diversity in the cerebral cortex. Although
the present sample size was sufficient for evaluating the statistical significance of the
structural difference, schizophrenia is a complex and heterogeneous psychiatric
disorder.1,9 Therefore, the differences observed in this study should be re-examined by
analyzing more three-dimensional structures of cerebral tissues of schizophrenia and
control cases. Such further analyses will lead to a better understanding of our mental
individuality and to better diagnosis and treatment of schizophrenia.
Conflict of interest
The authors declare no conflict of interest.
Acknowledgements
We are grateful to Prof. Motoki Osawa and Akio Tsuboi (Tokai University School of
Medicine) for their generous support of this study. We are also grateful to Prof. Yasuo
Ohashi (Chuo University; Statcom Co., Ltd.) for his helpful advice regarding the
statistical tests. We appreciate Prof. Yoshiro Yamamoto (Tokai University) for his
helpful advice regarding the data analysis. We also appreciate Dr. Jun Horiuchi (Tokyo
Metropolitan Institute of Medical Science) for his suggestions regarding the manuscript.
We thank Noboru Kawabe (Support Center for Medical Research and Education, Tokai
University) for assistance in preparing the histology sections. We also thank the
Technical Service Coordination Office of Tokai University for assistance in preparing
adapters for nanotomography. This work was supported by Grants-in-Aid for Scientific
Research from the Japan Society for the Promotion of Science (nos. 21611009,
25282250, and 25610126). The synchrotron radiation experiments at SPring-8 were
performed with the approval of the Japan Synchrotron Radiation Research Institute
(JASRI) given to R.M. (proposal nos. 2011B0034, 2012B0034, 2013A0034,
2013B0034, 2013B0041, 2014A1057, 2015A1160, 2015B1101, 2017A1143,
2018A1164, and 2018B1187). The synchrotron radiation experiments at the Advanced
Photon Source of Argonne National Laboratory were performed during the 2016-2 and
2017-3 runs under General User Proposal GUP-45781 by R.M. This research used
resources of the Advanced Photon Source, a U.S. Department of Energy (DOE) Office
of Science User Facility operated for the DOE Office of Science by Argonne National
19
Laboratory under Contract No. DE-AC02-06CH11357.
References
1. Carpenter WT Jr., Buchanan RW. Schizophrenia. N Engl J Med 1994; 330: 681 --
690.
2. Glantz LA, Lewis DA. Decreased dendritic spine density on prefrontal cortical
pyramidal neurons in schizophrenia. Arch Gen Psychiatry 2000; 57: 65 -- 73.
3. Sweet RA, Henteleff RA, Zhang W, Sampson AR, Lewis DA. Reduced dendritic
schizophrenia.
subjects with
density
spine
of
Neuropsychopharmacology 2009; 34: 374 -- 389.
auditory
cortex
in
4. Moyer CE, Shelton MA, Sweet RA. Dendritic spine alterations in schizophrenia.
Neurosci Lett 2015; 601: 46 -- 53.
5. Selemon LD, Goldman-Rakic PS. The reduced neuropil hypothesis: a circuit based
model of schizophrenia. Biol Psychiatry 1999; 45: 17 -- 25.
6. Wright IC et al. Meta-analysis of regional brain volumes in schizophrenia. Am J
Psychiatry 2000; 157: 16 -- 25.
7. Olabi B et al. Are there progressive brain changes in schizophrenia? A
meta-analysis of structural magnetic resonance imaging studies. Biol Psychiatry
2011; 70: 88 -- 96.
8. Haijma SV et al. Brain volumes in schizophrenia: a meta-analysis in over 18,000
subjects. Schizophr Bull 2013; 39: 1129 -- 1138.
9. Bakhshi K, Chance SA. The neuropathology of schizophrenia: A selective review
of past studies and emerging themes in brain structure and cytoarchitecture.
Neuroscience 2015; 303: 82 -- 102.
10. Helmstaedter M. Cellular-resolution connectomics: challenges of dense neural
circuit reconstruction. Nat Methods 2013; 10: 501 -- 507.
11. Peddie CJ, Collinson LM. Exploring the third dimension: volume electron
microscopy comes of age. Micron 2014; 61: 9 -- 19.
12. Hildebrand DGC, Cicconet M, Torres RM, Choi W, Quan TM, Moon J et al.
Whole-brain serial-section electron microscopy in larval zebrafish. Nature 2017;
545: 345 -- 349.
13. Briggman KL, Helmstaedter M, Denk W. Wiring
specificity
in
the
direction-selectivity circuit of the retina. Nature 2011; 471: 183 -- 188.
14. Hanslovsky P, Bogovic JA, Saalfeld S. Image-based correction of continuous and
section microscopy.
serial
in
discontinuous non-planar axial distortion
Bioinformatics 2017; 33: 1379 -- 1386.
20
15. Wilson T. Resolution and optical sectioning in the confocal microscope. J Microsc
2011; 244: 113 -- 121.
16. Ascoli GA. Mobilizing the base of neuroscience data: the case of neuronal
morphologies. Nat Rev Neurosci 2006; 7: 318 -- 324.
17. Watson KK, Jones TK, Allman JM. Dendritic architecture of the von Economo
neurons. Neuroscience 2006; 141: 1107 -- 1112.
18. Suzuki Y, Takeuchi A, Terada Y, Uesugi K, Mizutani R. Recent progress of hard
x-ray imaging microscopy and microtomography at BL37XU of SPring-8. AIP
Conference Proceedings 2016; 1696: 020013.
19. De Andrade V et al. Nanoscale 3D imaging at the Advanced Photon Source. SPIE
Newsroom 2016; doi: 10.1117/2.1201604.006461.
20. Takeuchi A, Uesugi K, Takano H, Suzuki, Y. Submicrometer-resolution
three-dimensional imaging with hard x-ray imaging microtomography. Rev Sci
Instrum 2002; 73: 4246 -- 4249.
21. Schroer CG et al. Nanotomography based on hard x-ray microscopy with refractive
lenses. Appl Phys Lett 2002; 81: 1527 -- 1529.
22. Chu YS et al. Hard-x-ray microscopy with Fresnel zone plates reaches 40nm
Rayleigh resolution. Appl Phys Lett 2008; 92: 103119.
23. Stampanoni M et al. Broadband X-ray full field microscopy at a superbend. J Phys
Conf Ser 2009; 186: 012018.
24. Kaira CS et al. Probing novel microstructural evolution mechanisms in aluminum
characterization. Adv Mater 2017; 29:
alloys using 4D nanoscale
doi:10.1002/adma.201703482.
25. Müller S et al. Quantification and modeling of mechanical degradation in
lithium-ion batteries based on nanoscale imaging. Nat Commun 2018; 9: 2340.
26. Mizutani R et al. Submicrometer tomographic resolution examined using a
micro-fabricated test object. Micron 2010; 41: 90 -- 95.
27. Beckmann F, Bonse U, Busch F, Günnewig O. X-ray microtomography (microCT)
using phase contrast for the investigation of organic matter. J Comput Assist
Tomogr 1997; 21: 539 -- 553.
28. Mizutani R, Takeuchi A, Hara T, Uesugi K, Suzuki Y. Computed tomography
imaging of the neuronal structure of Drosophila brain. J Synchrotron Radiat 2007;
14: 282 -- 287.
29. Mizutani R et al. Three-dimensional microtomographic imaging of human brain
cortex. Brain Res 2008; 1199: 53 -- 61.
30. Botvinick M, Nystrom LE, Fissell K, Carter CS, Cohen JD. Conflict monitoring
21
versus selection-for-action in anterior cingulate cortex. Nature 1999; 402: 179 -- 181.
31. Bush G, Luu P, Posner MI. Cognitive and emotional influences in anterior cingulate
cortex. Trends Cogn Sci 2000; 4: 215 -- 222.
32. Bouras C, Kövari E, Hof PR, Riederer BM, Giannakopoulos P. Anterior cingulate
cortex pathology in schizophrenia and bipolar disorder. Acta Neuropathol 2001;
102: 373 -- 379.
33. Fornito A, Yücel M, Dean B, Wood SJ, Pantelis C. Anatomical abnormalities of the
anterior cingulate cortex in schizophrenia: bridging the gap between neuroimaging
and neuropathology. Schizophr Bull 2009; 35: 973 -- 993.
34. Bush
et
al.
G
in
attention-deficit/hyperactivity disorder revealed by fMRI and the Counting Stroop.
Biol Psychiatry 1999; 45: 1542 -- 1552.
dysfunction
cingulate
Anterior
cortex
35. Rosenberg DR et al. Reduced anterior cingulate glutamatergic concentrations in
childhood OCD and major depression versus healthy controls. J Am Acad Child
Adolesc Psychiatry 2004; 43: 1146 -- 1153.
36. Mizutani R et al. Microtomographic analysis of neuronal circuits of human brain.
Cereb Cortex 2010; 20: 1739 -- 1748.
37. Takeuchi A, Uesugi K, Suzuki Y. Zernike phase-contrast x-ray microscope with
pseudo-Kohler illumination generated by sectored (polygon) condenser plate. J
Phys Conf Ser 2009; 186: 012020.
38. Mizutani R, Taguchi K, Takeuchi A, Uesugi K, Suzuki Y. Estimation of
radiation
presampling modulation
microtomography. Nucl Instrum Meth A 2010; 621: 615 -- 619.
synchrotron
function
transfer
in
39. Mizutani R et al. A method for estimating spatial resolution of real image in the
Fourier domain. J Microsc 2016; 261: 57 -- 66.
40. Cloetens P et al. Holotomography: Quantitative phase tomography with micrometer
resolution using hard synchrotron radiation x rays. Appl Phys Lett 1999; 75: 2912 --
2914.
41. Mizutani R, Saiga R, Takeuchi A, Uesugi K, Suzuki Y. Three-dimensional network
of Drosophila brain hemisphere. J Struct Biol 2013; 184: 271 -- 279.
42. Hering H, Sheng M. Dendritic spines: structure, dynamics and regulation. Nat Rev
Neurosci 2001; 2: 880 -- 888.
43. Arellano JI, Benavides-Piccione R, Defelipe J, Yuste R. Ultrastructure of dendritic
spines: correlation between synaptic and spine morphologies. Front Neurosci 2007;
1: 131 -- 143.
44. Arai M et al. Enhanced carbonyl stress in a subpopulation of schizophrenia. Arch
22
Gen Psychiatry 2010; 67: 589 -- 597.
45. Arai M et al. Pentosidine accumulation in the pathophysiology of schizophrenia:
overview of schizophrenia with carbonyl stress. IMARS Highlights 2014; 9: 9 -- 16.
46. Arai M et al. Carbonyl stress and schizophrenia. Psychiatry Clin Neurosci 2014;
68: 655 -- 665.
47. Kolluri N, Sun Z, Sampson AR, Lewis DA. Lamina-specific reductions in dendritic
spine density in the prefrontal cortex of subjects with schizophrenia. Am J
Psychiatry 2005; 162: 1200 -- 1202.
48. Jacobs B et al. Regional dendritic and spine variation in human cerebral cortex: a
quantitative golgi study. Cereb Cortex 2001; 11: 558 -- 571.
49. Tatetsu S. A histological study of schizophrenic brains: findings in telencephalon.
Psychiatria et Neurologia Japonica 1960; 62(10): 20 -- 43.
50. Blackwood DH et al. Schizophrenia and affective disorders -- cosegregation with a
translocation at chromosome 1q42 that directly disrupts brain-expressed genes:
clinical and P300 findings in a family. Am J Hum Genet 2001; 69: 428 -- 433.
51. Brandon NJ et al. Understanding the role of DISC1 in psychiatric disease and
during normal development. J Neurosci 2009; 29, 12768 -- 12775.
52. Kvajo M et al. A mutation in mouse Disc1 that models a schizophrenia risk allele
leads to specific alterations in neuronal architecture and cognition. Proc Natl Acad
Sci USA 2008; 105: 7076 -- 7081.
53. Nakao A et al. Immature morphological properties in subcellular-scale structures in
the dentate gyrus of Schnurri-2 knockout mice: a model for schizophrenia and
intellectual disability. Mol Brain 2017; 10: 60.
54. Coyle JT. NMDA receptor and schizophrenia: a brief history. Schizophr Bull 2012;
38: 920 -- 926.
55. Wozniak DF et al. Disseminated corticolimbic neuronal degeneration induced in rat
brain by MK-801: potential relevance to Alzheimer's disease. Neurobiol Dis 1998;
5: 305 -- 322.
56. Garey L. When cortical development goes wrong: schizophrenia as a
neurodevelopmental disease of microcircuits. J Anat 2010; 217: 324 -- 333.
57. Glausier JR, Lewis DA. Dendritic spine pathology in schizophrenia. Neuroscience
2013; 251: 90 -- 107.
58. Finn ES et al. Functional connectome fingerprinting: identifying individuals using
patterns of brain connectivity. Nat Neurosci 2015; 18: 1664 -- 1671.
59. Gerdes MJ et al. Emerging understanding of multiscale tumor heterogeneity. Front
Oncol 2014; 4: 366.
23
60. Natrajan R et al. Microenvironmental heterogeneity parallels breast cancer
progression: a histology-genomic integrationanalysis. PLoS Med 2016; 13:
e1001961.
61. Xu C, Prince JL. Snakes, shapes, and gradient vector flow. IEEE Trans Image
Process 1998; 7: 359 -- 369.
62. Al-Kofahi KA et al. Rapid automated three-dimensional tracing of neurons from
confocal image stacks. IEEE Trans Inf Technol Biomed 2002; 6: 171 -- 187.
24
|
1601.02974 | 2 | 1601 | 2016-03-14T17:10:11 | Memory Recall and Spike Frequency Adaptation | [
"q-bio.NC"
] | The brain can reproduce memories from partial data; this ability is critical for memory recall. The process of memory recall has been studied using auto-associative networks such as the Hopfield model. This kind of model reliably converges to stored patterns which contain the memory. However, it is unclear how the behavior is controlled by the brain so that after convergence to one configuration, it can proceed with recognition of another one. In the Hopfield model this happens only through unrealistic changes of an effective global temperature that destabilizes all stored configurations. Here we show that spike frequency adaptation (SFA), a common mechanism affecting neuron activation in the brain, can provide state dependent control of pattern retrieval. We demonstrate this in a Hopfield network modified to include SFA, and also in a model network of biophysical neurons. In both cases SFA allows for selective stabilization of attractors with different basins of attraction, and also for temporal dynamics of attractor switching that is not possible in standard auto-associative schemes. The dynamics of our models give a plausible account of different sorts of memory retrieval. | q-bio.NC | q-bio |
Memory Recall and Spike Frequency Adaptation
Neuroscience Graduate Program, University of Michigan
James P. Roach∗
Leonard M. Sander
Department of Physics, University of Michigan and
Center for the Study of Complex Systems, University of Michigan
Michal R. Zochowski
Department of Physics, University of Michigan
Center for the Study of Complex Systems, University of Michigan and
Biophysics Program, University of Michigan
(Dated: May 3, 2018)
The brain can reproduce memories from partial data; this ability is critical for memory recall.
The process of memory recall has been studied using auto-associative networks such as the Hopfield
model. This kind of model reliably converges to stored patterns which contain the memory. However,
it is unclear how the behavior is controlled by the brain so that after convergence to one configuration,
it can proceed with recognition of another one. In the Hopfield model this happens only through
unrealistic changes of an effective global temperature that destabilizes all stored configurations. Here
we show that spike frequency adaptation (SFA), a common mechanism affecting neuron activation
in the brain, can provide state dependent control of pattern retrieval. We demonstrate this in a
Hopfield network modified to include SFA, and also in a model network of biophysical neurons. In
both cases SFA allows for selective stabilization of attractors with different basins of attraction, and
also for temporal dynamics of attractor switching that is not possible in standard auto-associative
schemes. The dynamics of our models give a plausible account of different sorts of memory retrieval.
The brain stores memories as patterns of synaptic
strengths in the network of neurons. It can store mul-
tiple memories and retrieve them in a reliable way, and
can change from one to another as attention wanders.
However, there is no agreement in the neuroscience com-
munity of how this occurs. This paper offers a partial so-
lution to understanding the mechanism for retrieval and
switching based on a known physiological effect, spike
frequency adapation (SFA).
Decades of work on understanding storage and re-
trieval have focussed on versions of the Hopfield model (a
special form of the Ising model) [1 -- 3]. Hopfield networks
have many attractive features: they are auto-associative:
that is, memories are recalled from a fragment of their
data because the memories are stored in attractors, i.e.
metastable states.
However, as in any statistical model at zero temper-
ature there is no mechanism for escaping an attractor:
a single memory pattern would exist for all time. To
overcome this problem an artificial 'temperature' is in-
troduced in Hopfield models to allow switching. This
'temperature' (i.e fast random noise in synaptic current)
has no obvious biological origin. Thus, despite the ele-
gance of the model, and its utility in computer science, its
application to the brain is problematic. Previous efforts
to over come this limitation have used feedback input [4],
synaptic depression [5], and adaptive mechanisms [6]. As
we will see, SFA allows escape from attractors, and, in
some cases acts in the same way as the Hopfield tempera-
ture. In addition, it turns off particular memories rather
than globally smearing out all of them, as temperature
does.
SFA is an activity induced reduction in neural firing
rate induced by a hyperpolarizing current that grows as
a neuron fires -- neurons that fire a good deal tend to stop
firing after a delay. Thus SFA is a natural mechanism to
turn off activity. Further, SFA can be controlled by neu-
romodulators such as Acetylcholine (ACh), an important
regulator of neural excitability. ACh causes a reduction
in SFA and provides for its dynamic regulation [7, 8]. In
previous work [9] we have presented a network model of
Hodgkin-Huxley (HH) neurons with SFA and 'Mexican
Hat' coupling which reproduces many features of cortical
activity as ACh levels change between sleep and waking
states.
In the present work (below) we use this model
to concentrate on memory retrieval. However, we claim
that the essentials of our results are quite robust and in-
dependent of the details of the neuron model. To show
this, we first consider a version of the Hopfield model
[1 -- 3] which has SFA.
In our model we consider networks composed of
N=1000 spins, S = {si} where si = ± 1. The network
is fully connected with weights σi,j. As usual, spin up
corresponds to a neuron that fires, and spin down to a
silent one. Each spin gets an input:
hi(t) =
N
X
j=1
σi,jsj − θi(t),
(1)
where θi(t) is a local offset field at site i which changes
slowly in time. The first term is the usual Hopfield-Ising
term and the second represents SFA.
The dynamics of the spins are as follows: at each time
step a random spin is flipped with probability:
Ph(si) =
1
1 + e−2sihi/T ,
(2)
where T is the noise. In much of what follows we take T
to be very small so that Ph is essentially a step function.
The dynamics of θi is:
θi(si) =
A
1 + e−si(t−τ1)/τ2
.
(3)
Here, t is the time since the last state change of the spin,
and τ1,2 are time constants which govern the dynamics
of attractors. The field θ increases to A for up spins and
decreases to zero for down spins. The time constant τ1 is
the time to the half-maximum value of θi. We take τ1 = 5
(timesteps/N), except for the data in Figure 1 where τ1 =
1.5 (timesteps/N). The rate at which SFA activates/ de-
activates is controlled by τ2 which is set to 0.2, except for
the data in Figure 4 where τ2 = 0.6 (timesteps/N). This
implementations of adaptation is different than others in
the hopfield model [6]. because it integrates over a longer
time (i.e. considers more than the activity at the previ-
ous time step). This more closely resembles adaptation
in biophysical models.
In the Hopfield scheme memories are stored as attrac-
i }. We
tors, i.e. metastable configurations, Ξµ = {ξµ
encode attractors using a modified Hebbs rule [1]:
σi,j =
1
N W
p
X
µ
wµξµ
i ξµ
j .
(4)
Each attractor is given a weight, wµ and W = Pp
µ wµ.
Thus σi,j = 1 for two spins with correlated activity across
all attractors, Ξµ, and -1 for spins with anti-correlated
activity. We set w1/wp−1 = 0.5 and wp = 1, except for
the data reported in Figure 4D where all the weights wi =
1.0. The saturation is defined as α = p/N . Attractors
encoded with lower wµ are weaker attractors.
In order to determine if the dynamics has settled into
the various Ξµ we measure the overlap between the stored
memory and the current state:
mµ =
1
N
N
X
i
siξµ
i ,
(5)
2
1
0
m
s
t
r
o
n
g
−
m
w
e
a
k
−1
0.17
0.11
T
0.05
0.05 0.1 0.15
α
0.05
0.15
α
0.25
5
10
t i m e
s t e p s
15
x 1 0 4
20
0.11
0.07
A
0.03
A
i
m
1
0.5
0
1
0.5
0
1
0.5
B
i
m
C
i
m
0
0
FIG. 1. SFA and T control attractor stability in Hopfield net-
works. With noise Hopfield networks have three functional
states: stability of local attractors, stability of global attrac-
tors, and stability of no attractor. The stability of local versus
global versus no attractors is shown by the ability of the at-
tractor to move from a weak attractor to a strong one, which
is quantified by the difference of the mµ of the strong attractor
and the weak attractor in which the system was initialized.
(T op)The saturation of memories (α), noise (T ), and adapta-
tion (A) affect stability in a similar manner. For low levels of
noise local attractors are stable (black). For a given α either
increasing A (left) or T (right) leads to a strong attractors be-
ing stable (white). Further increase destabilizes all attractors
(gray). (Bottom) Example dynamics of memory overlap, mi,
for strongly (black) and weakly (gray) weighted memories. In
each case α = 0.01; adaptation levels are (A) A = 0.01 (B) A
= 0.05 (C) A = 0.3.
which is ±1 when S = ±Ξ and 0 when S⊥Ξ. In each
simulation S was always initialized to a random weak
attractor, Ξweak.
In the usual Hopfield model the preference for local,
global, or no attractors changes as the noise, T , increases
[1, 10]. This transition depends on the storage capacity
α [1]. We find very similar behavior as we increase the
magnitude of SFA, i.e. A. To show this we compare the
standard T versus α plot with a plot of A versus α in
Figure 1. The right panel shows how T and α interact to
affect the stability of the strong and weak attractors and,
eventually, to destabilize all attractors. For small T the
dynamics keeps the system in a weak attractor (black on
the colormap); for larger T the system enters a regime of
stability of stronger attractors (white). For large T no
attractors are stable (gray).
A
i
w
1
0.8
0.6
0.4
0.1
0.2
B
h
t
g
n
e
r
t
s
r
o
t
c
a
r
t
t
A
1
0.8
0.6
0.4
0
0.3
A
0.6
0.3
A
C
1.6
1.4
1.2
1
0
0.4
0.5
0.1
0.2
g K s
0.3
1
0.8
0.6
0.4
0.2
0
A
t
t
r
a
c
t
o
r
p
r
e
f
e
r
e
n
c
e
FIG. 2. Attractor stability varies as a function of A. For
SFA to induce the network to leave a attractor it must be
large enough to overcome the energy barrier of the attractor.
Mean field calculations predict a linear relationship between
the strength of an attractor and the amount of adaptation, A,
to destabilize it (A).This is best seen in the Hopfield model
(B). The threshold value of A increases linearly as the at-
tractor strength, wµ increases. A similar effect is seen in the
spiking network model (C ).
SFA has a very similar effect; see Figure 1, left. To see
how this comes about, consider the Hamiltonian:
E = −
1
2
N
X
i,j
σi,j sisj +
N
X
i
θi(t)si,
(6)
which defines a slowly varying energy landscape. The
first term is the ordinary Ising energy, and the second
can be thought of as a time-dependent magnetic field that
increases (for up spins) to A when t ≫ τ2. Increasing A
destabilizes minima in E; an attractor becomes unstable
for t → ∞ if A is large enough. In Figure 2 we show the
magnitude of SFA required to cause the system to leave
an attractor of a given strength. It increases linearly as
attractor strength increases. A similar effect occurs in
the HH model, below.
The stability of an attractor of a given weight can also
be investigated by mean field theory [10, 11]. The mean
field equations for the system are:
hsii = tanh(
β
N W X
j,µ
wµξµ
i ξµ
j hsji − 2θi),
(7)
where β = 1/T . By exploiting the fact that hsii = mξυ
i
the mean field equations can be rewritten as:
mξυ
i = tanh(
β
N W X
j,µ
wµξµ
i ξµ
j mξυ
i − 2θi).
(8)
3
1B
0.8
0.6
0.4
0.2
0
1
D
0.8
0.6
0.4
0.2
1
0
0
0.5
m
0.5
m
1
1
) A
A
2
1T
0.8
0.6
0.4
0.2
−
m
p
w
1T
(
h
n
a
t
0
1
C
)
m
p
w
1T
(
h
n
a
t
0.8
0.6
0.4
0.2
0
0
FIG. 3. SFA and T destabilize attractors of different strengths
in Hopfield networks. Solutions to mean field equations illus-
trate how adaptation and temperature destabilize weak at-
tractors. When A is low many memories are stabile (panel
A; A=0.1).
Increasing A destabilizes weak memories while
preserving strong memories (panel B; A=0.25). When adap-
tation is absent temperature has a similar effect where all
memories are stable for low T (panel C; T = 0.2), while only
strong memories are stable for high T (panel D; T = 0.5).
The dashed line shows m = m; the solid black line shows the
mean field equation for wυ = 0.45; the solid gray line shows
the mean field equation for wυ = 0.75.
If p ≪ N any overlap between memories is negligible so
the mean field equation and the system is in memory υ
for a time ≫ τ1 becomes:
mξυ
i = tanh(βwυmξυ
i −
1
T
2A),
(9)
which be simplified to m = tanh(βwυm − β2A). When
this equation has solutions beyond m = 0 a memory with
strength wυ is stable for a given T or A. Figure 3 shows
mean field solutions for memories with strengths wυ =
0.45 (solid black line) and wυ = 0.75 (solid gray line). As
in the numerical results adaptation (3 top panels) and
temperature (bottom panels) have similar effects on the
stability of memories. For low levels (A = 0.1, T = 0.2;
left panels) both strong and weak memories are stable
(i.e. both have solutions beyond m = 0), but moderate
increases in A or T destabilize weaker memories (A = .25,
T = 0.5; right panels)
Thus, changes in the strength of SFA can play the same
role as changes in T by destabilizing attractors of vary-
ing strength as A increases. Interesting time-dependent
effects occur for intermediate values of A when the τ 's
are not too large. Because θi is a function of t we can
A
i
m
B
i
m
C
i
m
1
0
−1
1
0
−1
1
0
−1
3
3
6
6
9
9
5
15
s t e p s
t i m e
x 1 0 4
25
φ
φ
1
0
1
0
0
4
Strong
Weak
0.5
g K s
( m S/ c m 2 )
1
FIG. 4. Examples of network dynamics in the modified Hop-
field model. For small A even a weak attractor is stable (panel
A; A = 0.01). A moderate increase leads to the strong at-
tractor becoming stable (B, A = 0.1). For larger A damped
oscillations with period ∼ 4τ1 emerge: C, A = 0.4. In panels
A and B the gray line corresponds to a weak attractor and
the black corresponds to a strong one. In panel C, the light-
est gray is the weakest attractor. All other lines are strong
attractors of equal weight.
FIG. 5.
Attractor preference and gKs. The quantity φ
is the fraction of time that activity is located within an at-
tractor. Black dashed line, control value. (Top) For initial
locations outside any attractor no clear preference emerges
for small gKs. For moderate gKs there is clear preference for
the strong attractor. (Bottom) For initial conditions within
the weak attractor activity never leaves for small gKs. There
is significant preference for the strong attractor for moderate
gKs.
generate chains of attractor preferences, as opposed to
stability in a deep attractor or a random walk (as in the
standard Hopfield model for large T ). These results are
shown in Figure 4. For small A local, weak attractors
are stable. A moderate increase leads to strong attrac-
tors being stable (Figure 4 A, B). Further increase of A
leads to oscillations of period ∼ 4τ1; Figure 4 C. This is
similar to the latching dynamics found in [6].
We next turn to a more realistic neuron model to com-
pare the effects of SFA for the two cases. The spik-
ing network model introduced previously [9] considers
NE = 1225 excitatory and NI = 324 inhibitory HH neu-
rons arrayed on two square lattices of size LE/I . The cou-
pling was of lateral inhibition (Mexican Hat) type where
short range excitation is balanced with global inhibi-
tion. All excitatory neurons were connected to neighbors
within radius Rxx = qL2
E/I kxx/πNE/I where kei = 16
is the degree of excitatory to excitatory connections,
kei = 4 is the degree of excitatory to inhibitory con-
nections. Neural dynamics were modeled by the current
balance equation [12]:
cm
= −gN am3
dVi
dt
−gKss(Vi − EK) − gL(Vi − EL) − Isyn,i + Iext (10)
∞h(Vi − EN a) − gKdirn4(Vi − EK )
In this equation, as we will see, gKs sets the magnitude
of the SFA; it corresponds to A in the model above.
The dynamics of the gating variables h, n and s is of
the form dx/dt = (x∞(V ) − x)/τx(V ) with additional
specific evolution of the two voltage dependent param-
eters x∞ and τx. The slow potassium current conduc-
tance, gKs controls the level of SFA (i.e. lower values of
gKs correspond to low SFA). The level of ACh modulates
gKs: the maximum (gKs = 1.5 mS/cm2) and minimum
(gKs = 0) correspond to the absence or maximum of
ACh, respectively. For more details see [9].
The synaptic current to neuron i is Isyn,i = gE(t)(Vi −
EE) + gI (t)(Vi − EI ) and the dynamics of gE/I (t) is:
gE/I (t) = K
∈E/I
X
j
σi,j (e
−(tj −τD )
τS
−(tj −τD )
τF
− e
)
(11)
where σi,j is the synaptic weight between neuron i and
neuron j, and τS,F are time constants equal to 3.0 and
0.3 ms respectively. σi,j is set to 0.02 mS/cm2 unless
otherwise stated, tj is the time since the last spike of
neuron j, and K is a normalization constant. Iext set so
that all neurons fire at 10 Hz in the absence of synaptic
input and gKs. The equations were integrated using the
4th order Runge-Kutta method at a 0.05 ms time step
to 20 s. Data points are averages of 20 sets of initial
conditions.
We have shown [9] that the nature of the dynamics is
that for small gKs (large concentrations of ACh) there is a
stationary, localized region (a 'bump') of spiking activity.
For larger gKs the bump travels through the lattice.
To consider memory we introduce spatial attractors by
increasing synaptic strength in certain locations [9, 13].
These attractors fix the location of the bump when the
dynamics is in the stationary regime. For a single at-
tractor, preference for the attractor falls as gKs increases
[9]. This model is quite different from the one discussed
above: the excitatory coupling is short-ranged in contrast
to the Hopfield σ's which are long-ranged. The attrac-
tors here are defined by local geometry. Nevertheless,
SFA gives common results for the two cases.
To consider multiple attractors of variable strength we
strengthened synaptic strength in two network regions
at opposite ends of the lattice. The strong attractor had
100% stronger excitatory connections and the weak at-
tractor had a 50% increase. To examine how network
preference changed as a function of gKs in multi-attractor
networks we did simulations where activity was initial-
ized by injecting a 0.25 µA/cm2 current to a region out-
side either of the attractors for the first 0.5 s of the sim-
ulation. Preference for a given attractor was quantified
by the measure φ which is the fraction of time that the
center of the bump, calculated according to [14], is lo-
cated within the attractor. For low levels of SFA there
is no clear preference indicating that activity localizes to
attractors randomly; Figure 5, top. Increasing gKs leads
to a clear preference for the strong attractor. The pref-
erence for any attractor disappears for large gKs as the
network enters the regime of traveling bump dynamics.
To further test the stability of the two attractors we
put activity initially on the weak attractor. For small
gKs activity remained localized there: Figure 5, bottom.
For larger gKs activity moved to the stronger attractor.
This confirms that the strength of SFA can control the
stability of attractors having different depths.
We also considered how the relative stability of the
attractors depends on the ratio of inhibitory to excita-
tory coupling, we/i). Figure 6 is a phase plot of final
attractor preference for different we/i) and gKs. The rel-
ative preference for the two attractors was measured by
φstrong − φweak, which ranges between 1 (preference for
the strong attractor) and -1 (preference for the weak). In-
terestingly, weakening inhibition abolishes any preference
for the strong attractor at intermediate levels of gKs.
Thus, for both models SFA can selectively destabi-
lize attractors effectively controlling attractor preference.
With SFA we can have long-term preferential activation
of attractors of different strengths and also non-trivial
time dependence of attractor sequences. This provides
a biologically plausible mechanism for switching between
encoded patterns [10, 15].
In the biophysical model SFA depends on ACh [8, 12,
16]. Our results imply that ACh controls memory re-
trieval dynamics. Note the relevance of our results to
context dependent release of ACh and its role in atten-
tion [17]. During tasks requiring a high degree of fo-
cus, low SFA allows the brain to fix on the memory that
closely fits the current sensory input. On the other hand,
with high ACh the attractor can be reinforced by synap-
tic plasticity. As attention requirements are relaxed, and
5
1
0
φ
s
t
r
o
n
g
−
φ
w
e
a
k
−1
0.016 0.018 0.02 0.022 0.024 0.026 0.028
w e i
( m S / c m 2)
)
2
m
c
/
S
m
(
s
K
g
1
0.8
0.6
0.4
0.2
0
FIG. 6. Strong attractor preference is controlled by inhibition
strength in spiking networks. We measure the differential
attractor preference for the two attractors by (φstrong −
φweak), which is 1 when all activity is located within the
strong attractor and -1 when all activity is within the weak.
The preference for the strong attractor at moderate SFA dis-
appears when inhibition is decreased.
ACh levels fall, moderate levels of SFA allow for sampling
of the memory space, see [18].
The largest variation in cortical ACh levels occurs be-
tween sleep/ wake states. In this case the highest levels
occur during rapid eye movement (REM) sleep and the
lowest during slow wave sleep [19]. We argue that in-
termediate levels of ACh during wake states allow for
memory recall when externally driven network states are
allowed to wander to find the optimal state. REM sleep
is thought to be important for memory consolidation,
where retrieval of weakly stored attractors of previous ex-
perience is essential to their consolidation. NREM sleep
associated with low ACh levels is characterized by slow
waves and may play a role in synaptic rescaling [20].
JPR was supported by an NSF Graduate Research Fel-
lowship Program under Grant No. DGE 1256260 and a
UM Rackham Merit Fellowship. MRZ and LMS were
supported by NSF PoLS 1058034.
∗ [email protected]
[1] D. J. Amit, H. Gutfreund, and H. Sompolinsky, Annals
of Physics 173, 30 (1987).
[2] M. E. Hasselmo, B. P. Anderson, and J. M. Bower, Jour-
nal of Neurophysiology 67, 1230 (1992).
[3] J. J. Hopfield, Proceedings of the National Academy of
Sciences 79, 2554 (1982).
[4] S. Recanatesi, M. Katkov, S. Romani, and M. Tsodyks,
Frontiers in computational neuroscience 9, 275 (2015).
[5] I. Lerner and O. Shriki, Frontiers in psychology 5, 314
(2014).
[6] A. Akrami, E. Russo,
and A. Treves, Brain research
6
1434, 4 (2012).
(2003).
[7] S. P. Aiken, B. J. Lampe, P. A. Murphy,
and B. S.
Brown, British journal of pharmacology 115, 1163
(1995).
[14] L. Bai and D. Breen, Journal of Graphics, GPU, and
Game Tools 13, 53 (2008).
[15] R. Monasson and S. Rosay, Physical Review Letters 115,
[8] A. C. Tang, A. M. Bartels, and T. J. Sejnowski, Cerebral
098101 (2015).
Cortex 7, 502 (1997).
[9] J. P. Roach, E. Ben-Jacob, L. M. Sander, and M. Zo-
chowski, PLoS Computational Biology 11, e1004449
(2015).
[16] Y. Tsuno, N. W. Schultheiss, and M. E. Hasselmo, The
Journal of Physiology 591, 2611 (2013).
[17] M. E. Hasselmo and M. Sarter, Neuropsychopharmacol-
ogy 36, 52 (2011).
[10] M. Lewenstein and A. Nowak, Physical Review Letters
[18] J. J. Hopfield, Proceedings of the National Academy of
62, 225 (1989).
Sciences 107, 1648 (2010).
[11] D.J. Amit, H. Gutfreund, and H. Sompolinsky, Physical
review. A 32, 1007 (1985).
[12] K. M. Stiefel, B. S. Gutkin, and T. J. Sejnowski, Journal
of Computational Neuroscience 26, 289 (2008).
[19] J. Vazquez and H. A. Baghdoyan, American journal
of physiology. Regulatory, integrative and comparative
physiology 280, R598 (2001).
[20] G. Tononi and C. Cirelli, Brain Research Bulletin 62, 143
[13] A. Renart, P. Song, and X.-J. Wang, Neuron 38, 473
(2003).
|
1601.07574 | 1 | 1601 | 2016-01-27T21:30:06 | Hierarchical organization of functional connectivity in the mouse brain: a complex network approach | [
"q-bio.NC",
"physics.data-an"
] | This paper represents a contribution to the study of the brain functional connectivity from the perspective of complex networks theory. More specifically, we apply graph theoretical analyses to provide evidence of the modular structure of the mouse brain and to shed light on its hierarchical organization. We propose a novel percolation analysis and we apply our approach to the analysis of a resting-state functional MRI data set from 41 mice. This approach reveals a robust hierarchical structure of modules persistent across different subjects. Importantly, we test this approach against a statistical benchmark (or null model) which constrains only the distributions of empirical correlations. Our results unambiguously show that the hierarchical character of the mouse brain modular structure is not trivially encoded into this lower-order constraint. Finally, we investigate the modular structure of the mouse brain by computing the Minimal Spanning Forest, a technique that identifies subnetworks characterized by the strongest internal correlations. This approach represents a faster alternative to other community detection methods and provides a means to rank modules on the basis of the strength of their internal edges. | q-bio.NC | q-bio | Noname manuscript No.
(will be inserted by the editor)
Hierarchical organization of functional connectivity in the
mouse brain: a complex network approach
Giampiero Bardella, Angelo Bifone, Andrea Gabrielli, Alessandro Gozzi,
Tiziano Squartini
6
1
0
2
n
a
J
7
2
]
.
C
N
o
i
b
-
q
[
1
v
4
7
5
7
0
.
1
0
6
1
:
v
i
X
r
a
Received: date / Accepted: date
Abstract This paper represents a contribution to the
study of the brain functional connectivity from the per-
spective of complex networks theory. More specifically,
we apply graph theoretical analyses to provide evidence
of the modular structure of the mouse brain and to
shed light on its hierarchical organization. We propose
a novel percolation analysis and we apply our approach
to the analysis of a resting-state functional MRI data
set from 41 mice. This approach reveals a robust hier-
archical structure of modules persistent across different
subjects. Importantly, we test this approach against a
statistical benchmark (or null model) which constrains
only the distributions of empirical correlations. Our re-
sults unambiguously show that the hierarchical charac-
ter of the mouse brain modular structure is not triv-
ially encoded into this lower-order constraint. Finally,
we investigate the modular structure of the mouse brain
by computing the Minimal Spanning Forest, a tech-
nique that identifies subnetworks characterized by the
strongest internal correlations. This approach represents
a faster alternative to other community detection meth-
ods and provides a means to rank modules on the basis
of the strength of their internal edges.
Keywords brain networks · percolation analysis · null
models · minimal spanning forest
ISC-CNR, Universit´a
Giampiero Bardella, Andrea Gabrielli, Tiziano Squartini∗
Istituto dei Sistemi Complessi
"Sapienza" di Roma, P.le A. Moro 5, 00185 Rome, Italy.
IMT Institute for Advanced Studies Lucca, P.zza S. Ponziano
6, 55100 Lucca, Italy.
Alessandro Gozzi, Angelo Bifone
Italian Institute of Technology, Universit´a di Trento, C.so
Bettini 31, I-38068 Trento, Italy.
∗E-mail: [email protected]
1 Introduction
The brain can be represented as a network of con-
nected elements at different spatial scales, from indi-
vidual neurons to macroscopic, functionally specialized
structures [1,2,3, 4,5]. Interestingly, neuroimaging data,
like those obtained with functional Magnetic Resonance
Imaging (fMRI) techniques, naturally lend themselves
to a network representation, thus attracting the inter-
est of both graph-theorists and network scientists to-
wards a study of the topological properties of brain
connectivity structures [6]. Indeed, correlations between
fMRI signals arising from responses to stimuli or from
spontaneous fluctuations in the brain resting-state can
be interpreted as a measure of functional connectiv-
ity between remote brain regions and represented as
edges in a graph. Moreover, alterations in the strength
and structure of functional connectivity networks have
been observed in groups of patients suffering from sev-
eral brain diseases, including Alzheimer, Autism and
Schizophrenia, thus providing potential markers of neu-
ropsychiatric illness [1, 7,9,10, 11, 12].
Of particular interest is the study of the modular
structure of these networks, i.e. the presence of clus-
ters of nodes that are more tightly connected among
themselves than with nodes in other network substruc-
tures [1,2,13,14, 15]. A modular structure has been ob-
served for different types of brain networks (functional
and structural) and in different species, including hu-
mans, primates and rodents [7, 14, 15,16]. Functional
connectivity networks derived from fMRI experiments
in human subjects exhibit a hierarchical structure of
modules-within-modules [3,4]. It has been suggested
that hierarchical modularity may confer important evo-
lutionary and adaptive advantages to the human brain
by providing intermediate modules that can respond
2
Giampiero Bardella, Angelo Bifone, Andrea Gabrielli, Alessandro Gozzi, Tiziano Squartini
to the evolutionary or environmental pressure without
jeopardizing the function of the entire system [17]. A
similar hierarchical organization has been observed in
other species, e.g. non-human primates, but not in lower
species, like the worm C. Elegans, which seems to have
a modular network of neurons that is not hierarchi-
cally organized [18,19]. Here, we investigate the mod-
ular structure of the mouse brain and its hierarchical
organization using a graph theoretical approach.
Percolation analysis, a tool derived from statistical
physics, provides a powerful means to investigate the
hierarchical organization of networks [20,22]. This ap-
proach is based on the assessment of the fragmentation
of a network as weaker edges are gradually removed
from the graph. A striking demonstration of this hierar-
chical organization is the presence of multiple percola-
tion thresholds [15], whereby disaggregation of modules
occurs abruptly for critical values of the control param-
eter, pc. On the contrary, application of this analysis to
Erdos-Renyi random graphs [20,21, 22] shows a single
threshold value, separating two phases characterized by
different topological features. Below the threshold (i.e.
for p < pc) several tree-like components are observed
whose size is of the order of ln N - with N the total
number of nodes. Above the threshold (i.e. for p > pc),
instead, a single giant component appears, whose struc-
ture admits cycles and from which tree-like structures
(whose size is again of the order of ln N ) are excluded
[20,21].
Here we have analyzed functional connectivity net-
works constructed from a large resting state fMRI dataset
from mice to assess the presence of multiple percolation
thresholds. Specifically, we have applied standard per-
colation analysis and variations thereof to assess the
hierarchical modular structure in this species. Impor-
tantly, we have applied novel approaches to avoid some
of the pitfalls that may affect more conventional anal-
ysis of functional connectivity networks. Indeed, it will
be shown that traditional percolation detects a modu-
lar structure even in random networks, thus making it
necessary to introduce a null model in order to correctly
asses the statistical significance of the percolation anal-
ysis. Here we introduce a novel null model, independent
of the choice of a particular threshold and resting exclu-
sively on the information encoded into the correlation
matrix. Moreover, we propose the use of an algorithm
to calculate the closest correlation matrix to a given
symmetric matrix, thus ensuring that the proposed null
model has the peculiar features of a proper correlation
matrix.
We have complemented our percolation analysis by
computing the Minimal Spanning Forest (MSF). Al-
thought the MSF is not, by itself, a community detetc-
Fig. 1: Dendrogram and correlation matrix for the
average brain,
induced by the dissimilarity measure
Dij = 1 − Cij, ∀ i, j.
tion technique, it represents a faster alternative for the
identification of modules, defined by the strength of the
functional relations between nodes. Such modules can
be, in turn, linked to obtain the Minimal Spanning Tree
(MST), which provides the "backbone" of the mouse
brain functional connectivity.
131935374203638151674941131427284546293021312254434453472551335232481718263485042391140232451221096(cid:45)11 Hierarchical organization of functional connectivity in the mouse brain: a complex network approach
3
Fig. 2: Empirical CDF of the correlations for the aver-
age brain (blue trend) and CDF of a gaussian distribu-
tion whose means and standard deviations have been
estimated through the maximum-of-the-likelihood pro-
cedure (red trend).
These methodological developments make it possi-
ble to assess the presence of a hierarchically-organized
modular structure in the mouse brain, both at the level
of population and of individual subjects.
2 Methods
2.1 Data acquisition and data pre-processing
The data-set used for this analysis has been reported
in a recent paper [7,8], where experimental details are
extensively described. In short, MRI experiments were
performed on male 20-24 week old C57BL/6J (B6) mice
(n = 41, Charles River, Como, Italy). Mice were anaes-
thetised with isoflurane (5% induction), intubated and
artificially ventilated under 2% isoflurane maintenance
anesthesia. All experiments were performed with a 7.0
T MRI scanner (Bruker Biospin, Milan) using an echo
planar imaging (EPI) sequence with the following pa-
rameters: TR/TE 1200/15 ms, flip angle 30 degrees,
matrix 100 × 100, field of view 2 × 2 cm2, 24 coronal
slices, slice thickness 0.50 mm, 300 volumes and a total
rsfMRI acquisition time of 6 minutes. All experiments
were conducted in accordance with the Italian law (DL
116, 1992 Ministero della Sanit´a, Roma) and the recom-
mendations in the "Guide for the Care and Use of Lab-
oratory Animals" of the National Institutes of Health.
Animal research protocols were also reviewed and con-
sented to by the animal care committee of the Istituto
Italiano di Tecnologia (permit 07-2012). All surgical
procedures were performed under anesthesia.
The mouse brain was parcellated into 54 macro-
regions (27 per hemisphere) described in the Appendix.
Fig. 3: Comparison between the usual percolation anal-
ysis (top panel) and our modified percolation analysis
(bottom panel) run on the average brain (red trend), on
a randomized version of it, retaining the same empir-
ical distribution of correlations (brown trend) and on
the ensemble-averaged matrix (green trend). While the
usual percolation analysis detects a hierarchical mod-
ular structure even on the null model, thus making it
difficult to asses the statistical significance of the ob-
served patterns, our modified percolation analysis en-
ables discrimination between the real and the random
cases.
Resting state fMRI signals from individual image vox-
els were averaged across each region of interest (ROI)
to generate 54 time-series of approximately 300 s dura-
tion. The 54 collected time-series were pairwise corre-
lated calculating the Pearson coefficient and organized
in a 54 × 54 symmetric matrix describing the resting-
state connectivity network for each mouse.
Image preprocessing was carried out using tools from
FMRIB Software Library (FSL, v5.0.6 [23, 24]) and AFNI
(v2011 12 21 1014 [25]). RsfMRI time series were de-
spiked (AFNI/3dDespike), corrected for motion (AFNI/3dvolreg)
and spatially normalised to an in-house C57Bl/6J mouse
brain template [26] (FSL/FLIRT,12 degrees of freedom).
The normalised data had a spatial resolution of 0.2 ×
0.2 × 0.5 mm3 (99 × 99 × 24 matrix). Head motion
0.00.20.40.60.81.00.00.20.40.60.81.0rCDF0.00.20.40.60.81.00.020.050.10.20.51.r(cid:37)sizeofthelargestcomponent0.00.20.40.60.81.001020304050r(cid:240)connectedcomponents 4
Giampiero Bardella, Angelo Bifone, Andrea Gabrielli, Alessandro Gozzi, Tiziano Squartini
traces and mean ventricular signal (averaged fMRI time
course within a manually-drawn ventricle mask) were
regressed out of each of the timeseries (AFNI/3dDeconvolve).
To assess theeffectof global signal removal, separate rsfMRI
time series with the whole-brain average time course
regressed out were also generated. All rsfMRI time se-
ries were spatially smoothed (AFNI/3dmerge, Gaussian
kernel of full width at half maximum of 0.5 mm) and
band-pass filtered to a frequency window of 0.01-0.08
Hz (AFNI/3dBandpass) [26].
In order to create an average adjacency matrix de-
scribing brain functional connectivity at the population
level, subject-wise matrices were first Fisher-transformed,
averaged across subjects and then back-transformed.
2.2 Percolation analysis
The percolation analysis proposed by Makse et al. [15]
includes the following steps: a) a threshold parameter
p, ranging between 0 and 1 (and thus interpretable as a
probability), is chosen; b) the links corresponding to the
correlations below the threshold are removed and the
size of the giant component C (i.e. the largest connected
component) is computed; c) the parameter p is varied
and C is evaluated for different thresholds.
This procedure ignores the complex evolution of the
structure of the whole network, which is not captured
by the giant component only. This becomes a relevant
issue when the classical percolation is applied to small
networks, i.e. to networks for which no giant compo-
nent is clearly distinguishable: in this case the signal
provided by this kind of analysis may be rather noisy,
thus misrepresenting the modular structure of the brain
at the global level.
In order to overcome this drawback, we propose a
variation of the percolation analysis along the following
lines: a) all experimentally determined correlation co-
efficients are listed in increasing order; b) starting from
the lowest value, each entry in the list is chosen as a
threshold; c) all the links corresponding to the correla-
tions below the threshold are removed; d) the number
of connected components characterizing the remaining
part of the network is computed.
Beside providing a much more precise picture of the
dynamics of the brain at the global level, our variation
of the percolation analysis is also more robust, since our
signal results from the fragmentation of many different
components at the same time and is thus less prone
to the statistical noise which, instead, accompanies the
fragmentation of the giant component only.
Moreover, while each step of the classical perco-
lation analysis is always mappable into a step of our
Fig. 4: In order to assess the statistical significance of
the results of our modified percolation analysis, a test is
needed. Top panel represents the test statistics we have
chosen: the slope of the percolation plot of both the av-
erage brain (red trend) and of a randomized version of
it, retaining the same empirical distribution of correla-
tions (brown trend). Bottom panel: ensemble distribu-
tion of our test statistics; the red point represents the
(statistically significant) observed value of the latter.
method, the reverse is not true: the detection of a newly
disconnected module from a secondary component would
be missed by the classical percolation analysis (which
focuses on the giant component only).
2.3 A statistical benchmark for mice brains
In order to define to what extent the stepwise structure
highlighted by the percolation analysis is significant,
we need to compare the results with a proper statisti-
cal benchmark. In other words, in order to understand
whether the "stepwise behavior" is a mere consequence
of lower- order constraints or a genuine sign of self-
organization we need to define a proper null model.
As a first step, we have calculated the empirical
probability distributions of the entries of the correla-
tion matrix characterizing each subject in our sample
0.00.20.40.60.81.001020304050r(cid:240)connectedcomponents01002003004005006000.0000.0020.0040.006rPDF Hierarchical organization of functional connectivity in the mouse brain: a complex network approach
5
and fitted them to normal distributions, whose means
and standard deviations were estimated through the
maximum-of-the-likelihood procedure. We have also re-
peated this analysis for the average mouse, i.e. the brain
functional connectivity at the population level. In all
cases, the distributions of the elements of the correla-
tion matrices appeared to be well behaved, with nearly
Gaussian distributions.
Secondly, we have generated a "null brain", by draw-
ing correlations from the corresponding normal distri-
butions. However, this procedure does not guarantee
that true correlation matrices, which should be positive-
definite, are obtained: in fact, although the synthetic
matrices can be chosen to be sym-metric and with uni-
tary elements on the main diagonal, they may still have
negative eigenvalues. This problem can be solved by im-
plementing the procedure illustrated in [27] where a fast
algorithm for comput- ing the nearest correlation ma-
trix to a given, symmetric, one is described. The last
step of our method consists in the implementation of
this procedure.
3 Results
3.1 Average correlation matrix
We first focus on the average correlation matrix, de-
fined by the sample mean (i.e. over all individuals) of
each back-transformed pair-specific correlation coeffi-
cient. Fig. 1 shows the average correlation matrix whose
rows and columns have been reordered according to the
dissimilarity measure
Dij = 1 − Cij, ∀ i, j.
(1)
The algorithm we have adopted proceeds by com-
puting, at each step, the minimum dissimilarity be-
tween pairs of areas and clustering them together. In
other words, clusters are grouped according to the mini-
mum intercluster dissimilarity, a linkage rule also known
as "single-linkage" clustering [28]. The same algorithm
can be used to generate the corresponding dendrogram.
While negative correlations are pronounced in subject-
wise matrices they tend to be averaged-out in the population-
wise matrix, whose terms are all positive. Although this
confirms the larger inter-subject variability of negative
correlations with respect to the positive ones, it af-
fects the nested structure of the average matrix, which
is far less pronounced than for the single individuals:
nonetheless, nested red square-shaped patterns along
the diagonal are still clearly visible.
whose mean and variance have been estimated through
the maximum-of-the-likelihood procedure. The devia-
tion of the distribution of experimentally-determined
correlations form the normal distribution is larger for
this matrix than for the individual ones.
Percolation analysis. The results of the classical and
modified percolation analyses are shown in fig. 3.
Our method identifies multiple steps for increasing
threshold, corresponding to the stable partitions of the
network [1,15] highlighted in fig. 5. The plateaus in-
dicate the presence of connections whose removal does
not affect the number of connected components, indi-
cating that these links are not critical in determining
the structure of functional correlations.
Fig. 5, shows that each connected group of areas
detected in correspondence of a given correlation value
is composed by many nested modules, whose hierarchi-
cal organization emerges form the application of higher
thresholds. Two main groups of areas can be clearly
identified (colored in blue and green in fig. 5 and de-
tected for rth (cid:39) 0.45). The first group (colored in green)
regions include the cingulate cortex, the motor cor-
tex, the medial prefrontal cortex and the primary so-
matosensory cortex [1, 7,10]. The second group (colored
in blue) is constituted by areas 3, 4, 19, 20, 35 and
36 (i.e. anterio-dorsal hippocampus, the right dentate
gyrus and the right posterior gyrus), all parts of the
hippocampal formation.
Upon rising the threshold to rth = 0.52, sub-areas
appear: for example, the hippocampus splits into right
and left part - i.e. 3, 19, 35 and 4, 20, 36 (evidenced in
blue and purple); further rising the threshold to rth =
0.6, the two latter subgroups reveal a core structure de-
fined by the pairs 19, 35 and 20, 36. An analogous result
is found for the sensory system, confirming the hierar-
chical character of the mouse brain modular structure.
Interestingly, the percolation curve for the null model
shows a remarkably different trend. Indeed, drawing
a matrix (whose distribution of correlations coincides
with the observed one for the average mouse) from
our ensemble and repeating our percolation analysis
leads to a single, sharper transition, with basically no
plateaus. This indicates that rising the threshold value
leads to the sequential disconnection of individual nodes,
which are removed one after the other. This supports
the idea that the hierarchical structure observed for the
brain connectivity network is genuine, as the stepwise
behavior does not emerge in a null model with similar
distribution of correlations (the same conclusion holds
true for the individual-wise matrices as well).
The distribution of edge values of the average matrix
is shown in fig. 2, alongside with the normal distribution
While this is reassuring, a statistical test is needed
to quantify the significance of the experimental trend
6
Giampiero Bardella, Angelo Bifone, Andrea Gabrielli, Alessandro Gozzi, Tiziano Squartini
Fig. 5: Each group of areas detected by our percolation analysis in correspondence of a given correlation value
is composed by many sub-modules, whose presence is evidenced by rising the threshold value. A clear example
is provided by the blue area detected for rth = 0.51, comprising the anterio-dorsal hippocampus, the dentate
gyrus and the posterior dentate gyrus - i.e. areas 3, 4, 19, 20, 35, 36. Upon rising the threshold to rth = 0.52,
two subgroups appear, composed respectively by the right and left parts - i.e. 3, 19, 35 and 4, 20, 36 - of the
aforementioned areas (evidenced in blue and purple). Further rising the threshold to rth = 0.6, the two subgroups
reveal a core structure defined by the pairs 19, 35 and 20, 36. This finding confirms the hierarchical character of
the mouse brain modular structure. See also the map of the neuroanatomical ROI in Appendix.
Hierarchical organization of functional connectivity in the mouse brain: a complex network approach
7
olds in both the experimental network and in the net-
work generated according to our null model. Although
this finding provides a significant evidence of the struc-
tural differences between the observed average brain
and an Erdos-Renyi-like graph, for example, it also im-
plies that the claim according to which, in this "clas-
sical" version of percolation, revealing multiple thresh-
olds is, by itself, a proof of the hierarchical modular
structure of a network is arguable. Indeed, recovering
the presence of steps also in the null model seems to
suggest that the dynamics of the giant component is
(at least) partially encoded into the correlations distri-
butions, while this is no longer true when considering
also the remaining components, implying that one of
the genuine signatures of the brain self-organization lies
in their dynamics.
Minimal Spanning Forest. The MSF algorithm is de-
fined by two simple steps: a the observed correlations
are sorted in reverse order; b) starting from the largest
observed correlation, a link is drawn between the cor-
responding brain areas. This is done sequentially, with
the limitation that any new connection must link at
least one previously completely disconnected area.
At each step of the MSF algorithm, either a pre-
viously isolated area is assigned to an existing group
or two previously isolated areas are linked together. In
this way, "communities" remain naturally defined by
the strength of their internal correlations, while redun-
dant connections are discarded. In particular two dif-
ferent communities are eventually connected by edges
whose correlation value is smaller than all the links of
both communities. Althought the MSF is not, by itself,
a community detetction technique, it provides a means
to hierarchically order modules based on the strength
of their internal edges. Such modules are tree-shaped
and provide information on the structural importance
of each area (e.g. its betweeness centrality).
The MSF of average brain is shown in fig. 6. Our
analysis reveals that presence of both inter- and intra-
hemispheric modules. The module with the strongest
internal connectivity is the medial-prefrontal cortex,
consistent with the finding that this bilateral structure
persists in the percolation analysis at high values of
the threshold. The second and third modules in the
MSF rank are the right and left hippocampal formation.
Interestingly, larger, inter-hemisferic modules, like the
one comprising frontal and orbitofrontal cortices, cau-
date putamen and the amygdala, are characterized by
more numerous, but weaker links. Altogether, the MSF
structure reflects the hierarchical organization of con-
nectivity modules revealed by our percolation analysis.
Fig. 6: Result of the MSF algorithm mapped into the
average mouse brain areas. The algorithm works by first
sorting the observed correlations in decreasing order
and then linking pairs of areas sequentially, with the
only limitation that each new link must connect at least
one previously disconnected area. Colors correspond to
the average correlation value of the links defining each
tree composing the forest. See also the map of the neu-
roanatomical ROI in Appendix.
th, 53), with r(cid:48)
with respect to the null hypothesis. Our choice of such
test moves from the observation that the experimen-
tal trend is less steep than the one obtained by run-
ning the null model. For this reason, the test statistics
we have computed is the steepness of the experimental
trend, measured between two points: the pairs (r(cid:48)
th, 2)
and (r(cid:48)(cid:48)
th indicating the values of
correlations in correspondence of which we detect 2 and
53 communities respectively (we have deliberately ex-
cluded the trivial communities represented by the whole
brain and the single areas/nodes). The ensemble distri-
bution of our test statistics is shown in fig. 4, together
with the experimental point: the latter lies well outside
the 95% confidence intervals.
th and r(cid:48)(cid:48)
On the other hand, as evident upon inspecting fig.
3, classical percolation, in which only the size of the
largest component is monitored, detects multiple thresh-
8
Giampiero Bardella, Angelo Bifone, Andrea Gabrielli, Alessandro Gozzi, Tiziano Squartini
maining correlations in the list to build the Minimal
Spanning Tree (MST). As for the forest, only one lim-
itation exists: any new added link must connect a pair
of previously-disconnected trees (which become part of
the same tree afterwards). Naturally, the links between
trees are weaker than the links within trees and the
MSF can be recovered upon removing the weakest links.
The information provided by the MSF can be thus
complemented by the information provided by the MST,
which gives a clear picture of the mouse brain connec-
tivity skeleton. In particular, the structural role of each
area becomes evident and a classification of connec-
tor areas VS provincial areas becomes now possible.
Among the most prominent examples of the former are
the posterio-ventral hippocampus (i.e. 38) whose phys-
ical centrality is recovered as a functional centrality,
the parietal association cortex (i.e. 33) which connects
all the sensory areas (i.e. the rhinal, auditory and vi-
sual ones) and the orbitofrontal cortex (i.e.31) whose
physical connections are mirrored by a high degree of
functional (inter)-connectivity (e.g. it connects the tha-
lamus and the frontal association cortex).
4 Conclusions
In this paper we have presented the results of a network
theory-based analysis of a large mouse fMRI dataset,
aimed at assessing the hierarchical modular structure
of resting state functional connectivity networks in this
species. In order to overcome the limitations of cur-
rently available techniques, we propose a modified per-
colation analysis that retains the information on all the
connected components of a given network. Our varia-
tion of the percolation analysis takes into account neg-
ative correlations, and does not require the application
of a threshold to binarize the connectivity networks.
Our technique, straightforwardly applicable to ex-
perimental correlation matrices, reveals a hierarchically
organized modular structure that does not appear in a
null model defined by constraining the distribution of
the observed correlations. Notably, conventional perco-
lation analysis shows the presence of multiple percola-
tion thresholds also in the null model, thus suggesting
that results based on the giant connected component
alone maybe misleading.
Our percolation analysis represents a generalization
of the classical one. Indeed, while each step of the clas-
sical percolation analysis is always mappable into a step
of our method, the reverse is not true, since the detec-
tion of a newly disconnected module from a secondary
component would be completely missed.
We have also computed the Minimal Spanning For-
est (MSF) and the Minimal Spanning Tree (MST) for
Fig. 7: MST of our average mouse brain. The MST has
been built by connecting the trees of the MSF, with the
only limitation that any newly-added link must connect
a pair of previously-disconnected trees: a consequence
of the MST algorithm is that the correlations within
the trees are, on average, higher than the correlations
between the trees. The MST also allows us to distin-
guish between connector and provincial areas. See also
the map of the neuroanatomical ROI in Appendix.
Remarkably, the insular cortex and the secondary
somatosensory cortices are found within the same tree,
thus showing that the reciprocal structural connectiv-
ity among these areas results in a consistent pattern
of functional connectivity which has been recently de-
scribed also using voxelwise community detection ap-
proaches [7]. Similarly, the thalamus is found to be
strongly linked to the bed nucleus of stria terminals,
consistent with the reciprocal neuroanatomical links con-
necting these regions [30]. Interestingly, our MSF re-
veals a strong functional connection between the visual
cortex and the retrosplenial cortex (i.e. between areas
43, 44, 53 and 54), an area that has been recognized as
fundamental in tasks like orientation, head movement
and processing of visual cues [29]. As a last example,
the MSF suggests a role for the temporal association
cortex (i.e. 49, 50) in the coordination of the sensori
stimuli [31], receiving inputs from the auditory and the
rhinal corteces (i.e. 7, 8 and 41, 42).
Once the MSF has been built, we can use the re-
123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354 Hierarchical organization of functional connectivity in the mouse brain: a complex network approach
9
our population-wise mouse brain. The latter represents
a faster alternative to the usual community detection
techniques, since it identifies modules on the basis of
the strengths of their internal correlations. The MSF re-
veals both intra- and inter-hemispherical modules, and
the presence of small, tightly coupled modules along-
side with larger subnetworks characterized by weaker
internal links. The MST, on the other hand, enables
the classification of connector and provincial areas.
Our results indicate that the tools provided by net-
work theory indeed provide additional, non-trivial in-
formation on the topology of functional connectivity
networks from the mouse brain. This work can be straight-
forwardly extended to the study of the human brain.
Appendix
ROI - Regions of interest
The list of the neuroanatomical ROI considered for our
analysis, together with their abbreviation, is the follow-
ing (alphabetical order). Fig. 8 shows the ROI mapped
into a mouse brain.
1. Acb: accumbens nucleus dx;
2. Acb: accumbens nucleus sx;
3. AdHC: anterio-dorsal hippocampus dx;
4. AdHC: anterio-dorsal hippocampus sx;
5. Amy: amygdala dx;
6. Amy: amygdala sx;
7. Au: auditory cortex dx;
8. Au: auditory cortex sx;
9. BF: basal forebrain dx;
10. BF: basal forebrain sx;
11. BNST: bed nucleus of stria terminals dx;
12. BNST: bed nucleus of stria terminals sx;
13. Cg: cingulate cortex dx;
14. Cg: cingulate cortex sx;
15. Collicoli: collicoli dx;
16. Collicoli: collicoli sx;
17. Cpu: caudate putamen dx;
18. Cpu: caudate putamen sx;
19. DG: dentate gyrus dx;
20. DG: dentate gyrus sx;
21. FrA: frontal association cortex dx;
22. FrA: frontal association cortex sx;
23. Hypo: hypothalamus dx;
24. Hypo: hypothalamus sx;
25. Ins: insular cortex dx;
26. Ins: insular cortex sx;
27. M: motor cortex dx;
28. M: motor cortex sx;
29. mPFC: medial prefrontal cortex dx;
Fig. 8: The neuroanatomical ROI considered for our
analysis, mapped into a mouse brain.
30. mPFC: medial prefrontal cortex sx;
31. OFC: orbitofrontal cortex dx;
32. OFC: orbitofrontal cortex sx;
33. Parietal Ass: parietal association cortex dx;
34. Parietal Ass: parietal association cortex sx;
35. pDG: posterior dentate gyrus dx;
36. pDG: posterior dentate gyrus sx;
37. pHC: posterio-ventral hippocampus dx;
38. pHC: posterio-ventral hippocampus sx;
39. Pir: piriform cortex dx;
40. Pir: piriform cortex sx;
41. Rhinal: rhinal cortex dx;
42. Rhinal: rhinal cortex sx;
43. RS: retrosplenial cortex dx;
44. RS: retrosplenial cortex sx;
45. S1: primary somatosensory cortex dx;
and T is the total temporal length
In order to create an average adjacency matrix de-
scribing brain functional connectivity at the population
level, subject-wise matrices were first Fisher-transformed,
i.e.
(cid:80)T
t=1 X i
T
t
where mi =
of the series.
(cid:19)
(cid:18) 1 + Cij
1 − Cij
zij =
1
2
ln
10
Giampiero Bardella, Angelo Bifone, Andrea Gabrielli, Alessandro Gozzi, Tiziano Squartini
= arctanh(Cij),
(3)
then averaged across subjects
(cid:80)54
zij =
n=1 zn
ij
54
, ∀ i, j
(4)
(i.e. the generic entry of the average Fisher-transformed
matrix is the arithmetic mean of the corresponding indi-
vidual entries, z1
ij ) and then back-transformed:
ij . . . z54
ij, z2
C ij = tanh(zij), ∀ i, j.
(5)
Fig. 9: fMRI BOLD-signals corresponding to the right
cingulate cortex (top) and left cingulate cortex (bot-
tom) of the brain BE ag130207a.
Acknowledgements We acknowledge support from the EU
project FET-Open FOC (grant num. 255987), the FET project
SIMPOL (grant num. 610704) and the FET project DOLFINS
(grant num. 640772).
46. S1: primary somatosensory cortex sx;
47. S2: secondary somatosensory cortex dx;
48. S2: secondary somatosensory cortex sx;
49. TeA: temporal association cortex dx;
50. TeA: temporal association cortex sx;
51. Th: thalamus dx;
52. Th: thalamus sx;
53. Vctx: visual cortex dx;
54. Vctx: visual cortex sx.
From time series to correlation matrices
Our data consist of 41 sets of 54 fMRI BOLD-signals
each, collected as the time series shown in fig. 9.
The information carried by each mouse-specific set
of time series has been condensed into a correlation ma-
trix, whose generic entry Cij is the Pearson coefficient
between time series X i and X j, defined as
Cij =
=
Cov[X i, X j]
(cid:112)Var[X i] · Var[X j]
(cid:80)T
(cid:113)(cid:80)T
t − mi)2 ·(cid:80)T
t − mi)(X j
t=1(X i
t=1(X i
=
t − mj)
t=1(X j
t − mj)2
(2)
References
1. M. P. van den Heuvel, H. E. Hulshoff Pol, Exploring
the brain network: a review on resting-state fMRI func-
tional connectivity, European Neuropsychopharmacology
20, 519-534 (2010).
2. Z. Yao, Y. Xie, P. Moore, J. Zheng, A review of structural
and functional brain networks: small world and atlas, Brain
Informatics 2(9), doi:10.1007/s40708-015-0009-z (2015).
3. E. Bullmore, O. Sporns, Complex brain networks: graph
theoretical analysis of structural and functional systems, Na-
ture Reviews Neuroscience 10, 186-198 (2010).
4. O. Sporns, R. Betzel, Modular brain networks, Annual Re-
view of Psychology 67(19), 1-28 (2016).
5. C. Nicolini, A. Bifone, Modular structure of brain networks:
breaking the resolution limit by surprise, Scientific Reports
6(19250), doi:10.1038/srep19250 (2016).
6. C. Li, H. Wang, W. de Haan, C. J. Stam, P. Van
Mieghem, The correlation of metrics in complex networks
with applications in functional brain networks, Journal of
Statistical Mechanics: Theory and Experiment P11018,
doi:10.1088/1742-5468/2011/11/P11018 (2011).
7. A. Liska, A. Galbusera, A. J. Schwarz, A. Gozzi,
Functional connectivity hubs of mouse brain, NeuroImage,
doi:10.1016/j.neuroimage.2015.04.033 (2015).
8. F. Sforazzini, A.J. Schwarz, A. Galbusera, A. Bifone,
A. Gozzi, Distributed BOLD and CBV-weighted resting-
state networks in the mouse brain, NeuroImage 87, 403-415
(2014).
050100150200250300(cid:45)2(cid:180)106(cid:45)1(cid:180)10601(cid:180)1062(cid:180)106tfMRIsignal050100150200250300(cid:45)2(cid:180)106(cid:45)1(cid:180)10601(cid:180)1062(cid:180)106tfMRIsignal Hierarchical organization of functional connectivity in the mouse brain: a complex network approach
11
30. D. Choi, A. Furay, N. Evanson, M. Ostrander, Y. Ulrich-
Lai, J. Herman, Bed nucleus of the stria terminalis subre-
gions differentially regulate hypothalamic-pituitary-adrenal
axis activity: implications for the integration of limbic in-
puts. Journal of Neuroscience 27(8) (2007).
31. L. R. Squire, C. E. Stark, R. E. Clark, The medial temporal
lobe. Annual Review of Neuroscience 27, 279-306 (2004).
9. B. Biswal, F. Z. Yetkin, V. M. Haughton, J. S. Hyde,
Functional connectivity in the motor cortex of resting hu-
man brain using echo-planar MRI, Magnetic Resonance in
Medicine 34, 537-541 (1995).
10. C. Rosazza, L. Minati, Resting-state brain networks: liter-
ature review and clinical applications, Neurological Science
32, 773-785, doi:10.1007/s10072-011-0636-y (2011).
11. D. Zhang, M. E. Raichle, Disease and the brain's dark en-
ergy, Nature Reviews Neurology 6, 15-28 (2010).
12. M. D. Fox, M. Grecius, Clinical applications of resting state
functional connectivity, Frontiers in Systems Neuroscience
4(19) (2010).
13. D. Meunier, R. Lambiotte, E. Bullmore, Modular and hier-
archically modular organization of brain networks, Frontiers
in Neuroscience 4(200), doi:10.3389/fnins.2010.00200
(2010).
14. S. Achard, R. Salvador, B. Whitcher, J. Suckling, E. Bull-
more, A resilient, low-frequency, small-world human brain
functional network with highly connected association cortical
hubs, The Journal of Neuroscience 26(1), 63-72 (2006).
15. L. K. Gallos, H. A. Makse, M. Sigman, A small world of
weak ties provides optimal global integration of self-similar
modules in functional brain networks, Proceedings of the
National Academy of Science 109(8), 2825-2830 (2012).
16. A. Bifone, A. Gozzi, A. J. Schwarz, Functional connectivity
in the rat brain: a complex network approach, Magnetic Res-
onance Imaging, 1200-9, doi:10.1016/j.mri.2010.07.001
(2010).
17. H. Simon, The architecture of complexity, Proceedings
of the American Philosophical Society 106(6), 467-482
(1962).
18. N. Chatterjee, S. Sinha, Understanding the mind of a
worm: hierarchical network structure underlying nervous
system function in C. elegans, in Progress in Brain Re-
search 168(12), 145-153 (2008).
19. J. G. White, E. Southgate, J. N. Thomson, S. Brenner,
The structure of the nervous system of the nematode C. el-
egans, Philosophical Transactions of the Royal Society of
London B 314, 1-340 (1986).
20. R. Albert, A.-L. Barabasi, Statistical mechanics of complex
networks, Reviews of Modern Physics 74, 47-97 (2002).
21. B. Bollobas, Random Graphs, Cambridge University Press
(2001).
22. S. M. Hadi Hosseini, S. R. Kesler, Influence of choice of
null network on small-world parameters of structural corre-
lation networks, PLoS ONE 8(6): e67354 (2013).
23. http://fsl.fmrib.ox.ac.uk/fsl/.
24. M. Jenkinson, C. F. Beckmann, T. E. Behrens, M. W.
Woolrich, S. M. Smith, Fsl, NeuroImage 62, 782-790
(2012).
25. http://afni.nimh.nih.gov/afni/.
26. F. Sforazzini, A. Bertero, L. Dodero, G. David, A. Gal-
busera, M. L. Scattoni, M. Pasqualetti, A. Gozzi, Altered
functional connectivity networks in acallosal and socially im-
paired BTBR mice. Brain Structure and Function, 1-14
(2014).
27. N. J. Higham, Computing the nearest correlation matrix - a
problem from finance, IMA Journal of Numerical Analysis
22, 329-343 (2002).
28. F. Murtagh, P. Contreras, Algorithms for hierarchical clus-
tering: an overview, Wiley Interdisciplinary Reviews: Data
Mining and Knowledge Discovery 2(1), 86-97 (2012).
29. S. D. Vann, J. P. Aggleton, E. A. Maguire, What does
the retrosplenial cortex do?, Nature Reviews 10, 792-803
(2009).
|
1610.06360 | 3 | 1610 | 2017-09-29T13:42:05 | Accounting for the Complex Hierarchical Topology of EEG Phase-Based Functional Connectivity in Network Binarisation | [
"q-bio.NC"
] | Research into binary network analysis of brain function faces a methodological challenge in selecting an appropriate threshold to binarise edge weights. For EEG phase-based functional connectivity, we test the hypothesis that such binarisation should take into account the complex hierarchical structure found in functional connectivity. We explore the density range suitable for such structure and provide a comparison of state-of-the-art binarisation techniques, the recently proposed Cluster-Span Threshold (CST), minimum spanning trees, efficiency-cost optimisation and union of shortest path graphs, with arbitrary proportional thresholds and weighted networks. We test these techniques on weighted complex hierarchy models by contrasting model realisations with small parametric differences. We also test the robustness of these techniques to random and targeted topological attacks.We find that the CST performs consistenty well in state-of-the-art modelling of EEG network topology, robustness to topological network attacks, and in three real datasets, agreeing with our hypothesis of hierarchical complexity. This provides interesting new evidence into the relevance of considering a large number of edges in EEG functional connectivity research to provide informational density in the topology. | q-bio.NC | q-bio | Accounting for the complex hierarchical topology
of EEG phase-based functional connectivity in
network binarisation
7
1
0
2
p
e
S
9
2
]
.
C
N
o
i
b
-
q
[
3
v
0
6
3
6
0
.
0
1
6
1
:
v
i
X
r
a
Keith Smith1,2,∗, Daniel Ab´asolo3, & Javier Escudero1
ABSTRACT
Research into binary network analysis of brain function faces a methodological challenge in selecting an appropriate threshold
to binarise edge weights. For EEG phase-based functional connectivity, we test the hypothesis that such binarisation should take
into account the complex hierarchical structure found in functional connectivity. We explore the density range suitable for such
structure and provide a comparison of state-of-the-art binarisation techniques, the recently proposed Cluster-Span Threshold
(CST), minimum spanning trees, efficiency-cost optimisation and union of shortest path graphs, with arbitrary proportional
thresholds and weighted networks. We test these techniques on weighted complex hierarchy models by contrasting model
realisations with small parametric differences. We also test the robustness of these techniques to random and targeted topological
attacks.We find that the CST performs consistenty well in state-of-the-art modelling of EEG network topology, robustness to
topological network attacks, and in three real datasets, agreeing with our hypothesis of hierarchical complexity. This provides
interesting new evidence into the relevance of considering a large number of edges in EEG functional connectivity research
to provide informational density in the topology.
INTRODUCTION
Functional connectivity assesses the interdependent relationships between time-series recorded at spatially separated brain
regions. Estimating and analysing functional connectivity using network science is an established methodology for extracting
functional information from brain recordings taken using various platforms, most prominently the Electroencephalogram (EEG)
and the Magnetoencephalogram (MEG) for high temporal resolution and functional Magnetic Resonance Imaging (fMRI) for
high spatial resolution [1]. Networks are proficient in their ability to capture the interdependent activity which underlies brain
function [2], enabling powerful methods for classification of clinical patients [3] including in Alzheimer's disease (AD) [4]
and Schizophrenia [5]. Such analysis generally falls under the nomenclature of brain networks [1], which also includes studies
of the brain's physical connections referred to as structural connectivity.
Functional connectivity defined between all possible pairs of brain regions, whether sensors or cortical sources from the EEG
or the MEG or partitioning of spatial regions in fMRI, present the researcher with a full adjacency matrix whose entries are
only distinguished by relativity of weights. Since the most popular network science techniques are based on binary networks,
initial efforts on the analysis of brain network topologies were implemented via the binarisation of the weights using some
arbitrarily chosen thresholds with some good success [1], [6], [7]. Still, studying the original weighted networks holds appeal
in that it is more direct and has advantages of maintaining the information of relative strengths of connections [8]. However,
these computed weights are known to vary due to any number of different pre-processing choices or connectivity analyses
implemented, thus complicating comparisons and obfuscating results [3], [9]. Furthermore, since the weights of dependency
measures generally follow a non-scaling distribution between 0 and 1, many interesting topological considerations in binary
networks, such as concerning paths, become redundant in light of the fact that the shortest weighted path between any two
nodes is likely to be just the weighted edge connecting them [10]. Therefore, binarising the weights remains a more widely
used approach which can explain the main topology of the underlying activity while alleviating methodological biasing and
topological redundancy of weights.
Selecting a method to binarise the network is thus seen as a major step in network construction in which the researcher
is presented with a large degree of subjective choice [2], [9], [11], [12]. Because of this, recent research emphasises the
importance of solutions to the thresholding or binarisation problem in functional connectivity [12]–[20]. While some find
sparsity desirable based on the physiological hypothesis that function should be regarded as emerging through physically
connected regions explained by a low wiring cost [6], [15], [19], others propose that less strong connections do not necessarily
1Institute for Digital Communications, School of Engineering, University of Edinburgh, West Mains Rd, Edinburgh, EH9 3FB, UK
2Alzheimer Scotland Dementia Research Centre, Psychology Department, University of Edinburgh, 7 George Square, Edinburgh, EH8 9JZ, UK
3Centre for Biomedical Engineering, Department of Mechanical Engineering Sciences, Faculty of Engineering and Physical Sciences, University of Surrey,
∗PhD funded by EPSRC, e-mail: [email protected]
Guildford, United Kingdom
2
mean lower importance [21], [22] and, indeed, in real data analysis higher densities have been found to be as or more
relevant [7], [18], [23]. In this study we focus on a particular case of phase-based connectivity obtained from EEG signals
and hypothesise that the recently found complex hierarchical topology in this modality [24] contributes important information
found only in higher densities.
In the case of EEG recorded at the scalp, one measures the electromagnetic activity of the brain after its propogation
through the layers of bone and tissue of the head and is thus susceptible to noise and volume conduction effects. Nonetheless,
its combination of practicality, portability and high temporal resolution make it a very appealing methodology. It has recently
been uncovered that the hierarchical structure of EEG functional connectivity is a key aspect of its informational complexity
[24]. Indeed, this article also showed that functional connectivity is characterised by high degree variance which is indicative
of the large range of the general strength of network nodes. That is, one can expect that certain nodes have generally large
adjacent weights, while others may have generally small adjacent weights. For a given node, the relativity of weight magnitudes
of edges adjacent to the other nodes in the network is thus important to keep track of throughout the network and not just in
the largest connections and nodes in the highest hierarchy levels as would be promoted in sparse densities, see Fig 1. Thus
we propose that any useful binarisation technique for EEG phase-based functional connectivity should necessarily be able to
account for the density of information inherent in a broad complex hierarchical structure. It should follow that sparsity is not
necessarily a desirable feature of such brain functional networks and also that statistical thresholding on a case by case basis
of the connectivity computations does not necessarily translate to a topological advantage in the resulting networks.
Fig 1. Illustration of the likelihood of an edge appearing between nodes i and j, in hierarchy levels denoted by the x and y
axes, in the binarised form of a weighted hierarchical network. Left, the effects of increasing binarised network density
(strongest weights kept) on the hierarchical information of the network where black indicates 0% density and white indicates
100% density.
To provide rigorous simulation results for binarisation techniques, we implement the Weighted Complex Hierarchy (WCH)
model [24]. To our knowledge, this is the only generative weighted model of EEG functional connectivity to date which
approximates the topological characteristics of EEG functional connectivity over the full range of densities. Moreover, the
parameters of the model provide a fine-tuning of hierarchical topology which allows us to create a ground truth of subtly
different topologies. Here we exploit this to assess the ability of binarisation methods to correctly identify topological differences
between networks. This is intentionally done to echo the setup of a neuroscientific study and make the simulations as relevant
as possible to the research community. This follows since, for the lack of a ground truth, studies in network neuroscience are
generally based on contrasting conditions such as in cognitive tasks or contrasting populations where, for example, network
features of patients are compared against those of healthy, age-matched controls [4].
Using simulations, we seek to clarify how network size and density range may effect the ability to discern small topological
differences in network topology. In analysis, we compare state-of-the-art non-arbitrary binarisation techniques with a number of
arbitrary percentage thresholds. Here we provide an extensive and full comparison of state-of-the-art non-arbitrary binarisation
techniques- the Minimum Spanning Tree (MST) [15], [25], Union of Shortest Paths (USP) [17], the Cluster-Span Threshold
(CST) [10], [26], Efficiency Cost Optimisation (ECO) [19]. We also compare with weighted networks to directly compare
binary and weighted approaches, as well as a number of arbitrary density thresholds for distinguishing differences. The MST
and ECO create very sparse networks while the CST is notable in its production of consistently medium density networks
(30-50%) and the density of the USP depends on the distribution of edge weights [10].
We further analyse these techniques when the simulations are subject to random and targeted topological attack. We regard
these as random and targeted attacks [27] which preserve the network size. This is desirable given that many metric values
are dependent on network size [28]. By randomising a percentage of weights in populations of subtly different complex
3
hierarchical networks in parallel we can test how well the binarisation techniques can still uncover the differences between
these populations, under varying sizes of 'attack'. We implement these analyses to test the binarisation techniques' robustness
in representing true network characteristics in the face of noise and/or outliers in the estimation of coupling between brain
time series.
We go on to apply our non-arbitary binarisation techniques to three real EEG datasets. We compare our thresholds on
distinguishing the well known alpha activity existing between eyes open vs eyes closed resting state conditions in healthy
volunteers with a 129 channel EEG [29]. We then compare these techniques for distinguishing visual short-term memory
binding tasks in healthy young volunteers with a 30 channel EEG [26]. Finally, we compare our techniques in distinguishing
between AD patients and healthy control in a 16 channel EEG set-up [30]. The varying sizes of these networks provides evidence
for the translatability of the methods to different network sizes in the applied setting. The scripts, functions and data sets used
in this study are available at the University of Edinburgh's data depository: http://datashare.is.ed.ac.uk/handle/10283/2783.
METHODS
This section details the network simulations (A), binarisation techniques (B), network metrics (C), statistical tests (D) and
real datasets (E) used in this study. Let W be a weighted adjacency matrix for an undirected graph G = (E,V) such that wij
is the weight of the edge between vertices i and j in G and G has no multiple edges or self-loops. A simple graph is a graph
such that wij ∈ {0, 1} for all i and j and wij = 0 for i = j , where 1 indicates the existence of an edge. Then n = V is the
size of the network and m is the number of undirected edges so that 2m is the number of non-zero entries in the symmetric
adjacency matrix.
Simulated experiments
Complex hierarchy models: The Weighted Complex Hierarchy (WCH) network takes the existence probabilities of edges in
an Erdos-R´enyi random network [31] as the base weights of the edges. It then randomly chooses the number of hierarchical
levels (between 2 and 5) before randomly assigning nodes to these h levels, based on a geometric cumulative distribution
function with a default parameter of 0.6. All of the edges adjacent to nodes in a given level are then provided an addition
weight of (h − 1)s for some suitably chosen strength parameter, s. The weight of an edge in the WCH model is then
¯wij = wij + (h(i)− 1)s + (h(j)− 1)s, where wij is distributed uniformly in [0, 1] and h(i) is the hierarchical level of node i.
By fine-tuning the strength parameter, these networks have been shown to mimic the topology of EEG networks formed from
the weighted phase-lag index [24]. Small differences in parameter determine the strength of the hierarchical structure of the
network: s = 0 gives a random network; s= 1 gives a strict 'class-based' topology determined by the node hierarchical levels.
Thus, a larger strength parameter provides a network with a more rigid hierarchical structure.
We repeat this methodology for networks with 16, 32, 64 and 128 nodes, spanning a large range of network sizes as used
in current research, e.g. see [4].
Experimental design: For the simulations, we follow the procedure as illustrated in Fig 2. WCH models are generated as
a ground truth to test ability of binarisation techniques to distinguish subtly different populations of size 20. These different
populations are generated using realisations of the WCH model with small differences in the strength parameter, s. The
procedure thus follows that of a typical clinical study, where small populations of contrasted conditions are analysed using
network science techniques with statistical tests used to determine significance of the differences between the populations.
Fig 2. Methodological steps in the evaluation of binarisation techniques for determining ground truth topological differences.
4
Random and targeted topological attacks: We test the robustness of the given binarisation techniques by subjecting these
same simulated networks to random and targeted 'topological attacks' before implementing similar topological comparative
analysis as above. Random and targeted attacks were originally formulated by deleting entire nodes from the network [27]. We
implement a weight randomising approach, thus preserving network size which is important when comparing different metrics
[28]. Further, this is more relevant to brain networks where the network size is determined a-priori, but rather the information
recorded at the nodes are susceptible to attacks. These random topological attacks are implemented by substituting the WCH
models weights with corresponding non-zero entries of a sparse, randomly weighted adjacency matrix. We implement this
comparison by increasing the density of the sparse matrix, i.e. densities of 0, 0.05, 0.1, ..., 0.95, 1. Targeted topological attacks
are implemented similarly except the attacks are restricted only to those nodes whose average adjacent weight is over one
standard deviation above the mean, relating to those nodes with abnormally strong connectivity. Such strongly weighted nodes
are known as 'hub' nodes for their importance to the topology of the network.
Network binarisation
Minimum spanning tree: The MST is a construction based binarisation approach which obtains a tree by selecting the
strongest edges of the network such that the network is fully connected and no cycles are present. The algorithm for its
construction is well known [32] and included in popular toolboxes [33].
Union of shortest paths: The USP is another construction based approach to unbiased network binarisation [17]. The shortest
path between two nodes in a network is the set of edges with the minimum sum of weights connecting them. This can be
constructed using e.g. Dijkstra's [34] algorithm to find the shortest paths between each pair of nodes in the network, adding
all the edges of those paths to an initially empty binary network. Because connectivity has an inverse relation to distance,
the weights of the network must first be relationally inversed in order to construct the shortest paths. This inversion process
can take several forms which involves a certain amount of subjective discretion and depends largely on the distribution of the
original weights. For our study we choose W = − ln(W)/α, where
α = min{N} s.t. maxi,j( wij) < 1,
(1)
as it has been shown to offer a better spread of metric magnitudes which is important for shortest path problems [10].
Cluster-span threshold: The CST chooses the binary network at the point where open to closed triples are balanced [26].
This balance occurs when CGlob = 0.5, which is obvious from the definition. Importantly, the balancing of this topological
characteristic necessarily endows the binary network with a trade-off of sparsity to density of edges. We see this since a
network is sparse if most triples are open and dense if most triples are closed. It is hypothesised that this balance achieves an
informational richness useful for capturing different topologies of EEG functional connectivity [24].
The algorithm for the CST computes the binary networks for each possible number of strongest edges between 15% to 85%,
rounded to the closest real edge density. The clustering coefficient is then computed for each of these networks, obtaining a
vector C = {C15, C16, . . . , C84, C85}, where Ci is the clustering coefficient of the binary network at the ith % density, rounded
to the nearest number of edges. Then the network of the CST is the binary network corresponding to Z = argmini(Ci − 0.5),
i.e. the threshold achieving minimum value of the vector C minus the clustering coefficient value which obtains an equilibrium
between triangles and non-triangle triples, 0.5. The values of 15% and 85% are chosen as safe values based on experimental
evidence [10], [26]. Particularly, below 15%, real brain networks can have a tendency to fracture into more than one component,
thus making calculations of metric values, including the clustering coefficient, inconsistent and unreliable [7].
Efficiency-cost optimisation threshold: The ECO proposes a threshold to keep the strongest 1.5n edges (equivalent to a
density threshold of 3/(n-1)) which is an approximation based on consistent observations of simulated and real brain networks
of the maximum ratio of the combined local and global network efficiencies and density [19]. It is hypothesised that such
a trade-off of network efficiency and sparsity provides networks which are meaningful to the concept of economy in brain
function [35].
Arbitrary proportional thresholds: Arbitrary thresholds can be chosen by either choosing a weight above which edges are
kept and below which edges are discarded, or by choosing a percentage of strongest weighted edges to keep in the network.
The latter choice is more robust and easier to compare between different set-ups and subjects because it keeps the connection
density constant and thus is not affected by the values of the weights, which may vary wildly particularly when considering
the comparison of different connectivity measures. In order to cover the density ranges of both the sparsity hypothesis and the
hierarchical complexity hypothesis, we choose arbitrary thresholds which maintain the strongest 5%, 10%, 20%, 30%, 40%,
and 50% of edges to make sure we cover the relevant array of connection densities whilst reducing redundancy. Note, very
sparse densities are already covered by the MST and ECO thresholds.
Network metrics
To analyse the simulated and real EEG networks we use a variety of common metrics.
• The characteristic path length, L, is the average of the shortest path lengths between all pairs of nodes in the graph
[36].
5
• The efficiency, Eff of a weighted network is the mean of one over the shortest path lengths, thus inversely related to L
[33].
• The diameter, D, of a graph is the largest shortest path length between any two nodes in the graph [15].
• The clustering coefficient, C, is the mean over i of the ratio of triangles to triples centred at node i [36].
• The weighted clustering coefficient, CW, is a weighted version of the clustering co-efficient for binary networks.
• The leaf fraction, LF, of a tree is the fraction of nodes in the graph with degree one. Note, every path containing such
a node either begins or ends at that node [15].
• The edge density, P , is the ratio of the number of edges in the graph to the total possible number of edges for a graph
with the same number of nodes, i.e. P = 2m/n(n − 1) [36]. For the CST, P takes an inversely relational position
to C of proportional thresholds. This can be seen by considering two weighted networks whose values of C increase
monotonically with increasing P and such that one has higher values of C than the other, which is a working assumption
in our case. Then the network with the greater values of C will attain its CST at a lower density, P . In a similar vein, P
of the USP is inversely related to L of proportional thresholds- the higher the density of the USP, the shorter the average
shortest path in the weighted network.
• The degree variance, V , is the variance of the degrees of the graph which distinguishes the level of scale-freeness present
in the graph topology [24].
network [15].
• The maximum degree, MD, of a network is just as named- the degree of the node with the most adjacent edges in the
For each binarisation technique we choose three metrics to analyse the subsequent binary networks. These differ for each
technique because of the construction of the network. Particularly, the MST metrics are chosen based on a study of Tewarie
et al. [15]. Similarily, we choose three weighted metrics for analysing the original weighted networks. These choices can be
found in Table I. We try as much as possible to stick to three main categories of metrics for each binarisation technique:
segregation (M1 in Table I), efficiency (M2) and irregularity (M3) [1], [24]. This notably deviates for M3 in the weighted case
where the mean weight of the network edges, µW, is an appropriate and more obvious choice of metric than the variance of
those weights.
TABLE I. Grouped Topological Metrics- Three for Each Network Type
Metric
M1
M2
M3
CST MST
P
L
V
USP
C
LF
D
P
M D V
ECO Weight %T
C
L
V
CW
Eff
µW
C
L
V
Functional connectivity: For the real EEG datasets, FIR bandpass filters were implemented for α (8-13Hz), β (13-32Hz) or
both bands, as specified in the Materials. The filter order of 70 is used to provide a good trade-off between sharp transitions
between the pass and stop bands while keeping the filter order low. The filtered signals were then analysed for pairwise
connectivity using a suitable representative of the various phase-based methods [37], the Phase Lag Index (PLI) [38], to
account for the important phase-dependence information seen to underlie electrophysioliogical brain activity [39]. This avoids
the problems of volume conduction found in e.g. correlation or coherence by relying solely on the phase lead/lag information
of the signal to negate the redundant correlation effects of synchronous time-series analysis. The PLI between time series i
and j is defined as
(2)
where the instantaneous phase at time t, φk(t), is regarded as the angle of the Hilbert transform of the signal k at time t.
These values were averaged over trials to obtain one connectivity matrix per subject per condition.
P LIij = (cid:104)sgn(φi(t) − φj(t))(cid:105),
Statistical testing
In the simulations we undergo 50 simulated trials of two populations of 20 networks for each of a combination of populations.
A population of networks is selected from a bank of 1000 WCH networks with given strength parameter. These banks exist for
s = 0, 0.05, 0.1, 0.15, 0.2, 0.25 and 0.3. The other population in the trial then comes from a WCH bank with strength parameter
with 0.05 difference. We undergo such trials for all possible combinations of 0.05 differences. We binarise these networks using
each of our binarisation methods. We then compute the three metrics, M1, M2 and M3, for each of these networks (weighted
metrics are computed from the original weighted networks). We perform population t-tests of these metrics for the WCH
binarisations. The assumption of normality of these populations of values are validated with z-tests. We choose the best metric
of the three to represent the ability of the binarisation method to discern subtle topological differences where the 'best' metric
is chosen as that which attains the maximum number of significant p-values out of the 50 simulated trials which are less than
the standard α = 0.05 level. If two or more metrics obtain the maximum value, we then choose the one with the lowest mean
log of p-values. Choosing the log in this instance emphasises the importance of smaller p-values for distinguishing differences.
The number of differences discovered, taken as a percentage of the total number of trials run, then represents the 'accuracy'
of the binarisation technique at distinguishing the ground truth, i.e. that the topologies of the populations are different.
TABLE II. Grand average percentage of topological differences discovered between Weighted Complex Hierarchy models.
6
Method
Grand Average
CST
Weight
71.3% 22.4% 50.0% 53.38% 40.5%
MST
ECO
USP
50% T
70.4%
40% T
71.5%
30% T
67.3%
20% T
60.0%
10% T
41.3%
5% T
41.6%
MATERIALS
Eyes open - eyes closed resting state data: The eyes-open, eyes closed 129 node dataset is available online under an Open
Database License. We obtained the dataset from the Neurophysilogical Biomarker Toolbox tutorial [29]. It consists for 16
volunteers and is downsampled to 200Hz. We used the clean dataset which we re-referenced to an average reference before
further analysis. The data were filtered in alpha (8-13Hz), according to known effects [40], using an order 70 FIR bandpass
filter with hamming windows at 0.5Hz resolution. For each subject we take one arbitrarily long 1000 sample epoch (5s) from
the 1000-2000th samples, for each subject.
Visual short-term memory binding task data: We study a 30-channel EEG dataset for 19 healthy young volunteers partici-
pating in different VSTM tasks. Full details of the task can be found in [23]. Written consent was given by all subjects and
the study was approved by the Psychology Research Ethics Committee, University of Edinburgh. The task was to remember
objects consisting of either black shapes (Shape) or shapes with associated colours (Binding), presented in either the left or
right side of the screen. The sampling rate is 250 Hz and a bandpass of 0.01-40 Hz was used in recording. Based on relevant
results [23], [26], the data were filtered in beta (13-32Hz) using an order 70 FIR bandpass filter with hamming windows at
0.5Hz resolution. The epochs are 1s long and the number of trials is 65.7 ± 9.27 (mean ± SD). PLI adjacency matrices are
averaged over trials.
Alzheimer's disease data: The EEG recordings were taken from 12 AD patients and 11 healthy control subjects. The patients
-5 men and 7 women; age = 72.8 ± 8.0 years, mean ± standard deviation (SD)- were recruited from the Alzheimer's Patients'
Relatives Association of Valladolid (AFAVA). They all fulfilled the criteria for probable AD. EEG activity was recorded at
the University Hospital of Valladolid (Spain) after the patients had undergone clinical evaluation including clinical history,
neurological and physical examinations, brain scans and a Mini Mental State Examination (MMSE) to assess their cognitive
ability [41]. The ethics committee of the Hospital Clinico Universitario de Valladolid approved the study and control subjects
and all caregivers of the patients gave their written informed consent for participation. The 16 channel EEG recordings were
made using Profile Study Room 2.3.411 EEG equipment (Oxford Instruments) in accordance with the international 10-20
system. Full details can be found in [30]. The data were filtered both in alpha (8-13Hz) and beta (13-32Hz) as in [10], for
separate analysis, using an order 70 FIR bandpass filter with hamming windows at 0.5Hz resolution. Recordings were visually
inspected by a specialist physician who selected epochs with minimal artefactual activity of 5s (1280 points) from the data for
further analysis. The average number of these epochs per electrode per subject was 28.8 ± 15.5 (meanSD).
The EEG PLI adjacency matrices used in this study are available at the University of Edinburgh's data depository at
http://datashare.is.ed.ac.uk/handle/10283/2783.
Sensitivity to subtle topological differences in synthetic EEG connectivity
RESULTS
Fig 3 shows the results for differences discovered between WCH models with differences in strength parameter of 0.05. The
grand averages are shown in Table II The CST is shown to outperform other non-arbitrary methods in general. In testing the
comparisons of the WCH model with varying strength parameter, s, it discovers significant differences at the α = 5% level
71.3% of the time over all strength comparisons and network sizes, Table II. On the other hand, the MST discovers just 22.4%
of the differences, the USP discovers 50% of the differences and the weighted metrics discover just 40.5% of the differences.
Out of these methods, in fact, it discovers the most differences in all but two cases- those being the 0.1 vs 0.15 comparison in
the 16 node networks and the 0.25 vs 0.3 node comparison in the 128 node cases, of which the USP is best on both occasions.
In comparison with arbitrary percentage thresholds the CST appears to perform approximately the same as the 40%
proportional threshold which discerns a slightly higher rate of 71.5% of differences over all cases. The 50% threshold also
appears to be good at discerning differences here with an overall rate of 70.4% of differences discovered. These results agree
with the hypothesis that complex hierarchical structures are best captured by larger density ranges. It is important to recall that
we need non-arbitrary solutions rather than simply to find the best possible threshold for these specific simulations. With this
in mind we can see that, compared to the other techniques, the CST outperforms the field in this study.
Robustness to random and targeted topological attacks
The robustness to random and targeted topological attacks is evaluated by comparing the metrics from the attacked WCH
models over all non-arbitrary binarisation techniques using population t-tests as before. For these analyses we look at the case
in Table II with the maximum ratio between the mean and standard deviation of accuracy over binarisation techniques, i.e.
the case which maximises the ratio of average performance and comparability of performances. This happens in the 32 node
case (6.3538) with differences in strength parameter of s = 0.1 and s = 0.15. The grand percentage over all sizes of attack
7
Fig 3. Percentage of topological differences discovered from population t-tests between WCH models. The y-axis shows
s = a vs. s = b for WCH populations with strength parameter, s. The x-axis shows the binarisation method used where W is
the weighted approach and percentages indicate arbitrary density thresholds.
for p-values below 0.05 for each metric is presented in Fig 4 for both random and targeted topological attacks. Generally,
the binarisation methods as well as the weighted metrics are more robust to targeted attacks than to random attacks. Notably,
the CST maintains the highest average accuracy of distinguishing topological differences for all metrics and both random and
targeted attacks. This is particularly evident in the targeted attacks. For both kinds of attack, the weighted metrics come in
second best while the ECO, the USP and MST perform relatively poorly.
The metric achieving highest accuracy for each binarisation technique and for each size of attack is shown in Fig 5 for both
random and targeted topological attacks.
Strictly in terms of robustness, as opposed to best accuracy, the weighted networks prove the best, with the least detriment
noted by increasing the size of attack in its topological acuity (green). The CST also does well here. For the random topological
attacks, even at 50% of connections attacked, the CST notes an accuracy of 70% (blue line). The USP (yellow), MST (red)
and ECO (purple) networks are not at all robust to random topological attacks in this scenario with immediate drop offs on
the implementation of attacks.
For the targeted topological attacks (Fig 5, right), the CST network (blue) shows the most resilience with no noticeable
depreciation of accuracy. The other methods, in contrast, show a notable decrease in accuracy as more weights are randomised.
Real dataset results
We maintain our focus on comparing non-arbitrary methods since arbitrary approaches are inappropriate for neurophysio-
logical studies where one can pick from an order of n(n − 1)/2 thresholds.
Table III shows the results for distinguishing the difference in α activity well known to exist between eyes-closed and
eyes-open conditions [40]. The CST finds a significant difference in V of eyes open and eyes closed resting state activity
indicating that the phase-dependent topology of EEG activity is less scale-free in the eyes-open condition implying greater hub
8
Fig 4. Percentage of topological differences discovered from population t-tests via binarisation methods (CST,MST, USP,
ECO and weighted network (W)) between WCH models (s = 0.1 vs. s = 0.15) with random and targeted network attacks.
M1, M2 and M3 as in Table I
Fig 5. Percentage accuracy of method for distinguishing topological differences between attacked WCH models against the
size of those attacks for random topological attacks (left) and targeted topological attacks (right). The values plotted are the
maximum from the three metrics, M1, M2 and M3 (as in Table I), for the corresponding technique as indicated in the legend.
dominance in the eyes-closed condition, see Fig 6, left. All of the weighted metrics also find significant differences. Neither
the MST or USP find any differences between these conditions. Probing further, ρ(M 1, M 2) being the correlation coefficient
of metric values across subjects of metrics M1 and M2 as defined in Table I, the weighted metrics in this case are all very
highly correlated (all > 0.95 Pearson correlation coefficient, ρ, Table IV) within condition. Therefore they cannot be seen
to provide any distinct topological information. The corresponding correlations of the CST show a more distinct topological
characterisation, see Table IV.
TABLE III. The p-values from paired t-tests between eyes open (EO) and eyes closed (EC) in α-band 129-channel EEG
PLI Networks. Underline: Best value for each method. Bold: Significant values. M1, M2 and M3 as in Table I
Metric
M1
M2
M3
CST
0.7504
0.9319
0.0006
MST
0.4178
0.4513
0.9616
USP
0.5063
0.9942
0.6577
ECO
0.6034
0.5281
0.6805
Weight
0.0016
0.0034
0.0016
Table V shows the results for distinguishing the difference in β activity existing between Shape and Binding tasks when tested
in the Left and Right sides of the screen separately. A significant difference is found in V of the CST networks in the Right
condition. This indicates that the phase-dependent topology of EEG activity is less scale-free in the Binding condition implying
greater hub dominance in the Shape condition, see Fig 6, right. On the other hand, a significant difference (p = 0.0059) is
found in C for the ECO networks in the Left condition. This indicates sparse density topology of EEG activity is less integrated
in the Binding condition.
TABLE IV. Pearson correlation coefficients of Metrics in Eyes Open (EO) - Eyes Closed (EC) Dataset for the CST and
Weighted Metrics (wgt)
Metric corr.
ρ(M1,M2)
ρ(M1,M3)
ρ(M2,M3)
EC (CST)
0.7662
-0.0458
-0.1695
EO (CST)
0.9692
-0.7301
0.6677
EC (wgt)
0.9993
0.9999
0.9994
EO (wgt)
0.9512
0.9996
0.9576
9
Fig 6. Scatter plots of degree variance for CST networks of Eyes Open vs Eyes closed resting state conditions in α, left, and
degree variance for CST networks of Shape only vs Shape-colour binding conditions in the Right screen in β, right.
Noticeably, the USP failed to find meaningful network information in this task because, even after transformation, all the
weight magnitudes were in a range such that the shortest weighted path between each pair of nodes was the weight of the
single edge joining them.
Table VI shows the results for distinguishing the difference in both α and β activity existing between AD patients and
healthy age matched control. For the CST, an effect is noticed in V of α activity (Fig 7, right) and a larger effect is found
in the P of β activity (Fig 7, left). Since P of CST networks is inversely relational to C of arbitrary threshold networks, this
tells us that β of AD patients is less segregated than control. Contrasting with this, the activity in α suggest a more scale-free
network in the alpha band of AD patients than in age matched control.
Fig 7. Box plots of connection density, left, and degree variance, right, for CST networks of AD and control in β and α,
respectively.
Density
The densities for the datasets used in this study are as in Table VII. For the USP we see both a dependency on network
size, where the WCH networks are less dense with increasing size, and on the distribution of weights, where analysis of real
datasets becomes implausible since connectivity, averaged over trials, creates a smaller spread and so redundant shortest paths.
The MST and ECO are dependent on network size as is obvious from their formulations. The CST binarises the network
consistently with a density between 0.3-0.5. From the models we notice that the higher the paramater, s, the less dense are
the resulting CST networks. Network size appears to have much less effect, which provides evidence to suggest that the CST
is dependent on topology, but not on network size.
TABLE V. The p-values from paired t-tests between Shape only and Shape Colour Binding tasks in β-band 30-channel EEG
PLI networks. Formatting as in Table III
10
Hemifield Metric
Left
Right
M1
M2
M3
M1
M2
M3
CST
0.5128
0.0898
0.8997
0.5877
0.9196
0.0088
MST
0.7186
0.1383
0.0911
0.1919
0.5716
0.8146
USP
-
-
-
-
-
-
ECO
0.0059
0.1870
0.0238
0.5504
0.5038
0.2138
Weight
0.1007
0.1010
0.1010
0.7742
0.7733
0.7733
TABLE VI. The p-values from population t-tests of network measures AD and control in 16-channel EEG PLI Networks
Band Metric
M1
Alpha M2
M3
M1
M2
M3
Beta
CST
0.0852
0.3634
0.0406
0.0062
0.0529
0.1782
MST
0.3468
0.2630
0.7324
0.4618
0.6245
0.5437
USP
0.1167
0.1081
0.0942
0.1500
0.1485
0.1397
ECO
0.7496
0.0582
0.2089
0.9775
0.4946
0.2575
Weight
0.6736
0.4189
0.5570
0.7080
0.4215
0.5564
DISCUSSION
From the simulation results of complex hierarchy models we see from proportional thresholds that a larger density range
is more effective than sparse models. This agrees with our hypothesis that complex hierarchical models contain a density
of information beyond what sparse levels of binarisation can reveal. The fact that the real results for EEG datasets confirm
the results in simulations provides further strength to the argument that EEG functional connectivity is highly hierarchically
complex and so that sparsity is not always the best working hypothesis for functional brain networks. Other evidence in EEG
studies alluding to the benefit of medium density ranges has been documented [7], [18], [23]. For a physiological rationale for
medium densities we can, for instance, regard function as not only emerging through physical connections as hypothesised in
the sparsity hypothesis, but through globally interdependent synchronisations via rapid interregional communications.
Binarised networks generally outperformed weighted approaches in our simulations. Furthermore, weighted network metrics
should be used with caution. Particularly, we advise checking their correlations with the mean connection strength.
Other efforts looking to study the role of less strong connections in brain networks have also considered intermediate
thresholding by considering networks constructed from connections within different ranges of connectivity strength [21], [22].
However, studying such topologies is hindered by the fact that the true overlying hierarchical structure becomes hidden. Nodes
having more edges in an intermediate level does not, for instance, indicate that that node is a hub, but rather that most of its
connections lie within the given range. That is to say that intermediate connections maybe interesting to study, but constructing
network topologies from them for analysis is rather obscure. The role of the weakest connections, or, if you will, topological
gaps of brain connectivity is also an active area of research [22]. One can consider, in fact that medium density binarisation
does much more to account for such features than sparse binarisations since these gaps become much more defined in higher
densities.
As an important addition, the results show that the random topological attacks, rather than targeted topological attacks, are
the most effective at deconstructing the topology of our simulations. This perhaps seems counter-intuitive, but can be explained
by the fact that only the very top levels of the hierarchy are attacked in the targeted setting, whereas the topology in the
remaining levels remains largely intact and, in fact, a new 'top level' emerges, maintaining the differences exhibited in the
strength parameter, s, between the two sets of topology. In fact this agrees with previous functional connectivity studies which
detailed the greater resilience of functional brain networks to targeted attacks [42], [43]. These simulations thus provide the
clues as to how the hierarchical structure of functional brain networks play a vital role in this resilience.
The results for V in both the Eyes open vs closed and Shape vs Binding datasets combined can explain that more intensive
stimulation (eyes open and Binding) leads to a drop in network efficiency where more localised activity is required for higher
functional processing [1], [3]. The results for the AD dataset indicate both the increased power in binarisation with the CST
compared to other approaches and highlights the importance of binarisation itself for distinguishing dysfunctional AD topology.
AD network studies, over varying platforms, network sizes and density ranges, have been found to show seemingly contrasting
results [4]. Particularly, Tijm's et al. [4] reported that these studies were at different density ranges, and in many cases the
density range was simply not recorded. Importantly, no functional studies reported density ranges over 25%. Nonetheless, we
note that our results are in agreement with a 149 node MEG PLI study by Stam et al. [44], showing lower clustering in AD
than control (density not recorded). This is indicative of a move to a more random topology [45].
In terms of network size our simulations suggest that the larger the networks are, the more likely it is that topological
differences will be picked up by commonly used metrics. This trend is bucked by the MST for which there is a marked drop
off from 32 nodes to 128 nodes. This, however may be explained by the fact that at 32 nodes, the MST makes up 2/n = 6.25%
of all possible connections whereas at 64 nodes this percentage is 3.12% and for 128 nodes it is just 1.56% which is in line
with the previous discussion that lower network densities can inhibit the ability to find topological differences.
TABLE VII. Mean and standard deviation (M ± SD) of densities of CST, USP, MST and ECO networks. WCH## =
Weighted Complex Hierarchy model of size ##; s = strength parameter of WCH model; PLI## = Phase-Lag Index networks
of size ##.
11
Dataset
WCH16
WCH32
WCH64
WCH128
PLI16
PLI30
PLI129
Condition
s = 0.1
s = 0.15
s = 0.2
s = 0.25
s = 0.3
s = 0.1
s = 0.15
s = 0.2
s = 0.25
s = 0.3
s = 0.1
s = 0.15
s = 0.2
s = 0.25
s = 0.3
s = 0.1
s = 0.15
s = 0.2
s = 0.25
s = 0.3
Patient
Control
Shape Left
Shape Right
Bind Left
Bind Right
Eyes closed
Eyes open
CST
0.49 ± 0.04
0.46 ± 0.05
0.44 ± 0.06
0.41 ± 0.06
0.40 ± 0.06
0.48 ± 0.02
0.45 ± 0.03
0.42 ± 0.04
0.39 ± 0.04
0.37 ± 0.04
0.48 ± 0.01
0.45 ± 0.03
0.41 ± 0.04
0.38 ± 0.04
0.36 ± 0.04
0.47 ± 0.01
0.45 ± 0.02
0.40 ± 0.04
0.37 ± 0.04
0.35 ± 0.03
0.47 ± 0.04
0.41 ± 0.05
0.41 ± 0.07
0.40 ± 0.06
0.42 ± 0.06
0.41 ± 0.08
0.35 ± 0.07
0.36 ± 0.07
ECO
2/16
2/32
3/31
MST
3/15
= 0.20
USP
0.53 ± 0.08
0.59 ± 0.08
0.62 ± 0.08
0.64 ± 0.09 = 0.125
0.64 ± 0.05
0.45 ± 0.07
0.50 ± 0.06
0.52 ± 0.07
0.53 ± 0.07 = 0.0625 = 0.0968
0.54 ± 0.08
0.38 ± 0.05
0.42 ± 0.05
0.44 ± 0.05
0.45 ± 0.05 = 0.0312 = 0.0476
0.45 ± 0.06
0.33 ± 0.04
0.35 ± 0.03
0.37 ± 0.03
0.37 ± 0.04 = 0.0156 = 0.0236
0.37 ± 0.06
1 ± 0
0.98 ± 0.05 = 0.125
1 ± 0
1 ± 0
1 ± 0
1 ± 0
0.98 ± 0.04
0.98 ± 0.06 = 0.0155 = 0.0234
2/30
= 0.0667 = 0.1034
3/15
= 0.20
2/129
3/128
2/128
3/127
2/64
3/63
2/16
3/29
The CST is presented here as a sensitive and powerful binarisation technique for network modelling of EEG phase-based
functional connectivity. In simulations it performed to a high standard in all network sizes and topological comparisons as
well as in robustness to topological attacks. This was echoed in the results of the real data sets where it was consistently able
to identify differences in the presented conditions with not obviously correlated metrics. From the simulation results we can
infer a large part of this ability to the density range in which the CST binarises the network. We must note, of course, that
all of our real data were from EEG recordings and thus we are cautious of similar comparisons for e.g. fMRI. Indeed, we
must acknowledge the limitations and narrow focus of this study for EEG PLI networks. Further, although we conjecture that
hierarchical complexity of network topology may be behind the CSTs success, there remain unanswered questions and there is
certainly scope for better topological thresholds to be developed based on such hypotheses. We hope this study will stimulate
interesting discussions and inspire future research in this direction.
The MST is seen to be robust to fluctuations of the underlying network [15]. It holds appeal in studies where sparsity is
desirable, showing utility in a number of other studies [46]–[49], although it must be noted that in these studies it was not
compared with other binarisation methods. In our study, however it appears ineffective and particularly so in larger networks
which noticeably corresponds to the MST making up less and less of the connection density as the network grows. This concurs
with a recent study where we argued that the robustness to fluctuations also means a poverty of information, supported by
evidence from an EEG dataset of cognitive tasks [23].
The USP is the set union of those edges which form the shortest paths between all possible pairs of electrodes. Since, in
general, all weights of a functional connectivity network lie between 0 and 1, it is likely that a large percentage of the shortest
paths in the network will be constituted of just the single edge joining those nodes. Thus, we can expect very high density
networks which only differ in topology by the weakest connections. Indeed, in the original paper [17] the authors did not
implement any transformation of the weights and reported densities above 90%. To try and counter this unwanted outcome
we used a negative exponential transform of the weights before extracting the union of shortest paths, however, in the end it
appeared that this was limited in its ability to mitigate the flaws of this method. This was most apparent in the VSTM tasks
where it turned out that every shortest path was just the edge between each pair of nodes, redundantly returning complete
networks. We believe that further work would need to be done regarding the reliability of the USP in order to make it of use
to the neuroscience community.
Although generally outperformed by the CST, we note an agreement with the introductory paper of the ECO threshold
[19] that it generally outperforms the MST. Therefore, we would recommend it over the MST in cases where sparse densities
appear more important. It was also able to detect differences in the left hemifield condition of the VSTM dataset. The fact that
a mutually exclusive difference was found in the right hemifield condition with the CST suggests the interest in considering
how different density ranges may reveal different topological traits of conditions. For example, one may conclude from these
12
results that the backbone of the network is effected by binding in the left hemifield, but that in the right hemifield the binding
effect is notable rather in the 'fleshed out' regions of the network.
We note that research is also undertaken into statistical methods on expected values of connectivity methods to threshold the
network [2]. However, rather than resolving the arbitrary choice problem, it merely diverts it towards the statistical significance
paradigm, where arbitrary standards (e.g. α = 0.05) have long been adopted to mitigate an intractable problem. In our case
we can rely on graph topological techniques so the problem is not intractable. Moreover, as we have seen, it can be more
beneficial to include a large number of edges since it allows for richer information coming from the hierarchical relationships
of lower degree nodes. Further problems with a statistical approach relate to difficulties in finding the correct solutions for the
numerous new connectivity measures available in a way that is consistent and reliable, biases from the size of available data,
and, in the case of data surrogate methods, biases due to network size [2]. Other researchers look towards the integration of
different density ranges, however such an approach will have a tendency towards diluting the potency of potential differences
[13] or falling prey to the multiple comparison problem. We note that, without a ground truth, we rely on the assumption
that contrasting conditions provokes a contrast in functional connectivity. Further, although small sample sizes are common in
the literature, we fully encourage the move towards larger sample sizes from which more powerful and robust results can be
obtained. We find the PLI to be reliable and straightforward to interpret but we recognise that finding appropriate connectivity
measures is a much debated topic with many considerations including hypotheses of how brain function takes place; the part
and frequencies of the signals that should be used for a given paradigm; whether the measure should provide directedness;
and whether the signals should be orthogonalised or relocated to the source space.
CONCLUSION
The hierarchical topologies of simulated weighted complex hierarchy models and several real datasets of EEG functional
connectivity assessed from phase dependencies are found to be well characterised by non-arbitrary binarisations using the CST
and arbitrary density binarisations in the range of 40%-50%. It is conjectured that this is due to their topologies in this range
accounting for a wider range of hierarchical structure, i.e. not just the connectivity in the largest degree nodes. The CST and
weighted networks were shown to be robust to random and targeted topological attacks when compared with MST, USP and
ECO graphs. In three real datasets constituting varied neuroscientific questions, the medium density range which the CST
occupies does indeed appear to be useful with other evidence showing that the ECO is useful in sparse densities. Considering
both sparse and larger densities in tandem may prove a more effective way forward than either on their own. We were able
to successfully identify different topologies in resting states, in VSTM cognitive tasks and in AD patients compared with
control with a notable performance from the CST. This study also validates the WCH model as a sensitive tool for topological
comparisons of great relevance to the EEG.
We would like the thank Dr. Mario A. Parra (MAP) for providing the VSTM data. This data was supported by Alzheimer's
Society, Grant # AS-R42303 and by MRC grant # MRC-R42552. We would like to thank Dr. Pedro Espino (Hospital Clinico
San Carlos, Madrid, Spain) for his help in the recording and selection of EEG epochs of the AD dataset.
ACKNOWLEDGEMENTS
REFERENCES
1. Bullmore E, Sporns O. Complex brain networks: graph theoretical analysis of structural and functional systems. Nature Reviews Neuroscience.
10. Smith K, Abasalo D, Escudero J. A Comparison of the Cluster-Span Threshold and the Union of Shortest Paths as objective thresholds of EEG
functional connectivity networks from Beta activity in Alzheimer's disease. IEEE Proc EMBC 2016. 2016;2826-2829.
11. Papo D, Zanin M, Pineda-Pardo JA, Boccaletti S, Buld´u JM. Functional brain networks: great expectations, hard times, and the big leap forward. Phil.
Trans. R Soc. B. 2014;369(1653):20130525.
12. van Diessen E, Numan T, van Dellen E, van der Kooi AW, Boersma M, Hofman D, van Lutterveld R, van Dijk BW, van Straaten EC, Hillebrand A,
Stam CJ. Opportunities and methodological challenges in EEG and MEG resting state functional brain network research. Clinical Neurophysiology.
2015;126(8):1468-1481.
13. Ginestet CE, Nichols TE, Bullmore ET, Simmons A. Brain Network Analysis: Separating Cost from Topology Using Cost-Integration. PLOS One.
14. Schwarz AJ, McGonigle J. Negative edges and soft thresholding in complex network analysis of resting state functional connectivity data. NeuroImage.
2011; http://dx.doi.org/10.1371/journal.pone.0021570.
2011;55(3):1132-1146.
2009;10:186-198.
Soc. B. 2014;369(1653): 20130521.
2. Fallani F, Richiardi J, Chavez M, Achard S. Graph analysis of functional brain networks: practical issues in translational neuroscience. Phil Trans. R
3. Stam CJ. Modern network science of neurological disorders. Nature Reviews Neuroscience. 2014;15:683695.
4. Tijms B, Wink AM, de Haan W, van der Flier WM, Stam CJ, Scheltens P, et al. Alzheimer's disease: connecting findings from graph theoretical studies
of brain networks. Neurobiology of Ageing. 2013;34:2023-2033.
5. Calhoun VD, Eichele T, Pearlson G. Functional brain networks in schizophrenia: a review. Front. Hum. Neurosci. 2009;doi:10.3389/neuro.09.017.2009.
6. Achard S, Bullmore E. Efficiency and Cost of Economical Brain Functional Networks. PLOS One. 2007;doi:10.1371/journal.pcbi.0030017.
7. Stam CJ, Jones BF, Nolte G, Breakspear M, Scheltens P. Small World Networks and Functional Connectivity in Alzheimers Disease. Cerebral Cortex.
8. Rubinov M, Sporns O. Weight-conserving characterization of complex functional brain networks. NeuroImage. 2011;56(4):2068-2079.
9. van Wijk BCM, Stam CJ, Daffertshofer A. Comparing Brain Networks of Different Size and Connectivity Density Using Graph Theory. PLOS One.
2007;17:92-99.
2010;5(10):e13701.
13
15. Tewarie P, van Dellen E, Hillebrand A, Stam CJ. The minimum spanning tree: An unbiased method for brain network analysis. NeuroImage. 2014;
16. Garrison KA, Scheinost D, Finn ES, Shen X, Constable RT. The (in)stability of functional brain network measures across thresholds. NeuroImage.
doi:10.1016/j.neuroimage.2014.10.015.
2015;118:651-661.
17. Meier J, Tewarie P, Van Mieghem P. The Union of Shortest Path Trees of Functional Brain Networks. Brain Connectivity. 2015;5(9):575-581.
18. Jalili M. Functional Brain Networks: Does the Choice of Dependency Estimator and Binarization Method Matter? Sci Rep. 2016;6:29780.
19. Fallani F, Latora V, Chavez M. A Topological Criterion for Filtering Information in Complex Brain Networks. PLOS Comp Biol. 2017;13(1):e1005305.
20. van den Heuvel M, de Lange S, Zalesky A, Seguin C, Yeo T, Schmidt R. Proportional thresholding in resting-state fMRI functional connectivity networks
and consequences for patient-control connectome studies: Issues and recommendations. NeuroImage. 2017;doi: 10.1016/j.neuroimage.2017.02.005.
21. Schwarz AJ, McGonigle J. Negative edges and soft thresholding in complex network analysis of resting state functional connectivity data. Neuroimage.
22. Santarnecchi E, Galli G, Polizzotto NR, Rossi A, Rossi S. Efficiency of weak brain connections support general cognitive functioning. Human Brain
2011;55(3):1132-46.
Mapping. 2014;35(9):4566-4582.
23. Smith K, Azami H, Escudero J, Parra MA, Starr JM. Comparison of network analysis approaches on EEG connectivity in beta during Visual Short-term
Memory binding tasks. IEEE Proc. EMBC 2015. 2015;doi:10.1109/EMBC.2015.7318829.
24. Smith K, Escudero J. The complex hierarchical topology of EEG functional connectivity. J Neurosci Methods. 2017;276:1-12.
25. Stam CJ, Tewarie P, Van Dellen E, van Straaten ECW, Hillebrand A, Van Mieghem P. The Trees and the Forest: Characterization of complex brain
networks with minimum spanning trees. International Journal of Psychophysiology. 2014;92:129-138.
26. Smith K, Azami H, Parra MA, Starr JM, Escudero J. Cluster-Span Threshold: An Unbiased Threshold for Binarising Weighted Complete Networks in
Functional Connectivity Analysis. IEEE Proc. EMBC 2015. 2015;doi:10.1109/EMBC.2015.7318983.
27. Tanizawa T, Paul G, Cohen R, Havlin S, Stanley HE. Optimization of network robustness to waves of targeted and random attacks. Phys Rev E.
2005;71:047101.
28. Dormann CF, Frund J, Bluthgen N, Gruber B. Indices, graphs and null models: analyzing bipartite ecological networks. Open Ecol J. 2009;2:724.
29. Neurophysiological Biomarker Toolbox, available from www.nbtwiki.net.
30. Escudero J, Ab´asolo D, Hornero R, Espino P, L´opez M. Analysis of electroencephalograms in Alzheimer's disease patients with multiscale entropy.
Physiological Measurement. 2016;27(11):1091-1106.
31. Erdos P, R´enyi A. On Random Graphs. Publicationes Mathematicae Debrecen. 1959;6:290-297.
32. Kruskal JB, On the Shortest Spanning Subtree of a Graph and the Traveling Salesman Problem. American Mathematical Society. 1956;7(1):48-50.
33. Rubinov M, Sporns O, Complex network measures of brain connectivity: Uses and interpretations. NeuroImage. 2010;52:1059-1069.
34. Dijkstra EW. A note on two problems in connexion with graphs. Numerische Mathematik. 1959;1(1):269-271.
35. Bullmore E, Sporns O, The economy of brain network organization. Nat Rev Neurosci. 2012;13(5):336349.
36. Newman MEJ, Networks. Oxford University Press, Oxford. 2010.
37. Dauwels J, Viallate F, Musha T, Cichocki A. A comparative study of synchrony measures for the early diagnosis of Alzheimer's disease based on
EEG. NeuroImage. 2010;49(1):668-693.
38. Stam CJ, Nolte G, Daffertshofer A. Phase lag index: Assessment of functional connectivity from multi channel EEG and MEG with diminished bias
from common sources. Human Brain Mapping. 2007;28(11):11781193.
39. Stam CJ, van Straaten ECW. The organization of physiological brain networks. Clinical Neurophysiology. 2012;123(6):10671087.
40. Barry RJ, Clarke AR, Johnstone SJ, Magee CA, Rushby JA. EEG differences between eyes-closed and eyes-open resting conditions. Clinical
Neurophysiology. 2007;118(12):2765-2773.
Res. 1975;12:189198.
41. Folstein MF, Folstein SE, McHugh PR. Mini-mental state: a practical method for grading the cognitive state of patients for the clinician. J Psychiatr
42. Achard S, Salvador R, Whitcher B, Suckling J, Bullmore E. A Resilient, Low-Frequency, Small-World Human Brain Functional Network with Highly
Connected Association Cortical Hubs. The Journal of Neuroscience. 2006;26(1):63-72.
43. Joyce KE, Hayasaka S, Laurienti PJ. The Human Functional Brain Network Demonstrates Structural and Dynamical Resilience to Targeted Attack.
PLOS Comp Biol. 2013;9(1): e1002885.
44. Stam CJ, de Haan W, Daffertshofer A, Jones BF, Manshanden I, van Cappellen van Walsum AM, et al. Graph theoretical analysis of magnetoen-
cephalographic functional connectivity in Alzheimer's disease. Brain. 2009;132(1):213-224.
45. Watts DJ, Strogatz SH, Collective dynamics of small-world networks. Nature. 1998;393:440-442.
46. Lee U, Oh G, Kim S, Noh G, Choi B, Mashour GA. Brain networks maintain a scale-free organization across consciousness, anesthesia, and recovery:
evidence for adaptive reconfiguration. Anesthesiology, 2010;113:1081-1091.
47. Schoen W, Chang JS, Lee U, Bob P, Mashour GA, The temporal organization of functional brain connectivity is abnormal in schizophrenia but does
not correlate with symptomatology. Conscious Cogn., 2011;20:050-1054.
48. Boersma M, Smit DJ, Boomsma DI, De Geus EJ, Delemarre-van de Waal HA, Stam CJ. Growing trees in child brains: graph theoretical analysis of
electroencephalography-derived minimum spanning tree in 5- and 7-year-old children reflects brain maturation, Brain Connect, 2013;3:50-60.
49. Engels MM, Stam CJ, van der Flier WM, Scheltens P, de Waal H, van Straaten EC, Declining functional connectivity and changing hub locations in
Alzheimers disease: an EEG study. BMC neurology, 2015;15(1):145.
|
1611.08913 | 2 | 1611 | 2017-09-27T13:40:49 | A Simple Model of Attentional Blink | [
"q-bio.NC",
"eess.SY"
] | The attentional blink (AB) effect is the reduced ability of subjects to report a second target stimuli (T2) among a rapidly presented series of non-target stimuli, when it appears within a time window of about 200-500 ms after a first target (T1). We present a simple dynamical systems model explaining the AB as resulting from the temporal response dynamics of a stochastic, linear system with threshold, whose output represents the amount of attentional resources allocated to the incoming sensory stimuli. The model postulates that the available attention capacity is limited by activity of the default mode network (DMN), a correlated set of brain regions related to task irrelevant processing which is known to exhibit reduced activation following mental training such as mindfulness meditation. The model provides a parsimonious account relating key findings from the AB, DMN and meditation research literature, and suggests some new testable predictions. | q-bio.NC | q-bio |
A Simple Model of Attentional Blink
Nadav Amir
The Edmond and Lily Safra Center for Brain Sciences
The Hebrew University of Jerusalem
[email protected]
Israel Nelken
Department of Neurobiology, Institute for Life Sciences &
The Edmond and Lily Safra Center for Brain Sciences
The Hebrew University of Jerusalem
[email protected]
Naftali Tishby
The Rachel and Selim Benin School of Computer Science and Engineering &
The Edmond and Lily Safra Center for Brain Sciences
The Hebrew University of Jerusalem
[email protected]
September 28, 2017
Abstract
The attentional blink (AB) effect is the reduced ability of subjects to report a second target stimuli (T2) among
a rapidly presented series of non-target stimuli, when it appears within a time window of about 200-500 ms
after a first target (T1). We present a simple dynamical systems model explaining the AB as resulting from
the temporal response dynamics of a stochastic, linear system with threshold, whose output represents the
amount of attentional resources allocated to the incoming sensory stimuli. The model postulates that the
available attention capacity is limited by activity of the default mode network (DMN), a correlated set of
brain regions related to task irrelevant processing which is known to exhibit reduced activation following
mental training such as mindfulness meditation. The model provides a parsimonious account relating key
findings from the AB, DMN and meditation research literature, and suggests some new testable predictions.
I.
Introduction
Questions regarding the characteristics, and
limitations, of attention allocation over time
have attracted considerable interest over the
last decades (see [1] for an overview). More
recently, it has been shown that intensive men-
tal training, in the form of mindfulness med-
itation, can modulate the control of attention
allocation over time [2, 3, 4], yet the computa-
tional and neural mechanisms underlying this
process remain unclear. A possible clue lies
in the recent deluge of studies concerning the
so called default mode network (DMN) (see
[5, 6] for reviews): an anatomically defined, in-
terconnected system of cortical regions which
are preferentially activated when individuals
are involved in self referential processing and
mind wandering, rather than paying attention
to the external environment [7, 8, 9]. Various
meditation practices have been shown to be
associated with reduced activity levels of the
DMN [10, 11, 12, 13], as well as improved sus-
tained [14] and selective [3] attention capabili-
1
ties.
Here we describe a simple dynamical sys-
tems model suggesting that changes in DMN
activity levels can modulate attentional capac-
ity. The proposed mechanism is quite gen-
eral in nature but we demonstrate it using the
well-known attentional blink (AB) paradigm
[15, 16], which is the reduced ability ("blink-
ing") of subjects asked to report a second tar-
get stimuli in a rapid serial visual presenta-
tion (RSVP) task, when it appears within a
time window of about 200-500 ms after the
first one. This AB effect has played a central
role in studying the limits of humans' ability
to allocate attention over time (for reviews see
[17, 18]) and more recently has also been used
to study the effects of mental training (in the
form of mindfulness meditation) on the tempo-
ral distribution of attentional resources [2, 3, 4].
While several theoretical accounts have been
suggested over the last decade for explaining
the AB (see discussion below), the model pro-
posed here has the advantage of being sim-
ple and parsimonious in parameters, while ac-
counting for many empirical findings relating
the AB, DMN and mental training research
literature. Additionally, it provides testable
hypotheses relating changes in DMN activity
levels with performance in attention demand-
ing tasks and associated ERP amplitudes.
II. Methods
i. Basic description of the model
Formally, we model the "attentional channel"
as a linear time-invariant (LTI) system [19] with
additive Gaussian noise and an upper response
threshold (Eq. 1). The system's output, or re-
sponse, y(t), corresponds to the amount of at-
tentional resources currently used by the brain.
When this response reaches a predetermined
blinking threshold, y(t) = yB, the attentional
resources are exhausted and the system can-
not respond to incoming stimuli for a short
"refractory" period, resulting in decreased iden-
tification of the subsequent target stimuli - the
AB effect.
2
The system's response consists of determin-
istic and stochastic components. The former
is given by a convolution of the input stim-
uli's cognitive representation uc(t) (see next sub-
section), with the system's impulse response
function h(t), which describes the time course
of the system's response to a pulse shaped
input. The stochastic part of the response is
given by the additive Gaussian random vari-
able nDMN(t) „ N (µDMN, σ2
DMN), with µDMN
representing the DMN activity baseline (mean)
and σDMN its fluctuations (standard deviation).
Conceptually, DMN activity is represented by
the model as task irrelevant noise loading the
attentional system and thus modulating blink-
ing propensity.
To summarize, we present a model describ-
ing the temporal capture of attention using a
dynamical system described by Eq. 1, which
is specified by the cognitive representation of
the input stimuli uc(t), the blinking threshold
yB, the impulse response function h(t) and the
DMN noise nDMN.
#
y(t) = min
uc(t) h(t) + nDMN(t)
yB
(1)
ii. Cognitive representation of input
stimuli
In line with other two-stage accounts of the
AB (e.g.
[20, 16] and see discussion below),
the model assumes that sensory stimuli un-
dergo an initial cognitive representation stage
before they can be further processed, or consol-
idated into working memory. The input sen-
sory stimuli stream is represented by a discrete
(10Hz) signal of impulses us(t), with ampli-
tudes corresponding to stimulus saliency. Tar-
get stimuli are represented as unit impulses
and non-targets as zeros, i.e. us(t) = 1 if
one of the target stimuli (T1 or T2) appears
at time t and us(t) = 0 otherwise. This for-
mulation can be easily adjusted to account for
various target/distractor saliency relationships
by modulating the impulse amplitudes of the
different stimuli (see discussion section below).
Concretely, the representation stage is imple-
mented by resampling the sensory stimulus
signal us(t) at a higher sampling rate (1KHz),
and clipping it at a threshold value uc(t) = 1
(not to be confused with the blinking threshold
yB in Eq. 1). The resulting cognitive represen-
tation signal uc(t), is then used as input to the
attentional system described by Eq. 1. Con-
ceptually, resampling represents the temporal
resolution of the attentional system and clip-
ping corresponds to sub-linear combination of
temporally close sensory stimuli. Importantly,
the sub-linearity is most pronounced when the
two target stimuli appear successively, in which
case their combined cognitive representation
signal exhibits maximal overlap and clipping.
As discussed below, this sub-linear summation
of temporally proximal target stimuli can ex-
plain various effects related to the somewhat
counterintuitive finding that blinking is signif-
icantly attenuated when T2 appears directly
after T1 - the so called "lag-1 sparing" effect
[21].
iii. DMN activity as attentional noise
A central feature of the model is its interpre-
tation of DMN activity as stochastic noise in
the attentional system. This is formally de-
scribed by the additive noise term nDMN(t) in
Eq. 1 which is a Gaussian random variable,
nDMN(t) „ N (µDMN, σ2
DMN), with mean and
variance corresponding to the DMN's baseline
activity and moment-by-moment fluctuation
levels respectively. Conceptually, the mean
DMN activity level represents the underlying
mental state (e.g. mind wandering vs task en-
gagement), whereas the variance represents
momentary attentional fluctuations around the
baseline level. A lower DMN activity level re-
sults in more attentional resources available for
task performance and less probability of cross-
ing the blinking threshold. This mechanism en-
ables the model to relate DMN activity and AB
task performance in a quantitative way. The
recognition probability of T2 can be directly
computed from the blinking gap, defined as the
difference between yB (the blinking threshold),
and the maximal value of y(t) during the sys-
tem's excitation by the stimulus inputs.
iv. The impulse response function of
the attentional system
The impulse response function h(t), deter-
mines the temporal profile of attentional re-
source allocation.
It is modeled here using
the Gamma (or Erlang) distribution function
(Eq. 2):
#
h(t) =
st(n´1)e´t/τ
0
t ą 0
t ď 0
(2)
Here n and τ are the shape and scale parame-
ters which fix the half duration and maximum
response time of h(t) (Fig. 2) and the scaling
factor s controls the system's gain. While the
precise form of h(t) is not of major significance,
the Gamma distribution function is known to
provide a reasonably good fit for attention re-
lated responses such as pupillary dilation (a
correlate of attentional effort) [22, 23] , tempo-
ral sensitivity in the visual system [24] and
even Blood-oxygen-level dependent (BOLD)
signal hemodynamic responses [25].
v. The P3b brain potential
The P3b is a positive ERP, peaking at around
300ms, associated with updating of working
memory [26, 27, 28] and attentional resources
allocation [29, 30]. To interpret the P3b ampli-
tude in terms of the model's parameters, we
first define the Resource Allocation Index (RAI),
as the ratio between attentional resources allo-
cated to the stimuli under consideration and
the total resources available (Eq. 3).
maxt y(t) ´ µDMN
RAI =
yB ´ µDMN
(3)
The maximum is taken over the time period
during which the system is responding to the
stimuli under consideration. Note that the RAI
is a dimensionless number taking values be-
tween 0, corresponding to no excitation of the
attentional system and 1, representing maxi-
mal excitation or exhaustion of attentional re-
sources. Since P3b amplitude ranges typically
3
differ between subjects, even when performing
the same task, we define the model correlate of
the P3b amplitude as the RAI divided by the
subject specific DMN activity standard devia-
tion σDMN, which serves as a natural signal to
noise scaling factor in this context (Eq. 4).
P3b =
RAI
σDMN
(4)
How are changes in DMN activity levels re-
flected in the P3b amplitude defined by equa-
tion 4? For non-blinking trials, i.e. when
RAI ă 1, the RAI and thus the P3b amplitude
increases as µDMN is decreased and σDMN re-
mains fixed. However, when both µDMN and
σDMN are decreased, the expected P3b ampli-
tude may increase or decrease depending on
their specific values, since reducing σDMN re-
duces both the numerator in equation 3 and
the denominator in 4. For blinking trials, i.e.
when RAI = 1, the situation is simpler since
in this case P3b = σ´1
DMN regardless of µDMN's
value.
III. Results
We first tested whether the model can repro-
duce the basic AB effect, namely a reduction in
detection accuracy of T2 within a time window
of approximately 200-500 ms after T1 presen-
tation. Due to the resampling of stimuli at
the cognitive representation stage, and the fi-
nite rise and fall time of the system's impulse
response function (Fig. 2), two targets which
appear close to each other will have overlap-
ping responses. This overlap can cause the
system's response to the combined signal to
reach the blinking threshold, even though the
individual response to each one of the targets
does not do so (Fig. 1, top). As the time in-
terval between T1 and T2 increases, their over-
lap decreases and the system's response to the
combined signal no longer reaches the blinking
threshold (Fig. 1, bottom). Importantly, when
T2 appears immediately after T1, the combined
response may not cross the blinking threshold
since in this case a significant portion of the
overlap between targets is clipped at the cogni-
4
Figure 1: Attentional Blink model Top: When
T2 (in green) appears close (but not directly, see
Fig. 3) after T1 (blue), the blinking threshold
(red line) is reached and blinking occurs. Bot-
tom: When the time interval between T1 and
T2 is longer, both targets can be processed with-
out causing the system to reach the blinking
threshold.
020406080100120up(t)01T1T2020040060080010001200uc(t)-0.500.51time(msec)020040060080010001200y(t)0204060020406080100120up(t)01T1T2020040060080010001200uc(t)-0.500.51time(msec)020040060080010001200y(t)0204060 Figure 2: Impulse response function
The temporal response of the attentional sys-
tem to an impulse stimuli follows the Gamma
distribution function (Eq. 2). The values of
the shape and scale parameters used here are
n = 1 and τ = 0.05 respectively.
tive representation stage (Fig. 3), resulting in a
sub-linearly combined signal.
Next, we tested whether the effects of men-
tal training on AB performance and the corre-
sponding P3b ERP can be reproduced by the
model. In [2] it was shown that mental training,
in the form of a 3 month mindfulness medi-
tation retreat, resulted in significantly lower
blinking probability and larger reduction in
T1 elicited P3b amplitudes for non-blinking vs
blinking trials in expert meditators compared
to a group of matched novice controls who did
not attend the retreat. Furthermore, the medi-
tators' reduction in T1 evoked P3b amplitudes
was significantly correlated with improvement
in T2 detection accuracy, suggesting that the
ability to detect T2 depends on efficient deploy-
ment of attentional resources to T1 [2]. Using
the theoretical P3b amplitude defined in Eq. 4,
the model was able to reproduce the three-way
interaction (meditators vs. novices, before vs.
after retreat, blinking vs. non-blinking trials),
Fig. 4, (cf. figure 3 in [2]), and the relationship
between T2 detection accuracy and reduction
in T1 elicited P3b amplitude. Fig. 5 bottom
right, (cf. figure 4 in [2]).
Figure 3: Lag-1 sparing Top: When T2 (green)
appears immediately after T1 (blue), their
larger overlap at the cognitive representation
stage results in more clipping which in turn re-
duces the maximal response of the system and
decreases blinking probability. Bottom: Proba-
bility of T2 detection as a function of the time
lag between T1 and T2. This u-shaped profile
is typical in AB experiments ([15]).
5
Time050100150200250300Amplitude00.10.20.30.40.50.60.70.80.91020406080100120up(t)01T1T2020040060080010001200uc(t)-0.500.51time(msec)020040060080010001200y(t)0204060T2-T1 lag12345678T2 detection probability0.20.30.40.50.60.70.80.91 The model reproduced these results only for
certain values of DMN activity levels before
and after training. We thus wanted to iden-
tify all pairs of (µDMN, σDMN) points, repre-
senting DMN activity at times 1 and 2, which
reproduce the main findings reported in [2],
namely a reduction of about 25% in T1 elicited
P3b amplitude and an increase of about 20%
(from 60% to 80%) in T2 detection accuracy. In
addition to reproducing these empirical find-
ings, we required that these points reproduce
the lag-1 sparing effect, which we defined as
the condition of crossing the blinking thresh-
old for T2-T1 lags of 2 but not for lags 1 or 5
and higher. We plotted the probability of lag-1
sparing occurrence, as well as the model P3b
amplitude (Eq. 4) and T2 detection probability,
both at a T2-T1 lag of 4 (Fig. 5, top left, bottom
left and top right respectively). We then tested
which pairs of (µDMN, σDMN) points, repre-
senting DMN activity at time 1 and 2, yield
a substantial probability (ą 0.2) of lag-1 spar-
ing occurrence at both times and reproduce
the main empirical findings of [2] mentioned
above (Fig. 5, blue and red crosses, top left
and right and bottom left). For these pairs
of points, we plotted the change in T2 detec-
tion probability a function of the reduction in
P3b amplitude (Fig. 5, bottom right). This
yielded a set of DMN activity parameter pairs
which reproduce the empirical correlation be-
tween improvement in T2 detection accuracy
and reduction in T1 elicited P3b amplitude as
reported in figure 4 of [2]. Thus, the model
provides a quantifiable relationship between
DMN activity levels, AB task performance and
T1 evoked P3b amplitudes, which can presum-
ably be tested empirically.
IV. Discussion
i. Meditation, P3b and the default
mode network
Differences in attentional processing between
expert meditators and novices have been re-
ported in several studies over the last years
(for an overview see [3]). Of particular inter-
Figure 4: Effects of mental training on T1
elicited P3b Top: Modeled T1 evoked P3b
amplitude as a function of T2 accuracy: no-
blink vs. blink, session: Time 1 (before the
meditation retreat) vs. Time 2 (after the re-
treat), and group (practitioners vs. novices).
Meditation practitioners show a greater reduc-
tion in T1 evoked P3b amplitude compared
to novices in no-blink vs blink trials at time
2 vs time 1. Bottom: Selective reduction in
T1 evoked P3b amplitude in no-blink trials in
the practitioner group. DMN parameter levels
(µDMN, σDMN), for practitioners: (3.65, 28) at
time 1 and (3.5, 25) at time 2. For novices:
(3.7, 30) at both times. These results, and
choice of colors, follow figure 3 in [2].The larger
T1 evoked P3b amplitude for practitioners at
time 1, for blinking vs. non-blinking trials and
at time 2 vs. time 1, for blinking trials are novel
predictions of the model.
6
BlinkNo Blink0.20.250.3NovicesTime1BlinkNo Blink0.20.250.3PractitionersBlinkNo Blink0.20.250.3Time2BlinkNo Blink0.20.250.3Time 1Time 20.20.250.3NovicesBlinkTime 1Time 20.20.250.3PractitionersTime 1Time 20.20.250.3NoblinkTime 1Time 20.20.250.3 est here is a study dealing with the effect of
meditation on the AB and the related P3b po-
tentials evoked during the task performance
[2]. The P3b wave is an event related potential
(ERP) component linked with attention alloca-
tion and memory encoding of targeted stim-
uli [29, 26]. The results reported in [2] show
that experienced meditators exhibit a smaller
attentional blink effect (higher T2 recognition
accuracy), and larger amplitudes of T1 evoked
P3b compared to a matched control group of
meditation novices. In addition, after attend-
ing a 3 month intensive meditation retreat, the
meditators showed a significant reduction of T1
evoked P3b amplitudes but only during non-
blinking trials. The novice control group, who
did not attend the meditation retreat, showed
a significantly smaller reduction in both at-
tentional blink size and T1 evoked P3b ampli-
tude. The magnitude of reduction in T1 evoked
P3b amplitudes during no-blink trials was cor-
related with the improvement in T2 recogni-
tion accuracy for both meditators and novices.
These results suggest that the ability to identify
T2 depends on efficient attentional resource al-
location to T1 and that mental training, such as
mindfulness meditation, can improve this re-
source allocation process. The model suggests
a concrete mechanism, namely the reduction in
DMN noise levels, through which this process
can take place. This also provides a quanti-
tative interpretation of the widespread claim
that mindfulness meditation reduces ongoing
mental noise in the brain.
Interestingly, the T1 evoked P3b amplitudes
were larger for the experienced meditators com-
pared to the novices in all experimental con-
ditions (blinking and non-blinking trials, be-
fore and after the retreat), with the possible
exception of non-blinking trials after the re-
treat, where they were of comparable magni-
tude (figure 3. in [2]). The model suggests that
these differences may be due to smaller fluctu-
ations (noise variance) in the DMN activity of
experienced meditators compared to novices.
Over the last few years, several studies have
shown that different types of meditation are
linked with reduced DMN activity levels com-
7
Figure 5: Model behavior in DMN parameter
space Top left: lag-1 sparing probability as a
function of DMN activity parameters (mean
and variance). The color indicates the prob-
ability of crossing the blinking threshold at
lag 2 but not at lags 1 and 5. Top right: T2
detection probability for a T2-T1 lag of 4. Bot-
tom left: The model P3b amplitude defined in
Eq. 4, for a T2-T1 lag of 4. Blue and red crosses
correspond to possible (µDMN, σDMN) values
at time 1 and time 2 respectively, for which
the model reproduces lag-1 sparing as well as
the main findings of [2], namely an increase
of T2 detection probability from about 60% to
about 80% (ibid. figure 2) and a reduction of
about 25% in P3b amplitude (ibid. figure 3)
and. Bottom right: Change in T2 detection ac-
curacy as a function of the change in P3b for
(µDMN, σDMN) pairs reproducing lag-1 sparing
and the effects of meditation (cf. figure 4. in
[2]).
µDMN253035σDMN33.13.23.33.43.53.63.73.83.940.050.10.150.20.250.30.350.40.45µDMN253035σDMN33.13.23.33.43.53.63.73.83.940.10.20.30.40.50.60.70.80.91µDMN253035σDMN33.13.23.33.43.53.63.73.83.940.250.260.270.280.290.30.310.320.33P3b amplitude, Time2-Time1-0.095-0.09-0.085-0.08-0.075-0.07-0.065-0.06%T2 accuracy, Time2-Time1-0.2-0.100.10.20.30.40.5 pared to rest [11, 14] and other active tasks [10].
It has also been shown recently that experi-
enced mindfulness meditators exhibit reduced
resting state DMN activity and fluctuations
amplitudes during a visual memory task [13]
compared to novice controls. By proposing that
DMN activity reduction is the neural mecha-
nism by which meditation increases attentional
capacity, our model relates these studies with
the finding that meditation affects performance
and brain potentials in the AB task [2, 3, 4].
In a 2004 study measuring BOLD activation
levels during AB task performance [32], the
only brain area reported to exhibit decreased
activation for non-blinking vs. blinking trials
was the right temporoparietal junction (TPJ), a
DMN associated region known to deactivate
during attention demanding tasks [33]. Other,
non DMN related regions, showed the oppo-
site effect, namely increased activation during
non-blinking compared to blinking trials [32].
This finding is also in line with the model's
hypothesis regarding the relationship between
DMN activity and blinking probability.
that blinking is due to top-down inhibition of
attention while T1 is being processed.
The model proposed here can be broadly
categorized as symbolic and capacity limited.
However, it is based on first-principles (lin-
ear dynamical-systems) and is essentially de-
scribed by a single equation (Eq. 1) assum-
ing only a few free parameters. The model
also incorporates a top-down control element
by modulating the available attentional capac-
ity through changes in DMN activity. As dis-
cussed below, this element enables the model
to explain certain findings which challenge the
notion of a purely capacity-limited account of
the AB, such as the"spreading" of lag-1 sparing
(see below).
While simple, the model can explain many
findings from the AB literature and, in par-
ticular, reproduce the results reported in [2]
relating mental training with improved AB per-
formance and reduced P3b amplitudes. We
discuss below the central findings relating to
the AB and describe how they are accounted
for by the model.
ii. Comparison with existing models
ii.1 Reversal of T1 and T2
Several mathematical models for the AB effect
have been suggested over the past decade or
so (for an overview see the section "Formal
Theories" in [17]). Formally, these models can
be divided into two classes: connectionist and
symbolic [34]. Connectionist, also known as
neural-network models (e.g [35, 36, 37, 38, 39]),
usually rely on tuning a large number of pa-
rameters without providing much insight re-
garding the underlying mechanisms. On the
other hand, symbolic, or computationalist mod-
els (e.g.
[40, 41]), are often described using
complex box-and-arrow type rules which seem
rigid and ad-hoc. Another way of categoriz-
ing AB models is by the underlying cause for
blinking. The two pravelant mechanisms are ca-
pacity limitations and attentional control ([18]).
Models emphasizing capacitiy limits attribute
blinking to some resource bottleneck at the at-
tentional or working memory systems, while
those emphasizing a control mechanism posit
It has been reported that when T2 appears im-
mediately after T1 (lag 1 trials), identification
of T2 is often superior to T1 and report order is
often reversed [42, 20]. The model is consistent
with these effects since it posits that during
the cognitive representation stage, T1 may be
partially occluded by T2 and both are merged
together into a single "cognitive" trace. This
process of occlusion and merging may explain
degredation of T1 detection probability and
T1,T2 order confusion.
ii.2 Spreading of the sparing
This refers to a set of findings showing that
AB can be attenuated or even eliminated as
long as a target is not followed by a distractor.
Thus, when subjects were presented with RSVP
streams containing three consecutive targets
(T1,T2,T3), there was no deficiency in report-
ing T3 even though it appeared at the temporal
8
position in which blinking is typically maxi-
mal (lag 2) [43]. This phenomenon was called
"spreading of lag-1 sparing" [44] since the atten-
uation of blinking at lag-1 spreads to additional
(target) stimuli. A related finding is the atten-
uation of AB when subjects were required to
give a whole report of a six letter sequence, i.e.
when all stimuli were considered targets, com-
pared to the standard case in which only two
letters had to be reported [45]. Such findings
pose a challenge for limited-capacity accounts
of the AB and seem to indicate the workings of
a top-down attentional control mechanism [18].
However, the model presented here can pro-
vide an interesting explanation of these find-
ings by assuming that a consecutive sequence
of targets temporarily reduces DMN fluctua-
tions, perhaps by focusing the subject's atten-
tion on the task, thus reducing blinking prob-
ability until appearance of the next distractor.
This hypothesis explains "spreading" effects in
terms of an interaction between capacity limita-
tion (the blinking threshold), and a top-down
control mechanism (temporary reducing DMN
noise). The partial occlusion of earlier by later
targets at the cognitive representation stage
may also explain the poorer identification of
T1 typically observed when three targets are
presented sequentially [43, 46].
ii.3 No T2 evoked P3b during blinking tri-
als
In [47] it was shown that missed T2 targets
do not evoke a P3b ERP. In the model, these
trials correspond to cases in which the atten-
tional resources were depleted by the T1, or the
combined representation of T1 and T2, signal.
This results in a large P3b amplitude elicited
by the T1, or combined T1,T2, signal and no
attentional resource capture, and thus no P3b
signal, evoked in response to T2 itself.
ii.4 Distractor and target saliency
The experiments reported in [48, 49] showed
that salient distractors which match features
with the target set (having the same color),
can also trigger an AB for a subsequent target.
The model can account for these findings by
representing salient distractors as pulses with
smaller amplitudes (compared to targets) at the
sensory signal level. These distractor pulses
will also contribute towards triggering an AB
but to a lesser extent than an additional tar-
get pulse would. Using a similar amplitude
modulation mechanism, the model can explain
why targets which capture the attention more
powerfully (e.g. by switching their color com-
pared to pre-target distractors) have a higher
propensity for causing AB [50].
ii.5 Effect of inter-target blanks
There are mixed results in the AB literature
regarding the effects of inserting a blank stim-
ulus between T1 and T2. In some cases blink-
ing was reported although at attenuated lev-
els compared to the case of a (non-blank) dis-
tractor (e.g. [15, 51, 52]), while other studies
showed no blinking when inserting inter-target
blanks (e.g. [20, 53, 54]). Such blanks can be
represented in the model by small inter-target
impulses whereas (non-blank) distractors can
be represented by somewhat larger, yet still
smaller than target, pulses. While this particu-
lar analysis is beyond the scope of the current
work, we wish to point out that this and similar
questions can be addressed by the model using
slight modifications of the basic framework.
ii.6 Task-irrelevant mental activity
A somewhat counterintuitive series of findings
reported in [55, 56] suggest that certain seem-
ingly task irrelevant stimuli, or mental activi-
ties (background music or visual motion, per-
forming a concurrent memory task etc.) can
significantly attenuate the AB. These interven-
tions presumably load the attentional channel,
and thus seem at odds with a capacity lim-
ited account of the AB. However, the model
provides an interesting explanation for these
findings by hypothesizing that such activities
may reduce DMN activity levels, and this in-
crease attentional capacity, perhaps by focusing
subjects on their immediate environment thus
reducing mind wandering.
9
V. Conclusion
This paper proposed a model of attentional
blink which is simple and parsimonious while
providing explanatory and predictive power.
The model's main features are a dynamical sys-
tem's account of attentional capacity with a
top-down control mechanism in the form of
DMN activity, represented by stochastic noise
loading the system's capacity. The model gen-
erates quantifiable predictions relating DMN
activity levels, AB task performance and the
P3b ERP, which can be tested in future experi-
ments.
VI. Acknowledgments
The authors thank the ICRI-CI, the Israel Sci-
ence Foundation, the Gatsby Charitable Foun-
dation and the European Research Council. We
wish to thank Leon Deuoell and Tamar Regev
for helpful discussions.
References
[1] K. Shapiro and K. (Ed), The Limits of
Attention: Temporal Constraints in Human
Information Processing. Oxford University
Press, oct 2001. [Online]. Available: http:
//www.oxfordscholarship.com/view/
10.1093/acprof:oso/9780198505150.001.
0001/acprof-9780198505150
[2] H. A. Slagter, A. Lutz, L. L. Greischar,
J. M.
A. D. Francis, S. Nieuwenhuis,
Davis, and R.
J. Davidson, "Mental
Training Affects Distribution of Limited
Brain Resources," PLoS Biology, vol. 5,
no. 6, p. e138, may 2007.
[Online].
Available: http://dx.plos.org/10.1371/
journal.pbio.0050138
[3] A. Lutz, H. A. Slagter, N. B. Rawl-
ings, A. D. Francis, L. L. Greischar,
and R. J. Davidson, "Mental Training
Enhances Attentional Stability: Neu-
ral and Behavioral Evidence," Journal
of Neuroscience, vol. 29, no. 42, pp.
10
13 418–13 427, oct 2009. [Online]. Avail-
able: http://www.jneurosci.org/cgi/doi/
10.1523/JNEUROSCI.1614-09.2009
[4] M. K. van Vugt and H. A. Slagter, "Con-
trol over experience? Magnitude of the
attentional blink depends on meditative
state," Consciousness and Cognition, vol. 23,
pp. 32–39, 2014.
[5] M. E. Raichle, A. M. MacLeod, A. Z.
Snyder, W. J. Powers, D. A. Gusnard,
and G. L. Shulman, "A default mode of
brain function." Proceedings of the National
Academy of Sciences of the United States of
America, vol. 98, no. 2, pp. 676–82, jan 2001.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/11209064http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC14647
[6] M. E. Raichle, "The Brain's Default
Mode Network," Annual Review of
Neuroscience,
pp.
433–447,
[Online]. Available:
http://www.annualreviews.org/doi/10.
1146/annurev-neuro-071013-014030
2015.
no.
vol.
38,
1,
and
[7] M. F. Mason, M. I. Norton, J. D. Van Horn,
D. M. Wegner, S. T. Grafton, and C. N.
Macrae, "Wandering minds: the default
network
stimulus-independent
thought." Science (New York, N.Y.), vol.
315, no. 5810, pp. 393–5,
jan 2007.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/17234951http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC1821121
[8] K. Christoff, A. M. Gordon, J. Smallwood,
R. Smith, and J. W. Schooler, "Experience
sampling during fMRI reveals default
network and executive system contribu-
tions to mind wandering," Proceedings
of the National Academy of Sciences, vol.
106, no. 21, pp. 8719–8724, may 2009.
[Online]. Available: http://www.pnas.
org/cgi/doi/10.1073/pnas.0900234106
[9] R. L. Buckner, J. R. Andrews-Hanna, and
D. L. Schachter, "The Brain's Default
and
Network: Anatomy, Function,
Relevance to Disease," Annals of
the
New York Academy of Sciences, vol. 1124,
no. 1, pp. 1–38, mar 2008.
[Online].
Available: http://doi.wiley.com/10.1196/
annals.1440.011
[15] J. Raymond, K. Shapiro, and K. Arnell,
"Temporary suppression of visual pro-
cessing in an RSVP task: An attentional
Journal of experimental, 1992.
blink?"
[Online]. Available: http://psycnet.apa.
org/journals/xhp/18/3/849/
[10] K. A. Garrison, T. A. Zeffiro, D. Scheinost,
R. Todd Constable, and J. A. Brewer,
"Meditation leads to reduced default mode
network activity beyond an active task,"
2015.
[11] J. A. Brewer, P. D. Worhunsky, J. R. Gray,
Y.-Y. Tang, J. Weber, and H. Kober, "Med-
itation experience is associated with dif-
ferences in default mode network activity
and connectivity."
[12] N. A. S. Farb, Z. V. Segal, H. Mayberg,
J. Bean, D. McKeon, Z. Fatima, and
A. K. Anderson, "Attending to the
present: mindfulness meditation reveals
distinct neural modes of self-reference."
Social cognitive and affective neuroscience,
vol. 2, no. 4, pp. 313–22, dec 2007.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/18985137http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC2566754
[13] A. Berkovich-Ohana, M. Harel, A. Ha-
hamy, A. Arieli, and R. Malach, "Alter-
ations in task-induced activity and resting-
state fluctuations in visual and DMN
areas revealed in long-term meditators,"
NeuroImage, vol. 135, pp. 125–134, 2016.
[14] G. Pagnoni,
"Dynamical properties
of BOLD activity from the ventral
posteromedial cortex associated with
meditation and attentional skills." The
Journal of neuroscience
the official
journal of
the Society for Neuroscience,
vol. 32, no. 15, pp. 5242–9, apr 2012.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/22496570http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC3362741
:
[16] D. E. Broadbent and M. H. P. Broad-
"From detection to identifica-
bent,
tion:
Response to multiple targets
in rapid serial visual presentation,"
Perception & Psychophysics,
vol. 42,
no. 2, pp. 105–113, mar 1987.
[On-
line]. Available: http://www.springerlink.
com/index/10.3758/BF03210498
[17] P. Dux, "The attentional blink: A review
of data and theory," vol. 71, no. 8, pp.
1683–1700, 2009.
[18] S. Martens and B. Wyble, "The attentional
blink: past, present, and future of a
in perceptual awareness."
blind spot
Neuroscience and biobehavioral
reviews,
vol. 34, no. 6, pp. 947–57, may 2010.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/20025902http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC2848898
[19] C.-T. Chen and C.-T. Chen, Linear system
theory and design. Holt, Rinehart, and
Winston, 1984.
[20] M. M. Chun and M. C. Potter, "A Two-
Stage Model for Multiple Target Detection
in Rapid Serial Visual Presentation," Jour-
nal of Experimental Psychology: Human Per-
ception and Performance, vol. 21, no. 1, pp.
109–127, 1995.
[21] M. C. Potter, M. M. Chun, B. S. Banks,
and M. Muckenhoupt, "Two Attentional
Deficits in Serial Target Search: The Visual
Attentional Blink and an Amodal Task-
Switch Deficit," Journal of Experimental Psy-
chology: Learning, Memory, and Cognition,
vol. 24, no. 4, pp. 979–992, 1998.
[22] B. Hoeks and W. J. M. Levelt, "Pupillary
dilation as a measure of attention: A quan-
titative system analysis," Behavior Research
11
Methods, Instruments, & Computers, vol. 25,
no. l, pp. 16–26, 1993.
[23] S. M. Wierda, H. van Rijn, N. A.
Taatgen, and S. Martens, "Pupil dilation
deconvolution reveals the dynamics of
attention at high temporal resolution."
Proceedings of the National Academy of
Sciences of the United States of America,
vol. 109, no. 22, pp. 8456–60, may 2012.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/22586101http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC3365158
[24] A. Watson, "Temporal sensitivity," in
Handbook of perception and human per-
formance. Vol 1: Sensory processes and
perception., B. et Al., Ed. Wiley New York,
1986, vol. 18, no. 4, p. 340. [Online]. Avail-
able:
http://www.sciencedirect.com/
science/article/pii/000368708790144X
[25] G. M. Boynton, S. A. Engel,
the
fMRI
and
D. J. Heeger, "Linear systems analysis
signal." NeuroImage,
of
vol. 62, no. 2, pp. 975–84, aug 2012.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/22289807http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC3359416
[26] J. Polich, "Updating P300: an integrative
theory of P3a and P3b." Clinical neurophys-
iology : official journal of the International
Federation of Clinical Neurophysiology,
vol. 118, no. 10, pp. 2128–48, oct 2007.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/17573239http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC2715154
[27] J. Polich and J. R. Criado, "Neuropsychol-
ogy and neuropharmacology of P3a and
P3b," International Journal of Psychophysiol-
ogy, vol. 60, no. 2, pp. 172–185, 2006.
[28] E. Donchin and M. Coles, "Is the P300
component a manifestation of context up-
dating?" Behavioral and Brain Sciences,
1988.
12
[29] C. Wickens, A. Kramer, L. Vanasse, and
E. Donchin, "Performance of concurrent
tasks: a psychophysiological analysis of
the reciprocity of information-processing
resources." Science (New York, N.Y.), vol.
221, no. 4615, pp. 1080–2, sep 1983.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/6879207
[30] J. Johnson, R., "The amplitude of the P300
component of the event-related potential:
Review and synthesis," in Advances in Psy-
chophysiology, 3, 1988, pp. 69–137.
[31] H. A. Slagter, A. Lutz, L. L. Greischar,
S. Nieuwenhuis, and R. J. Davidson,
"Theta phase synchrony and conscious
target perception:
impact of intensive
training." Journal of cognitive
mental
neuroscience,
pp.
1536–49,
[Online]. Available:
http://www.ncbi.nlm.nih.gov/pubmed/
18823234$\delimiter"047A648$nhttp:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC2698032
2009.
vol.
21,
no.
8,
[32] R. Marois, D.-J. Yi, and M. M. Chun, "The
Neural Fate of Consciously Perceived and
Missed Events in the Attentional Blink,"
Neuron, vol. 41, pp. 465–472, 2004.
functional
[33] G. L. Shulman, S. V. Astafiev, M. P.
McAvoy, G. D'Avossa, and M. Corbetta,
"Right TPJ deactivation during visual
search:
significance and
support for a filter hypothesis." Cerebral
cortex (New York, N.Y. : 1991), vol. 17,
no. 11, pp. 2625–33, nov 2007. [Online].
Available: http://www.ncbi.nlm.nih.gov/
pubmed/17264254
[34] M. L. Minsky, "Logical Versus Analogical
or Symbolic Versus Connectionist or Neat
Versus Scruffy," AI Magazine, vol. 12, no. 2,
p. 34, 1991.
[35] B. Wyble, H. Bowman, and M. Nieuwen-
stein, "The Attentional Blink Provides
Sparing at
Episodic Distinctiveness:
a Cost." [Online]. Available:
http:
//dx.doi.org/10.1037/a0013902.supp
[36] S. Nieuwenhuis, M. S. Gilzenrat, B. D.
Holmes, and J. D. Cohen, "The Role of
the Locus Coeruleus in Mediating the At-
tentional Blink: A Neurocomputational
Theory."
[37] S. Chartier, D. Cousineau, and D. Char-
bonneau, "A Connectionist Model of the
Attentional Blink Effect During a Rapid
Serial Visual Presentation Task," 2004.
[38] N. Fragopanagos, S. Kockelkoren, and J. G.
Taylor, "A neurodynamic model of the
attentional blink," 2005.
[39] C. N. L. Olivers and M. Meeter, "A Boost
and Bounce Theory of Temporal Atten-
tion," 2008.
[40] S.-I. Shih, "The attention cascade model
and attentional blink," 2007.
[41] N. A. Taatgen, I. Juvina, M. Schipper, J. P.
Borst, and S. Martens, "Too much control
can hurt: A threaded cognition model of
the attentional blink," Cognitive Psychology,
vol. 59, no. 1, pp. 1–29, 2009.
[42] B. Hommel and E. G. Akyürek, "Lag-1
sparing in the attentional blink: benefits
and costs of
integrating two events
into a single episode." The Quarterly
journal
experimental psychology. A,
Human experimental psychology, vol. 58,
no. 8, pp. 1415–33, nov 2005. [Online].
Available: http://www.ncbi.nlm.nih.gov/
pubmed/16365947
of
[43] V. DiLollo, J.-I. Kawahara, S. M. Shahab,
G. Ae, and J. T. Enns, "The attentional
blink: Resource depletion or temporary
loss of control?"
[44] C. N. L. Olivers, S. Van Der Stigchel,
and J. Hulleman, "Spreading the sparing:
against a limited-capacity account of the
attentional blink."
[45] M. R. Nieuwenstein and M. C. Potter,
"Temporal limits of selection and memory
encoding: A comparison of whole versus
in rapid serial visual
partial
report
presentation." Psychological science, vol. 17,
no. 6, pp. 471–5,
[Online].
Available: http://www.ncbi.nlm.nih.gov/
pubmed/16771795
jun 2006.
[46] J.-i. Kawahara,
J. T. Enns, and V. D.
Lollo, "The attentional blink is not
a unitary phenomenon," Psychological
Research, vol. 70, no. 6, pp. 405–413, oct
2006. [Online]. Available: http://link.
springer.com/10.1007/s00426-005-0007-5
[47] E. K. Vogel, S. J. Luck, K. L. Shapiro,
S. J. Luck, E. K. Vogel, K. L. Shapiro,
L. Berglan, M. Chun, V. Di Lollo, H. Egeth,
J. Hoffman, B. Maki, lane Raymond, and
E. K. Vogel or Steven J Luck, "Electrophys-
iological Evidence for a Postperceptual Lo-
cus of Suppression During the Attentional
Blink," Journal of Experimental Psychology
Nature, vol. 24, no. 382, pp. 616–618, 1998.
[48] Folk L. Charles, Leber B. Andrew, and
Egeth E. Howard, "Made you blink! Con-
tingent attentional capture produces a
spatial blink," Perception & Psychophysics,
2002.
[49] W. Maki and M. Mebane, "Attentional
capture triggers an attentional blink,"
Psychonomic Bulletin & Review, vol. 13,
no. 1, pp. 125–131,
[Online].
Available:
http://dx.doi.org/10.3758/
BF03193823
2006.
presented
sequentially
[50] P. E. Dux, C. L. Asplund,
"An attentional
and
blink
R. Marois,
for
targets:
evidence in favor of resource depletion
accounts." Psychonomic bulletin & review,
vol. 15, no. 4, pp. 809–13, aug 2008.
[Online]. Available: http://www.ncbi.
nlm.nih.gov/pubmed/18792508http:
//www.pubmedcentral.nih.gov/
articlerender.fcgi?artid=PMC3644218
[51] F. K. Chua,
"The effect of
target
contrast on the attentional blink," Per-
ception & Psychophysics, vol. 67, no. 5,
13
pp. 770–788,
able:
index/10.3758/BF03193532
jul 2005. [Online]. Avail-
http://www.springerlink.com/
[52] T. A. W. Visser and T. A. W., "Masking T1
difficulty: Processing time and the atte-
nional blink." Journal of Experimental Psy-
chology: Human Perception and Performance,
vol. 33, no. 2, pp. 285–297, 2007. [Online].
Available: http://doi.apa.org/getdoi.cfm?
doi=10.1037/0096-1523.33.2.285
[53] T. D. Grandison, T. G. Ghirardelli,
and H. E. Egeth, "Beyond similarity:
masking of the target is sufficient to
cause the attentional blink." Perception &
psychophysics, vol. 59, no. 2, pp. 266–74, feb
1997. [Online]. Available: http://www.
ncbi.nlm.nih.gov/pubmed/9055621
[54] B. G.
Breitmeyer, A.
Ehrenstein,
K. Pritchard, M. Hiscock, and J. Crisan,
"The roles of location specificity and
masking mechanisms in the attentional
blink." Perception & psychophysics, vol. 61,
no. 5, pp. 798–809,
jul 1999. [Online].
Available: http://www.ncbi.nlm.nih.gov/
pubmed/10498996
[55] C. N. L. Olivers and S. Nieuwenhuis,
"The beneficial effect of concurrent task-
irrelevant mental activity on temporal
attention." Psychological science, vol. 16,
no. 4, pp. 265–9, apr 2005. [Online].
Available: http://www.ncbi.nlm.nih.gov/
pubmed/15828972
[56] I. Arend, S. Johnston, and K. Shapiro,
"Task-irrelevant visual motion and flicker
attenuate the attentional blink." Psy-
chonomic bulletin & review, vol. 13,
no. 4, pp. 600–7, aug 2006. [Online].
Available: http://www.ncbi.nlm.nih.gov/
pubmed/17201358
14
|
1603.06879 | 1 | 1603 | 2016-03-22T17:23:15 | A Unifying Framework for the Identification of Motor Primitives | [
"q-bio.NC"
] | A long-standing hypothesis in neuroscience is that the central nervous system accomplishes complex motor behaviors through the combination of a small number of motor primitives. Many studies in the last couples of decades have identified motor primitives at the kinematic, kinetic, and electromyographic level, thus supporting modularity at different levels of organization in the motor system. However, these studies relied on heterogeneous definitions of motor primitives and on different algorithms for their identification. Standard unsupervised learning algorithms such as principal component analysis, independent component analysis, and non-negative matrix factorization, or more advanced techniques involving the estimation of temporal delays of the relevant mixture components have been applied. This plurality of algorithms has made difficult to compare and interpret results obtained across different studies. Moreover, how the different definitions of motor primitives relate to each other has never been examined systematically. Here we propose a comprehensive framework for the definition of different types of motor primitives and a single algorithm for their identification. By embedding smoothness priors and specific constraints in the underlying generative model, the algorithm can identify many different types of motor primitives. We assessed the identification performance of the algorithm both on simulated data sets, for which the properties of the primitives and of the corresponding combination parameters were known, and on experimental electromyographic and kinematic data sets, collected from human subjects accomplishing goal-oriented and rhythmic motor tasks. The identification accuracy of the new algorithm was typically equal or better than the accuracy of other unsupervised learning algorithms used previously for the identification of the same types of primitives. | q-bio.NC | q-bio |
A Unifying Framework for the Identification of
Motor Primitives
Enrico Chiovetto1, Andrea s'Avella2,3, and Martin Giese1
1Section for Computational Sensomotorics, Department of
Cognitive Neurology, Hertie Institute for Clinical Brain Research,
Centre for Integrative Neuroscience, University Clinic Tuebingen,
Tuebingen, Germany.
2Department of Biomedical and Dental Sciences and
Morphofunctional Images, University of Messina, Messina, Italy.
3Laboratory of Neuromotor Physiology, Santa Lucia Foundation,
Rome, Italy.
Abstract
A long-standing hypothesis in neuroscience is that the central nervous
system accomplishes complex motor behaviors through the combination
of a small number of motor primitives. Many studies in the last couples
of decades have identified motor primitives at the kinematic, kinetic, and
electromyographic level, thus supporting modularity at different levels of
organization in the motor system. However, these studies relied on het-
erogeneous definitions of motor primitives and on different algorithms for
their identification. Standard unsupervised learning algorithms such as
principal component analysis, independent component analysis, and non-
negative matrix factorization, or more advanced techniques involving the
estimation of temporal delays of the relevant mixture components have
been applied. This plurality of algorithms has made difficult to compare
and interpret results obtained across different studies. Moreover, how the
different definitions of motor primitives relate to each other has never
been examined systematically. Here we propose a comprehensive frame-
work for the definition of different types of motor primitives and a single
algorithm for their identification. By embedding smoothness priors and
specific constraints in the underlying generative model, the algorithm can
identify many different types of motor primitives. We assessed the iden-
tification performance of the algorithm both on simulated data sets, for
which the properties of the primitives and of the corresponding combina-
tion parameters were known, and on experimental electromyographic and
kinematic data sets, collected from human subjects accomplishing goal-
oriented and rhythmic motor tasks. The identification accuracy of the
new algorithm was typically equal or better than the accuracy of other
unsupervised learning algorithms used previously for the identification of
the same types of primitives.
1
Introduction
A fundamental challenge in neuroscience is to understand how the central ner-
vous system (CNS) controls the large number of degrees-of-freedom (DOF) of
the musculoskeletal apparatus to perform a wide repertoire of motor tasks and
behaviors. A long-standing hypothesis is that the CNS relies on a modular ar-
chitecture in order to simplify motor control and motor learning [1-3]. Many
studies in recent years have indeed shown that kinematic [4-5], kinetic [6-7] and
electromyographic (EMG) patterns [8-11] underlying complex movements can
be approximated by the combinations of a small number of components, usually
referred to as motor primitives or motor synergies. The identification of such
components has typically been carried out by applying unsupervised learning
algorithms, including principal component analysis (PCA), independent compo-
nent analysis (ICA) [5, 12-15], non-negative matrix factorization (NMF) [15, 16]
or other methods inspired by such algorithms [17]. While these classical meth-
ods are based on instantaneous mixture models, that linearly combine a set
of basis vectors time-point by time-point, more advanced techniques have also
been proposed that involve the estimation of temporal delays between relevant
mixture components [15, 18-21]. This multitude of underlying mathematical
models complicates the comparison of results from different studies on motor
primitives. In addition, even for the same mathematical models often multiple
algorithms for the estimation of motor primitives have been proposed, and it is
not always clear if their results are comparable. This further complicates the
comparison of the results. Finally, how the different definitions of motor primi-
tives relate to each other has never been systematically examined. We propose
in this article a new comprehensive framework for the definition of motor prim-
itives and a new algorithm for their identification. We show that many different
definitions of spatial, temporal and spatiotemporal primitives given in the liter-
ature can be derived from a single generative model that is known as "anechoic
mixture" and relies on the combination of components that can be shifted in
time. When the delays of all primitives are constrained to be zero, the anechoic
model reduces to the instantaneous linear combination model, which underlies
the definition of spatial or temporal synergies, usually identified by PCA, ICA
or NMF. Similarly, when specific equality and non-negativity constraints are
imposed on its parameters, the model can describe spatiotemporal synergies [9,
15]. In addition to this unification of models, we present a new identification
algorithm that estimates motor primitives, according to the different definitions,
with an accuracy that is equal or even better than the standard techniques that
are commonly used for the identifications of these motor primitives. The robust-
ness of this new algorithm results from an integration of smoothness priors and
appropriate constraints in the underlying generative model. The new algorithm
has been validated by assessing its identification performance both on simulated
data sets, for which the properties of the primitives and of the corresponding
combination parameters were known, and on experimental EMG and kinematic
data sets, collected from human participants accomplishing goal-oriented and
rhythmic motor tasks. The new algorithm is publically available, is provided as
a toolbox in MATLAB (The Mathworks, Natick, MA) and can be downloaded
for free from www.compsens.uni-tuebingen.de. In this way, we aim to provide
the field of motor control with a new usable and robust tool for the identification
of motor primitives, helping to reduce the inconsistencies and incompatibilities
2
between the different generative models.
Methods
Generative models for the description of motor primitives
We give in this section a brief survey of the definitions of motor primitives and
of the corresponding generative models that have been used in the literature
for the investigation of the modular organization of motor behavior. The differ-
ent approaches can be subdivided into different groups, according to the model
features that are assumed to be invariant across conditions. In the following,
a matrix Xl indicates the data corresponding to a specific trial l (0 ≤ l ≤ L),
where L is the total number of trials collected during an experiment. Each row
of Xl represents a specific degree of freedom (DOF) of the system under investi-
gation (for instance an angular trajectory associated with a specific joint in the
case of kinematic data, or the electrical signals associated with the contraction
of a specific muscle in the case of EMG data). Each column of Xl contains the
values assumed by the different DOF at a particular point in time. Unless the
size of the matrix is explicitly mentioned, from now on Xl will be assumed to
have M rows (number of DOF) and T columns (equivalent to the number of
time samples in one trial). Signals are supposed to be sampled at constant sam-
pling frequency and to have duration Ts. In the following, an individual column
that corresponds to the time point t of Xl will also be signified by the column
vector xl(t), so that Xl = [xl(1), ..., xl(T )]. The components of these vectors
will be indicated by the variables xl
n(t). In the following, we give an overview
of different models for motor primitives that have been proposed previously in
the literature.
Spatial primitives
One classical definition of motor primitive is based on the idea that groups of
DOF might show instantaneous covariations, reflecting a coordinated recruit-
ment of multiple muscles or joints. This implies the assumption that the ratios
of the signals characterizing the different DOF remain constant over time. This
type of movement primitive has been applied in particular in muscle space, where
muscle synergies have been defined as weighted groups of muscle activations [3,
10, 22]. Such synergies have also been referred to as "synchronous" synergies,
since the different muscles are assumed to be activated synchronously without
muscle-specific time delays. Consistent with this definition is the following gen-
erative model that, from now on, will be referred to as 'spatial decomposition':
xl(t) =
P
X
p=1
wp · cl
p(t) + residuals
(1)
In this equation the vectors xl(t) indicate the values of the individual DOF
at time point t (assuming discrete time steps, 1 ≤ t ≤ T ) in trial number l. The
column vectors wp define the 'spatial patterns' of the muscle synergies, which
are assumed to be invariant over trials. The number of primitives is P , and
the scalars cl
p(t) are the time-dependent mixing weights of the primitives. The
mixing weights, as well as the residuals, are different in every trial. Processed
3
EMG data typically consists of time series of non-negative signals, i.e. xp
m(t) ≥
0, for 1 ≤ t ≤ T and 1 ≤ m ≤ M . In these models it is typically also assumed
that the components of the mixture model (1) (except for the residuals) are
non-negative, i.e. cl
p(t) ≥ 0 and wp,m ≥ 0 (where the subscript m indicates the
m-th element of the vector wp).
Temporal primitives
An alternative way to characterize motor primitives is based on the idea that
they express invariance across time, defined by basic temporal patterns or func-
tions sp(t) that are combined or superposed in order to reconstruct a set of
temporal signals. Temporal components based on this definition have been
identified in kinematic [4-5, 18], dynamic [6] and EMG [8, 11-12] space. The
underlying generative model (which from now on we will refer to as 'temporal
decomposition') is mathematically described as:
xl
m(t) =
P
X
p=1
cl
mp · sp(t) + residuals
(2)
In this equation xl
m(t) is the value of the m-th DOF at time t in trial number
l, and the corresponding scalar mixing weights cl
mp change between trials of dif-
ferent types (experimental conditions). The temporal primitives sp(t), however,
are assumed to be invariant over trials. P signifies the total number of temporal
primitives. Another more elaborated model of this type has been proposed in
[19, 21-23]. This model allows for temporal shifts between the temporal basis
functions for different DOF. This can be interpreted as reflecting, for example,
delays between the activation of different muscles within the same primitive.
Mathematically, this model is characterized by the equations:
xl
m(t) =
P
X
p=1
cl
mp · sp(t − τ l
mp) + residuals
(3)
The time shifts between the basis functions for the different degrees of free-
dom are captured by the variables τ l
mp. The time delays and linear mixing
weights are typically assumed to vary over trials, while it is assumed that the
basis functions sp(t) are invariant, as in model (2). Like for model (1), inequal-
ity constraints can be imposed on the mixing weights in models of type (2) and
(3), for example to account for the non-negativity of EMG signals.
Spatiotemporal (time-varying) primitives
Spatiotemporal (or time-varying) primitives have been proposed as a way to
model EMG components that are invariant in both, space and time [9, 15, 24].
Moreover, for each primitive additional temporal delays are admitted, similar
to model (3). This results in the following generative model (referred to as
'spatiotemporal decomposition'), where xl(t) signifies again the time-dependent
column vector of the DOF as function of time:
xl(t) =
P
X
p=1
cl
p · wp(t − τ l
p) + residuals
(4)
4
p and the delays τ l
Again, the mixing weights cl
p change between different trial
types while the functions wp(t) are assumed to be invariant, defining the prim-
itives or muscle synergies. The time-varying synergies and the corresponding
mixing weights have typically been assumed to be non-negative [15], although
also models with unconstrained parameters have been applied to model phasic
EMG activity [9].
Space-by-time primitives
Recently, Delis and colleagues [25] proposed a new synergy model for EMG
data, which they named 'space-by-time decomposition'. This model merges
the definitions of spatial and temporal components into a new definition of
primitives that is given by the following equation:
xl(t) =
Ptp
Psp
X
p=1
X
q=1
s(t − τ l
pq) · cl
pq · wq + residuals
(5)
In this model, wq and sp(t) define the trial-independent spatial and temporal
components as in models (1) and (2), while the mixing weights cl
pq and time
delays τ l
pq are trial-dependent. The constants Ptp and Psp indicate the numbers
of temporal and spatial components. Since the model was originally designed
to account for EMG data, Delis and colleagues assumed all parameters of the
model equation (5) to be non-negative (except for the time delays).
Unifying model
All previously discussed models can be derived as special instantiations of a sin-
gle model, called 'anechoic mixture model'. This type of model is known from
acoustics, where it is applied for modeling of acoustic mixtures in reverberation-
free rooms [26-29]. This model assumes typically a set of R recorded acoustic
signals yr(t) that are created by the superposition of U acoustic source functions
fu(t), where time-shifted versions of these source functions are linearly super-
posed with the mixing weights aru. The time shifts are given by the time delays
τru. This models the fact that for a reverberation-free room the signals from
the acoustic sources arrive receiver with different time delays and attenuated
amplitudes, which are dependent on the distances between the acoustic sources
and the receivers. The corresponding generative model has the following form
(for 1 ≤ r ≤ R ):
yr(t) =
U
X
u=1
aru · fu(t − τru) + residuals
(6)
Equivalence between the unifying model and the other models
By addition of appropriate constraints, the anechoic mixture model (6) can be
made equivalent to all previously discussed models for motor primitives. This
becomes obvious by the following considerations:
a) Identifying the signals of type yr(t) with the components of the vectors
xl(t), i.e. yr(t) = xl(r)
m(r)(t) (where the integer functions l(r) and m(r) define
a one-to-one mapping between the m-th degree of freedom in trial l and the
5
corresponding signal yr(t) (with 1 ≤ r ≤ M ·L), and constraining the time delays
τru to be zero, one obtains the model (1). Since in this model the weight vectors
wp are assumed to be invariant over trials, all mixing weights arp belonging to
the same DOF and primitive number P have to be equal and independent of
the trial number, so that arp = wp,m(r), where the function m(r) returns the
number of the DOF that belongs to index r independent of the trial number.
The time-dependent mixing coefficients cl
p(t) of the model (1) correspond to the
source functions fu of the model (6), thus fu(t) = cl(u)
p(u)(t) where here the index
u runs over all combinations of the indices p and l, thus 1 ≤ u ≤ U = P · L
and where the integer functions l(u) and p(u) establish mappings between the
number of the source function in model (6) and the time-dependent mixing
weights in model (1). Non-negativity constraints can be added for the model
parameters arpand the functions fu(t), e.g. for the modeling of EMG data.
b) If one identifies the source functions in model (6) with the temporal
primitive functions sp(t), i.e. fp(t) = sp(t), 1 ≤ p ≤ P and again constrains
the delays τru to be zero, equation (6) becomes equivalent to model (2).
In
this case, the mixing weights arpare equated with the mixing coefficients cl
mp in
model (2), where the index r runs over all combinations of m and l, formally
arp = cl(r)
m(r),p, with appropriately chosen integer functions m(r) and l(r). Like
for model (1), the components of the data vector have to be remapped over DOF
and trials according to the relationship yr(t) = xl(r)
m(r)(t). Again, non-negativity
constraints can be added for the parameters arpand to the source functions f .
c) Dropping the constraints τru = 0 in the equivalences described in b), and
m(r),p,
equating the delays in model (3) according to the relationship τrp = τ l(r)
makes model (6) equivalent to model (3).
d) Introducing individual sets of basis functions for the different DOF, group-
ing them into vectors and equating the mixing weights and temporal delays for
the components of each vector, transforms model (6) into the model (4). On the
level of the time-dependent basis functions, this equivalence can be mathemat-
ically described by the equation fu(t) = wp(u),m(u)(t), where wp,m corresponds
to the component of the basis function vector wp(t) that belongs to the m-th
DOF, and where the integer functions m(u) and p(u) establish a one-to-one
mapping between the indices of the basis functions in the two models and the
number of the associated DOF. This assignment is independent of the trial index
l. The index r in (6) runs over all combinations of DOF and trial numbers, thus
1 ≤ r ≤ M · L. The integer functions m(r) and l(r) assign the corresponding
trial number and DOF to the index r in the model (6). Thus, the assignment
equation for the data vector is again given by yr(t) = xl(r)
m(r)(t) for the m-th
DOF in the l-th trial. The requirement that all mixing weights and temporal
delays belonging to the same basis function vector wp are equal is equivalent
to a set of equality constraints, which can be captured by the equation systems
aru = cl(r)
p(r) . Again, non-negativity constraints can be added, if
necessary.
p(r) and τru = τ l(r)
e) In order to establish equivalence with the model (5), the data vectors of
the models are mapped onto each other according to the relationship yr(t) =
xl(r)
m(r)(t), where again l(r) and m(r) are integer mapping functions that assign
the r-th element of the data vector of the model (6) to the m-th DOF of the
data vector xl for the l-th trial in (5) with 1 ≤ r ≤ M · L. Model (5) has a total
6
of Psp · Ptp temporal basis functions, where however the functional forms of the
basis functions for different indices q (i.e. different spatial components) for the
same p (i.e. same temporal component) just differ by time shifts. This is equiv-
alent to an equality constraint for these functions, which can mathematically
be characterized in the form fu(t) = sp(u)(t), with 1 ≤ u ≤ Ptp and the index
functions p(u) and q(u) that map the index u in the model (6) onto the indices
of the temporal and spatial primitive in (5). Since all indices with the same
p(u) are mapped onto the same basis function sp the last equation specifies an
equality constraint. With the same integer mapping functions, finally, also the
relationship between the mixing weights can be established, which is given by
the equation aru = cl(r)
p(u),q(u) · wq(r),m(u), where wq,m is the m-th element for
the vector wq. The last equation specifies a bilinear constraint for the weight
parameters of the model (6). Using the same notation, the equivalence between
the delays is established by the equation system τru = τ l(r)
p(u),q(u). A summary
of the established equivalences between the general model (6) and the other
models is given in Table 1.
An efficient algorithm for the identification of motor prim-
itives within the unified framework
[35], Arberet et al.
The solution of anechoic demixing problems is a well-known topic in unsuper-
vised learning, with close relationship to methods such as ICA and blind source
separation (cf. e.g. [30-31]). Numerous algorithms have been proposed to solve
this problem for the most general case where the functions f are assumed to be
elements of relatively general function spaces. For the under-determined case
(in which the number of signals/sensors is smaller than the number of sources)
well-known algorithms include information maximization approaches [32] and
frequency, or time-frequency methods [33-34], such as the DUET algorithm
[29]. Other work for the under-determined case is summarized in Ogrady et
al.
[36] and Cho and Kuo [37]. The over-determined case
(where the signals outnumber the sources) is much more interesting for dimen-
sionality reduction applications, but has been addressed more rarely. Harshman
and colleagues [38] developed an alternating least squares (ALS) algorithm for
this problem (Shifted Factor Analysis). Their method was later revised and im-
proved by Mørup and colleagues [39], who exploited the Fourier shift theorem
and information maximization in the complex domain (SICA, Shifted Indepen-
dent Component Analysis). More recently, Omlor and Giese [19] developed a
framework for blind source separation, starting from stochastic time-frequency
analysis that exploited the marginal properties of the Wigner-Ville spectrum.
The discussed algorithms solve the anechoic demixing problem for the most gen-
eral case, at the cost that they are computationally expensive. All algorithms
for blind source separation require the identification of a large number of pa-
rameters. Given model (6), T · U parameters need to be identified to represent
all sources fu(t), and for each trial l, M · U weights aru and M · U delays τru.
Given a whole data set, this results in a total number of parameters to be iden-
tified that is (T + 2M · L) · U , where typically T ≫ M, U, L (with T, M, U and
L indicating the total numbers of time samples, DOF, sources, and trials). For
applications in motor control, the relevant signals are subject to additional con-
straints, which can be exploited for the derivation of more efficient algorithms.
7
Spatial (1)
Temporal (2) or (3)
Spatiotemporal (4)
Space-by-time (5)
xl(t) =
P
Pp=1
wp · cl
p(t)
xl
m(t) =
P
Pp=1
cl
mp · sp(t − τ l
mp) xl(t) =
P
Pp=1
cl
p · wp(t − τ l
p) xl(t) =
Ptp
Pp=1
Psp
Pq=1
s(t − τ l
pq) · cl
pq · wq
Anechoic (6)
yr(t) = xl(r)
m(r)(t)
yr(t) = xl(r)
m(r)(t)
yr(t) = xl(r)
m(r)(t)
yr(t) = xl(r)
m(r)(t)
8
yr(t) =
U
Pu=1
aru · fu(t − τru)
fu(t) = cl(u)
p(u)(t)
fp(t) = sp(t),
fu(t) = wp(u),m(u)(t)
fu(t) = sp(u)(t),
arp = wp,m(r)
arp = cl(r)
m(r),p
τru = 0
τru = 0 or τrp = τ l(r)
m(r),p
aru = cl(r)
p(r)
τru = τ l(r)
p(r)
aru = cl(r)
p(u),q(u) · wq(r),m(u)
τru = τ l(r)
p(u),q(u)
Table 1: Constraints that make the primitive models (1), (2), (3), (4) and (5) equivalent to the general anechoic model (6). See text for
details.
Signals in motor control are typically smooth. This allows to reduce consider-
ably the complexity of the anechoic demixing problem and to devise algorithms
that are more robust than those developed for general purposes. We present in
this section a unifying algorithm for standard anechoic demixing, which can be
used for the identification of the parameters associated with the unconstrained
model (6). The general version of this algorithm, which from now on we will re-
fer to as FADA (Fourier-based Anechoic Demixing Algorithm), was introduced
in a previous study to identify primitives defined according to eq. (3) [18]. Here
we describe how this algorithm can be extended by inclusion of additional con-
straints that make it suitable for the identification of the parameters associated
with different models for primitives. The time-courses of signals related to body
movements (trajectories as well as EMG traces) often are relatively smooth and
thus can be approximated well by anechoic mixtures of smooth signals [18].
This smoothness of the source functions f (u) can be expressed by appropriate
priors that help to stabilize the source separation problem. Smooth temporal
sources can be approximated by truncated Fourier expansions. Consequently,
each source can be approximated by K complex Fourier coefficients, whereK is
typically far below the Nyquist limit (K ≪ T /2). Consequently, the number of
parameters to identify drops remarkably to (K + 2M · L) · U . This decreases
substantially the computational costs of the parameter estimation and make it
more robust. When the temporal signals yr(t) and sources fu(t) are assumed
to be band-limited they can be approximated by truncated Fourier expansions
of the form:
and
yr(t) =
K
X
k=−K
crke
2πikt
Ts
fu(t − τru) ∼=
K
X
k=−K
νuke−ikτru e
2πikt
Ts
(7)
(8)
where crk and νuk are complex constants (crk = crk eiϕcrk and νuk =
νuk eiϕνuk ), and where i is the imaginary unit. The positive integer K is
determined by Shannon's theorem according to the limit frequency of the sig-
nals, and Ts is the temporal duration of the signal. The source separation algo-
rithm tries to ensure that the source functions fu(t) are uncorrelated over the
distributions of the approximated signals. This implies E {fu(t) · fu′(t′)} = 0
for u 6= u′ and any pair t 6= t′. For the corresponding Fourier coefficients this
implies E {νuk · νu′k′ } = 0 for u 6= u′ and any pair k 6= k′ . Combining equation
(6), (7) and (8) we obtain by comparison of the terms for the same frequency
crk =
U
X
u=1
aru · νuke−ikτru
(9)
From this follows with E {νuk · ν ∗
u′k′ } = E nνuk2o · δuu′ the equation:
9
crk2 = E {crk}
=
=
=
U
U
X
u=1
X
u′=1
aruaru′ E {νuk · ν ∗
u′k′ } e−ik(τru−τru′ )
U
X
u=1
U
X
u=1
ruE nνuk2o
a2
aru2 νuk2
(10)
The symbol ∗ indicates the conjugate of a complex number. The derivation of
this equation replaces the expectations of the Fourier coefficients crk with their
deterministic values and treats the source weights ark as deterministic trial-
specific variables. This can be justified if these mixture weights are estimated
separately from the sources in an EM-like procedure. Empirically, however, we
obtain reasonable estimates of the model components based on equation (10)
also using other methods (see below). Since the signals fu(t) and yr(t) are real
the corresponding Fourier coefficients fulfil crk = c∗
u,−k. Thus
the demixing problem needs to be solved only for parameters with k ≥ 0.
r,−k and νuk = ν ∗
The previous considerations motivate the following iterative algorithm for
the identification of the unknown parameters in model (6). After random ini-
tialization of the estimated parameters, the following steps are carried out iter-
atively until convergence:
1. Compute the absolute values of the coefficients crk and solve the following
equations:
crk2 =
U
X
u=1
aru2 νuk2
(11)
with r = 0, 1, . . . R and k = 0, 1, . . . K. In our study we exploited non-
negative ICA [40] to solve this equation.
In the distributed version of
the software equation (10) can also be solved via non-negative matrix
factorization [34, 41].
2. Initialize for all pairs and iterate the following steps:
(a) Update the phases of the Fourier coefficients of the sources, defined
as ϕνuk = angle(νuk) = arctan(Im(νuk)/Re(νuk)) by solving the
following non-linear least square problem
min
Φ
kC − Z(Φ)k2
F
(12)
where (C)rk = crk, (Z)rk =
U
Pu=1
arue−ikτuk νuk eiϕνuk and indicates
the Frobenius norm. (In order to avoid cluttered notation, for the
function Z(.) only the arguments with relevance for the optimization
are explicitly written).
10
(b) Keeping the identified source functions fu(t) constant, identify for
each signal yr(t) the weights aru and delays τru by minimization of
the following cost functions:
arg min
ar,τ r
kyr(t) − f (t, τ r)′ark2
F
(13)
Optimization with respect to ar and τ r is feasible, assuming uncorrelatedness of
the functions fu and independence of the time delays [42]. The column vector
ar concatenates all weights associated with DOF r, i.e ar = [ar1, . . . , arU ]′.
The vector function fr(t, τ r) = [f1(t − τr1), . . . , fU (t − τrU )] concatenates the
functions fu, shifted by the time delays associated with the r-th DOF.
The original version of the FADA algorithm was designed to solve the source
separation problems without constraints. Additional constraints, such as the
non-negativity of the parameters or additional equality constraints for the weights
and delays can be easily added, due to the modular structure of the algorithm.
The following sections briefly describe the additional constraints that were in-
troduced in order to implement the identification of the parameters of models
(1), (2), (3), (4) and (5).
Non-negativity of the primitives
For the case where the primitives fu can assume only non-negative values, equa-
tion (10) cannot be derived in the way discussed above, and the expression of the
non-negativity constraints in the Fourier representation is not straightforward.
We decided, instead to estimate the time-dependent values of fu(t) directly,
taking the inequality constraint fu(t) ≥ 0 for discretely sampled values of into
account. This results in the following algorithm: Starting from random values
of the parameters, the following three steps are iterated until convergence:
1. Update of the absolute values of the Fourier coefficients νuk of the primi-
tives fu, assuming their phases ϕνuk and the mixing weights aru are known,
by solving the non-linear constrained optimization problem:
minimize
N
kC − Z(N)k2
F
subject to
fu(N, t) ≥ 0, u = 1, 2, . . . U and t = 1, . . . T.
(14)
In order to avoid cluttered notation, for the functions Z(.) and fu(.) only
the arguments with relevance for the optimization are explicitly written.)
The matrix is defined as in (12) and (Z)rk =
with (N)uk = νuk.
U
Pu=1
arue−ikτuk νuk eiϕνuk ,
2. Assuming the other parameters are fixed, update the phases ϕνuk of the
Fourier coefficients of the primitives by solving the non-linear constrained
optimization problem
minimize
Φ
kC − Z(Φ)k2
F
subject to fu(Φ, t) ≥ 0, u = 1, 2, . . . U and t = 1, . . . T.
(15)
Remind that the Fourier coefficients νu0 are real so that it is sufficient to
regard consider only k = 1, . . . , K.
11
3. Update weights and delays as in the unconstrained version of FADA by
solving the optimization problem (13).
Non-negativity of the mixing coefficients
Non-negativity of the scaling coefficients aru of the primitives can be easily im-
posed in the algorithm. In (13) the scaling coefficients are determined, assuming
that primitives and temporal delays are known, solving a least squares problem.
The same optimization problem can be solved adding the linear inequality con-
strains aru ≥ 0, ∀r, u, resulting in a non-negative least squares problem for the
weights aru.
Identification of spatial synergies
The FADA algorithm presented above can be used to identify not only temporal,
but also spatial primitives. This can be achieved simply by transposing the data
matrix X and constraining all the delays in the algorithm to be equal to 0. In
this way indeed, the FADA algorithm identifies a set of invariant spatial (instead
of temporal) vectors, interpreting the elements of each vector x(t) as a series
of time points. Although there is no theoretical evidence for the existence of
any smoothness relation between the values of the different DOF at a given
time instant t (so that the smoothness assumptions of FADA on the data are
satisfied), it will be shown in the next sections how the algorithm can however
still provide identification performance at least as good as those associated with
other standard machine learning techniques.
Identification of spatiotemporal synergies
For the identification of spatiotemporal synergies, constraints for the parameters
have to be set according to model (4). In a first step, for each DOF m in the data
set P source functions fp are assigned, resulting in a total of M · P independent
source functions. The following three steps are then carried out iteratively until
convergence:
1. The optimal delays τ l
p for each spatiotemporal primitive are found, for
each trial l, applying a matching pursuit procedure [43-44], consisting of
an iterative search for a set of time-shifted primitives that best match
the data. For each primitive, the scalar product between the original data
and the time-shifted primitive is computed, testing all possible time delays
between 0 and T -1. The primitive and delay associated with the highest
scalar product is then selected and its contribution is subtracted from the
data. Then the same procedure is repeated for the remaining primitives
on the residual of the data. This search is repeated until all delays have
been determined.
2. The combination coefficients cl
p are updated by minimizing, for each trial l,
the difference between the original data and the reconstruction, estimated
exploiting model (4) and assuming that the source functions fu and the
delays τru are known.
12
3. Assuming that the weights and the delays are known from the previous
steps, the functions fu, which correspond to the components of the spa-
tiotemporal primitives wp(t) are updated. The Fourier coefficients of the
corresponding source function are determined in the same fashion as for
the original FADA algorithm without constrains. Non-negativity con-
straints for the primitives and weights can be imposed in the same way as
described above.
Identification of space-by-time synergies
We developed a new algorithm for the identification of space-by-time primitives,
exploiting the core of FADA algorithm (the mapping onto the Fourier space)
for the identification of the temporal primitives associated with the space-by-
time factorization. Similar to Delis and colleagues [25], this algorithm was also
designed for the processing of EMG-like data and all the parameters in model (5)
(with the exception of the delays) are constrained to be non-negative. Given
the data matrix X, in the first step of the algorithm Psp spatial primitives
wq are identified, applying non-negative matrix factorization [41]. Then the
FADA algorithm is applied to X in order to identify Ptp non-negative temporal
primitives sp(t). In the second step of the algorithm, the spatial primitives are
kept constant, while temporal primitives, weights and delays are updated. The
algorithm consists of the iteration of the two following steps:
1. The Fourier coefficients of the functions sp(t) are updated as in the con-
strained FADA algorithm, by minimizing the difference between the Fourier
coefficients of the original data and the linear combination of the corre-
sponding Fourier coefficients.
2. Weights and delays are updated minimizing the difference between the
original data and the estimates provided by model (5). The optimal delays
τ l
qp are found for each trial l, following a matching pursuit procedure.
Similarly, the weights cl
qp are identified, solving for each trial a constrained
linear least-squares problem.
To minimize the risk of finding local minima, we always ran the FADA algorithm
10 times on the same data set with different random initial conditions and we
considered only the solutions that provided the lowest error in the reconstruction
of the original data. To test whether these solutions actually represented points
close to the global minimum, we computed the average similarity between the
sets of primitives identified at the end of each run of the algorithm (see below
for the definition of similarity). Indeed, a high level of similarity between these
solutions can be considered as a strong sign that, with very high probability,
these solutions are close to the optimal one.
In the case, for instance, of an
artificial mixture of non-negative temporal components based on model (2), we
found that the average similarity between the identified primitives was very
high (0.98 on a scale where 1 indicates perfect matching (see equation 16).
This high level of similarity allows to rule out the hypothesis that the solutions
provided by FADA represent local minima. For the identification of temporal,
spatiotemporal or space-by-time primitives, the number of harmonics K was
always set according to the following procedure: We computed the average
spectrum from all signals within the data and defined K as the closest integer
13
that approximates the product of the signal duration Ts and the average band-
width B of the data set. This number was alwayssmaller thanthe limit Kmax
imposed by the the Nyquist-Shannon theorem. Differently, in the case of spatial
primitives we always set K = Kmax.
Other identification algorithms
The FADA algorithm was benchmarked against other unsupervised learning
methods for the extraction of synergies. For data based on the synchronous
unconstrained generative models (1) and (2) we used the fastICA algorithm
[45-46] (function 'fastica.m' of the corresponding toolbox). We examined the
performance of fastICA after reducing the dimensionality of the data using
principal component analysis. For the fastICA algorithm we found the level
of similarity between original and identified synergies depended on the num-
ber of principal components and it reached the highest value when the num-
ber of principal components was equal to the number of synergies in the data.
Based on this observation we always set the number of principal components
to the number of identified synergies. Non-negative matrix factorization [34,
41] (NMF) was used to identify the model parameters for synchronous mixture
with non-negative components and mixing weights. We used the Matlab func-
tion "nnmf.m", implementing the matrix multiplication update rule version of
the algorithm introduced by Lee and Seung [34, 41]. For data relying on model
(3) we used the anechoic demixing algorithm (AnDem) developed by Omlor and
Giese [19] and the shifted ICA algorithm (SICA) by Morup [39] (downloaded
from http://www2.imm.dtu.dk/pubdb/views/publication details.php?id=5206
[47]). For anechoic demixing with non-negativity constraints we used an ane-
choic NMF algorithm (ANMF) developed by Omlor and Giese [19] and the
shifted NMF (sNMF) by Morup and colleages [48], who kindly provided us with
the Matlab implementation of the algorithm. To extract time-varying syner-
gies we used the modified NMF algorithm developed by d'Avella and colleagues
[9, 15] (stNMF, standing for spatiotemporal non-negative matrix factorization).
Finally, we compared the performance of the FADA algorithm for the identi-
fication of temporal and spatial primitives from the space-by-time model with
the performance of the sample-based non-negative matrix tri-factorization algo-
rithm (sNM3F) developed by Delis and colleagues [25].
Generation of the simulated data
For the quantitative assessment of algorithm's performance we simulated kine-
matic and EMG data sets that were compatible with equations (1), (2), (3), (4)
and (5). Each of these data sets approximated coarsely the properties of real
biological signals. Each data set consisted of M -dimensional trajectories with
T time steps and Lrepeated trials. Synthesized EMG signals were constrained
to be non-negative, like real EMG signals after rectification and filtering. All
generative models were based on a set of statistically independent temporal
waveforms. These waveforms (source functions, or synergies) corresponded to
the time-dependent combination coefficients cl
p(t) in model (1), to the tem-
poral signals sp(t) in models (2) and (3) and (5), and to the components of
the vector function in model in (4). For the generation of the unconstrained
sources we drew 100 random samples from a normal distribution (Matlab func-
14
tion "randn.m") and low-pass filtered with a Butterworth filter with normalized
cut-off frequency equal to 0.15 (Matlab functions "butter" and "filtfilt"). This
procedure allowed to generate band-limited, smooth sources mimicking the typ-
ical properties of real kinematic or kinetic trajectories with a length of T = 100
time samples. For the generation of EMG-like sources we produced spike trains
from a multi-dimensional stochastic renewal process [49], and convolved them
with a Gaussian function. The renewal process was a homogeneous Poisson pro-
cess characterized by random inter-spike intervals drawn from an exponential
distribution with mean 1/λ, where the rate parameter of the Poisson process
was given by λ = 40 Hz. Based on the random inter-spike intervals, spike trains
with length T = 100 were generated. Each spike train was then convolved with
a Gaussian filter kernel with a standard deviation of 8 discrete time steps. The
generated source signals were used to construct the synergies in the generative
models (2), (3), (4) and (5). The weight vectors w in (1) and (5) were obtained
by drawing M random samples from a uniform distribution over the interval
[-40 40] for the unconstrained case, and from an exponential distribution with
mean 20 for the cases with non-negativity constraints. Examples of generated
primitives are shown in Fig 1. For kinematic (unconstrained) data sets based on
model (2) and (3) the values of the coefficients cmp were drawn from a uniform
distribution over the interval [-20, 20]. For EMG-like data sets based on the
models (2), (3), (4) and (5) the scaling coefficients were drawn from exponential
distributions with mean 10. For all the models with time delays τ 6= 0, the
delays were drawn from exponential distributions with mean 20 and rounded to
the nearest integer. The time delays sampled from this distribution with values
larger than T = 100 were taken modulo to map them back to the interval [0,
T -1]. Noisy data was derived by adding signal-dependent noise [50-53] to the
generated data. The noise was drawn from a Gaussian distribution with mean
0 and standard deviation σ = α x(t), where α is a scalar and x(t) is the value
of the noiseless data at the time instant t. The slope α was computed though
an iterative procedure. Starting from α = 0, its value was iteratively increased
of a predefined increment until the level of the difference 1 − R2 (where the
parameterR2 describes the level of similarity between two data sets, see below)
reached a predefined value. For each noiseless data set, three data sets were
generated with1 − R2 levels equal to 0.05, 0.15, 0.25 and 0.35). For each gen-
erative model, 20 noiseless data sets were simulated that were consistent with
equations (1) to (5), randomly selecting synergies, scaling coefficients and time
delays. The number of synergies P was always set to 4 and the number of
simulated DOFs was 10. The number of simulated trials L was 25. The time
duration of each trial was assumed to be Ts = 1 and the sampling frequency
was set to 100 Hz.
Experimental kinematic and EMG data
We assessed the identification performance of each algorithm also on actual
experimental kinematic and EMG data. The kinematic data set consisted of
flexion angle trajectories of the body joints recorded from human actors walk-
ing with different emotional styles (neutral, happy and sad). These data were
used in previous work on emotional gaits [20, 23, 54]. From this data set, uncon-
strained temporal primitives were identified with the FADA and the anechoic
demixing algorithm. EMG data consisted of previously published recordings [9]
15
obtained from 16 arm muscles during arm reaching movements. These muscle
activation patterns were used to investigate the production of behaviors through
combination of muscle synergies. The recorded EMG raw signals were digitally
full-wave rectified, low-pass filtered (20 Hz cut-off) and integrated within time
bins of 10 ms. All EMGs signals in the data set were resampled to fit a 75-point
time window (0.75s).
Assessment of algorithm performance
For each algorithm we assessed three different performance measures, quantify-
ing the capability of each algorithm to identify the original movement primitives,
the original activation coefficients and the original delays in comparison to the
parameters used to generate the data. The similarity between original and
extracted primitives was quantified by computing the maximum of the scalar
products between original and identified primitives, taking the maximum over
all possible time delays in cases where model contained temporal shifts. Let
p1(t) and p2(t) signify the compared primitives or source functions (discretely
sampled in time) and that these signals are normalized so that their norm is
one. The similarity measure is defined by the scalar product of these normal-
ized signals, where one of them is time-shifted with time delay , where this delay
is optimized by maximizing the similarity measure. For models without time
shifts the time delays are constrained to be 0. Mathematically the correlation
measure is given by the equation:
S = max
τ X
p1(t) · p2(t − τ )
τ
(16)
For the case of time-varying synergies (model 4) the compared signals were
vector-valued. In this case the scalar product of the vectors in (16) was taken
for each (delayed) time step and the signals were normalized ensuring that for
pj((t)2 = 1, for j = 1, 2. The similarity measure takes values between -1
Pt
and 1, where the value 1 corresponds to the situation that both source function
have identical shape (except for maybe a time delay). In order to establish cor-
respondence between the individual primitives of the generative model and the
identified primitives, we first computed the similarity measure S for all possible
pairings of the primitives and chose then the pairing with the highest similar-
ity score. For this purpose, first the pairing with the highest similarity score
was determined and removed from the original and reconstructed model. Then
this procedure was repeated for the second-best matching pair of the remaining
set of primitives, and so forth. This procedure was iterated until all primitives
had been matched. The similarity between original and identified coefficients
(or time delays) was assessed by computing the correlation coefficients between
activation coefficients (temporal delays) of the matched primitives. These cor-
relation coefficients were then averaged across all matched pairs of primitives.
We also defined a measure of similarity between original and reconstructed data
sets. Since the generated and experimental kinematic or EMG patterns and the
residuals of the reconstruction of the patterns by synergy combinations were
multivariate time-series, a measure of the goodness of the reconstruction (typi-
cally a ratio of two variances) had to be defined. We used the "total variation"
[55], defined as the trace of the covariance matrix of the signals. A multivariate
16
measure R2 for the explained data variance is then given by the expression
R2 = 1 −
L
Pl=1(cid:13)(cid:13)Xl − Xl
rec(cid:13)(cid:13)
Pl=1(cid:13)(cid:13)Xl − ¯Xl(cid:13)(cid:13)
L
(17)
where each Xlwas the matrix of the actual data associated with trial l, Xl
rec
the reconstructed values by the fitted model, and where ¯X is the matrix of the
mean values of the data over trial l. The statistical distributions of all similarity
measures described above for randomized data were assessed for each algorithm
and each data set.
In order to calculate baseline levels for these similarity
measures (Σb), we first randomly generated 20 independent sets of synergies,
coefficients, and delays (where appropriate) using the corresponding generative
model. The similarities between the identified synergies (activation coefficients
or delays) between these randomly generated sets were then computed. The
obtained similarities were then averaged over all 20 simulations, resulting in
a baseline value Σb for the corresponding similarity measure. The similarity
measures Σ resulting from the comparison between the identified and the sim-
ulated primitives were than transformed into a normalized similarity measure
according to the formula:
Σnorm =
Σ − Σb
1 − Σb
(18)
The normalized similarity measure Σnorm takes the value one for perfect
similarity, and it is zero if the similarity matches the average similarity between
two randomly generated data sets.
Statistical analysis
All tested measures were normally distributed according to a Chi-square goodness-
of-fit test. Student's t-test was used to test whether the reconstructions accu-
racies and the levels of similarities were statistically different from chance level.
Differences between more than two groups were statistically tested by two-way
ANOVAs (with Algorithm and Noise Level as factors), where appropriate. Post-
hoc analysis was conducted with Tukey-Kramer test, when necessary and ap-
propriate. As level of significance for the rejection of the null hypotheses in this
study we chose 5
Results
Comparisons of algorithm performance on simulated data
sets
To assess algorithm performance we simulated ground-truth data sets based on
the mixture models described in equations (1), (2), (3) ,(4) and (5). The aim
of our comparison was to show that the FADA algorithm can identify mixture
parameters at least as well as other well-known unsupervised learning methods.
Fig 2A shows the average performance (±SD) of the FADA and the fastICA al-
gorithm applied to mixtures of unconstrained synchronous synergies, similar to
17
the ones illustrated in Fig 1A. The bar plots indicate the reconstruction accuracy
measure R2 and the normalized similarity measures for the extracted synergies
Snorm, averaged across 20 data sets for five different levels of signal -- dependent
noise. Asterisks indicate significant differences according to post-hoc testing
between average values, obtained with different algorithms for the same level
of noise. The figure shows that both algorithms provide a good level of recon-
struction accuracy and resulted in an accurate estimation of the original model
parameters. Normalized accuracy measures were typically larger than 0.5, and
the similarity measures for the recovered primitives and weighting coefficients
were always significantly larger than chance level (t(19)>9.93, p<0.001). The
two-factor ANOVAs revealed a significant main effect for the factor Noise Level
for both the reconstruction accuracy and the identification of the primitives
(F(4,190) ≥ 5.08, p<0.001). We found a significant main effect for the factor
Algorithm for all tested parameters (F(1,190)≥11.92, p<0.001). The interaction
between the two factors was significant for and the similarity of the primitives
(F(4,190)≥5.08, p<0.001). The post-hoc analysis revealed that, for the same
level of noise, the fastICA and FADA algorithm did not provide significantly dif-
ferent identification performance, neither for the identification of the primitives
nor for the weighting coefficients (p>0.05), although fastICA provided always
significantly higher reconstruction accuracy (p<0.05).
Similarly to Fig 2A, Fig 2B depicts the identification performance of the
algorithms applied to mixtures based on model (1), but synthesized with non-
negative parameters. In this case, we compared the FADA algorithm to non-
negative matrix factorization (NMF), as the fastICA does not provide a way to
constrain parameters to be non-negative. Even in this case both algorithms pro-
vided a good fit of the data and very accurate estimates of the original primitives
and mixture weights. Not surprisingly, performance of both algorithms degraded
with increasing noise, more remarkably than in Fig 2A. ANOVAs indicated a
significant main effect of the factor Algorithm on the similarity of the weight-
ing coefficients (F(1,190)=23.14, p<0.001). Also the main effect of the factor
Noise Level was significant for both, R2 and levels of normalized similarities
(F(4,190)≥20.85, p<0.001). The interaction of both factors was significant only
for (F(4,190)=5.51,p<0.001). The post-hoc analysis showed that only in one
case (25% of noise) NMF performed better than the FADA algorithm in terms
of the identification of the weight coefficients (p=0.74). Differentl;y, NMF pro-
vided significantly lower reconstruction accuracy (p=0.01) for the highest level
of noise (35%). All average values in Fig 2B were significantly above chance
level (t(19)≥6.85, p<0.001).
Fig 3 shows the identification performance of the FADA, fastICA and NMF
algorithms applied to synchronous mixtures based on model (2). Fig 3A shows
qualitatively that, for the case of unconstrained mixtures, the level of recon-
struction accuracy and the level of similarity of the primitives were modulated
by the level of noise. In contrast, noise seems to have no significant effect on
the estimation of the weighting coefficients. ANOVAs confirmed a significant
main effect for the factor Noise level for the accuracy of reconstruction and the
similarity of the estimated primitives with the original for the reconstruction ac-
curacy of the data and the estimation of the weights (F(1,190)≥9.67, p<0.001).
The interaction between Algorithm and Noise Level was significant only for R2
(F(4,190)=5.12, p<0.001). Post-hoc testing revealed that the FADA algorithm
performed significantly worse than fastICA in approximating the noisiest data
18
B
40
20
0
40
20
0
40
20
0
40
20
0
Prim. 1
Prim. 2
Prim. 3
Prim. 4
1 2 3 4 5 6 7 8 9 10
1 2 3 4 5 6 7 8 9 10
dof
dof
D
Prim. 1
Prim. 2
Prim. 3
Prim. 4
0.5
0.4
0. 3
0. 2
0.1
0
0
20
40
60
80
100
Time samples
20
40
60
80
100
Time samples
A
50
0
−50
50
0
−50
50
0
−50
50
0
−50
C
1
0
−1
0
E
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
Prim. 1
Prim. 2
Prim. 3
Prim. 4
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
0
20 40 60 80 100
0
20 40 60 80 100
0
20 40 60 80 100
0
20 40 60 80 100
Time samples
Time samples
Time samples
Time samples
0.25
0.2
0.15
0.1
0.05
0
Figure 1: Examples of artificial primitives used for the generation of the ground-
truth data sets. (A) Unconstrained spatial primitives, associated with model (1).
(B) Non-negative spatial primitives, associated with model (1) and (5). (C) Un-
constrained temporal components used in models (2) and (3). (D) Non-negative
(EMG-like) temporal components associated with model (5). (E) Time-varying
non-negative primitives associated with model (4).
sets (p<0.001) and in terms of the identification of the weighting coefficients
(p=0.004) for one tested noise level (15%). For all other cases the identifi-
cation performance of the FADA and fastICA algorithm did not significantly
differ (p>0.05). Fig 3B shows the results of the comparison between the FADA
algorithm and NMF applied to non-negative data. The differences in perfor-
mance between the two methods for the same noise levels were very small.
Correspondingly, ANOVAs showed that the factor Algorithm had a significant
main effect only for the reconstruction accuracy R2 (F(1,190)=25,99,p<0.001),
while the factor Noise had significant main effects for all three tested measures
(F(4,190)≥17.38 p<0.001). Post-hoc testing revealed that the FADA algorithm
approximated the original data with significantly higher reconstruction accu-
racy than fastICA, only for the data were corrupted with the two highest levels
19
Variance
Primitives
Weights
1
*
*
FADA fastICA
*
*
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
0.8
0.6
0.4
0.2
0
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
0.8
0.6
0.4
0.2
0
A
i
d
e
n
a
r
t
s
n
o
c
n
U
B
e
v
i
t
a
g
e
n
-
n
o
N
i
e
r
u
t
x
m
s
u
o
n
o
r
h
c
n
y
s
i
e
r
u
t
x
m
s
u
o
n
o
r
h
c
n
y
s
2
R
2
R
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
Level of noise
Level of noise
Level of noise
Variance
FADA
NMF
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
Primitives
Weights
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
Level of noise
Level of noise
0
0.05 0.15 0.25 0.35
Level of noise
Figure 2: Spatial primitives. Identification performance (mean ± SD) of the
Fourier-based anechoic demixing algorithm (FADA), fast independent compo-
nent analysis (factICA), and non-negative matrix factorization (NMF) applied
to artificial data sets, which are corrupted by different amounts of noise, where
data was simulated by combining spatial primitives according to the generative
model (1). The number of DOFs in each spatial primitive was set to M = 10.
(A) Level of variance explained with the extracted parameters, and similarities
between original and identified primitives, and for the corresponding combina-
tion coefficients for unconstrained data (see Methods for details). (B) Level of
variance accounted for with the extracted parameters, similarities between origi-
nal and identified spatial primitives, and corresponding combination coefficients
for non-negative data.
of noise (p<0.05). Taken together, Figs 2 and 3 show that, when applied to
data based on synchronous models, FADA was in general able to provide iden-
tification performance comparable to those provided by the fastICA and NMF
algorithms for the model (2). In terms of identification of the actual parame-
ters, Tthe FADA algorithm had worse performance than the fastICA algorithm
only in two single cases concerning the identification of the weighsfor large noise
only for the case of mixtures based on the combination of unconstrained spatial
primitives. Interestingly, the variability associated with the similarities between
original and identified parameters is higher in Fig. 3A than in Fig 3B. This is
most probably due to an increase of regularization in the algorithms introduced
by the non-negativity constraints imposed on the model parameters. For the
lowest levels of noise (25% and 35%) NMF provided significantly higher recon-
struction accuracy (p<0.05). All measures in Fig 3 were significantly above
chance level (t(19)≥8.86, p<0.001).
20
A
d
e
n
i
a
r
t
s
n
o
c
n
U
B
e
v
i
t
a
g
e
n
-
n
o
N
e
r
u
t
x
i
m
s
u
o
n
o
r
h
c
n
y
s
i
e
r
u
t
x
m
s
u
o
n
o
r
h
c
n
y
s
2
R
2
R
1
0.8
0.6
0.4
0.2
0
1
0.8
0.6
0.4
0.2
0
Variance
FADA
fastICA
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
Primitives
Weights
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
Level of noise
Level of noise
Level of noise
Variance
FADA
NMF
*
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
Primitives
Weights
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
Level of noise
Level of noise
0
0.05 0.15 0.25 0.35
Level of noise
Figure 3: Instantaneous mixtures (without delays). Identification performance
of the FADA, factICA, and NMF algorithm applied to artificial data sets gener-
ated as instantaneous mixtures of temporal synergies, defined by the generative
model (2). (A) Level of variance explained by the extracted parameters, source
similarities between original and identified primitives, and similarities of cor-
responding combination coefficients for unconstrained data. (B) Same plot for
non-negative data.
Fig 4 shows a comparison between the identification performance of the
FADA algorithm and the alternative methods AnDem, SICA, AnNMF, and
sNMF for data generated with model (3) assuming unconstrained (Fig 4A) and
non-negative anechoic (Fig 4B) mixtures.
In addition to the similarity mea-
sures assessed before, we also quantified the similarity between original and
identified delays. For the unconstraint case (Fig 4A), FADA performed qualita-
tively better than both AnDem and SICA. For all levels of noise it provided a
higher level of data approximation quality and higher normalized similarities for
primitives and weighting coefficients. FADA and AnDem provided comparable
performance in the identification of the delays. When comparing FADA and
AndDem, ANOVAs revealed a significant main effect of the factor Algorithm
on R2 and on the approximation quality for the primitives and the weights
(F(4,190)≥210.5, p<0.001). The noise level affected only R2 and the similarity
of the primitives (F(4,190)≥12.85, p≪0.001). The interaction between noise
and the type of algorithm was significant for the estimation of the primitives
(F(4,190)≥5.85, p≪0.001). Post-hoc analysis revealed that FADA provided
significantly higher R2 values, as well as higher similarities of primitives and
weighting coefficients (p<0.05). Compared to the SICA algorithm, the FADA al-
gorithm showed higher approximation quality of the original data, for almost all
levels of noise, and higher similarities between original and identified primitives,
weights and delays. ANOVAs confirmed a significant a significant main effect
of the factor Algorithm for the estimation of all parameters (F(4,190)≥10.57,
21
p<0.01), while the influence of the factor Noise Level was sikgnificant only for
R2 and the similarities of the weights (F(4,190)≥3.85, p<0.01). Post-hoc testing
revealed that the FADA algorithm provided significantly higher R2 values than
the SICA algorithm (p<0.001). For the three highest level of noise the FADA al-
gorithm also the estimation of the time delays was more accurate (p<0.001). All
measures in Fig 4A were significantly above chance level (t(19)≥3.23, p<0.001).
Fig 4B shows the results for the performances of the FADA, AnNMF and
sNMF algorithms for data that are derived from non-negative anechoic mixtures.
Even in this case the FADA algorithm qualitatively provided, for all levels of
noise, higher values of R2 and similarity between original and identified primi-
tives in comparison to the AnNMF algorithm. ANOVAs revealed a significant
main effect of the factor Algorithm for R2 and the similarities of primitives and
delays with the generative model parameters (F(4,190)≥6.64, p<0.05). The
ANOVAs revealed also a significant main effect of the factor Noise Level for R2
and all other estimated model parameters (F(4,190)≥2.69, p<0.05). The inter-
action between the Noise level Algorithm was significant only for R2 and the
estimation accuracy of the delays (F(4,190)≥4.29, p<0.01). The post-hoc test-
ing revealed that the FADA algorithm provided significantly higher R2 values
than the AnNMF algorithm (p<0.001), higher similarity of the primitives for
noisy data sets (p<0.05) as well as higher similarities between original and iden-
tified delays (p<0.05) for three level of noise (0%, 5% and 15%). Comparing the
FADA algorithm and sNMF, ANOVAs resulted in a significant main effect of the
factor Algorithm for all measures, except for the similarity between original and
identified delays (F(4,190)≥10.34,p<0.01). The main effect of the factor Noise
Level was significant for all parameters (F(4,190)≥4.41,p<0.01). The interac-
tion between both factors was significant for R2 and the identification accuracy
of the time delays (F(4,190)≥2.96,p<0.05). Post-hoc testing showed that the
FADA algorithm always resulted in higher reconstruction accuracy and more ac-
curate estimates of the primitives than sNMF (p<0.05), with the only exception
of one level of noise (15%). All similarity measures in Fig 4B were significantly
above chance level (t(19)≥30.8, p<0.01), except for the reconstruction accuracy
provided by sNMF for the most noisy data sets (t(19)=0.25, p=0.80).
Fig 5 shows the ability of then FADA and the stNMF algorithm to identify
spatiotemporal synergies and the corresponding weight coefficients and delays
from simulated non-negative mixtures, derived from model (4) and mimicking
EMG-like features. A significant main effect of the factor Algorithm was found
for the reconstruction performance R2 and the accuracies of the estimation of
the weighting coefficients and delays (F(4,190)≥13.34, p<0.001), but not for
the accuracy of the reconstruction of the primitives (F(4,190)=0.4 p>0.05).
The factor Noise Level had a significant main effect for R2 and the accuracy
of the identified parameters (F(4,190)≥5.15, p<0.001). The interaction of the
factors Noise level and Algorithm was significant for R2 and the accuracy of
the estimation of delays (F(4,190)≥2.84, p<0.05). Post-hoc testing revealed
that the FADA algorithm resulted in significantly higher R2 values than stNMF
for all noise levels (p<0.001). Contrasting with this result, the FADA and
stNMF algorithm provided indistinguishable identification performance for all
parameters (always p>0.05), except for the identification of the delays when
data were corrupted with the highest level of noise (p=0.03). For all tested
noise levels and algorithms, R2 and the normalized similarities were always
significantly above chance level (t(19)≥11.78,p<¡0.001).
22
A
Unconstrained
anechoic mixture
B
Non-negative
anechoic mixture
1
** **
AnNMF
** ** **
sNMF
Variance
**
**
AnDem
*
*
**
FADA
SICA
* *
0
0.05 0.15 0.25 0.35
Primitives
*
*
*
*
*
0 0.05 0.15 0.25 0.35
Weights
*
*
*
*
*
0 0.05 0.15 0.25 0.35
Delays
*
*
*
1
0.8
0.6
0.4
0.2
0
0.5
0.4
0.3
0.2
0.1
0
0.4
0.3
0.2
0.1
0
0.3
0.2
0.1
0
2
R
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
Variance
FADA
0 0.05 0.15 0.25 0.35
Primitives
* **
*
** **
0 0.05 0.15 0.25 0.35
Weights
0 0.05 0.15 0.25 0.35
Delays
*
*
*
0.8
0.6
0.4
0.2
0
0.8
0.6
0.4
0.2
0
0.8
0.6
0.4
0.2
0
0.3
0.2
0.1
0
0 0.05 0.15 0.25 0.35
0 0.05 0.15 0.25 0.35
Level of noise
Level of noise
Figure 4: Mixtures with time delays. Identification performance of the FADA
algorithm, the anechoic demixing algorithm by Omlor and Giese (AnDem),
anechoic demixing with non-negativity constraints (AnNMF), shifted indepen-
dent component analysis (SICA) and shifted non-negative matrix factorization
(sNMF) algorithm applied to artificial data sets, obtained by combining tem-
poral synergies linearly with time shifts as described by model (3). (A) Level
of variance that explained with the extracted parameters, similarities between
original and identified primitives, and between corresponding combination co-
efficients and delays for unconstrained data. (B) Level of variance accounted
for, similarities between original and identified primitives, and corresponding
combination coefficients and delays for non-negative data.
The identification performance on simulated data based on the space-by-
time generative model (5) is summarized in Fig 6. Qualitatively, the FADA
23
1
*
Variance
FADA
*
*
stNMF
*
*
Primitives
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
2
R
0.8
0.6
0.4
0.2
0
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
Level of noise
Level of noise
Weights
Delays
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
0
0.05 0.15 0.25 0.35
Level of noise
Level of noise
Figure 5: Time-varying primitives. Performance of the FADA algorithm and the
identification of time-varying synergies (stNMF) for the learning of the param-
eters of model (4) from ground-truth data, obtained by combining non-negative
spatiotemporal synergies. The top-left panel shows the explained variance of
the data for the two algorithms as function of the noise level. In addition, the
average similarities between original and identified spatiotemporal primitives
(top-right panel) and the similarities of the corresponding weights and delays
(bottom-left and bottom-right panels) are shown.
algorithm always provided better data fitting, and more precise identification of
the original temporal sources, weights and delays than the sNM3F method. The
two methods identified the spatial components with similar levels of accuracy.
In the ANOVAs the main effect of the factor Algorithm was significant for all
parameters (F(4,190)≥10.72, p<0.001). The factor Noise Level had a significant
main effect for R2, as well as on the identification accuracy of weights and delays
(F(4,190)≥2.74, p<0.05). A significant interaction of the two factors was found
for R2 and for the normalized similarities associated with spatial primitives
and weighting coefficients (F(4,190)≥2.46, p<0.05). Post-hoc testing showed
that the FADA algorithm always provided significantly better reconstruction
of the data for all noisy data sets (p<0.001). Regarding the primitives, the
FADA algorithm provided more accurate estimates of the temporal primitives
for the most extreme levels of noise (0% and 35%, p<0.01). The algorithms
24
identified the spatial primitives equally well (p>0.05). The FADA algorithm
always outperformed the sNM3F method with respect to the identification of
the weighting coefficients and temporal delays (p<0.05). t-tests showed that
FADA and sNM3F always provided estimates of the parameters that were better
than chance level (t(19)≥3.68, p<0.01), with the only exception of the sNM3F
estimation of the weighting coefficients identified from data corrupted with 5%
of noise (t(19)=1.91, p=0.07).
2
R
1
0.8
0.6
0.4
0.2
0
Variance
*
*
FADA
sNM3F
*
*
0
0.05 0.15 0.25 0.35
Level of noise
Temporal
primitives
Spatial
primitives
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
*
*
0
0.05 0.15 0.25 0.35
Level of noise
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
0.8
0.6
0.4
0.2
0
0
0.05 0.15 0.25 0.35
Level of noise
Weights
Delays
*
*
*
*
*
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
1
*
*
*
*
*
0.8
0.6
0.4
0.2
0
y
t
i
r
a
l
i
i
m
s
d
e
z
i
l
a
m
r
o
N
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0
0.05 0.15 0.25 0.35
Level of noise
0
0.05 0.15 0.25 0.35
Level of noise
Figure 6: Space-by-time primitives. Performance of the FADA algorithm
and sampled-based non-negative matrix tri-factorization algorithm (sNM3F) to
identify the parameters of model (5) from ground-truth data obtained combin-
ing space-by-time synergies. The top-left graph shows the variance explained by
the two algorithms as a function of the level of noise. In addition, the average
similarities between original and identified primitives, and the similarities of the
corresponding weights and delays are shown in the other panels.
Comparisons of algorithm performance on real experimental data
In addition to the validation on synthesized data, we tested the FADA algo-
rithm also using previously published real experimental data sets. In addition,
we compared the primitives extracted by the FADA algorithm with those iden-
tified with other techniques. The first real experimental data set consists of
kinematic joint angle trajectories of the body joints of participants perform-
ing emotional walks. Trajectories represented a single gait cycle, resampled
with 100 time steps [20, 54]. We tested the FADA algorithm against the ane-
choic demixing algorithm developed by Omlor and Giese [19]. Fig 7A shows
25
that for both algorithms the explained variance as a function of the number
of primitives. (Such plots were also used in order to determine the number of
synergies, similarly to the scree plot in statistics [9].) The number of primitives
was identified from the R2 curve, determining the point where the slope levels
off considerably, forming an "elbow". For both methods in Fig 7A this point
is reached for N = 3, indicating that three anechoic components are sufficient
for a reasonable approximation of the experimental data set. Fig 7B shows also
the three primitives extracted by the two algorithms, which explain the largest
amount of variance of the data. The sources extracted with the FADA algorithm
are almost identical (S = 0.96) with those extracted with the other anechoic
demixing method.
The second data set comprises EMG signals assessed during point-to-point
arm reaching movements, recording from 16 different muscles [9]. We used the
FADA and the stNMF algorithm to extract time-varying synergies. The most-
left panel in Fig 8A shows the curve obtained with the FADA algorithm. In
this case, both methods identify N = 4 as levelling-off point of the R2 curves.
The other panels in the figure show the five time-varying synergies that were
identified by the FADA algorithm. Fig 8B shows the results obtained applying
the stNMF algorithm. Similarly to the results obtained for the FADA algorithm,
the curve levels off for N = 4. The synergies of Fig 8B matched closely those in
Fig 8A according to their (not normalized) level of similarity. Average similarity
across the four pairs of synergies was S = 0.97 ± 0.01, indicating that, also
on real EMG data, the identification performance of the FADA algorithm was
comparable to the one of the time-varying synergies algorithm.
Discussion In this article, we have developed a new mathematical framework
that unifies, for the first time, many different definitions of motor primitives.
We have described how the different kinds of primitives can be derived from
a more general mixture model, which is known as anechoic mixture, by addi-
tion of appropriate constraints. Starting from this mathematical framework,
we have implemented a new efficient unsupervised learning algorithm for the
identification of motor primitives that achieves an identification performance
typically at least as good as the other standard methods used to study mod-
ularity in human motor control. Such framework simplifies the comparisons
between the results from different studies using different generative models for
the definition of motor primitives. In addition, our general and robust algorithm
(Fourier Anechoic Demixing Algorithm, FADA) allows to extract motor primi-
tives according to specific generative models as special cases. To promote wide
adoption of the algorithm by researchers in motor control and neurophysiology,
we provide a downloadable implementation as a MATAB toolbox. Our quan-
titative validation indicated that this new algorithm performs typically equal
or better than the established methods for the extraction of primitives using
different underlying mathematical models. In the following, we discuss in detail
some computational aspects associated with the FADA algorithm and the other
unsupervised learning techniques compared with this algorithm. Moreover, tak-
ing a broader perspective, we discuss to what extent the different definitions of
motor primitives can be really linked to a single model.
Computational considerations regarding the FADA and other unsupervised
learning algorithms As in previous studies [56-60], we compared FADA with
other unsupervised learning techniques and assessed identification performance
on both ground-truth and experimental data sets. Differently from all the
26
Variance
A
2
R
1
0.9
0.8
FADA
AnDem
0.7
3
Number of primitives
Primitives
FADA
AnDem
S = 0.98
S = 0.95
B
0.2
0.1
0
-0.1
0.2
0
−0.2
−0.2
0.2
0
−0.2
0
20
40
60
80
100
S = 0.94
Time samples
Figure 7: Kinematic primitives extracted from experimental data.
(A) Ex-
plained variance as function of the number of extracted synergies. The blue
curve refers to the FADA algorithm and the red one to the AnDem algorithm.
(B) Temporal synergies identified by the two algorithms applied to kinematic
data collected from human participants executing emotional walks.
other techniques, which are based on a single generative model and sets of
constraints for the corresponding parameters (e.g. statistical independence or
non-negativity), the FADA algorithm allows to test different types of constraints
27
A
D
A
F
2
R
1
0.8
0.6
0.4
0.2
0
A
B
6
0
0
2
a
l
l
e
v
A
d
'
2
R
0.8
0.6
0.4
0.2
0
Total variance
1
Variance
TV Syn. 1
TV Syn. 2
TV Syn. 3
TV Syn. 4
TV Syn. 5
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
5
Time
Time
Time
Time
Time
S = 0.97
S = 0.96
S = 0.94
S = 0.92
S = 0.85
TV Syn. 1
TV Syn. 2
TV Syn. 3
TV Syn. 4
TV Syn. 5
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
m1
m2
m3
m4
m5
m6
m7
m8
m9
m10
m11
m12
m13
m14
m15
m16
0.25
0.2
0.15
0.1
0.05
0
-0.05
5
Time
Time
Time
Time
Time
Figure 8: EMG primitives extracted from experimental data. Spatiotemporal
synergies extracted with the FADA (A) and the stNMF algorithm (B) applied
to muscle activations collected during point-to point arm reaching movements.
Synergies are grouped according to their similarity. Most left panels show the
explained variance as function of the number of extracted synergies for both
algorithms. The dependence of this measure on the number of extracted syn-
ergies is consistent with previously reported data, indicating that five synergies
are sufficient to account for the significant part of the variability in the data.
within the same class of generative models. In this way the proposed algorithm
provides a unifying framework for the extraction of motor primitives. A key
element of the FADA algorithm is the mapping onto a finite Fourier basis. This
mapping reduces remarkably the number of identified parameters in comparison
with more general anechoic demixing methods [19, 39], but at the cost that only
band-limited data can be adequately modelled. For almost all data in motor
control (including at least kinematic or electromyographic data) the informative
part of the frequency spectrum typically never exceeds 100Hz after appropriate
processing. The reduction of the dimensionality of the parameter space results
in a more reliable and robust estimation of the primitives (even in presence of
substantial levels of noise) and in a lower probability of getting stuck in local
minima during the optimization. Consequently, the FADA algorithm performed
better than other methods for the identification of anechoic primitives (Fig 4)
and of temporal components associated with space-by-time decompositions (Fig
6). The only case where it showed lower performance was the comparison with
the fastICA algorithm for the identification of unconstrained spatial primitives
using generative model (1) (Fig 2).
In this case, due to the structure of the
data matrix, the data was not smooth along the dimension that is smoothed
by the FADA algorithm. In this case the inherent smoothness prior might thus
have reduced approximation quality. In spite of this problem, the reconstruc-
28
tion accuracy of the FADA algorithms was high also in this case and so were
the similarity scores for the reconstructed sources and weight matrices. For all
models including temporal delays the FADA algorithm outperformed the other
algorithms in terms of approximation quality, potentially due the reduced num-
ber of estimated parameters. For the anechoic unconstrained model (Figs 4),
all tested algorithms achieve relatively low values of similarities between orig-
inal and identified weighting coefficients and temporal delays, while (with the
exception of Fig 6) the similarity is significantly above chance level. Opposed
to the other tested algorithms in [25, 39, 48] we allowed for large delays and did
not restrict the tested delays to a small interval. A more detailed investigation,
which is beyond the scope of this paper, shows that the low reconstruction accu-
racy is caused by ambiguities in the estimation of delays and source functions,
especially for sources with higher fundamental frequencies. For real data the
FADA algorithm provided estimates for the primitives that were consistent with
those obtained with other traditional techniques. Also the estimated numbers
of primitives for a good approximation of the data matched between the FADA
and other established algorithms (cf. Figs 7 and 8). In addition, the functional
forms of the estimated primitives were very similar for the FADA and stNMF al-
gorithms. Despite its good identification performances and flexibility, the FADA
algorithm also suffers from a number of limitations. In situations where the fre-
quency spectrum of the real sources is not band-limited the truncated Fourier
approximation can decrease the identification performance, as likely in the case
of synchronous mixtures with few time samples (see Fig 2). Moreover, the iden-
tification of the parameters presently is realized by a gradient descent procedure.
There exist faster optimization methods that could be integrated in the future.
For the simulations carried out for this study the cross-correlation procedure
for the identification of the delays was implemented using entirely Matlab built-
in functions. Due to the modular architecture of FADA algorithm, it should
be easy to replace different steps by more speed-optimized implementations and
optimization methods. In this study we focused on the design of a highly flexible
rather than of a speed-optimized algorithm.
Different definitions of motor primitives and the problem of model selection
The central mathematical contribution of this article is that we derived how
different models of motor primitives relate mathematically to each other and
how they can be derived from the anechoic mixture model (6) by addition of
appropriate constraints. This raises the question how for a given data set the
most appropriate model structure can be found. As solutions for this model
selection problem classical criteria, such as the Akaike or the Bayesian Infor-
mation criterion (BIC) can be applied [61-62]. Alternatively, one can also use
Bayesian model selection. For this purpose, all tested models are embedded in a
joint model space, and one marginalizes the prediction error (evidence) using an
uninformative prior distribution over all possible model architectures [63]. This
procedure typically finds automatically a good balance between the goodness-
of-fit and simplicity of the model. An implementation of this idea for automatic
model selection has been proposed in [63], where the resulting non-Gaussian
distributions were approximated using a Laplace approximation in order to ob-
tain an analytically tractable selection criterion that allows to compare different
demixing models, including ones with time delays. The same type of procedure
also allows to make inferences about the most suitable smoothness priors for a
given data set.
29
Conclusion Inspired by the idea of Bernstein [64], experimental investiga-
tions in the last couple of decades have put forward the hypothesis that the
CNS might simplify the control of movement by relying on a modular orga-
nization of control [1, 2]. The modules (primitives) underlying such a control
architecture have been defined in multiple ways [65], and by applying a variety
of unsupervised learning algorithms to kinematic, dynamic and EMG data sets
(see for instance [56]). This heterogeneity of approaches makes the comparison
of results across different studies very difficult. We have developed a unifying
mathematical framework for the identification of motor primitives that links
these approaches, and we have implemented a unifying identification algorithm
(FADA) that implements many different methods as special cases as a free Mat-
lab toolbox (FADA) that is available online. We demonstrated that the FADA
algorithm typically shows identification performance that is competitive with
other classical unsupervised learning techniques. In some cases, it even outper-
forms these techniques for data from motor control, especially in presence of
noise. We hope that the new Matlab toolbox will help to establish more solid
links between different definitions of motor primitives, helping neuroscientists
with the comparison between different theoretical models and their data.
Acknowledgments
We thank Lars Omlor for his help during the first stage of this project, and
Dominik Endres for many helpful discussions. The research leading to these re-
sults has received funding from the European Union Horizon 2020 Programme
(H2020/2014-2020) under grant agreement n H2020 ICT-23-2014 /644727 Cogi-
mon (www.cogimon.eu). Martin Giese and Enrico Chiovetto have been also sup-
ported by the following funding sources: FP7-ICT-2013-10 (Koroibot); DFG GI
305/4-1, DFG GZ: KA 1258/15-1; BMBF, FKZ: 01GQ1002A, FP7-PEOPLE-
2011-ITN (Marie Curie): ABC PITN-GA-011-290011; FP7/2007-2013/604102
(HBP).
Bibliography
1. Bizzi E, Cheung VCK, d'Avella A, Saltiel P, Tresch M. Combining modules
for movement. Brain Research Reviews. 2008 Jan; 57(1): p. 125-133.
2. Flash T, Hochner B. Motor primitives in vertebrates and invertebrates. Curr
Opin Neurobiol. 2005 Dec; 15(6): p. 660-666.
3. Tresch MC, Saltiel P, Bizzi E. The construction of movement by the spinal
cord. Nat Neurosci. 1999 Feb; 2(2): p. 162-167.
4. Berret B, Bonnetblanc F, Papaxanthis C, Pozzo T. Modular control of point-
ing beyond arm's length. J Neurosci. 2009 Jan; 29(1): p. 191-205.
5. Kaminski TR. The coupling between upper and lower extremity synergies
during whole body reaching. Gait Posture. 2007 Jul; 26(2): p. 256-262.
6. Thomas JS, Corcos DM, Hasan Z. Kinematic and kinetic constraints on arm,
trunk, and leg segments in target-reaching movements. J Neurophysiol. 2005
Jan; 93(1): p. 352-364.
7. Mussa-Ivaldi FA, Giszter SF. Vector field approximation: a computational
paradigm for motor control and learning. Biol Cybern. 1992; 67(6): p. 491-500.
30
8. Chiovetto E, Berret B, Pozzo T. Tri-dimensional and triphasic muscle orga-
nization of whole-body pointing movements. Neuroscience. 2010 Nov; 170(4):
p. 1223-1238.
9. d'Avella A, Portone A, Fernandez L, Lacquaniti F. Control of fast-reaching
movements by muscle synergy combinations. J Neurosci. 2006 Jul; 26(30): p.
7791-7810.
10. Torres-Oviedo G, Macpherson JM, Ting LH. Muscle synergy organization
is robust across a variety of postural perturbations. J Neurophysiol. 2006 Sep;
96(3): p. 1530-1546.
11. Ivanenko YP, Poppele RE, Lacquaniti F. Five basic muscle activation pat-
terns account for muscle activity during human locomotion. J Physiol. 2004
Apr; 556(Pt 1): p. 267-282.
12. Ivanenko YP, Cappellini G, Dominici N, Poppele RE, Lacquaniti F. Coor-
dination of locomotion with voluntary movements in humans. J Neurosci. 2005
Aug; 25(31): p. 7238-7253.
13. Chiovetto E, Patan`e L, Pozzo T. Variant and invariant features characteriz-
ing natural and reverse whole-body pointing movements. Exp Brain Res. 2012
May; 218(3): p. 419-431.
14. Santello M, Flanders M, Soechting JF. Postural hand synergies for tool use.
J Neurosci. 1998 Dec; 18(23): p. 10105-10115.
15. d'Avella A, Saltiel P, Bizzi E. Combinations of muscle synergies in the
construction of a natural motor behavior. Nat Neurosci. 2003 Mar; 6(3): p.
300-308.
16. Torres-Oviedo G, Ting LH. Muscle synergies characterizing human postural
responses. J Neurophysiol. 2007 Oct; 98(4): p. 2144-2156.
17. Cheung VCK, d'Avella A, Tresch MC, Bizzi E. Central and sensory contri-
butions to the activation and organization of muscle synergies during natural
motor behaviors. J Neurosci. 2005 Jul; 25(27): p. 6419-6434.
18. Chiovetto E, Giese MA. Kinematics of the Coordination of Pointing during
Locomotion. PLoS One. 2013; 8(11): p. e79555.
19. Omlor L, Giese MA. Anechoic blind source separation using Wigner marginals.
Journal of Machine Learning Research. 2011 Mar; 12: p. 1111-1148.
20. Roether CL, Omlor L, Christensen A, Giese MA. Critical features for the
perception of emotion from gait. J Vis. 2009; 9(6): p. 15.1-1532.
21. Omlor L, Giese MA. Blind source separation for over-determined delayed
mixtures.
In: B. Scholkopf, J. Platt, and T. Hoffman, editors. Advances in
Neural Information Processing Systems 19. Cambridge, MA: MIT Press; 2007.
p. 1049-1056.
22. Ting LH, Macpherson JM. A limited set of muscle synergies for force control
during a postural task. J Neurophysiol. 2005 Jan; 93(1): p. 609-613.
23. Omlor L, Giese MA. Extraction of spatio-temporal primitives of emotional
body expressions. Neurocomputing, 2007; 70(10-12), p. 1938-1942.
24. d'Avella A, Tresch MC. Modularity in the motor system: decomposition of
muscle patterns as combinations of time-varying synergies. In: Solla SA, editor.
Advances in Neural Information Processing Systems 14. Cambridge, MA: MIT
Press; 2002. p. 141-148.
25. Delis I, Panzeri S, Pozzo T, Berret B. A unifying model of concurrent spa-
tial and temporal modularity in muscle activity. J Neurophysiol. 111(3), p.
675-693.
26. Bofill P. Underdetermined blind separation of delayed sound sources in the
31
frequency domain. Neurocomputing. 2003; 55(3): p. 627-641.
27. Emile B, Comon P. Estimation of time delays between unknown colored
signals. Signal Processing. 1998; 69(1): p. 93-100.
28. Torkkola K. Blind separation of delayed sources based on information max-
imization. In: IEEE International Conference on Acoustics, Speech, and Signal
Processing. 1996; 6, pp. 3509-3512.
29. Yilmaz O, Rickard S. Blind separation of speech mixtures via time-frequency
masking. IEEE transactions on Signal Processing. 2004 Jul;52(7):1830-47.
30. Choi S, Cichocki A, Park HM, Lee SY. Blind Source Separation and In-
dependent Component Analysis: A Review. Neural Information Processing -
Letters and Reviews. 2005; 6(1): p. 1-57.
31. Comon P, Jutten C, editors. Handbook of Blind Source Separation: Inde-
pendent component analysis and applications. Academic press; 2010.
32. Bell AJ, Sejnowski TJ. An information-maximization approach to blind
separation and blind deconvolution. Neural Comput. 1995 Nov; 7(6): p. 1129-
1159.
33. Be'ery E, Yeredor A. Blind separation of superimposed shifted images using
parameterized joint diagonalization.
IEEE Trans Image Process. 2008 Mar;
17(3): p. 340-353.
34. Lee DD, Seung HS. Algorithms for non-negative matrix factorization. In
Advances in neural information processing systems. 2001: p. 556-562.
35. O'Grady PD, Pearlmutter BA, Rickard ST. Survey of sparse and non-sparse
methods in source separation. International Journal of Imaging Systems and
Technology. 2005; 15(1): p. 18-33.
36. Arberet S, Gribonval R, Bimbot F. A robust method to count and locate
audio sources in a stereophonic linear instantaneous mixture. In Independent
Component Analysis and Blind Signal Separation. Springer; 2006. p. 536-543.
37. Cho N, Kuo CC. Underdetermined audio source separation from anechoic
mixtures with long time delay.
In Acoustics, Speech and Signal Processing,
2009. ICASSP 2009. IEEE International Conference on 2009 Apr 19 p. 1557-
1560.
38. Harshman RA, Hong S, Lundy ME. Shifted factor analysis -- Part I: Models
and properties. Journal of chemometrics. 2003; 17(7): p. 363-378.
39. Mørup M, Madsen KH, Hansen LK. Shifted independent component anal-
ysis. In: Independent Component Analysis and Signal Separation 2007 Jan 1
(pp. 89-96). Springer Berlin Heidelberg.
40. Højen-Sørensen PA, Winther O, Hansen LK. Mean-field approaches to in-
dependent component analysis. Neural Computation. 2002 Apr;14(4):889-918.
41. Lee DD, Seung HS. Learning the parts of objects by non-negative matrix
factorization. Nature. 1999 Oct; 401(6755): p. 788-791.
42. Swindlehurst A. Time Delay and Spatial Signature Estimation Using Known
Asynchronous Signals. IEEE Trans. Signal Processing. 1997; 46: p. 449-462.
43. d'Avella A, Bizzi E. Shared and specific muscle synergies in natural mo-
tor behaviors. Proceedings of the National Academy of Sciences of the United
States of America. 2005; 102(8): p. 3076-3081.
44. Mallat SG, Zhang Z. Matching pursuits with time-frequency dictionaries.
Signal Processing, IEEE Transactions on. 1993; 41(12): p. 3397-3415.
45. Hyvarinen A. Fast and robust fixed-point algorithms for independent com-
ponent analysis. IEEE Transactions on Neural Networks. 1999 May; 10(3): p.
626-634.
32
46. Hyvarinen A, Oja E. A Fast Fixed-Point Algorithm for Independent Com-
ponent Analysis. Neural Computation. 1997; 9(7): p. 1483-1492.
47. Mørup M, Madsen KH. SICA, 2007. 48. Mørup M, Madsen KH, Hansen
LK. Shifted non-negative matrix factorization. In: IEEE Workshop on Machine
Learning for Signal Processing. 2007; pp. 139-144.
49. Kass RE, Ventura V. A Spike-Train Probability Model. Neural Computa-
tion. 2001; 13(8): p. 1713-1720.
50. Harris CM, Wolpert DM. Signal-dependent noise determines motor plan-
ning. Nature. 1998 Aug; 394(6695): p. 780-784.
51. Schmidt RA, Zelaznik H, Hawkins B, Frank JS, Quinn JT. Motor-output
variability: a theory for the accuracy of rapid motor acts. Psychol Rev. 1979
Sep; 47(5): p. 415-451.
52. Sutton GG, Sykes K. The variation of hand tremor with force in healthy
subjects. J Physiol. 1967 Aug; 191(3): p. 699-711.
53. van RJ, Haggard P, Wolpert DM. The role of execution noise in movement
variability. J Neurophysiol. 2004 Feb; 91(2): p. 1050-1063.
54. Endres D, Chiovetto E, Giese MA. Model selection for the extraction of
movement primitives. Frontiers in Computational Neuroscience. 2013; 7(185).
55. Mardia KV, Kent JT, Bibby JM. Multivariate Analysis: Academic Press;
1979.
56. Tresch MC, Cheung VCK, d'Avella A. Matrix factorization algorithms for
the identification of muscle synergies: evaluation on simulated and experimental
data sets. J Neurophysiol. 2006; 95(4): p. 2199-2212.
57. Cashero Z, Anderson C. Comparison of EEG blind source separation tech-
niques to improve the classification of P300 trials. Conf Proc IEEE Eng Med
Biol Soc. 2011; 2011: p. 7183-7186.
58. Caulo M, Esposito R, Mantini D, Briganti C, Sestieri C, Mattei PA, et al.
Comparison of hypothesis- and a novel hybrid data/hypothesis-driven method
of functional MR imaging analysis in patients with brain gliomas. AJNR Am J
Neuroradiol. 2011; 32(6): p. 1056-1064.
59. Erhardt EB, Rachakonda S, Bedrick EJ, Allen EA, Adali T, Calhoun VD.
Comparison of multi-subject ICA methods for analysis of fMRI data. Hum
Brain Mapp. 2011 Dec; 32(12): p. 2075-2095.
60. Virtanen J, Noponen T, Merilainen P. Comparison of principal and inde-
pendent component analysis in removing extracerebral interference from near-
infrared spectroscopy signals. J Biomed Opt. 2009; 14(5): p. 054032.
61. Akaike H. A new look at the statistical model identification. Automatic
Control, IEEE Transactions on. 1974 dec; 19(6): p. 716-723. 62. Schwarz G.
Estimating the dimension of a model. The Annals of Statistics. 1978; 6: p.
461-464.
63. Bishop CM. Pattern Recognition and Machine Learning: Springer; 2007.
64. Bernstein N. The coordination and regulation of movements.: Oxford: Perg-
amon; 1967.
65. Chiovetto E, Berret B, Delis I, Panzeri S, Pozzo T. Investigating reduction
of dimensionality during single-joint elbow movements: a case study on muscle
synergies. Front Comput Neurosci. 2013; 7: p. 11.
33
|
1601.04253 | 1 | 1601 | 2016-01-17T05:56:02 | Topological mapping of space in bat hippocampus | [
"q-bio.NC"
] | Mammalian hippocampus plays a key role in spatial learning and memory, but the exact nature of the hippocampal representation of space is still being explored. Recently, there has been a fair amount of success in modeling hippocampal spatial maps in rats, assuming a topological perspective on spatial information processing. In this paper, we use the topological model to study $3D$ learning in bats, which produces several insights into neurophysiological mechanisms of the hippocampal spatial mapping. First, we demonstrate functional importance of the cell assemblies for producing accurate maps of the $3D$ environments. Second, the model suggests that the readout neurons in these cell assemblies should function as integrators of synaptic inputs, rather than detectors of place cells' coactivity and allows estimating the integration time window. Lastly, the model suggests that, in contrast with relatively slow moving rats, suppressing $\theta$-precession in bats improves the place cells capacity to encode spatial maps, which is consistent with the experimental observations. | q-bio.NC | q-bio | Topological mapping of space in bat hippocampus
Kentaro Hoffman1, Andrey Babichev1,2 and Yuri Dabaghian1,2
1Jan and Dan Duncan Neurological Research Institute,
2Department of Computational and Applied Mathematics, Rice University, Houston, TX 77005
Baylor College of Medicine, Houston, TX 77030,
(Dated: October 14, 2018)
Mammalian hippocampus plays a key role in spatial learning and memory, but the exact nature
of the hippocampal representation of space is still being explored. Recently, there has been a fair
amount of success in modeling hippocampal spatial maps in rats, assuming a topological perspective
on spatial information processing. In this paper, we use the topological model to study 3D learning
in bats, which produces several insights into neurophysiological mechanisms of the hippocampal
spatial mapping. First, we demonstrate functional importance of the cell assemblies for producing
accurate maps of the 3D environments. Second, the model suggests that the readout neurons in these
cell assemblies should function as integrators of synaptic inputs, rather than detectors of place cells'
coactivity and allows estimating the integration time window. Lastly, the model suggests that, in
contrast with relatively slow moving rats, suppressing θ-precession in bats improves the place cells
capacity to encode spatial maps, which is consistent with the experimental observations.
6
1
0
2
n
a
J
7
1
]
.
C
N
o
i
b
-
q
[
1
v
3
5
2
4
0
.
1
0
6
1
:
v
i
X
r
a
2
I.
INTRODUCTION
The principal neurons in mammals' hippocampus -- the place cells -- fire in discrete locations within the
environment -- their respective place fields [1, 2]. The spatial layout of the place fields -- the place field
map -- is commonly viewed as a representation of the animal's cognitive map of space, although the exact
link between them remains unclear [3]. Experiments in "morphing" 2D environments demonstrate that the
place field maps recorded in rats are "flexible": as the environment is deformed, the place fields change
their shapes, sizes and locations, while preserving their relative positions [4 -- 7], which suggests that the
sequential order of place cells' (co)activity induced by the animal's moves through the place fields remains
invariant within a certain range of geometric transformations. Moreover, the temporal patterns of place cell
coactivity is preserved [8, 9], which implies that the place cells' spiking encodes a coarse framework of
qualitative spatiotemporal relationships, i.e., that the hippocampal map is topological in nature [9 -- 12].
Recently, there appeared a few topological models of the hippocampal map [13 -- 15]. In particular, the
approach proposed in [16, 17] allows integrating the local spatial information provided by the individual
place cells into a large-scale topological representation of the environment. The idea of such integration is
based on the Cech's theorem, according to which the pattern of overlaps between regular spatial domains --
the regions -- U1, U2, UN, covering a space X, encodes the topological structure of X [18]. Specifically, the
covering regions are used to construct the nerve of the cover, N -- a simplicial complex whose 0D vertices
correspond to the covering regions Ui, the 1D links -- to pairwise overlaps Ui ∩ U j, the 2D facets -- to triple
overlaps Ui ∩ Ui ∩ Ui and so forth. The Cech's theorem ascertains that if all of the overlaps between the Uis
are contractible in X, then N is topologically equivalent to X. An implication is that, if the place fields cover
the environment sufficiently densely, then their overlaps should encode its topology. Moreover, since these
overlaps are represented by the place cells' coactivities, a similar construction can be carried out in temporal
domain [13, 16, 17]: if the animal enters a location in which several place fields overlap, the corresponding
place cells produce (with a certain probability) temporally overlapping spike trains, which can be received
and processed by the downstream brain areas (Suppl. Fig. 1). In other words, place cells' spiking encodes a
"temporal" analogue of N -- a coactivity complex, T , the vertices of which correspond to active place cells
ci, 1D links -- to pairs of coactive cells [ci, c j], 2D facets -- to coactive triples [ci, c j, ck], etc.
By construction, the coactivity complex incorporates, at any given moment of time t, the entire pool of
coactivities produced by the place cell ensemble. Hence, it provides a framework for representing spatial
information encoded by the place cells. For example, a sequence of the place cell combinations ignited
along a particular path γ corresponds to a sequence of "coactivity simplexes" -- a simplicial path Γ that
represents γ in T . It was shown in [16, 17] that if the coactivity complex is sufficiently large (i.e., includes
a sufficient number of the coactivity events) and if the parameters of the place cell activity fall into the
biological range, then T correctly captures the topology of the physical environment, E. For example, a
non-contractible simplicial path corresponds to a class of the physical paths that enclose unreachable or yet
unexplored parts of the environment. Similarly, two topologically equivalent simplicial paths Γ1 ∼ Γ2 in
T represent physical paths γ1 and γ2 that can be deformed into one another. However, such information
is not produced immediately: as the animal begins to navigate a new environment, the coactivity complex
T is small and provides an incomplete and fragmented representation of space; moreover, an undeveloped
complex T typically consists of several disjoint components, each one of which may contain gaps and holes
that do not correspond to physically inaccessible spatial domains. As the animal continues to navigate, more
combinations of coactive place cells are detected, and the "spurious" gaps disappear, yielding a coactivity
complex that faithfully represents the topological structure of the environment. Thus, the progress of spatial
learning can be quantified in terms of the evolving coactivity complex's structure, e.g., a list of the holes that
it contains at a given moment of time t, or a list of surfaces -- loops -- that encapsulate these holes [16, 18].
Both the holes and the loops are counted up to topological equivalence. For example, if two such loops can
be deformed into one another, then they are counted as different shapes of the same topological object --
a topological loop. The number of inequivalent topological loops of a dimensionality n is known as the
3
FIG. 1: Topological map of the bat's environment. (A) Timelines of 1D loops (colored horizontal lines) encoded
in a coactivity complex T . For as long as a given loop 1D persists, it indicates a noncontractible hole in T . In this
case, the 1D topological loops begin to appear (and hence the holes start to form) in about 2 minutes, and disappear
in about 17 minutes, when all the holes in T close up. (B) A view into a simulated 3D environment (290 × 280 × 270
cm, sizes taken from [20, 21]) that contains one vertical column and two protrusions -- a stalactite hanging over 50
cm from the ceiling, and a 50 cm tall stalagmite. A portion of the simulated trajectory is shown by dashed line. (C)
Simulated spikes produced by a virtual bat form 3D spatial clusters -- the place fields. Spikes produced by different
cells are marked by different colors.
nth Betti number, bn, and the list of all Betti numbers, (b0, b1, b2, ...), provides a convenient "barcode" of
the space's topological structure [19]. Methods of the Persistent Homology theory allow identifying the
topological loops in the hippocampal representation of the environment and comparing the resulting Betti
numbers bn(T ) with the Betti numbers of the underlying environment, bn(E). In addition, it is possible to
compute the minimal time, Tmin, after which the low-dimensional topological structure of T matches the
topology of the environment, which serves as a theoretical estimate of the time required to learn a given
space (Figure 1C).
Below we apply this model to investigate spatial learning in bats that learn 3D representations of their
environments [20, 21], which produces a number of neurophysiological insights. Specifically, we simulated
place cell spiking activity in a bat navigating a small cave with one vertical column and two vertical protru-
sions, representing a stalactite and a stalagmite (Figure 1A). The topological barcode of this environment
is the same as in the 2D open field studied in [16, 17]: b0 = 1 (the cave is connected), b1 = 1 (one 1D
topological loop represents paths encircling the column, but not the contractible stalactite and stalagmite)
and bn>1 = 0 (no loops in higher dimensions). However, the increased complexity of the space mapping
task elucidates several neurophysiological properties of the hippocampal network that were not explicitly
addressed in the previous models.
II. RESULTS
Simplicial coactivity complexes. Our first observation was that, in contrast with the 2D environment in
which the correct topological signature emerged in a matter of minutes, the simulated place cell coactivity
in 3D failed to represent the cave's topology. In particular, the first Betti number of the coactivity complex
often vanished, b1(T ) = 0, i.e., the place cells did not capture the most salient feature of the navigated
space -- the central column, although the bat's trajectory encircled it multiple times.
We reasoned that the discrepancy between b1(T ) and b1(E) was due to the relatively high speed of the
bat's movements, which caused a mismatch between the temporal pattern of place cell coactivities and the
spatial pattern of the place field overlaps. Indeed, if the bat is flying at the speed v, then place cells cofire
within a coactivity time window w, even if their place fields are up to d ≈ vw apart. If detected by the
downstream neurons, these coactivities may lead to an inaccurate representation of space by linking the
place cell representations of the physically separated spatial domains. In our simulations, the speed v of the
4
FIG. 2: Learning a topological map by detecting place cell coactivity. (A) A simplicial coactivity complex, built
by detecting high order coactivity events, produces large numbers of persisting topological loops in low dimensions.
In particular, there are five persistent 0D topological loops (top panel), which implies that hippocampal map of the
environment is fragmented in pieces. In addition, there are many persistent loops in 1D (middle panel) which implies
that the coactivity complex encodes many noncontractible paths, whereas there is only one class of noncontractible
loops in the physical environment. (B) The number of topological loops in the clique coactivity complex is much
smaller. In fact, after the Tmin = 28 minutes (red vertical dashed line), most topological loops disappear, leaving only
one topological loop in 0D (which correctly represents the caves connectivity) and one loop in 1D, which represents
physical paths circling around the central column. The loops in higher dimensions contract (only the 2D loops are
shown, bottom panel), and thus the correct topological barcode (b0 = 1, b1 = 1, bn>1 = 0) emerges. (C) A particular
connection may be identified in two ways: either as a clique (completing at tc, the moment when all the pairs have
been accumulated), or simultaneously, as a simplex all cells of which are observed at the moment ts. The figure shows
the probability distribution of the differences between the first appearance times, ∆ = tc − ts. Since all ∆s are positive,
cliques always appear sooner than simplexes.
bat reached at times 2 m/sec, while the place cell inputs were integrated over the coactivity window w = 0.25
sec [17, 22, 23]. As a result, place cell coactivities could falsely encode overlaps between place fields that
are physically up to 50 cm apart. In particular, place fields across the column may appear "connected," in
which case the column will fail to produce a hole in T .
Cell assembly constraints. The result outlined above suggests that the pool of place cell coactivities
must be additionally constrained to prevent the appearances of "faulty" connections, which, in fact, appeals
to well-known neurophysiological phenomenon observed in the hippocampal network. Electrophysiological
studies suggest that certain select groups of place cells form functionally interconnected assemblies, which
drive their respective "reader-classifier" or "readout" neurons in the downstream networks [25, 26]. Spiking
of the readout neurons "actualizes" the information provided by the place cells: if the coactivity of a place
cell assembly does not elicit a response of a readout neuron, the corresponding connectivity information
does not contribute to the hippocampal map [26]. Thus, a particular selection of the admissible place cell
combinations is determined not only by the place coactivities but also by the architecture of the hippocampal
cell assembly network, which constrains the pool of place cell coactivities.
It is widely believed that the readout neurons function as "all or none" coactivity detectors, i.e., that they
respond to nearly simultaneous activity of the presynaptic cells [26]. Curiously, the topological approach
based on the Cech's theorem corroborates with this point of view: the fact that the nerve complex N is
derived from the spatial overlaps between the regions suggests that a readout neuron should identify these
overlaps by detecting nearly simultaneous place cell activity at the times when the animal visits them.
These considerations suggest a simple phenomenological solution to the constraint selection task: to
exclude the "spurious" connections, we built the coactivity complex T , using only the coactivities produced
by place cells with the overlapping place fields. However, our subsequent simulations revealed that the
cell assembly constraint is too restrictive, because the resulting coactivity complex broke into pieces (on
average, b0 = 2.3) and produced a large number of topological loops in 1D and 2D (on average, b1 = 36
and b2 = 410, Figure 2A).
5
From a biological perspective, this implies that the hippocampal map of the environment remained
fragmented and riddled with holes, most of which do not correspond to the actual topological obstacles
encountered by the bat. In other words, the simulated cell assembly network, wired to detect place cells
coactivities, fails to learn the correct path connectivity of space, which suggests that the system may employ
an alternative mechanism of reading out the coactivity information.
Clique coactivity complexes. From a mathematical perspective, the higher order overlaps between
regions can be not only empirically detected, but also derived from the lower order overlaps. According
to Helly's theorem [27], a collection of N > D + 1 convex regions in D-dimensional Euclidean space
will necessarily have a nonempty common intersection if every D + 1 of them intersect. For example, a
collection of N ≥ 4 planar regions has a common intersection if any three of them overlap and a collection
of N ≥ 5 regions have a common intersection in 3D if any four of them overlap (Suppl. Fig. 2). From the
perspective of the Cech's theorem, this suggests that high-dimensional simplexes in N can be deduced from
their low-dimensional simplexes, which opens new possibilities for constructing coactivity complexes.
Numerical simulations demonstrate that, in fact, the approach of Helly's theorem can be extended be-
yond its strict mathematical validity. For example, high order overlaps between place fields in 2D envi-
ronments can be reliably identified by detecting graph-theoretic cliques of pairwise overlaps between them
(Suppl. Fig. 2A,B).
Moreover, this information is captured by the place cell coactivity: in the case of the triple place field
overlaps, the three pair coactivities, [ci, c j], [c j, ck] and [ci, ck], mark an existing triple overlap of the cor-
responding place fields in over 90% of cases, and for the higher order overlaps this percent is even higher.
This implies that most cliques in the coactivity graph G, defined by the connectivity matrix
1 if cells ci, c jare coactive and their respective place fields overlap
0 otherwise.
Ci j =
(1)
correspond to the simplexes of the restricted coactivity complex T , i.e., that the topological structure of the
"clique complex," Tcq, can approximate the topological structure of T .
In [28] we demonstrated that in 2D, the clique coactivity complexes do not only capture the topology
of the environment, but often perform better than simplicial coactivity complexes. Similar effects are also
observed in the 3D case. In our simulations, the number of topological loops in Tcq was much smaller than
the number of loops in T . Moreover, Tcq produced the correct topological signature on average in about
Tmin = 28 minutes -- a biologically plausible period of time (Figure 2B).
The success of the clique coactivity complex has a simple intuitive explanation. First, the pairwise
coactivities of the place cells are produced when the animal enters the domains where at least two place
fields overlap. Since these domains are bigger than the domains of the higher order overlaps (Suppl. Fig.
2C), the pairwise coactivities are produced and detected more reliably than the high-order coactivities.
Second, the process of detecting the pairwise coactivities is spread over time. For example, in order to
identify a third order coactivity clique, [ci, c j, ck], one can first detect the coactive pair [ci, c j], then the
pair [c j, ck], and then [ci, ck], whereas in order to encode a coactivity simplex [ci, c j, ck], all three cells must
become active within the same coactivity window w. As shown on Figure 2C, the higher order combinations
take longer to appear than the matching collection of pairs -- in fact, most coactivity cliques that can be
"assembled" over an extended observation process, are never observed as simultaneous coactivity events
(Suppl. Fig. 3). This implies that the clique coactivity complex is typically bigger and forms faster, and
hence the transient topological loops in Tcq contract sooner.
Integration times. From a physiological perspective, the qualitative difference between the results
produced by the "simplicial" and the "clique" approaches to counting the place cell coactivities suggests
that the readout networks should build their spiking responses not by detecting rare high-order events, but
by integrating low order coactivity inputs. The physiological mechanism for such integration may be based
on complex subthreshold summation of the action potentials impinging on the dendritic tree of the readout
6
FIG. 3: Timelines of the topological loops in the restricted clique simplicial complexes, obtained for different
dendritic integration times. (A) For = w (in which case the clique complex reduces to the simplicial coactivity
complex) spatial learning fails due to numerous spurious topological loops that persist indefinitely in 0D, 1D and in
2D. However, as the dendritic integration time increases to = 4 min, the number of topological loops drops, and
only occasional spurious loops remain for = 8 minutes. At = 12 minutes the last 2D loops disappear, and only
one persistent loop remains in 1D and 0D after 28 mins, indicating emergence of the correct topological signature.
(B) The distribution of the time intervals between the pairs impinging on the readout neurons (see formula (2)). (C)
Distribution of the same intervals, averaged for each clique.
neuron. Once a sufficient number of low order coactivity inputs has been received, the readout neuron may
produce an action potential, thus actualizing the information about the nth order connection between the
regions encoded by the place cells.
Although our modeling approach does not directly address spike integration mechanisms, it allows opti-
mizing parameters in a particular readout algorithm, based on the frequency of the place cell (co)activations
produced by the animal's movements through the environment. For example, the model predicts that the
coactivity window used by the coincidence detector readout neurons in 2D environments should be about
w ∼ 200 msec wide (smaller ws lead to a rapid increase of the learning times and larger ws lead to instability
of learning [17]), which falls into the physiological range of values [22, 23].
What would be then the model's estimate of the time required by the "integrator" neurons to accumulate
the place cell coactivities -- the clique integration time, -- to produce a complex that reliably represents
space? On the one hand, longer clique integration windows allow collecting more coactivities for assem-
bling the cliques, which improves the structure of the coactivity complex (compare Figure 2B and Fig-
ure 2A). On the other hand, the larger the is, the longer the intervals between the consecutive coactivity
inputs τ can be, so the information about the presynaptic inputs has to be retained for longer. In contrast, the
smaller is , the tighter the coactivities are "packed" in the integration window. In cases when ≈ w, the
7
FIG. 4: Suppression of the θ-phase precession improves learning. (A) If the θ-precession in the place cells produc-
ing the simplicial coactivity complex is supressed, then the number of the topological loops in it drops significanlty in
all dimensions (compare to Figure 2A,B in which θ-precession is present). (B) Similar effect is observed in the clique
coactivity complex, if the place cells do not θ-precess. Moreover the learning time Tmin = 18 minutes is smaller than
in the θ-precessing place cell ensembles. (C) The average time it took the coactivity complex, T to converge to a
stable signature, as a function of the bat's mean speed. The times computed for the θ-driven cell assemblies are shown
in red and those without θ-precession are colored blue, statistically significant differences are marked by the asterisks
over the error bars.
integrator neuron turns into a coincidence detector, and the clique coactivity complex reduces to the simpli-
cial coactivity complex (i.e., when ≈ w, the results shown on Figure 2B reduce to Figure 2A). By testing
several values of , we found that the place cell map converges if (cid:38) 8 minutes (Figure 3A), making
this value the model's estimate of the clique integration window. Physiologically, this value corresponds
naturally to the timescale of working memory (the memory functions responsible for the transient holding,
processing and retaining the partial results of learning and memory updating [24]).
We evaluated the combined statistics of the time intervals, τ, between the pairs of action potentials re-
ceived by the readout neurons within an integration window, across the cell assembly network. As shown
on (Figure 3B), the distribution of these intervals consists of two parts: a sharp delta-like peak, ∆(τ), con-
centrated at small values of τ, and an exponential tail,
P(τ) = C1∆(τ) + C2e−µτ.
(2)
The exponential rate µ is higher for short s and decreases towards a stable value of about µ ≈ 6.6 minutes
as grows. The fact that building a typical clique requires detecting several coactivities (e.g, a fourth order
clique contains six parwise connections) while mean interval µ between coactivities is comparable to the
size of the integration window , suggests that most coactive inputs are detected "on the spot," within a
small number of consecutive coactivity windows, at the time when the bat crosses a domain where several
place fields overlap. This contribution is described by the first part of the distribution, ∆(τ), whereas the
second part -- the exponential tail -- represents the connections accumulated over time.
To better understand the statistics of the inputs provided by the cells within the individual cell assemblies
we computed, for each , the mean inter-activity interval, ¯τσ, for each clique σ and studied the statistics of
these values. The results shown on (Figure 3C) suggest that the mean inter-activity intervals are distributed
much more uniformly, i.e., the population of cell assemblies in which the coactivity integration occurs
rapidly is about the same as the population of assemblies that take intermediate or longer times to integrate
inputs.
Theta precession is a coupling between the timing of neuronal spikes and the phase of the theta (θ, 4-12
Hz) component of the extracellular field potential, observed in the rat's hippocampus. As the rat proceeds
through a place field, the corresponding place cell spikes at progressively earlier phases of θ-oscillations
[29, 30]. This phenomenon produces a strong effect on rats' spatial learning [31]: suppressing θ-precession
with cannabinoids correlates with reduced ability to solve spatial tasks and poorer spatial memory [32, 33].
8
In contrast, θ-precession in bats is not clearly manifested: according to [34], less than 4% of the bat's place
cells exhibit significant θ-modulated firing.
From a biological perspective, the functional importance of θ-precession for rats navigating 2D environ-
ments and lack of thereof in bats navigating 3D spaces might suggest that in the latter case, θ-precession
may not produce a similar positive effect on spatial learning. To test this hypothesis, we numerically sup-
pressed the θ-precession in the simulated place cells, which lead to a noticeable decrease of the number of
spurious topological loops both in the simplicial and in the clique coactivity complexes (Figure 4). Second,
the learning time produced by the clique complex, Tmin = 18 minutes, became shorter than in the θ-driven
coactivity complex.
To explain these results, we reasoned as follows. One can view the effect of θ-precession from two
perspectives: on the one hand, it synchronizes place cells' spiking and hence increases their coactivity rate
[29, 30], which speeds up map learning. On the other hand, the fact that the place cells can spike only at
specific phases of θ can be viewed as a constraint that reduces the probability of cells' spiking at every given
moment. In [17] we demonstrated that in relatively slow moving rats, the probability of producing spikes
during a typical passage through the place field is sufficiently high despite the θ-constraint, so that the main
effect of θ-oscillations is spike-synchronization. In contrast, the results of Figure 4 suggest that in bats, the
place cells have time to produce only a few spikes during a high speed flight through the 3D place fields, so
that the θ-precession only further reduces the cells' chances of being (co)active.
We tested these results by simulating bat's movements at different mean speeds: (25, 50, 75, 100, 125,
and 150 cm/sec). For each case, 10 simulations were conducted using θ-precessing place cells and 10
simulations with the θ- precession turned off. The results on the Figure 4C show that at no point θ-on
ensembles learned faster than the θ-off ensembles: average time until a stable signature was reached in the
θ-off case was about 30% less than the time required by the θ-modulated place cell ensembles.
III. DISCUSSION
The topological model of the place cell map of space proposed in [16, 17] provides a framework for
bringing together different scales of spatial learning: the macroscale, i.e., the topological and geometrical
parameters of the encoded loops, paths, holes, and the microscale, i.e., the parameters of the hippocampal
network's neurophysiology. Applying this model to describe a bat navigating a 3D space reveals several in-
terrelationships between these scales that are implicit in the 2D case, but which have a number of important
physiological implications.
First, the failure of the naıve counting of the place cell cofirings indicates a functional necessity of
"thinning out" the pool of coactivities using a cell assembly network. Qualitatively, this result is based
on simple observation: if the readout neurons respond unrestrictedly to the place cell coactivities then the
animal's rapid moves across the environment will necessarily encode false connections between remote
spatial regions, which will result in an incorrect map of space. Second, a striking difference between the
simplicial and the clique complexes' ability to capture the topology of the 3D environment suggests that the
information about the high order spatial relationships should be deduced from low order coactivity events,
rather than instantaneously detected. The qualitative reasons for this effect are also transparent: the high
order coactivity events, represented by the simplexes, are rarer and harder to detect than the matching col-
lections of pairwise coactivities, represented by the cliques. Third, the amount of time over which readout
neurons integrate inputs should be longer than observed in previous computational or in vitro studies, in
which neurons were studied individually and independently from the task solved by the net neuronal en-
semble [35 -- 40]. In contrast, our approach provides a basic contextual description of neuronal activity: the
frequency with which the place cells' action potentials impinge on the readout neuron depends on the fre-
quency of the animal's visits to specific spatial locations. This frequency would not change significantly if
the number of simulated place cells would increase, which suggests that our estimates of the "clique inte-
9
gration windows" may give a correct qualitative estimate of the physiological dendritic integration times
required by the downstream networks.
Lastly, the model provides a functional insight into why the θ-precession is physiologically suppressed
in bats' hippocampi. Thus, our model provides an example of a "top-down" approach, in which neurophys-
iological properties of the network are deduced from the task solved by the network.
IV. METHODS
The navigational parameters: the mean speed of the bat (vmean = 66 cm/s, vmax = 150 cm/s), and the
dimensions of the environment (290 × 280 × 270 cm) were taken from [21]. For increased reliability of the
results, the duration of the navigation session (120 min) was longer than reported in [21].
Place fields. The centers of the place fields, rc = (xc, yc, zc), were uniformly distributed over the en-
vironment, to simulate non-preferential representation of locations. Given the typical sizes of the place
fields (Lc = 95 cm), the size of the simulated place cell ensemble was chosen to create enough 2D and 3D
simplexes in the coactivity complex, sufficient to build a 3D map. Computationally, we could afford about
seven place fields per dimension, i.e., Nc = 343 place cells total, which corresponds to about Nc = 50 cells
in a the planar 1 × 1 m environment studied in [16, 17]. The Poisson spiking rate of a place cell c at a point
r(t) = (x(t), y(t), z(t)) is
λc(r) = fce
− (r−rc)2
2s2
c
.
(3)
The peak firing rates fc of different cells were log-normally distributed around the typical experimental
value, f = 8 Hz and sc = Lc/3 [21].
θ-phase precession. As the animal enters the field of a cell c, and moves over a distance l towards the
center, the preferred spiking phase is ϕθ,c ≈ 2π(1 − l/Lc) [30, 41]. To simulate the coupling between the
firing rate and the θ-phase, we modulated the original Gaussian firing rate by a θ-factor Λθ,c(ϕ),
− (ϕ−ϕθ,c)2
2ε2
c
.
Λθ,c = e
(4)
The width, ε, of the Gaussian was defined in [17], as the ratio of the mean distance that the animal travels
during one θ-cycle to the size of the place field, ε = 2πv/Lωθ, where v is the rat's speed and ωθ/2π is the
frequency of the θ-signal.
Cell assembly constraint. The functional connectivity between place cells in the hippocampal network
is described by a graph G, the vertices of which correspond to the active place cells and links to coactive
pairs, restricted by the constraints (see formula (1)). The connectivity matrix Ai j of G is defined as follows.
First, we define the relational matrix Ci j: Ci j = 1 if the cells ci and c j exhibit coactivity during the navigation
period and Ci j = 0 otherwise. The place field map's connectivity matrix, Pi j, is based on the place fields'
spatial overlap: Pi j = 1 if the distance between the place field centers is smaller than the sum of their half-
sizes, d(ci, c j) ≤ (Li +L j)/2, and Pi j = 0 otherwise. In principle, the matrix Pi j can be deduced directly from
the temporal pattern of place cell coactivity [28]; however we used the place field information to simplify
our analyses. To constrain the pool of coactivities by the place field map structure, we used the Hadamard
product, Ai j = Ci jPi j. The simplexes of the clique coactivity complex, Tcq, correspond to "cliques" of the
relational graph G.
Coactivity. A group of cells c1, c2, ..., cn, counts as coactive, if each one of them fires at least two spikes
within the coactivity window w ≈ 250 msec, i.e., in less than two θ-periods. In [17] we demonstrated this
value of w is optimal both in presence and in absence of the θ-precession.
V. ACKNOWLEDGMENTS
The work was supported in part by Houston Bioinformatics Endowment Fund, the W. M. Keck Founda-
tion grant for pioneering research and by the NSF 1422438 grant.
10
VI. REFERENCES
11
[1] O'Keefe J, Dostrovsky J (1971) The hippocampus as a spatial map. Preliminary evidence from unit activity in
the freely-moving rat, Brain Res, 34, pp. 171-175.
[2] Best PJ, White AM, Minai A (2001) Spatial processing in the brain: the activity of hippocampal place cells,
[3] O'Keefe J, Nadel L (1978) The hippocampus as a cognitive map, New York: Clarendon Press; Oxford University
Annu Rev Neurosci., 24, pp. 459-486.
Press. xiv, 570 pp.
[4] Gothard KM, Skaggs WE, McNaughton BL (1996) Dynamics of mismatch correction in the hippocampal en-
semble code for space: interaction between path integration and environmental cues, J Neurosci., 16, pp. 8027-
8040.
[5] Leutgeb JK, Leutgeb S, Treves A, Meyer R, Barnes CA, et al. (2005) Progressive transformation of hippocampal
neuronal representations in "morphed" environments, Neuron, 48, pp. 345-358.
[6] Wills TJ, Lever C, Cacucci F, Burgess N, O'Keefe J (2005) Attractor dynamics in the hippocampal representation
of the local environment, Science, 308, pp. 873-876.
[7] Touretzky DS, Weisman WE, Fuhs MC, Skaggs WE, Fenton AA, et al. (2005) Deforming the hippocampal map,
Hippocampus, 15, pp. 41-55.
[8] Diba K, Buzsaki G (2008) Hippocampal network dynamics constrain the time lag between pyramidal cells
across modified environments, J Neurosci., 28, pp. 13448-13456.
[9] Dabaghian Y, Brandt VL, Frank LM (2014) Reconceiving the hippocampal map as a topological template, eLife
[10] Alvernhe A, Sargolini F, Poucet B (2012) Rats build and update topological representations through exploration,
10.7554/eLife.03476.
Anim. Cogn., 15, pp. 359-368.
coding, Behav. Processes, 53, pp. 155-162.
ment, J Neurosci., 34, pp. 6459-6469.
[11] Poucet B, Herrmann T (2001) Exploratory patterns of rats on a complex maze provide evidence for topological
[12] , Wu X, Foster DJ (2014) Hippocampal Replay Captures the Unique Topological Structure of a Novel Environ-
[13] Curto C, Itskov V (2008) Cell groups reveal structure of stimulus space, PLoS Comput. Biol., 4, e1000205.
[14] Chen Z, Gomperts SN, Yamamoto J, Wilson MA (2014) Neural representation of spatial topology in the rodent
hippocampus, Neural Comput., 26, pp. 1-39.
[15] Chen Z, Kloosterman F, Brown E, Wilson M (2012) Uncovering spatial topology represented by rat hippocampal
population neuronal codes, J Comput. Neurosci., 33, pp. 227-255.
[16] Dabaghian Y, Mmoli F, Frank L, Carlsson G (2012) A Topological Paradigm for Hippocampal Spatial Map
Formation Using Persistent Homology, PLoS Comput. Bio.,l 8: e1002581.
[17] Arai M, Brandt V, Dabaghian Y (2014) The Effects of Theta Precession on Spatial Learning and Simplicial Com-
plex Dynamics in a Topological Model of the Hippocampal Spatial Map, PLoS Comput. Biol., 10: e1003651.
[18] Hatcher A (2002) Algebraic topology, Cambridge; New York: Cambridge University Press.
[19] Ghrist R (2008) Barcodes: The persistent topology of data, Bulletin of the American Mathematical Society, 45,
pp. 61-75.
Nat. Neurosci., 10: pp. 224-233.
Bats, Science, 340, pp. 367-372.
[20] Ulanovsky N, Moss CF (2007) Hippocampal cellular and network activity in freely moving echolocating bats,
[21] Yartsev MM, Ulanovsky N (2013) Representation of Three-Dimensional Space in the Hippocampus of Flying
[22] Mizuseki K, Sirota A, Pastalkova E, Buzsaki G (2009) Theta oscillations provide temporal windows for local
circuit computation in the entorhinal-hippocampal loop, Neuron, 64, pp. 267-280.
[23] Huhn Z, Orbn G, rdi P, Lengyel M (2005) Theta oscillation-coupled dendritic spiking integrates inputs on a long
time scale, Hippocampus, 15, pp. 950-962.
[24] Goldman-Rakic PS (1995) Cellular basis of working memory, Neuron 14: 477-485.
[25] Harris KD, Csicsvari J, Hirase H, Dragoi G, Buzsaki G (2003) Organization of cell assemblies in the hippocam-
pus, Nature, 424, pp. 552-556.
[26] Buzsaki G (2010) Neural syntax: cell assemblies, synapsembles, and readers, Neuron, 68, pp. 362-385.
12
[27] Avis D, Houle ME (1995) Computational aspects of Helly's theorem and its relatives, International Journal of
Computational Geometry & Applications, 05, pp. 357-367.
[28] Babichev A, Memoli F, Ji D, Dabaghian Y (2015) Combinatorics of Place Cell Coactivity and Hippocampal
Maps. in submition; arXiv:1509.01677).
[29] Skaggs WE, McNaughton BL, Wilson MA, Barnes CA (1996) Theta phase precession in hippocampal neuronal
populations and the compression of temporal sequences, Hippocampus, 6, pp. 149-172.
[30] Buzsaki G (2002) Theta oscillations in the hippocampus, Neuron, 33, pp. 325-340.
[31] Buzsaki G (2005) Theta rhythm of navigation: link between path integration and landmark navigation, episodic
and semantic memory, Hippocampus, 15, pp. 827-840.
[32] Robbe D, Buzsaki G (2009) Alteration of theta timescale dynamics of hippocampal place cells by a cannabinoid
is associated with memory impairment, J Neurosci., 29, pp. 12597-12605.
[33] Robbe D, Montgomery SM, Thome A, Rueda-Orozco PE, McNaughton BL, et al. (2006) Cannabinoids reveal
importance of spike timing coordination in hippocampal function, Nat Neurosci., 9, pp. 1526-1533.
[34] Heys JG, MacLeod KM, Moss CF, Hasselmo ME (2013) Bat and Rat Neurons Differ in Theta-Frequency Reso-
nance Despite Similar Coding of Space, Science, 340, pp. 363-367.
[35] London M, Husser M (2005) Dendritic Computation, Annu Rev. Neurosci., 28, pp. 503-532.
[36] Brody CD, Romo R, Kepecs A (2003) Basic mechanisms for graded persistent activity: discrete attractors,
continuous attractors, and dynamic representations, Curr. Opin. Neurobiol., 13, pp. 204-211.
[37] Magee JC (2000) Dendritic integration of excitatory synaptic input, Nat. Rev. Neurosci, 1, pp. 181-190.
[38] Spruston N (2008) Pyramidal neurons: dendritic structure and synaptic integration, Nat. Rev. Neurosci., 9, pp.
206-221.
MIT Press. pp. 9-92.
[39] Rall W (1989) Cable theory for dendritic neurons. In: Christof K, Idan S, editors. Methods in neuronal modeling,
[40] Jarsky T, Roxin A, Kath WL, Spruston N (2005) Conditional dendritic spike propagation following distal synap-
tic activation of hippocampal CA1 pyramidal neurons, Nat. Neurosci., 8, pp. 1667-1676.
[41] Huxter JR, Senior TJ, Allen K, Csicsvari J (2008) Theta phase-specific codes for two-dimensional position,
trajectory and heading in the hippocampus, Nat. Neurosci., 11, pp. 587-594.
VII. SUPPLEMENTARY FIGURES
13
FIG. S1: Spatial overlap encoded by temporal coactivity. (A) Place cells produce spike trains as the bat is flying
through their respective place fields. (B) Coactivity of two cells (indicated by the dashed box) marks the domain
where two place fields overlap.
FIG. S2: A 2D illustration of Helly's theorem. (A) If every three out of four convex planar regions overlap each
other, then they necessarily have a common fourth order overlap, marked by the red cross. (B) For three regions,
pairwise overlappings may not yield a common third order intersection -- the hollow spot in the middle is marked
by the red circle. However, this configuration is statistically rare. (C) Typically, a set of three pairwise overlapping
regions is a reliable signature of a triple overlap. In other words, a 2D clique almost always is a 2D simplex (and any
simplex is always a clique).
14
FIG. S3: The histograms of appearance numbers of 3D cliques. (A) and 3D simplexes (B). Note that the highest
occupancy bin on the right panel corresponds to zero appearance rate, i.e., most configurations that observed as
cliques, never appear as simplexes, and as a result the total number Ntot of cliques and simplexes differ by an order of
magnitude.
|
1905.07474 | 1 | 1905 | 2019-05-17T20:51:31 | Decision-Making in a Social Multi-Armed Bandit Task: Behavior, Electrophysiology and Pupillometry | [
"q-bio.NC"
] | Understanding, predicting, and learning from other people's actions are fundamental human social-cognitive skills. Little is known about how and when we consider other's actions and outcomes when making our own decisions. We developed a novel task to study social influence in decision-making: the social multi-armed bandit task. This task assesses how people learn policies for optimal choices based on their own outcomes and another player's (observed) outcomes. The majority of participants integrated information gained through observation of their partner similarly as information gained through their own actions. This lead to a suboptimal decision-making strategy. Interestingly, event-related potentials time-locked to stimulus onset qualitatively similar but the amplitudes are attenuated in the solo compared to the dyadic version. This might indicate that arousal and attention after receiving a reward are sustained when a second agent is present but not when playing alone. | q-bio.NC | q-bio | Decision-Making in a Social Multi-Armed Bandit Task: Behavior,
Electrophysiology and Pupillometry
Julia Anna Adrian1, Siddharth Siddharth2, Syed Zain Ali Baquar1, Tzyy-Ping Jung3, & Gedeon Deák1
Department of Cognitive Science1, Electrical Engineering2, Bioengineering3, UC San Diego
9500 Gilman Drive, La Jolla, CA 92093
Abstract
Understanding, predicting, and learning from other people's
actions are fundamental human social-cognitive skills. Little
is known about how and when we consider other's actions
and outcomes when making our own decisions. We developed
a novel task to study social influence in decision-making: the
social multi-armed bandit task. This task assesses how people
learn policies for optimal choices based on their own
outcomes and another player's (observed) outcomes. The
majority of participants integrated information gained through
observation of their partner similarly as information gained
through their own actions. This lead to a suboptimal decision-
making strategy. Interestingly, event-related potentials time-
locked
the
amplitudes are attenuated in the solo compared to the dyadic
version. This might indicate that arousal and attention after
receiving a reward are sustained when a second agent is
present but not when playing alone.
Keywords: Decision-Making; Uncertainty; Multi-Armed
Bandit; Social Interaction; Dyadic EEG
Introduction
to stimulus onset qualitatively similar but
For successful social interaction it is useful to represent and
predict other people's actions and the consequences of those
actions. Joint action is defined as the ability to coordinate
one's actions with others to achieve a goal (Vesper et al.,
2016). Although it occurs in many sorts of human activities,
it can be conveniently studied using the social or two-player
versions of standardized cognitive
tasks
modified for social interactions can reveal complex dynamic
in people's use of social information for judgments and
action-planning.
tasks. Such
task
For example, the joint Simon task (Sebanz, Knoblich, &
Prinz, 2003) is a modified, two-player version of the
standard Simon
that measures stimulus-response
compatibility. Participants learn to respond with left or right
button press for visual or auditory cues and show a longer
reaction times are shorter when the cue location is
compatible with the response hand, than when the cue
occurs contralaterally. This Simon effect (Simon, 1990),
interestingly, remains in the two-player version in which
each participant is only responsible for one stimulus-
response pair (Sebanz et al., 2003). This can be interpreted
as evidence that human action planning is automatic and is
elicited by processing another person's actions as well as
planning and executing our own actions. The propensity to
develop this ability might have evolved to enable efficient
social learning (Kilner, Friston, & Frith, 2007; Liao, Acar,
Makeig, & Deák, 2015). Particularly under conditions of
uncertainty, the capacity to observe, encode, and imitate
others' actions can be beneficial (Laland, 2004), permitting a
sort of vicarious embodied modeling.
tasks have shown
joint Simon and other
However, it is not always adaptive to generalize from
other's actions and outcomes to one's own. The findings
from
that
representation of other's internal states occurs even when it
is unnecessary, or disadvantageous, for optimal
task
completion. To study the extent to which people use
observation of other's actions and outcomes to influence
their own choices, even when it is unfavorable, we
developed a novel task: the social multi-armed bandit task.
The standard multi-armed bandit is a single-player paradigm
to study decision-making under uncertainty. Named after
the 'one-armed bandit' slot machines of casinos that have a
fixed reward probability, multi-armed bandit tasks present
several different options ('arms') of different, unknown
reward probabilities. They manifest a classic exploitation/
exploration problem (Cohen, McClure, & Yu; Gittins,
Glazebrook, & Weber, 2011). Commonly, after an initial
phase of exploration players employ one of two strategies:
maximizing or matching. Maximizing, or consistently
choosing
(based on prior
observations), is the optimal strategy for problems with
static reward probabilities. By contrast, matching, or
choosing each arm in proportion to its relative reward
probability, is suboptimal but nevertheless seen in humans
and other animals (Sugrue, Corrado, & Newsome, 2004).
the most-rewarding arm
Notably, although a great deal of problem solving and
prediction updating occurs in social or joint tasks, only a
few studies have included multiple decision-makers in
social versions of prediction tasks such as multi-armed
bandit, and even these have not investigated effects of social
interaction or observation on decisions (Liu & Zhao, 2010).
In addition to studying behavior, we recorded participants'
electroencephalogram (EEG) and pupil size as physiological
metrics of cortical and neuromodulatory concomitants of
social decision-making. These bio-sensing methods may
provide insights into the underlying neural dynamics of
decision-making with high temporal resolution. Both of
these physiological measures are common in affective
computing (Partala, 2003) to measure valence and arousal,
and cortical changes (Fink, 2009).
The present study
The present study aims to address a "key question of today's
cognitive science: how and to what extent do individuals
mentally represent their own and others' actions, and how
do these representations influence, shape, and constrain an
individual's own behavior when interacting with others?"
(Dolk et al., 2014).
To do so, we converted a classic three-armed bandit
paradigm into a turn taking game. Reward probabilities for
the three arms were different for each of two participants,
allowing us to estimate the distinct effects of their own and
their partner's action and outcome history on their ongoing
decision-making. We studied three outcome measures:
(1) decision-making behavior, (2) event-related EEG
potentials, and (3) pupil dilation. Details are described
below.
In the multi-player version, the probabilities remain
constant for each player, however, they differ between the
two players (see Table 1). This allows us to examine to what
extent each participant takes into account their own and
their partner's choices and outcomes. We expected to
observe two different core strategies:
Egocentric strategy: Participants might make
their
decisions only based on their own outcome history and
ignore information from their partner's outcomes. Players
using this strategy should converge on choosing their own
highest gaining arm (90% reward probability) most of the
time.
Joint strategy: Participants might take into account
information from their partner's outcomes to the same extent
as information from their own outcomes. Players using this
strategy should not converge on choosing one arm, because
all arms average the same reward probability if both
participants' outcomes are encoded equally. Alternately, in
an intermediate strategy, participants might take into
account their partner's outcomes but weigh them less than
their own outcomes, and then more slowly converge on their
own optimal choice.
Table 1: Reward probabilities for the different arms of the
social multi-armed bandit
(social) multi-armed bandit was described
One pair was excluded from EEG and eye-tracking
analysis due to recording failure; another pair was excluded
because one player chose the same arm on every trial. This
left behavioral data from 26 participants, EEG data from 12
participants, and pupillometry data from 12 participants.
Experimental Design
The
to
participants as 'the ice-fishing game' and presented on a
touchscreen. They were shown three 'ice holes' (arms)
distinguished by shape, at approximately equal distances
from each other (see Figure 2). The arms were associated
with discrete and constant reward probabilities (30%, 60%,
and 90%) unknown to the players. Upon choosing and
touching a hole, participants heard and saw differential
reward feedback. Participants had 100 trials each (200 total)
to catch as many fish as possible, choosing one ice hole per
trial. Each participant played the game once on their own
(solo version) and once as a turn-taking game (dyadic
version). For each dyad of participants, EEG and
pupillometry data were collected from one player, and
behavioral data were recorded from both.
Figure 1: The game screen when a player has won a reward
(green
two players'
accumulated rewards as well as which player's turn it is.
fish). The display shows
the
Player 1
Player 2
Arm 1
30%
90%
Arm 2
60%
60%
Arm 3
90%
30%
Experiment 1
Participants
Participants were 28 female undergraduate students (14
dyads) recruited through the university's SONA system.
They received course credit for participation in addition to a
small monetary reward based on performance in the social
multi-armed bandit (0.05 USD per reward).
Figure 2: Two participants playing the social multi-armed
bandit. The player on the left is wearing an Emotiv EEG
headset and PupilLabs eye-tracker.
Data Acquisition
The game was presented on a table-mounted capacitive
touch screen monitor (diagonal: 66cm). During the dyadic
turn-taking game the participants sat facing each other
(figure 2). An Emotiv headset (www.emotiv.com) recorded
14-channel EEG data, and a PupilLabs headset (pupil-
labs.com/pupil/) captured pupillometry data. These sensors
were chosen for participants' comfort and natural movement
during a social interaction. EEG was sampled at 128 Hz,
and the eye-facing camera sampled at 120 Hz. PupilLabs
Software was used to detect the pupil in each frame and
calculate its diameter. Lab Streaming Layer (LSL) (Kothe,
2015) was used to synchronize all of the data streams (i.e.
EEG, eye-gaze video, and game events) by time stamping
each event and each sample.
Synchronized EEG and pupillometry data were locked to
participants' game choices in LSL-created XDF files so that
behavioral and physiological data were epoched to trials. On
each turn the 2 sec of data following the outcome stimulus
presentation (win/loss) was used for further analysis.
Data Analysis
The first 20 trials of each game for each player were
considered training trials, to teach the participant the game.
These trials were not considered in the current analyses.
Decision-Making Behavior Participants' ice hole choice
patterns were analyzed via Kullback-Leibler divergence
(KLD), a measure of relative entropy.
𝐷!"(𝑝,𝑞) =
𝑝(𝑥!)
!
!!!
∗𝑙𝑜𝑔𝑝(𝑥!)
𝑞(𝑥!)
This quantifies
the divergence of one probability
distribution to another one. In our experiment we use the
KLD to measure the difference between a participant's
observed choices, and expected choices according to
potential strategies. We hypothesize the employment of four
different strategies with the expected choice probabilities as
summarized in Table 2. For the egocentric-maximizing
strategy, the player chooses the highest gaining arm (arm 3)
at every trial. In the egocentric-matching strategy, each arm
is chosen in proportion to their reward probability. The
joint-equal strategy assumes that the outcomes of both
players are weighted equally, resulting in an apparent
reward probability of 60% for each arm. In that case each
arm is chosen ⅓ of the trials. The joint-social strategy
assumes a social value of the arm that has an equal,
relatively high reward probability for both of the players
(arm 2) and is thus chosen most often.
KLD of observed vs. expected probability distribution for
each of the hypothesized strategies was calculated and
compared to classify each participant's preferred strategy.
As the joint strategy is not applicable in the single-player
version, only the two egocentric strategies were compared.
Table 2: Expected choice probabilities for player 1 for each
of the four hypothesized strategies. Reward probabilities for
Strategy of player 1
Arm 1
Arm 2
Arm 3
Ego-
centric
Joint
maximizing
0%
0%
100%
matching
16.7%
33.3%
50%
equal
33.3%
33.3%
33.3%
social
0%
100%
0%
Game Data The game data was an 8 x 200 matrix which
included the turn number, player number, reward state,
choice, time taken, player 1 reward and player 2 reward. In
the single player case the last value was set as -1 and
disregarded.
EEG Data The EEG data was cleaned using EEGLAB's
Artifact Subspace Reconstruction (ASR) noise removal
pipeline (Delorme, 2004; Mullen, 2013). Region of interest
was the occipital cortex (channel O1).
Eye-Tracking Data The current analyses only consider a
single channel containing pupil diameter
information.
Samples with abnormally high or zero pupil diameter values
(due to detection errors or eye blinks) were ignored and data
was interpolated by adjacent values.
After interpolating, the data was normalized to range
between 0 and 1 to account for discrepancies in pupil
diameter across subjects.
Further Analysis Epochs for each trial containing the
response and 2 seconds of subsequent data (including the
reward outcome). Our goal was to illustrate the pupil
dilation (indicating autonomic response) and cortical
dynamics (focusing on updating responses) upon perceiving
a reward stimulus after choosing a specific action.
Pupil and EEG data for each type of choice and reward
combination were then averaged across all subjects. For
EEG data, each channel was averaged independently, to
facilitate Event Related Potential (ERP) analyses. The 0.2
sec of data before the event were used as baseline for the
normed succeeding EEG data, to control variance in EEG
amplitude across subjects.
Results
Decision-Making Behavior Over the course of the game,
participants received information through trial and error and
could learn that different arms were associated with
different probabilities of receiving a reward. In the solo
version, participants chose the highest gaining arm more
often than the other two. Table 3 summarizes the decision-
making behavior via mean total scores and mean number of
choices for each arm. Participants distributed their choices
more equally and scored lower during the dyadic game.
Table 3: Means (SD) of each decision type in the single- and
multi-player games. All measures differ significantly
between single and multi-player version (p<0.001).
Reward
probability
No. of
choices
by game
version
single
multi
30%
60%
90%
7
(5.0)
18
(9.1)
17
(10.3)
32
(15.1)
56
(13.4)
30
(19.5)
Mean total score in the single-player game was 76 (SD: 7.0)
compared to 63 (SD: 10.1) in the multi-player game
(p<0.001).
For each version, participants were categorized based on
the strategy employed. For each strategy, KLD was
calculated between the observed and the expected choices.
70% (18/26) of participants employed a maximizing
strategy in the solo version. In the dyadic version, strategies
were more varied (see Figure 3). Most common was the
joint-social strategy, used by 32% (9/28). There was no
correlation between individuals' strategies in the solo and
dyadic versions.
Strategy use affected overall scores in the dyadic version
(F=6.083, p=0.004) and had a marginally significant effect
in the solo version (p=0.073).
Brain Dynamics ERP locked to outcome stimuli for each of
the differentially rewarding arms were compared between
three conditions: (1) solo: a player's responses in the solo
version of the game, (2) dyad (self): a player's responses to
an outcome of their own action, and (3) dyad (other): a
player's responses to an outcome of the partner's action.
Figure 4 illustrates the findings. The ERP displayed a
prominent positive potential around 300ms (P3) after dyadic
self-reward events in the dyadic version, but not for
partner's reward or for reward in the solo game. We also
note that the ERP response for partner's rewards as well as
for own-reward in the solo game is attenuated but follows
the same profile as that of the self-reward condition.
Figure 3: Distribution of strategies being employed by the
participants in the single- and multi-player game, as
determined via KLD.
Figure 4: ERP after reward at arm 3 (90%/30%) averaged
across participants from channel. Dyad (Self): player
receives reward in dyadic version, Dyad (Other): player
observes partner receive a reward in dyadic version, Solo:
player receives a reward in solo version.
Figure 5: Pupillometry data after reward at arm 3
(90%/30%) averaged across participants from channel.
Dyad (Self): player receives reward in dyadic version, Dyad
(Other): player observes partner receive a reward in dyadic
version, Solo: player receives a reward in solo version.
Pupillometry Pupil dilation significantly increases initially
(Figure 5) when a reward is obtained for the player
themselves when playing the social multi-armed bandit, but
not in the single-player version and not when observing the
partner receive a reward. After this initial response, we see
that after 0.8 seconds of the reward onset, pupil diameter
increases for the Dyad (Other) condition whereas it
decreases for the self-reward one.
Discussion
The Social Multi-Armed Bandit revealed that adults take
into account
information from others' when making
decisions. In consistency with previous studies, the majority
of participants employed maximizing strategy in the solo
version of the task. With this novel paradigm we found that
this is not the case when there is a second player present,
and when the reward probabilities for the arms are not the
same for the two players. Instead, more than half of the
participants employed a 'joint'-strategy, in which the
actions and outcomes from the other player are integrated
with their own when making decisions. One explanation for
this phenomenon is a high prior belief of the same
underlying probability structure for both players. This is
likely the case because the visual representation remains
constant throughout the game aside from updating the score
count and the display whose turn it currently is. This issue is
addressed in Experiment 2.
The ERP time-locked to stimulus onset showed a
qualitatively similar pattern after receiving a reward at arm
3 (90%/30% reward probability) for the three conditions
analyzed. However, the amplitude is highest when receiving
a reward in the dyadic version, and attenuated when
observing the partner receive a reward. When receiving a
reward in the solo version of the game, the amplitude is also
attenuated in comparison to receive a reward in the presence
of a second player. We believe that this is due to the higher
stakes and reward scenario attached with the dyadic version
of the game.
Interestingly, pupil dilation increases drastically at about
0.6 seconds after stimulus onset when receiving a reward in
the dyadic version of the game. In contrast, pupil dilation
after observing the partner receive a reward has a longer
latency of about 0.8 second. This likely reflects differential
activation of the parasympathetic nervous system (PNS) for
self/other reward scenarios.
Experiment 2
Participants
Participants were 32 undergraduate students (16 dyads, 10
female-female, 4female-male, 2 male-male)
recruited
through the university's SONA system. They received
course credit for participation in addition to a small
monetary reward based on performance in the social multi-
armed bandit (0.05 USD per reward).
Experimental Design
The experimental design of Experiment 2, is very similar to
experiment 1. The modifications of the experiment are:
(1) Whereas in experiment 1, one person in each dyad
played the solo version of the game before the dyadic
version and the other person played in the reverse order, in
experiment 2 both played the solo version either before or
after the dyadic version. In other words, game order was
randomly assigned by dyad. This ensured that the game
process was not driven by prior knowledge of only one of
the players.
(2) To reduce the prior belief of a constant underlying
reward structure of the game, we changed the background
color of the game after each turn, such that there was a
distinct visual cue to signal each player's turns.
(3) EEG data was recorded from both participants in
Experiment 2 (vs. only one participant per dyad in Exp. 1).
Pupillometry data was not recorded.
Data Acquisition
See Experiment 1.
Data Analysis
As in Experiment 1, the first 20 trials were excluded from
analysis.
Decision-making behavior
The analysis performed in Experiment 1 is based upon the
assumption
that participants make use of particular
strategies. In this experiment, a different type of analysis
was performed, considering choice behavior 'bottom-up'
without assumptions of specific strategies.
Participants' choices were analyzed via the Jensen-Shannon
Divergence (JSD). The JSD is a distance metric between
two probability distributions and based on the KLD:
𝐷!"𝑝,𝑞 = !!𝐷!"𝑝,𝑥 + !!𝐷!"𝑞,𝑥
𝑥=(𝑝+𝑞)/2
with
The JSD between the relative choice distribution of the last
80 trials of the participants' empirical behavior and the
relative choice distribution if all choices were made towards
the highest gaining arms (=(0,0,1), maximizing strategy)
was used to analyze the data. Hence, the decision-making
behavior for each participant could be characterized by their
JS divergence in the solo and the dyadic version. As
reference, the JS divergence of relative choice distribution
between matching behavior
(0.17, 0.33, 0.5) and
maximizing (0,0,1) is 0.31. This value was used to further
cluster participants into 'learners' (JSD < 0.31) and 'non-
learners' (JSD ≥ 0.31).
could
Results
Decision-Making Behavior Participants
be
categorized into four groups, depending on if they learned
which option was the highest gaining in the solo and/or
dyadic version of the game. 50% (16/32) of participants
were grouped into 'learners' in the solo version, but into
'non-learners' in the dyadic version. As shown in Figure 6,
in the solo version, they choose the 90% reward arm
significantly more often than the other two arms (F = 179.1,
p < 0.0005) and significantly more often than in the dyadic
version (F = 59.7, p < 0.0005). 22% (7/32) of participants
were clustered into 'learners' in both versions of the game,
and 19% (6/32) were clustered into 'non-learners' in both
versions of the game. 9% (3/32) were clustered into
'learners' in the dyadic version of the game, but not in the
solo version.
Figure 6: Choice behavior of 50% of participants who
learned which arm has the highest reward probability in the
solo but not the dyadic version of the multi-armed bandit.
Figure 7: ERP after reward at arm 3 (90%/30%) averaged
across participants from channel. Dyad (Self): player
receives reward in dyadic version, Dyad (Other): player
observes partner receive a reward in dyadic version, Solo:
player receives a reward in solo version.
Brain Dynamics ERPs were examined at channel O1 of
participants who chose the highest gaining arm most of the
time in the solo version, but not in the dyadic version
(Figure 7). We observed a high increase in amplitude for the
Dyad (Self) condition. In comparison, there was a negative
deflection in ERP for the Dyad (Other) and Solo condition.
Discussion
Even when the prior belief of a common underlying reward
structure was decreased, half of the participants integrated
information gained through observation of their partner
similarly as information gained through their own actions.
In Experiment 2, our goal was to combine the subjects'
decision-making behavior with their physiology. Similarly
as in Experiment 1, the ERP time locked to stimulus onset
when receiving a reward at arm 3 showed a high increase in
amplitude for the Dyad (Self) condition. In comparison,
there was a negative deflection in ERP for the Dyad (Other)
and Solo condition. We consider this a good starting point
to move towards extracting more high-level features such as
EEG power spectrum density, mutual information and pupil
diameter-based fixations and saccades in the future.
General Discussion
We developed a Social Multi-Armed Bandit task to examine
the influence of social interaction on decision-making. We
found that while some individuals do figure out that the
other player's information does not apply to them, the
majority of participants converged to a suboptimal decision-
making strategy. We termed this strategy 'joint' as it most
likely emerges through averaging the reward probabilities
for both players. Measurement of electrophysiology showed
a distinct P3 when the player receives a reward in the dyadic
version of the multi-armed bandit but not the solo version.
P3 is thought to emerge through stimulus-driven 'top-down'
processes when the participant pays focused attention to a
task. The distinct presence of the P3 in the dyadic task
might thus hint towards heightened attention, particularly
towards own rewards, in the presence of another player.
Interestingly, the pupillometry data revealed a similar
pattern as the ERP.
This task has considerable possibilities for further studies
of social interaction. Next steps include a similar experiment
with participants are previously acquainted with each other,
e.g. friends, and children with their parents. It is likely that
having a prior relationship with the other partner will alter
the joint strategy for one or both partners. It would also be
interesting to test how an asymmetric relationship (e.g.,
parent-child) would influence decision-making strategies,
compared to a more symmetric relationship. It is possible
that less-experienced participants (e.g., children) are more
likely to follow, or match, the behavior of a 'reliable'
person, as is the case in imitation (Poulin-Dubois, Brooker,
& Polonia, 2011). Lastly, this task might give interesting
insights into decision-making processes of neuro-divergent
people, particularly those with potential differences in social
behaviors (Montague, 2018).
Sebanz, N., Knoblich, G., & Prinz, W. (2003). Representing
others' actions: just like one's own?. Cognition, 88(3), B11-
B21.
Sugrue, L. P., Corrado, G. S., & Newsome, W. T. (2004).
Matching behavior and the representation of value in the
parietal cortex. science, 304(5678), 1782-1787.
Vesper, C., Abramova, E., Bütepage, J., Ciardo, F., Crossey,
B., Effenberg, A., ... & Schmitz, L. (2017). Joint action:
mental representations, shared information and general
mechanisms for coordinating with others. Frontiers in
psychology, 7, 2039.
Wellman, H. M., Cross, D., & Watson, J. (2001). Meta‐
analysis of theory‐of‐mind development: The truth about
false belief. Child development, 72(3), 655-684.
Acknowledgements
This study was funded with an Innovative Research Grant
from the Kavli Institute for Brain and Mind.
References
Cohen, J. D., McClure, S. M., & Yu, A. Y. (2007). Should I
stay or should I go? How the human brain manages the
trade-off
exploration.
and
Philosophical Transactions of the Royal Society of London
B: Biological Sciences, 362(1481), 933-942.
exploitation
between
Delorme, A., & Makeig, S. (2004). EEGLAB: an open source
toolbox for analysis of single-trial EEG dynamics including
independent component analysis. Journal of neuroscience
methods, 134(1), 9-21.
Dolk, T., Hommel, B., Colzato, L. S., Schütz-Bosbach, S.,
Prinz, W., & Liepelt, R. (2014). The joint Simon effect: a
review and theoretical integration. Frontiers in Psychology,
5, 974.
Fink, A., Grabner, R. H., Benedek, M., Reishofer, G.,
Hauswirth, V., Fally, M., ... & Neubauer, A. C. (2009). The
creative brain: Investigation of brain activity during creative
problem solving by means of EEG and fMRI. Human brain
mapping, 30(3), 734-748.
bandit allocation indices. John Wiley & Sons.
Gittins, J., Glazebrook, K., & Weber, R. (2011). Multi-armed
Kilner, J. M., Friston, K. J., & Frith, C. D. (2007). Predictive
coding: an account of the mirror neuron system. Cognitive
Processing, 8(3), 159-166.
(2014). Lab
(LSL).
https://github.com/sccn/labstreaminglayer. Accessed on
February, 1, 2019.
Kothe, C.
streaming
layer
Liao, Y., Acar, Z. A., Makeig, S., & Deak, G. (2015). EEG
imaging of toddlers during dyadic turn-taking: Mu-rhythm
modulation while producing or observing social actions.
NeuroImage, 112, 52-60.
Liu, K., & Zhao, Q. (2010). Distributed learning in multi-
armed bandit with multiple players. IEEE Transactions on
Signal Processing, 58(11), 5667-5681.
Montague, P. R. (2018). Computational Phenotypes Revealed
In Computational
Interactive Economic Games.
by
Psychiatry.
Mullen, T., Kothe, C., Chi, Y. M., Ojeda, A., Kerth, T.,
Makeig, S., ... & Jung, T. P. (2013, July). Real-time
modeling and 3D visualization of source dynamics and
connectivity using wearable EEG. In 2013 35th annual
the IEEE engineering
international conference of
in
medicine and biology society (EMBC) (pp. 2184-2187).
Partala, T., & Surakka, V. (2003). Pupil size variation as an
indication of affective processing. International journal of
human-computer studies, 59(1-2), 185-198.
and P3b. Clinical neurophysiology, 118(10), 2128-2148.
Polich, J. (2007). Updating P300: an integrative theory of P3a
Poulin-Dubois, D., Brooker, I., & Polonia, A. (2011). Infants
prefer to imitate a reliable person. Infant Behavior and
Development, 34(2), 303-309.
|
1212.3470 | 1 | 1212 | 2012-12-14T13:40:43 | A Behavioural Perspective on the Early Evolution of Nervous Systems: A Computational Model of Excitable Myoepithelia | [
"q-bio.NC",
"nlin.PS",
"q-bio.PE",
"q-bio.TO"
] | How the very first nervous systems evolved remains a fundamental open question. Molecular and genomic techniques have revolutionized our knowledge of the molecular ingredients behind this transition but not yet provided a clear picture of the morphological and tissue changes involved. Here we focus on a behavioural perspective that centres on movement by muscle contraction. Building on the finding that molecules for chemical neural signalling predate multicellular animals, we investigate a gradual evolutionary scenario for nervous systems that consists of two stages: A) Chemically transmission of electrical activity between adjacent cells provided a primitive form of muscle coordination in a contractile epithelial tissue. B) This primitive form of coordination was subsequently improved upon by evolving the axodendritic processes of modern neurons. We use computer simulations to investigate the first stage. The simulations show that chemical transmission across a contractile sheet can indeed produce useful body scale patterns, but only for small-sized animals. For larger animals the noise in chemical neural signalling interferes. Our results imply that a two-stage scenario is a viable approach to nervous system evolution. The first stage could provide an initial behavioural advantage, as well as a clear scaffold for subsequent improvements in behavioural coordination. | q-bio.NC | q-bio |
1
A Behavioural Perspective on the Early Evolution of Nervous
Systems: A Computational Model of Excitable Myoepithelia
Ronald A. J. van Elburg1,∗,†, Oltman O. de Wiljes1,2,†, Michael Biehl3, Fred A. Keijzer2
1 Institute of Artificial Intelligence, Faculty of Mathematics and Natural Sciences, University of
Groningen, Groningen, The Netherlands
2 Department of Theoretical Philosophy, Faculty of Philosophy, University of Groningen, Groningen,
The Netherlands
3 Johann Bernoulli Institute for Mathematics and Computer Science, Faculty of Mathematics and
Natural Sciences, University of Groningen, Groningen, The Netherlands
∗ E-mail: [email protected]
† These authors contributed equally.
Abstract
How and why the first nervous systems evolved remain open questions. One influential scenario casts
excitable myoepithelia (epithelia that combine conductive and contractile properties) as a plausible proto-
nervous system. We argue that while modern myoepithelia rely on gap junctions, early myoepithelia
had to rely on paracrine signalling or equivalently local chemical transmission and constitute a crucial
step towards modern nervous systems. Our main questions concern the coordinative possibilities and
limitations of such excitable myoepithelia and their potential relevance as an intermediate step to nervous
systems.
We used conductance based model cells to create artificial myoepithelia with various shapes and
electrophysiological characteristics. We developed a measure for whole body coordination in the myoep-
ithelium activity pattern. Using this measure we show that excitable myoepithelia relying on paracrine
signalling can exhibit body-scale patterns of activation. Relevant factors determining the extent of pat-
terning are the noise level for spontaneous vesicle release, relative body dimensions, and body size. In
small myoepithelial models whole-body coordination emerges from cellular excitability and excitatory
chemical transmission alone. At larger body sizes the intrinsic noise of chemical transmission limits
whole-body coordination.
We speculate that while proto-neural myoepithelia could have provided a solution for basic forms
2
of muscle-based movement, there would have been a strong evolutionary pressure to improve on this
mechanism by (a) the development of non-chemical transmission mechanisms and (b) a switch to nervous
systems proper by including axodendritic processes. Our study supports a two-step evolutionary scenario
for nervous systems. In the first step chemical receptors and the unicellular machinery for exocytosis
evolve into paracrine signalling, thus providing multicellulars with whole-body coordination.
In the
second step, axodendritic processes evolve under the evolutionary pressure towards larger body sizes.
Introduction
The evolution of the first nervous systems raises fundamental questions. Nevertheless, it has received
comparatively little scientific attention and there is no consensus as to how nervous systems first evolved
[Miller(2009), Moroz(2009), Lichtneckert and Reichert(2007)]. One reason for this situation has been the
dearth of relevant scientific data that can be made to bear on these early events. Nowadays, molecular
and genetic techniques have changed the situation at a molecular level, although less progress has been
made concerning the macroscopic features of early nervous systems. In the present study, we introduce
a computational approach to investigate the potential behavioural relevance of an early (proto-)neural
organization: excitable myoepithelia.
Evolution of the earliest nervous systems
Modern molecular genetic approaches to evolutionary biology yield time estimates for the first ap-
pearance of the different molecular building blocks of modern neurons.
It is now increasingly clear
that key molecular components like ion channels, neurotransmitters and synaptic protein families must
have been present in precursor organisms without nervous systems and even in single-celled organ-
isms [Greenspan(2007), Ryan and Grant(2009)]. It is also clear that the outlines of the genetic signalling
devices involved in nervous system development have been present from a very early start, some com-
ponents also being present in single-celled precursors [Arendt et al.(2008)], while genomic studies have
uncovered that Cnidaria (e.g. jellyfish), which is the earliest divergent phylum to have a nervous system,
possess an almost complete set of signalling molecules that have critical roles in bilaterian neurodevelop-
ment [Watanabe et al.(2009)].
However, progress at the level of animal behaviour and nervous system anatomy has been much slower
3
and proposed ideas remain inconclusive [Miller(2009), Mackie(1990), Lichtneckert and Reichert(2007)].
Work here so far consists mostly of different evolutionary scenarios explicating how excitable cells devel-
oped processes and synapses, became connected and evolved into basic forms of nervous systems that still
exist today. Lichtneckert and Reichert [Lichtneckert and Reichert(2007)] discuss the main proposals for
the evolution of nerve cells and nervous systems and, like Mackie [Mackie(1990)] before them, conclude
that it seems impossible to rate any of these proposals as more plausible or relevant than the others.
The problem is that on the basis of the fossil record alone such proposals must remain speculative
and cannot be tested or validated in any conclusive way. It is unknown whether the organisms involved
looked or behaved like any modern animal. Even the most 'primitive' nervous systems in existence to-
day -- those of cnidaria -- are highly evolved [Mackie(2004b)] and cannot be directly taken as a model
for the morphological and organizational properties of the organisms that first evolved nervous systems.
According to conservative molecular estimates, the event of the evolution of the first nervous system
-- as indicated by the first divergence wthin the eumetazoa -- happened more than 600 million years
ago [Peterson et al.(2008)], while other molecular estimates push the event back to a 1000 million years
ago [Blair(2009)].
In either case, we do not have solid -- or indeed any -- fossil evidence concerning
morphology, life style or behaviour [Valentine(2007), Erwin(2006)], except for the fact that the animals
involved must have been below the millimetre range [Brasier(2009)]. It is even possible that nerve cells
evolved several times independently [Moroz(2009)], involving different organisms. Given these uncertain-
ties, we think that developing additional sources of evidence concerning this crucial initial evolutionary
step should receive a high priority.
In this paper we add computational modelling to the existing techniques of gathering evidence on
the early evolution of nervous systems. Computational modelling allows for a systematic investigation of
the possible macroscopic operation of the early organisms that evolved nervous systems. Most notably,
computational models enable us to investigate how electrophysiological, cell-signalling and biomechanical
properties of simple multicellular organisms can self-organize into whole-body behaviour and coordination.
While these models do not directly shed light on what actually happened, they allow us to say more
about the possibilities available for organization and coordination during early nervous system evolution.
Additionally, these models help clarify the relevance of nervous systems and their precursors for the
behaviour of simple multicellular organisms. In particular, the identification of behavioural options and
limitations of evolutionary precursors to the nervous system might help answer the question: What drove
4
the early evolution of the nervous system?
Excitable myoepithelia as potential proto-nervous systems
As regards the possible role of early nervous systems in organizing behaviour, the first question to
ask is in what context early nervous systems may have developed. While the most notable feature of
modern nervous systems is their involvement in sensory information processing and subsequent action
execution [Watanabe et al.(2009), J´ekely(2011)], the setup and operation of early nervous systems and
any immediate precursor systems may have been different. In the present paper we build on the general
finding that the evolution of nervous systems is intertwined with the evolution of muscle as the primary
source of motility in eumetazoa [Seipel and Schmid(2005)]. Muscles and nervous systems come together
in evolution and it seems plausible that nervous systems are deeply implicated in the evolutionarily
important switch from movement based on cilia to muscle-based movement.
A highly relevant idea in this context comes from Pantin [Pantin(1956)], who argued that the operation
of the first nervous system was to organize and coordinate body movement based on muscle contractions
giving rise to what he called the 'metazoan behaviour machine'. Pantin claimed that this new effector,
consisting of extended surfaces that had to contract in a coordinated way to enable movement, required
a mechanism of large-scale coordination, providing a clear proximate reason for the evolution of nervous
systems as such coordinators. Passano [Passano(1963)] added that such a neural organization also re-
quired endogenous pacemakers to keep it going. Direct sensorimotor connections, such as the reflex arc,
were only a later development in this view.
However, the later discovery of excitable epithelia [Mackie(1965)] seemed to undercut this explanation.
The epithelium of an animal can act as an excitable sheet where action potentials, once initiated, travel in
all directions. In modern cases, these epithelial cells are linked by direct cytoplasmic connections or gap
junctions [Anderson(1980), Josephson(1985)]. Seemingly, whole-body coordination of extensive muscle
sheets can be accomplished by an excitable epithelium closely linked to a sheet of contractile tissue, or
even by a single excitable myoepithelium that combines excitable and contractile (myoid) functions. For
example, in the hydrozoan species Sarsia and Euphysa, the 'subumbrellar ectoderm is a single layer of
cells having striated fibres running circularly' forming 'the swimming muscle' [Mackie and Passano(1968)],
in this epithelial tissue electrical impulses travel from cell to cell via gap junctions and are able to induce
effector responses at points distant from the signals' origin. Similar forms of excitable tissue are also
5
present in muscle [Josephson(1985), Mackie(2004a), Brink et al.(1996)], plant fibres [Masi et al.(2009)],
sponge syncytial tissues [Leys et al.(1999)] and myocardial tissues [Nash and Panfilov(2004),Ten Tusscher
and Panfilov(2006),Ten Tusscher et al.(2007)]. For basic muscle coordination, it seems, there is no obvious
need for a nervous system.
Nevertheless, the claim that large scale muscle coordination could evolve without any form of nervous
system can be challenged. Two issues are important. What is a nervous system in the first place? Second,
are modern excitable epithelia a primitive condition or a later development? We will discuss these two
points in turn.
How do we define nervous systems? Current views take neurons as a package deal: 'a typical neuron
has four morphologically defined regions: cell body, dendrites, axon and presynaptic terminals' [Kandel
et al.(1995)]. However, it is plausible to separate the evolution of chemical transmission from the evolution
of long distance processes [Mackie(1990), Lichtneckert and Reichert(2007)]. In this case the evolution of
the first nervous systems consists of at least two separate evolutionary steps:
i. The evolution of chemical transmission, allowing cells to pass electrical signals to adjacent cells.
ii. The evolution of axodendritic processes that enable electrical signals to be sent to nonneighbouring
cells.
Subsequently, the evolution of the first (proto-)nervous systems can be taken as the evolution of the
first excitable epithelia based on electrical signalling through chemical transmission but, so far, without
long-distance axodendritic processes. From these evolutionary considerations one can draw the conclusion
that an excitable system based on chemical transmission but without processes is already a proto-nervous
system. Thus, the origins of the first nervous systems can be traced to the evolution of such excitable
epithelia themselves.
Current molecular studies suggest that the molecular machinery for chemical transmission predates
multicellulars.
In fact already in early unicellular animals genes coding for glutamate receptors and
intracellular calcium signalling were found [Ryan and Grant(2009)]. In addition, the SNARE protein,
an important molecule for transmembrane transport, was available in unicellulars as well [Kloepper
et al.(2008)]. Furthermore, voltage-gated ion channels, crucial for excitability, are thought to be derived
from a prokaryotic ancestor [Anderson and Greenberg(2001), Derst and Karschin(1998)] and thus are
much older than multicellular animals and therefore older than neural systems.
6
In contrast, the molecular basis of gap junctions (connexins and pannexins) seems to have arrived
after multicellularity, and even today starlet sea anemones (Nematostella vectensis) have no genes coding
for these proteins [Shestopalov and Panchin(2008)]. Additionally, electrophysiology could not establish
gap junctions in certain other cnidarian species like anthozoans (sea anemones like Aiptasia pulchella)
and scyphozoans (jellyfish like Cyanea capillata) although similar methods revealed gap junctions in
hydrozoans [Mackie et al.(1984), Satterlie(2011)].
In view of the above we postulate that excitable (myo)epithelia relying on local chemical transmis-
sion may very well have been the first major step towards modern nervous systems. Hence, to clarify
the possible operation of such a proto-neural setup we need to know how excitation spreads through
local chemical transmission. In this paper, we describe a computational model of such an excitable my-
opepithelium aimed at describing this spread of excitation. Our model is comparatively abstract at the
macroscopic scale (making as few assumptions as possible concerning the animal involved) while incor-
porating as much of the biomolecular and physiological details as necessary. In this way, we can study
the spreading of activity on such an excitable myoepithelium which acts both as a signalling device and
a primitive muscle-based effector. In particular, we aimed to understand the influence that animal size
and shape exert on whole-body coordination.
To our knowledge, no earlier modelling studies addressing the behaviour of excitable myoepithelia
exist. A somewhat similar model was used to describe travelling waves on coral nerve networks that
extend across several polyps [Chen et al.(2008)]. However, that study targeted modern nerve nets rather
than the basic myoepithelium of the present study. On a more abstract level our model is related to
experimental and theoretical work on Mexican waves and related phenomena [Farkas et al.(2002)], which
highlights the self-organized patterns of activation across an excitable medium. Our study provides one
of the first computational models of early nervous system functioning and evolution.
In this respect,
we regard this study as a groundbreaking application of computational neuroscience to evolutionary
neurobiology.
Methods
We developed our model to establish under which conditions a surface of excitable cells can generate
coordinated patterns of activity suitable for the control of body movement through muscle contractions.
7
Figure 1. Model Network A. Three dimensional organization of our model network, resembling a
tube-shaped animal. The myoepithelium model consists of excitable cells arranged into a triangular
lattice wrapped around a cylinder. B. Colour coding used to refer to the three differently oriented wave
fronts on this lattice. The size of a wave front is established by counting the number of neighbouring
cell pairs on the corresponding wave front.
In myoepithelia coordinated electrical activity across the epithelium necessarily involves coordinated con-
tractions, we modeled the initiation and spreading of these electrical excitations. To analyse the model
outcomes we developed two indicators of whole-body coordination. These indicators can summarize simu-
lation outcomes in a form suitable for visualizing the outcomes of parameter scans over the circumference
and the length of the networks to be introduced shortly.
Model
Our cell network model is built from single-compartmental cell models endowed with standard Hodgkin-
Huxley dynamics [Hodgkin and Huxley(1952)]; for parameter details see our Nordlie-Gewaltig-Presser-
style [Nordlie et al.(2009)] model summary tables 1 to
8. These model cells were arranged in a
triangular lattice generated by a purpose-built weight generator. This generator was designed for a
network building library (Available upon request from the corresponding author) developed by one
of us for the NEURON simulation environment [Carnevale and Hines(2006)]. To facilitate parame-
ter scanning, the multiple run control [van Elburg and van Ooyen(2010)] made available in ModelDB
(http://senselab.med.yale.edu/ModelDB, accession number 114359) was used. The network consists of a
sheet of cells on a triangular lattice rolled up to a cylinder. We decided to use a triangular lattice because
it is more isotropic than other two-dimensional lattices and therefore introduces the least amount of
directional bias. Excitatory chemical transmission between nearest neighbours including the dynamics of
8
messenger molecules exocytosis and receptor channel kinetics are modelled as double exponential conduc-
tance changes in receptor channel populations [Destexhe et al.(1994)]. Only a single type of excitatory
chemical transmission is included. An important aspect of this model is the rate of spontaneous vesicle
release. Data on spontaneous release frequencies of vesicles are still scarce and we found only one such
study [Mackenzie et al.(2000)]. Experimentally, the maximum spontaneous vesicle release rate found was
0.25 Hz. In our model, which has integrated six chemical transmission sites into a single model synapse,
this yields a total spontaneous vesicle release rate of 1.5 Hz per model synapse. A reliable lower bound,
other than zero, on the vesicle release rate is not available because many synapses failed to show sponta-
neous vesicle release during the experiment. From the biophysics of vesicle release we further know that
vesicle release is calcium-concentration dependent [Augustine(2001)]. As calcium concentration dynam-
ics is known to vary with surface-to-volume ratio and with the concentration, mobility and kinetics of
the endogenous calcium binding proteins [Cornelisse et al.(2007)], we can expect a large range of vesicle
release rates. We have therefore chosen to vary model vesicle release rates over 5 orders of magnitude
around a value of 0.1 Hz. Thus we included the maximum directly observed vesicle release frequency in
our range, but have a strong bias towards lower vesicle release rates.
We also investigated the possibility to introduce spontaneous network activity through spontaneous
spiking resulting from stochastic ion channel gating. However, a short exploration using the model
developed by [Linaro et al.(2011)] (available from ModelDB accession number: 127992) showed that this
would lead to spike rates much lower than those induced with the vesicle release rates included in the
model. With this in mind and in view of the computational cost we refrained from including ion channel
gating in our model.
Analysis
As whole-body coordination is not a well-defined mathematical concept at present, it is crucial that we
should choose good indicators of it. The triangular lattice supports three possible wave-front orientations;
as indicators we have chosen the relative amounts with which these orientations appear in our simulations.
Subpanel A1 of figure 2 illustrates our analysis method and shows that pairs of neighbouring cells come
in three different orientations: North -- South, North East -- South West, and South East -- North West. For
each of these orientations we count the number of neighbouring pairs that fire within 2 ms of each other.
In subpanel A2 of the same figure we show how these raw counts (left) are translated into percentages
9
(middle left), which are then used to set the diameter of the circles in the oriented circle pairs (middle left
and right). This presentation, which we call relative wave-front orientation prevalence, is suited
for the analysis of parameter scans, e.g. figures 5 and 7. We used this representation to present averages
over all runs at a specific parameter setting.
Wave fronts propagate roughly perpendicularly to their own orientation. This idea leads to a second
representation. Instead of showing the percentages directly, we add up the vectors normal to the wave
fronts weighed by the same percentages used in our relative wave-front orientation prevalence represen-
tation. For the North -- South oriented wave-front, propagation is to the East or the West, i.e. parallel
to the normal vector pointing West, similarly propagation is parallel to a South East-pointing vector for
the North East -- South West-oriented wave front, and parallel to North East pointing vector for the South
East -- North West-oriented wave front. The choice of these normal vectors is not unique, we selected them
in such a way that they point from the center to the oriented pair in the relative wave-front orientation
prevalence representation.
In figure 2:A2 middle right we show the vector addition and the resulting
vector. We call this representation wave-front propagation orientation. This representation is also
suitable for the presentation of parameter scans and additionally allows us to show both the individual
simulation runs and the average over simulation runs in a single figure. Supplementary figure S. 1 uses
this representation to show that we obtain similar results over twenty runs in which only the random
number generator initialization is changed.
For the purpose of this study, visual inspection of our earlier simulations showed that both relative
wave-front orientation prevalence and average wave-front propagation orientation are reasonably good
indicators of the effects of body size and chemical transmission noise on whole-body coordination.
Results
In subpanel B of figure 3 we see the temporal development of wave-front patterns on the short cylindrical
network also shown three dimensionally in subpanel A of the same figure. After initial excitation the wave
fronts grow in size uniformly in all directions until the wave fronts propagating longitudinally reach the
edge of the cylinder and disappear. The remaining wave fronts propagate in both transverse directions.
Provided no other noise-induced wave fronts interfere, these wave fronts eventually annihilate each other
on the side opposite the wave front initiation point.
In subpanel C of figure 3, where the wave-front
10
Figure 2. Visual representations used. (A1) Yellow disks represent active cells at a given time; grey
cells represent inactive cells. Neighbouring pairs of active cells come in three different orientations and
are arbitrarily labelled with red, green and blue. To establish which orientation is dominant we simply
count the occurrence of the orientation labels. (A2) The orientation label counts (left) are translated
into relative wave-front orientation prevalences (graph left of middle) and represented by the diameters
of the disks with the corresponding colour labelling. In addition these counts are translated into an
average wave-front propagation orientation by adding the normal vectors to these wave fronts with a
weight proportional to their label count (diagram right of middle). Relative wave-front orientation
prevalences and propagation direction are combined into a single representation for use in parameter
scans (right). (B,C) Like A2 with different orientation label counts.
11
Figure 3. Wave patterns on a short tube (length: 8 cells, circumference: 32 cells, noise rate: 0.1 Hz).
(A) Network geometry: the scissors indicate the line at which the tube is cut for presentation in B. (B)
Snapshots of network activity during 4 ms intervals in an illustrative phase of the dynamics. The wave
fronts propagating longitudinally die out at the network edges, causing the North -- South oriented wave
fronts to disappear. As a result, wave-fronts propagate predominantly transversely to the tube and the
wave-front orientations are South East -- North West and South West -- North East. (C) Temporal
development of different wave-front orientations, including all snapshot times shown in B.
12
orientation counts are shown for the time interval depicted in subpanel B, we clearly see the phenomena
we just described reflected. Initially all wave-front counts grow at the same rate, then the North -- South
oriented wave front, propagating longitudinally, dies out and the remaining wave fronts continue to grow
in size. Wave-front counts are approximately stable during transverse propagation. The double wave
fronts visible at later times occur because the data in this figure are binned in 4 ms bins, while a pair of
neighbouring cells is considered to be part of a wave front if they fire within 2 ms of each other. Also
visible is the breaking up of an initial wave front into smaller subfronts. We consistently observe slower
spreading of the wave fronts at points where the wave front changes orientation and the two differently
oriented parts move apart. Similarly we see a faster spreading of the wave fronts at points where the wave
front changes orientation and the two differently oriented parts move towards each other. This dynamics
tends to flatten wave fronts.
In subpanel B of figure 4 we see the temporal development of wave-front patterns on the elongated
cylindrical network also shown in 3D in subpanel A of the same figure. After initial excitation, the wave
front grows uniformly in all directions until the wave fronts propagating transversely annihilate each other
opposite the wave front initiation point. What remains are two wave fronts propagating longitudinally.
Provided no other noise induced wave fronts interfere with these wave fronts they will eventually reach
the edge of the cylinder and disappear.
In subpanel C of figure 4, where the wave-front orientation
counts are shown for the time interval depicted in subpanel B, we clearly see the phenomena we just
described reflected. Initially all wave-front counts grow at the same rate, then the South East -- North
West and South West -- North East oriented wave fronts, that is, the wave fronts propagating transversely,
annihilate each other. Subsequently, the North -- South oriented wave fronts continue to grow in size until
they become approximately stable during longitudinal propagation.
To establish whether and under which conditions a surface of excitable cells generates coordinated
patterns, we simulate our network at different network sizes and different noise rates. Our analysis
methods allow us to screen a large parameter space for patterned activity. At an intermediate noise rate
of 0.1 Hz figure 5 shows the relative wave-front orientation prevalences (represented by the diameters
of the coloured disks) and average propagation orientations (indicated by the orientation of the black
bars) for various body lengths and circumferences of the excitable myoepithelium.
In the corners of
this figure we drew rectangles to illustrate the shape of the model network at the parameter settings
used for the simulation in the corresponding corner. The ratio of circumference to length used in these
13
Figure 4. Wave patterns on a long tube (length: 32 cells, circumference: 8 cells, noise rate: 0.1 Hz).
(A) Network geometry: the scissors indicate the line at which the tube is cut for presentation in B. (B)
Snapshots of network activity during 4 ms intervals in an illustrative phase of the dynamics. The wave
fronts propagating transversely collide with each other, causing extinction due to the refractory period.
Subsequently, the remaining wave-fronts propagating longitudinally dominate the dynamics and the
North -- South wave-front orientation dominates. (C) Temporal development of different wave-front
orientations. Snapshot time markings are consistent with those in B.
14
Figure 5. Analysis of whole-body coordination for various body shapes. Relative wave-front
orientation prevalences (represented by coloured disk diameter) and average propagation orientations
(indicated with an oriented black bar) are shown for various body lengths and circumferences of the
excitable myoepithelium. In the corners, the corresponding body shape is indicated with a rectangle.
Bottom left: small 'square' cylinder, top left: long cylinder with small circumference, top right: large
'square' cylinder, bottom right: short cylinder with large circumference. This parameter scan shows
three effects: (i) more elongated networks show better developed longitudinal wave fronts, (ii) whereas
shorter networks show better developed transverse wave fronts, (iii) however for fixed
length-to-circumference ratios (visible on the diagonals running from bottom left to top right) we can
see that with increasing size preference for transverse or longitudinally moving wave fronts is lost.
15
drawings are understated with 1:1 (left bottom corner), 1:5 (left top corner), 5:5 (right top corner) and
5:1 (right bottom corner), while the actual ratios in the model network are 4:4, 4:256, 256:256 and 256:4,
respectively. From the average propagation orientations in figure 5 we see that there is a strong preference
for longitudinal wave-front propagation if the model network axis is long compared to its circumference
(upper left). In contrast, if the model network circumference is large compared to its axis (lower right),
then we see a strong preference for transverse wave-front propagation. This is also visible from the
relative wave-front orientation prevalences. Hence, we see that for these networks there is significant
pattern formation. To extract how pattern formation scales with size we can study the change in pattern
formation on the diagonals running parallel to the main diagonal (bottom-left corner to top-right corner).
Along these diagonals body size varies while the ratio between axis length and circumference remains
constant; the smallest body size is at the lower left side of these diagonals and the large body size at
the upper right side. As we move along these diagonals to larger scale networks we observe that relative
wave-front orientation prevalences equalize and average propagation orientation diminishes. This shows
that pattern formation fails to reach network size when we move to large networks.
Noise, which represents spontaneous vesicle release, drives activity in our network, but we also expect
it to interfere with emerging patterns. The reduction of the noise rate might therefore rescue patterning
in large-scale networks, and an increase in noise rate might destroy patterning in small-scale networks.
Figure 6 illustrates the influence of the noise rate on whole-body coordination at a single fixed combination
of length and circumference. In this figure each subpanel shows, for a specific noise rate, the temporal
development of wave-front orientation prevalences. At the top of each subpanel we find the wave-front
counts plotted versus time, followed by snapshots of the activity in the network. Subpanel A shows the
low noise rate situation in which almost every excitation grows to network scale and induces coordinated
activity. Subpanel B shows the intermediate noise rate situation in which many wave fronts grow to the
short scale of the network but spreading on the long scale is often interrupted by collision with other wave
fronts. Subpanel C shows the high noise-rate situation in which, in general, wave fronts are disrupted
before reaching network scale and whole-body coordination at the network level is absent.
Figure 7 shows two parameter scans over network sizes, performed at the highest and the lowest noise
rates used in this study. Both panels show relative wave-front orientation prevalences (represented by
the diameters of the coloured disks) and average propagation orientations (indicated by the orientation
of the black bars) for various body lengths and circumferences of the excitable myoepithelium. The two
16
Figure 6. Development of wave-front orientation prevalences over time on a long tube (length: 32
cells, circumference: 8 cells) at three different noise rates. A. Low noise condition (0.01 Hz/cell). The
graph at the top shows wave-front counts for all three orientations. It is clearly visible how activity first
grows equally for all three orientations until the scale of activation pattern equals the tube's
circumference, at which point in time the wave front moving longitudinally starts to dominate. B. At
the intermediate noise rate (0.1 Hz/cell) we still observe growth of wave-front patterns to the scale of
the animal, but occasionally several wave fronts are initiated in close succession leading to destructive
interference. C. At a high noise rate (1 Hz/cell) wave fronts are initiated at a high rate and due to
destructive interference with each other these wave fronts often fail to grow to the scale of the animal.
As a result we no longer observe whole-body coordinated activity.
17
Figure 7. Analysis of whole-body coordination for the extreme noise rates used in this study. Relative
wave-front orientation prevalences (represented by the diameters of the coloured disks ) and average
propagation orientations (indicated with the oriented of the black bar) are shown for various body
lengths and circumferences of the excitable myoepithelium. The two subpanels are organized as in
figure 5. (A) Low noise rate: 0.001 Hz, (B) High noise rate: 10 Hz. Compared to figure 5 we find
slightly stronger relative wave-front orientation prevalences and average propagation orientations at the
low noise rate in panel (A). In contrast we clearly see the loss of whole-body coordination with
increasing noise rates, as relative wave-front orientation prevalences become uniform and average
propagation orientations are almost absent in panel (B).
18
subpanels are organized as in figure 5 but the simulations were performed at a low noise rate of 0.001 Hz
(A) and a high noise rate of 10 Hz (B). Compared to figure 5, we find slightly stronger relative wave-front
orientation prevalences and average propagation orientations at the low noise rate in panel (A). In panel
(B), in contrast, we see a loss of whole-body coordination with high noise rate, since relative wave-front
orientation prevalences become uniform and average propagation orientations are almost absent at most
parameter values in this scan. This shows that a reduction of chemical transmission noise can enhance
the range of network size over which we find whole-body coordination, whereas an increase will reduce
this range.
Discussion
To find answers to our main question, namely, how and why the first nervous systems evolved, we used
a computational approach to investigate the possible behavioural relevance of basic forms of neural or
proto-neural organization.
We argued first that two separate consecutive evolutionary steps introduced two key components of
functional neuron anatomy:
i. Chemical transmission for intercellular electrical signalling to adjacent cells.
ii. Axodendritic processes that enable electrical signals to be sent to nonneighbouring cells.
In addition, the origins of the first nervous systems are plausibly linked to the evolution of muscle, the key
effector innovation that allowed animals to move even when they had grown to a large size [Pantin(1956),
Seipel and Schmid(2005)].
The reasoning we used to construct our model consists of three main parts. First, we turned to modern
myoepithelia as the most basic form of muscle coordination in existence today. There, coordination is
brought about by an excitable epithelium which doubles as a contractile tissue, while the cells of this
epithelium are directly connected by gap junctions. Second, we took into consideration that there is
good evidence that gap junctions are a later evolutionary development compared to chemically induced
action potentials. Third, we drew the conclusion that without electrical connections there would be a
good evolutionary reason to develop chemical transmission as a first step towards the evolution of a basic
proto-neural organization consisting of a myoepithelium relying on such chemical transmission.
19
Thus, our model allowed us to investigate the evolutionary possibility that such a proto-neural orga-
nization, which relies on chemical transmission but no axodendritic processes, played a role in organizing
a basic form of muscle-based behaviour. We think that such a primitive form of muscle coordination is an
essential intermediate stage preceding the next step towards full nervous systems, i.e. the development
of axodendritic processes.
The emergence of whole-body coordination
The simulation results showed that a proto-neural organization based on chemical transmission alone can
lead to self-organized patterned activity at a whole-body scale, depending on the values of three main
features: noise, body dimensions and body size. Let us briefly consider each of these features.
Noise: Noise is required to initiate electrical activity. The cells included in our model epithelium
showed spontaneous vesicle release that initiated action potentials which local chemical transmission
subsequently spread out across the epithelium. This spontaneous activity could give rise to patterned
activity at the body scale if the noise level was sufficiently low. At high noise levels, the subsequent
spontaneous firing of action potentials disrupts already evolving large-scale patterns.
Body dimensions: At sufficiently low spontaneous vesicle release rates the relative myoepithelium
dimensions, i.e. the length to width ratio, determines the type of whole-body coordination. Wave fronts
travelling along the short dimension die out either through collision with the edge or through annihilation
with a wave front travelling in the opposite direction from the same initiation site. Wave fronts travelling
along the long dimension then travel on and finally die out through the before-mentioned mechanism.
At such vesicle release rates wave fronts travelling along the long dimension dominate the dynamics and
lead to a primitive form of whole-body coordination. Thus body dimensions are a key feature for the
emergence of body coordination under these circumstances.
Body size: In our model the scale of these patterns is determined by the rate of spontaneous vesicle
release. We see whole-body coordination emerge only when the scale of these patterns matches roughly
with one of the dimensions of the animal, i.e. matches with length or circumference. Consequently, we
see a reduction of whole-body coordination with the increase of animal size.
While the model remains rather basic, its significance lies in the way it shows the generic properties
of a proto-neural myoepithelium without any additional features. The important message here is that
such a system enables basic coordinated patterning on its own, without any sensory input, without any
central pattern generators and without requiring any specific wiring or particular connections between
the cells. Coordination can be cast as an ingrained self-organized feature of such an organization. At
this stage there is no intrinsic need for particular neural 'circuits' or wirings that require a specialized
20
evolutionary route to explain their presence.
Evolutionary implications
Our model shows how the first of the two evolutionary steps leading to full nervous systems could
have resulted in an excitable myoepithelium that is roughly comparable to modern forms of excitable
myoepithelia, but without direct electrical coupling. Given that current evidence suggests that such direct
electrical coupling was a later evolutionary development, enabling early excitable epithelial systems can
be cast as the primary evolutionary reason for the rise of the chemically induced transmission of electrical
signals between neighbouring cells. Importantly, despite its similarity to modern non-neural myoepithelia,
our model relies on local chemical transmission, which makes it a definite example of a proto-neural
system. As such it is halfway between other excitable biological systems and modern nervous systems.
In addition, the generic capacity for patterning in this proto-neural organization remains limited to
smaller body sizes and does not scale up to larger ones. If this limitation proves to be a genuine and
general characteristic of this kind of organization, it will provide a major constraint and influence on our
understanding of subsequent evolutionary developments such as the origins of full nervous systems with
axodendritic processes and modern (myo)epithelia with gap junctions.
Given the present focus on the evolution of nervous systems the second option is perhaps less relevant.
Still, our simulations can potentially explain why electrical transmission became an important feature of
modern excitable epithelia such as those in jellyfish: to reach whole-body or whole-organ coordination, a
noise free coupling between the cells in these tissues is required, thus allowing the spreading of wave-fronts
to the scale of the tissue.
More central to present purposes, the model contributes to our understanding of the evolution of
early nervous systems and their precursors by suggesting new relevant details for their operation as well
as a more gradual evolutionary path towards the modern neural organization. It will be interesting and
necessary to look at additional features in follow-up research to investigate how the basic myoepithelial
proto-neural organization can be evolutionarily improved upon to expand its coordinative features and
reaction to stimuli. In this way, it also provides a good starting point for a further investigation of the
21
other main step in nervous system evolution, i.e. axodendritic processes.
To conclude, the model suggests an evolutionary trajectory in which the first step of nervous system
evolution provided an initial advantage for coordinating an extended contractile tissue in the absence of
direct electrical connections between cells. At the same time, the model suggests that there are significant
limitations to this initial organization, thus evolutionary pressure towards larger body size may have given
rise to the next step in nervous system evolution.
Follow-up questions
The research described above suggest a number of general as well as specific questions and topics for
further research. Here we will describe some of the most urgent and important follow-up questions.
These will be divided into two methodological and two biological questions.
A first methodological issue concerns the problem of analyzing and measuring whole-body coordination
as an emergent feature. So far, the primary metric has been the coherence of wave fronts with the width
of a single cell. While this generally works for the sort of behaviour the current model exhibits, this is
not a direct measure of activation relative to body size: it is merely a measure of wave-front coherence.
To be able to say something about what activation does for the organism, it will be necessary to devise
a measure that equates activity in areas relative to topology.
A second methodological issue concerns the need and possibility of adding movement and environ-
mental dynamics to the basic model as presented here. Work on cardiomyocytes suggests that mechanics
can be integrated into the model [Nash and Hunter(2000), Nash and Panfilov(2004)]. This would make
it feasible to study interaction effects between wave propagation and changes in body shape. When
combined with, for example, a fluid dynamics model for the environment this could provide insight into
how early multicellulars may have moved in an aquatic environment.
From a biological perspective, the first important question concerns the various ways in which biolog-
ical features can be added to a proto-neural myoepithelial organization. We will briefly mention a few
possibilities of enriching such a myoepithelium:
• Central Pattern Generators (CPGs): Activity in the current model is initiated by random noise
that sets individual cells firing and activity spreads from there. Patterns of activity derived from
the interaction of such spontaneous wave fronts and the form of the animals are key factors for the
22
global activity patterns. However, pacemaker cells or structures that initiate firing at a specific place
and with a specific frequency might be able to entrain activity across the whole myoepithelium. It
would be important to investigate how such CPGs can arise in a basic set up as described above,
and how they impact on global myoepithelial activity.
• Excitation and inhibition: Modern neurotransmitters can have either excitatory or inhibitory effects
on the post-synaptic cell. The present model relies solely on excitation, but it would be interesting
to see how adding different interaction dynamics can change the pattern dynamics.
• Mechanosensory feedback: Modern cells contain mechanosensitive ion channels. These are ion
channels which are activated by membrane stretch, and which in addition to several other functions,
play a role in volume regulation and in the functioning of healthy muscle and cardiac cells [Hamill
and Martinac(2001)]. Despite their general presence in modern cells their molecular identity is
hard to establish electrophysiologically [Sachs(2010)], although recent developments show some
progress [Gottlieb and Sachs(2012), Coste et al.(2012), Kim et al.(2012)]. Sufficient insight into
the molecular identity of these mechanosensitive channels is a prerequisite for the kind of genetic
analysis which is needed to hypothesize about their role in the evolution of neural systems. Including
mechanosensitive channels in an excitable myoepithelium model which also includes the mechanics,
could reveal their impact on whole-body coordination.
• Sensory stimuli and feedback: Adding a way to initiate activity on the basis of external stimuli would
also be an obvious way to extend the model. Note, however, that from the present perspective it
is not sensory initiation but rather coordination of activity which is a prerequisite for whole body
coordination and thus for behaviour.
Another major biological question concerns the second step involved in the evolution of nervous sys-
tems namely the origin of elongated axodendritic processes. The proto-neural organization discussed
so far produces limited coordinative abilities that can presumably be extended by some additional re-
finements. However, the evolution of processes provides the transition to modern nervous systems. A
key question that must be answered is in what specific ways extended processes change the coordinative
behaviour of such myoepithelia. While this question has so far proven to be extremely difficult to answer
by more direct evolutionary approaches, the current modelling approach provides an excellent starting
point for further investigation.
23
References
Anderson(1980). P. A. V. Anderson. Epithelial conduction: its properties and functions. Progress in
Neurobiology, 15:161 -- 203, 1980.
Anderson and Greenberg(2001). P. A. V. Anderson and R. M. Greenberg. Phylogeny of ion channels:
clues to structure and function. Comparative Biochemistry and Physiology Part B, 129(1):17 -- 28,
May 2001.
Arendt et al.(2008). D. Arendt, A. S. Denes, G. J´ekely, and K. Tessmar-Raible. The evolution of
nervous system centralization. Philosophical Transactions Of The Royal Society B, 363:1523 -- 1528,
2008.
Augustine(2001). G. J. Augustine. How does calcium trigger neurotransmitter release? Current
Opinion in Neurobiology, 11(3):320 -- 326, Jun 2001.
Blair(2009). J. E. Blair. The timetree of life, chapter Animals (Metazoa), pages 223 -- 230. Oxford
University Press, Oxford, 2009.
Brasier(2009). M. Brasier. Darwin's lost world: The hidden history of animal life. Oxford University
Press, Oxford, 2009.
Brink et al.(1996). P. R. Brink, K. Cronin, and S. V. Ramanan. Gap junctions in excitable cells.
Journal of Applied Psychology Bioenergetics and Biomembranes, 28(4):351 -- 358, Aug 1996.
Carnevale and Hines(2006). N. T. Carnevale and M. L. Hines. The NEURON Book. Cambridge Uni-
versity Press, Cambridge, UK, 2006.
Chen et al.(2008). E Chen, K M Stiefel, T J Sejnowski, and Bullock T H. Model of traveling waves in
a coral nerve network. Journal of Comparative Physiology A, 194(2):195 -- 200, 2008.
Cornelisse et al.(2007). L. N. Cornelisse, R. A. J. van Elburg, R. M. Meredith, R Yuste, and H. D.
Mansvelder. High speed two-photon imaging of calcium dynamics in dendritic spines: consequences
for spine calcium kinetics and buffer capacity. PLoS One, 2(10):e1073, 2007. doi: 10.1371/journal.
pone.0001073. URL http://dx.doi.org/10.1371/journal.pone.0001073.
24
Coste et al.(2012). B. Coste, B. Xiao, J. S. Santos, R. Syeda, J. Grandl, K. S. Spencer, S. E. Kim,
M. Schmidt, J. Mathur, A. E. Dubin, M. Montal, and A. Patapoutian. Piezo proteins are pore-
forming subunits of mechanically activated channels. Nature, 483(7388):176 -- 181, MAR 8 2012.
ISSN 0028-0836. doi: {10.1038/nature10812}.
Derst and Karschin(1998). C. Derst and A. Karschin. Evolutionary link between prokaryotic and
eukaryotic K + channels. Journal of Experimental Biology, 201(Pt 20):2791 -- 2799, Oct 1998.
Destexhe et al.(1994). A. Destexhe, Z. F. Mainen, and T. J. Sejnowski. Synthesis of models for ex-
citable membranes, synaptic transmission and neuromodulation using a common kinetic formalism.
Journal of Computational Neuroscience, 1:195 -- 230, 1994.
Erwin(2006). D H Erwin. Evolutionary contingency. Current Biology, 16(19):R825 -- R826, Oct 2006.
doi: 10.1016/j.cub.2006.08.076. URL http://dx.doi.org/10.1016/j.cub.2006.08.076.
Farkas et al.(2002). I. Farkas, D. Helbing, and T. Vicsek. Social behaviour: Mexican waves in an
excitable medium. Nature, 419(6903):131 -- 132, 2002. ISSN 0028-0836.
Gottlieb and Sachs(2012). P. A. Gottlieb and F. Sachs. Cell biology: The sensation of stretch. Nature,
483(7388):163 -- 164, MAR 8 2012. ISSN 0028-0836.
Greenspan(2007). R. J. Greenspan. An introduction to nervous systems. Cold Spring Harbor Labora-
tory Press, Cold Spring Harbor, NY, 2007. ISBN 0879698217.
Hamill and Martinac(2001). O P Hamill and B Martinac. Molecular basis of mechanotransduction in
living cells. Physiological Reviews, 81(2):685 -- 740, APR 2001. ISSN 0031-9333.
Hodgkin and Huxley(1952). A. L. Hodgkin and A. F. Huxley. A quantitative description of membrane
current and its application to conduction and excitation in nerve. Journal of Physiology, 117(4):
500 -- 544, Aug 1952.
J´ekely(2011). G. J´ekely. Origin and early evolution of neural circuits for the control of ciliary locomo-
tion. Proceedings of the Royal Society B, 278(1707):914 -- 922, Mar 2011. doi: 10.1098/rspb.2010.
2027. URL http://dx.doi.org/10.1098/rspb.2010.2027.
Josephson(1985). R. J. Josephson. Comparative Neurobiology: Modes of communication in the nervous
system, chapter Communication by conducting epithelia., pages 133 -- 148. Wiley (New York), 1985.
25
Kandel et al.(1995). E. R. Kandel, J. H. Schwartz, and Jessell, editors. Essentials of of neural science
and behavior. McGraw-Hill, 1995.
Kim et al.(2012). S. E. Kim, B. Coste, A. Chadha, B. Cook, and A. Patapoutian. The role of drosophila
piezo in mechanical nociception. Nature, 483(7388):209 -- 212, MAR 8 2012. ISSN 0028-0836. doi:
10.1038/nature10801.
Kloepper et al.(2008). T. H. Kloepper, C. N. Kienle, and D. Fasshauer. SNAREing the basis of
multicellularity: consequences of protein family expansion during evolution. Molecular Biol-
ogy and Evolution, 25(9):2055 -- 2068, Sep 2008.
doi: 10.1093/molbev/msn151. URL http:
//dx.doi.org/10.1093/molbev/msn151.
Leys et al.(1999). S P Leys, G O Mackie, and R W Meech. Impulse conduction in a sponge. Journal
of Experimental Biology, 202:1139 -- 1150, 1999.
Lichtneckert and Reichert(2007). R. Lichtneckert and H. Reichert. Origin and evolution of the first
nervous system, pages 289 -- 315. Elsevier, Amsterdam, 2007.
Linaro et al.(2011). D. Linaro, M. Storace, and M. Giugliano. Accurate and fast simulation of
channel noise in conductance-based model neurons by diffusion approximation. PLoS Com-
putational Biology, 7(3):e1001102, Mar 2011. doi: 10.1371/journal.pcbi.1001102. URL http:
//dx.doi.org/10.1371/journal.pcbi.1001102.
Mackenzie et al.(2000). P. J. Mackenzie, G. S. Kenner, O. Prange, and T. H. Murphy. Vesicle number
does not predict postsynaptic measures of miniature synaptic activity frequency in cultured cortical
neurons. Neuroscience, 98(1):1 -- 7, 2000.
Mackie(1965). G. O. Mackie. Conduction in the nerve-free epithelia of siphonophores. American
Zoologist, 5:439 -- 453, Aug 1965.
Mackie(1990). G O Mackie. The elementary nervous system revisited. Integrative and Comparative
Biology, 30(4):907, 1990.
Mackie(2004a). G O Mackie. Epithelial conduction: recent findings, old questions, and where do we
go from here? Coelenterate Biology 2003, pages 73 -- 80, 2004a.
26
Mackie(2004b). G O Mackie. Central neural circuitry in the jellyfish Aglantha: A model 'simple
nervous system'. Neurosignals, 13:5 -- 19, 2004b.
Mackie and Passano(1968). G. O. Mackie and L. M. Passano. Epithelial conduction in hydromedusae.
Journal of General Physiology, 52(4):600, 1968.
Mackie et al.(1984). G. O. Mackie, P. A. V. Anderson, and C. L. Singla. Apparent absence of gap-
junctions in 2 classes of cnidaria. Biological Bulletin, 167(1):120 -- 123, 1984. ISSN 0006-3185. doi:
{10.2307/1541341}.
Masi et al.(2009). E. Masi, M. Ciszak, G. Stefano, L. Renna, E. Azzarello, C. Pandolfi, S. Mugnai,
F. Baluska, F.T. Arecchi, and S. Mancuso. Spatiotemporal dynamics of the electrical network
activity in the root apex. Proceedings of the National Academy of Sciences, 106(10):4048, 2009.
Miller(2009). G. Miller. On the origin of the nervous system. Science, 325:24 -- 26, 2009.
Moroz(2009). L L Moroz. On the independent origins of complex brains and neurons. Brain, Behavior
and Evolution, 74(3):177 -- 190, 2009.
Nash and Hunter(2000). M. P. Nash and P. J. Hunter. Computational mechanics of the heart. Journal
of Elasticity, 61:112141, 2000.
Nash and Panfilov(2004). M. P. Nash and A. V. Panfilov. Electromechanical model of excitable tissue
to study reentrant cardiac arrhythmias. Progress in Biophysics and Molecular Biology, 85(2-3):
501 -- 522, 2004.
Nordlie et al.(2009). E. Nordlie, M-O Gewaltig, and H. E. Plesser. Towards reproducible descriptions
of neuronal network models. PLoS Computational Biology, 5(8):e1000456, 2009. URL doi:10.
1371/journal.pcbi.1000456.
Pantin(1956). C. F. A. Pantin. The origin of the nervous system. Pubblicazioni della Stazione Zoologica
di Napoli, 28:171 -- 181, 1956.
Passano(1963). L. Passano. Primitive nervous systems. Proceedings of the National Academy of Sci-
ences, 50(2):306 -- 313, 1963.
27
Peterson et al.(2008). K. J. Peterson, J. A. Cotton, J. G. Gehling, and D. Pisani. The ediacaran emer-
gence of bilaterians: congruence between the genetic and the geological fossil records. Philosophical
Transactions of the Royal Society B, 363(1496):1435, 2008. ISSN 0962-8436.
Ryan and Grant(2009). T. J. Ryan and S. G. N. Grant. The origin and evolution of synapses. Nature
Reviews Neuroscience, 10(10):701 -- 712, 2009.
Sachs(2010). F. Sachs. Stretch-activated ion channels: What are they? Physiology, 25(1):50 -- 56, FEB
2010. ISSN 1548-9213. doi: 10.1152/physiol.00042.2009.
Satterlie(2011). Richard A Satterlie. Do jellyfish have central nervous systems?
Journal of Ex-
perimental Biology, 214(Pt 8):1215 -- 1223, Apr 2011.
doi: 10.1242/jeb.043687. URL http:
//dx.doi.org/10.1242/jeb.043687.
Seipel and Schmid(2005). K. Seipel and V. Schmid. Evolution of striated muscle:
jellyfish and the
origin of triploblasty. Developmental Biology, 282(1):14 -- 26, Jun 2005. doi: 10.1016/j.ydbio.2005.
03.032. URL http://dx.doi.org/10.1016/j.ydbio.2005.03.032.
Shestopalov and Panchin(2008). V. I. Shestopalov and Y. Panchin. Pannexins and gap junction protein
diversity. Cellular And Molecular Life Sciences, 65(3):376 -- 394, FEB 2008. ISSN 1420-682X. doi:
{10.1007/s00018-007-7200-1}.
Ten Tusscher and Panfilov(2006). K. H. W. J. Ten Tusscher and A. V. Panfilov. Alternans and spiral
breakup in a human ventricular tissue model. American Journal of Physiology, 291(3):H1088, 2006.
Ten Tusscher et al.(2007). K. H. W. J. Ten Tusscher, R. Hren, and A. V. Panfilov. Organization of
ventricular fibrillation in the human heart. Circulation Research, 100(12):e87, 2007.
Valentine(2007). J. W. Valentine. Seeing ghosts: Neoproterozoic bilaterian body plans (in The rise
and fall of the Ediacaran biota). Geological Society Special Publication, 286:369 -- 375, 2007.
van Elburg and van Ooyen(2010). R. A. J. van Elburg and A. van Ooyen. Impact of dendritic size and
dendritic topology on burst firing in pyramidal cells. PLoS Computational Biology, 6 (5):e1000781,
2010.
Watanabe et al.(2009). H. Watanabe, T. Fujisawa, and T. W. Holstein. Cnidarians and the evolu-
tionary origin of the nervous system. Development, Growth & Differentiation, 51(3):167 -- 183, Apr
2009. doi: 10.1111/j.1440-169X.2009.01103.x. URL http://dx.doi.org/10.1111/j.1440-169X.
2009.01103.x.
28
29
Model Overview
Single excitatory
Triangular lattice on a cylinder
Nearest-neighbour
Single-compartmental
Hodgkin-Huxley sodium and potassium
A
Cell populations
Topology
Connectivity
Cell
Channels
Chemical transmission Double-exponential and conductance-based
Noise
Measurements
Simulator
Independent fixed-rate Poisson processes of spontaneous vesicle release
Wave-front orientation
NEURON (www.neuron.yale.edu)
B
Name
HHCellList HHCell
StimList
Elements
Poisson process (NetStim)
Table 1. Model Summary A: Overview
Cell and stimulus populations
Size
length ∗ circumf erence
length ∗ circumf erence
Table 2. Model Summary B: Cell and stimulus populations
C
Name
NC HHCellList
NC NS HHCellList
Connectivity
Pattern
Target
Source
HHCellList HHCellList Nearest neighbour
StimList
HHCellList One-to-one
Table 3. Model Summary C: Connectivity
D
Name
Type
Dynamics
Parameters
30
Cell Model
HHCell
Single compartmental Hodgkin-Huxley model
Cm
dVm
dt
= −Agl(Vm − El) − A¯gKn4(Vm − EK)
−A¯gN am3h(Vm − EN a) + Isyn
= −(x − x∞(Vm))/τx
dx
dt
τx = 1/(αx + βx)
with x = m, h, n
x∞ = αx/(αx + βx)
A = 400π µm2 = 1257 µm2
Cm = Acm = 12.57pF
¯gN a = 0.12 S cm−2, EN a = 50 mV
−0.1(Vm + 40)
exp(−(Vm + 40)/10) − 1)
αm(Vm) =
βm(Vm) = 4 exp(−(Vm + 65)/18)
αh(Vm) = .07 exp(−(Vm + 65)/20)
βh(Vm) = 1/(exp(−(Vm + 35)/10) + 1)
¯gK = 0.036 S cm−2,
EK = −77 mV
−0.01(Vm + 55)
exp(−(Vm + 55)/10) − 1
αn(Vm) =
βn(Vm) = .125 exp(−(Vm + 65)/80)
gl = 0.0003 S cm−2,
El = −54.3 mV
Table 4. Model Summary D: Single Cell Model
E
Type
Poisson process Fixed rate νspontaneous = 103−noiseparameter Hz generator for each cell, the noise
Noise Model
Description
parameter is varied from 2 to 6.
Table 5. Model Summary E: Noise Model
F
Name
Type
Dynamics
Parameters
Chemical Transmission Model
Exp2Syn
Double exponential conductance based
(cid:88)
pre
(cid:88)
31
G(t − tpre − dpre)
Isyn = −(Vm − Esyn)
wpre
tpre+dpre≤t
exp(−t/τdecay) − exp(−t/τrise)
exp(−tpeak/τdecay) − exp(−tpeak/τrise)
τriseτdecay
τdecay − τrise
G(t) = G0
tpeak =
G0 = 1 µS
wpre = 0.001 or 0
dpre = 0.75 ms
τrise = 0.05 ms
τdecay = 2 ms
Esyn = 0 mV
Table 6. Model Summary F: Chemical Transmission Model
Network Structure
G
Triangular lattice on cylinder. The open ends are arbitrarily labelled East and West and both edges
are aligned with the same primitive vector of the triangular lattice. Cell indices start at zero on a
cell on the West edge. Indices are incremented by one for each lattice-constant-sized step on the edge
in a direction arbitrarily marked as North. After labelling all the cells on this and subsequent rings,
indexing continues stepping North from the first cell located North East of the last labelled cell until
all cells are labelled. The number of cells on a single ring is specified by the circumference parameter,
while the number of rings is specified by the length parameter.
Table 7. Model Summary G: Network Structure
H
Wave-front orientation preference is measured by counting how often two neighbouring cell pairs of a
single orientation fire within 2 ms of each other. In figure 2 we develop several visual representations
which capture this information. In addition, spike activities are shown as snapshots capturing 4 ms of
activity and as line graphs showing wave-front size as a function of time.
Analysis
Table 8. Model Summary H: Analysis
32
Figure S. 1. Analysis of propagation direction orientation for various body shapes and noise rates.
Propagation direction orientation, averaged over a unique combination of length, circumference and
noise rate is indicated by an oriented black bar. There are 19 individual runs per unique combination of
length, circumference and noise rate. Propagation direction of a single run is indicated with an oriented
grey bar. These grey bars usually largely overlap with the black bar, indicating that these experiments
are highly reproducible. The orientation calculation is explained in figure 2. (A) Noise rate: 0.001 Hz,
(B) Noise rate: 0.01 Hz, (C) Noise rate: 0.1 Hz, (D) Noise rate: 1 Hz, (E) Noise rate: 10 Hz.
|
1802.00217 | 1 | 1802 | 2018-02-01T09:50:29 | AMPA, NMDA and GABAA receptor mediated network burst dynamics in cortical cultures in vitro | [
"q-bio.NC",
"q-bio.CB"
] | In this work we study the excitatory AMPA, and NMDA, and inhibitory GABAA receptor mediated dynamical changes in neuronal networks of neonatal rat cortex in vitro. Extracellular network-wide activity was recorded with 59 planar electrodes simultaneously under different pharmacological conditions. We analyzed the changes of overall network activity and network-wide burst frequency between baseline and AMPA receptor (AMPA-R) or NMDA receptor (NMDA-R) driven activity, as well as between the latter states and disinhibited activity. Additionally, spatiotemporal structures of pharmacologically modified bursts and recruitment of electrodes during the network bursts were studied. Our results show that AMPA-R and NMDA-R receptors have clearly distinct roles in network dynamics. AMPA-Rs are in greater charge to initiate network wide bursts. Therefore NMDA-Rs maintain the already initiated activity. GABAA receptors (GABAA-Rs) inhibit AMPA-R driven network activity more strongly than NMDA-R driven activity during the bursts. | q-bio.NC | q-bio | AMPA, NMDA AND GABAA RECEPTOR MEDIATED NETWORK BURST
DYNAMICS IN CORTICAL CULTURES IN VITRO
H. Teppola1,2, S. Okujeni2, M.-L. Linne1, and U. Egert2
1 Department of Signal Processing, Tampere University of Technology, Tampere, Finland
2 Bernstein Center Freiburg & Department of Microsystems Engineering, IMTEK, Albert-
Ludwig University of Freiburg, Freiburg, Germany
[email protected], [email protected], [email protected], [email protected]
freiburg.de
ABSTRACT
In this work we study the excitatory AMPA, and NMDA,
and inhibitory GABAA receptor mediated dynamical
changes in neuronal networks of neonatal rat cortex in
vitro. Extracellular network-wide activity was recorded
with 59 planar electrodes simultaneously under different
pharmacological conditions. We analyzed the changes of
overall network activity and network-wide burst
frequency between baseline and AMPA
receptor
(AMPA-R) or NMDA receptor (NMDA-R) driven
activity, as well as between the latter states and
disinhibited
spatiotemporal
structures of pharmacologically modified bursts and
recruitment of electrodes during the network bursts were
studied. Our results show that AMPA-R and NMDA-R
receptors have clearly distinct roles in network dynamics.
AMPA-Rs are in greater charge to initiate network wide
bursts. Therefore NMDA-Rs maintain
the already
initiated activity. GABAA receptors (GABAA-Rs) inhibit
AMPA-R driven network activity more strongly than
NMDA-R driven activity during the bursts.
activity. Additionally,
1.
INTRODUCTION
Synchronous patterns of activity, accompanied with
intracellular Ca2+-transients, are considered to play an
essential role in the development of neuronal networks in
a wide range of brain structures [4], including networks
of dissociated cortical neurons in vitro [3]. Synchronous
activity in form of network-wide bursts (NB) is thought
to be generated by recurrent excitatory pathways. The
inhibitory pathways suppressing excitation are also
considered important in network burst dynamics.
in native cortical
tissue, glutamatergic
AMPA-Rs (expressing fast ion channel kinetics) and
NMDA-Rs (expressing slow ion channel kinetics, and
both voltage and Mg2+-dependency) are
the main
mediators of excitatory synaptic transmission among
neurons in vitro. Fast inhibition is mediated through
GABAergic transmission via GABAA-Rs. Despite the
solid biophysical characterization of AMPA, NMDA,
and GABAA receptors at the monosynaptic level, their
As
in
for
complex interplay on the network level is still not fully
understood.
recordings
Simultaneous multi-unit
from
neuronal networks of an intact brain in vivo are
challenging to conduct with the existing technology.
Therefore, microelectrode array (MEA) measurement
systems have been developed
simultaneous
extracellular recordings of many neurons in in vitro
networks. These recordings allow spontaneous and
stimulus evoked NB activity to be recorded over hours of
time [5]. Previous spontaneous recordings of neural
activity at 21 days in vitro (DIV) with MEA have shown,
similarly to this work, that networks posses stable
dynamics following synapse formation and maturation.
The characteristic network-wide bursting
these
previous studies consists of around 0.5 to 1 second long
periods of simultaneous high spiking activity on many
electrodes followed by 0.5 to 20 seconds quiescence
periods [3], [6], [7].
Pharmacological studies of stimulus evoked
network activity in cortical cultures have shown that the
late component of NBs (25ms after stimulus, lasting
several hundred milliseconds) was reduced or abolished
by NMDA-R antagonist, such as APV and high Mg2+,
without affecting to the early component (first 25ms after
stimulus). When inhibitory GABAA-Rs were blocked by
bicuculline (BIC) the opposite effect was obtained, the
late component of evoked responses was increased [1].
Another study in dissociated spinal cord cultures has
shown that the duration of AMPA-R driven bursts
decreased compared to the baseline, suggesting the role
of NMDA-Rs in the maintenance of high network
activity during bursting.
In addition, NMDA-R
antagonist has been found to reduce the burst rate
indicating a contribution of NMDA-Rs in network
activation. By silencing AMPA-R at the presence of
disinhibition (BIC), the burst rate was reduced and burst
onset was slowed down and in some experiments
bursting ceased completely by AMPA-R antagonist [2].
In this work, we study for first time the
coordinated interplay between excitation and inhibition in
network dynamics in dissociated neurons of rat neonatal
cortex. First, we obtained spontaneous baseline activity
with MEA from six different densely cultured networks
at 21 DIV (Figure 1). Then, we silenced each of the
excitatory pathways by blocking either AMPA-Rs or
NMDA-Rs with NBQX or D-AP5, respectively, both in
three different cultures. Finally, we disinhibited the
networks with GABAA-R antagonist PTX. This protocol
allows us to investigate the temporal and spatial patterns
of network dynamics, produced by fast and slow
excitatory and fast inhibitory transmission.
rats'
lobes
frontal
2. MATERIALS AND METHODS
Figure 1. Left: Cultured network of cortical neurons on
MEA plate at 21 DIV. Right: Spontaneous synchronized
patterns of activity. The upper plot shows array wide
firing rate (Hz) and the raster plot at the bottom shows
the activity (spikes) recorded with 59 electrodes during a
time period of 50 seconds.
Primary cell cultures: Cortical cells were extracted
from neonatal Wistar
and
enzymatically and mechanically dissociated. 200 000
cells were seeded on 0.1% polyethyleneimine coated
microelectrode arrays (60 TiN electrodes Ø 30µM,
500µM electrode pitch, Multi Channel Systems Ltd.).
Cells were cultured in 2ml minimum essential medium
supplemented with 5% heat inactivated horse serum,
20mmol/l glucose, 1% gentamycine and 0.5mmol/l L-
glutamine in 5% CO2 humidified incubator at 37°C. One
third of the medium were changed twice a week and
recordings were done between 21-23 DIV.
Pharmacology: The following agents were used for
suppressing ionotropic receptors: 10µM D-(-)-2-Amino-
5-phosphonopentanoic acid (D-AP5) to competitively
block the glutamate site of NMDA-Rs, 10µM 2.3-Dioxo-
6-nitro-1,2,3,4
tetrahydrobenzo-[f]quinoxaline-7-
sulfonamide (NBQX) to competitively block AMPA-Rs
and 10µM Picrotoxin (PTX) to non-competitively block
GABAA-R. All drugs were applied to medium with a
pipette.
analysis:
Electrophysiological
Spontaneous network activity was recorded with MEA-
1060 amplifier (Multi Channel Systems Ltd.) in normal
culture conditions and after applying each of the drugs
for one hour. Raw signals from each electrode were
digitally high-pass filtered at 200Hz and extracellular
recordings
and
activity of neurons (spikes) were detected based on
voltage threshold, -5 times of standard deviation from
mean noise baseline of each electrode using MC-Rack
software (Multi Channel Systems Ltd.).
For further analysis the initial 600s of recorded
data after drug application were discarded and the
following 3000s were analyzed. Artifacts and electrodes
with firing rates (FR) lower than 9% of the average FR
on electrodes with spike activity were removed. NBs
were detected according to following criteria: first inter
spike interval < 100ms defines onset time of NB and first
inter spike interval > 100ms defines offset of NB, in
addition, minimum 8 electrodes should be active during
the NB.
3. RESULTS
to
We pharmacologically blocked the excitatory AMPA-R
and NMDA-R separately. Upon monitoring network
activity, we additionally blocked GABAA-R to observe
the interplay between inhibition and each excitatory
pathway.
(n=3) were compared
First, we calculated the overall network activity
(spikes/sec) and burst frequency (bursts/min) over whole
analysis period of 3000s. Recordings from networks
under AMPA-R blockage (n=3) and under NMDA-R
blockage
the baseline
recordings. As expected, overall firing rate decreased
upon suppression of either AMPA-R or NMDA-R driven
excitation and increased when networks were disinhibited
as shown in Table1. When comparing the ratio of overall
firing rate and burst frequency between baseline and
NMDA-R suppressed excitation the results showed that
the overall firing rate and burst frequency decreased by a
factor of 10 and 2, respectively. However, blocking
AMPA-Rs decreased overall firing rate only 3 fold and
burst frequency 21 fold (Table1).
inhibitory
When the excitatory pathway was driven by
AMPA-Rs
pathway was
simultaneously suppressed, the overall network activity
increased 14 fold (n=3) but burst frequency decreased 2
fold (n=2) in two out of three cultures and increased 3
fold (n=1) in one out of three cultures. Disinhibition
during NMDA-R driven activity caused overall network
activity to increase only 2 fold and burst frequency to
increase 5 fold (Table1).
Burst firing rate profiles in Figure 2 show how
the temporal structure of bursts are modified following
each drug application at single electrode level and at
average over all electrodes. AMPA-R driven bursts
became shorter, contained fewer spikes, the peak FR was
reached earlier (~20-40 ms) in the bursts and late phase
was diminished. Contrarily, NMDA-R driven bursts
became longer, contained more spikes, the peak FR was
achieved later in the course of the burst (~ 90-200 ms)
and
the
and the late phase of NBs was more pronounced.
Disinhibition increased the early and late phases of bursts
in comparison to control and AMPA-R driven bursts but,
interestingly, in comparison to NMDA-R driven bursts
the late phase squeezed after disinhibition (Figure 2).
The recruitment of electrodes during NBs was
significantly decreased during AMPA-R driven activity.
In contrast, networks were widely recruited over 200ms
time periods with disinhibition of AMPA-R driven
activity as shown in Figure 3. Approximately the same
number of electrodes was recruited when NMDA-R
driven NBs were compared to control case. However,
after disinhibition, the recruitment of electrodes a bit
increased but more importantly recruitment was much
faster than during NMDA-R driven or baseline activity,
indicating that NMDA-R driven inhibition slows down
the propagation of network-wide activity at the beginning
of bursts (Figure 3).
4. CONCLUSIONS
for
role
lowered
Our results show that AMPA-R driven NBs are initiated
more often than NMDA-R driven ones contributing over
ten times more to the initiation of network activity
compared to NMDA-Rs. By blocking NMDA-Rs, the
overall firing rate was
three times more
compared to AMPA-R blockage, which indicates the
importance of NMDA-Rs in maintaining network activity
during the bursts.
Nearly ten times higher overall firing rates were
observed during disinhibited AMPA-R driven activity
compared to the NMDA-R driven one. This indicates the
GABAA-Rs primary
inhibiting especially
AMPA-R driven network activity. The connection
between the roles of GABAA-Rs and AMPA-Rs is also
indicated by the fact that NMDA-R driven activity
displays higher firing rates compared to the control case,
where both NMDA-Rs and AMPA-R are active (Figure
2).
Contradictory behavior of both decreasing and
increasing burst frequencies during disinhibition of
AMPA-R driven activity could be explained by two
different underlying mechanisms: 1. There are more
bursts, due to lower firing threshold, to initiate AMPA-R
driven NBs when inhibition is suppressed; 2. There are
less bursts because AMPA-R driven neurons are more
excited, thus lengthening network bursts and, hence,
recovery times.
We additionally found out that disinhibition, in
general, increases the early and late phase of bursts due
to decreased threshold for excitation. Suprisingly, the late
phase of NMDA driven bursts shortened and burst
frequency
increased by disinhibition. This can be
explained by our finding that GABAA-R slows down the
recruitment of NMDA driven NBs, thus by blocking
GABAA-Rs in such a network, the recruitment of NBs
speeds up and further, the duration of bursts shortens.
Our results further indicate that the dynamics of
cortical networks in vitro incorporate a complex interplay
between excitatory and inhibitory transmission. NMDA-
R and AMPA-R driven activities are both able to initiate
network-wide bursting. However, AMPA-Rs role for
initiation is much more important. The main role of
NMDA-Rs is to maintain already initiated activity.
GABAA-Rs inhibit AMPA-R driven network excitation
more strongly than the NMDA-R driven during bursts.
However, GABAA-R mediated inhibition slows down the
propagation of NMDA-R mediated network activity
more
than AMPA-R mediated. Experimentally
investigated interplay between excitatory and inhibitory
transmission in cortical networks in vitro can further be
studied using biologically plausible computational
network models.
5. ACKNOWLEDGEMENTS
The authors would like to acknowledge the following funding:
Graduate School of Tampere University of Technology,
Tampere Graduate School
Information Science and
Engineering (TISE), Academy of Finland Grant (213462,
129657, 106030, and 124615), German BMBF
(FKZ
01GQ0830, 01GQ0420), and EC (NEURO, 12788).
in
6. REFERENCES
[1] Y. Jimbo, A. Kanawa, P. Parodi and V. Torre, "The
dynamics of a neuronal culture of dissociated cortical
neurons of neonatal rats", Biol Cybern, vol. 83, pp. 1-20,
2000.
[2] J.- C. Legrand, P. Darbon and J. Streit, "Contributions of
NMDA receptors to network recruitment and rhythm
generation in spinal cord cultures", Eur J Neurosci, vol.
19, pp. 521-532, 2004.
[3] H. P. C. Robinson, M. Kawahara, Y. Jimbo, K. Torimitsu,
Y. Kuroda and A. Kawana, "Periodic synchronized
bursting and intracellular calcium transients elicited by
low magnesium
J
Neurophys, vol. 70, pp. 1606-1616, 1993.
in cultured cortical neurons",
[4] C. J. Shatz, "Impulse activity and the patterning of
connections during CNS development", Neuron, vol. 5, pp.
745-756, 1990.
[5] Stett, U. Egert, E. Guenther, F. Hofmann, T. Meyer, W.
Nisch and H. Haemmerle, "Biological applications of
microelectrode arrays
in drug discovery and basic
research", Anal Bioanal Chem, vol. 377, pp. 486-495,
2003.
[6] D. A. Wagenaar, J. Pine and S. M. Potter, "An extremely
rich repertoire of bursting patterns during the development
of cortical cultures", BMC Neurosci vol.7, no. 11, 2006.
[7] D. Eytan, and S. Marom,"Dynamics and effective
topology underlying synchronization
in networks of
cortical neurons", J Neurosci, vol. 26, no. 33, pp. 8465-
8476, 2006.
Figure 2. Burst firing rate profiles. Upper row corresponds to AMPA driven experiment and lower row to NMDA
driven experiment. Columns from left: firing rate profiles during control, suppression of excitation, suppression of
inhibition, and the average profiles for all three cases. D-AP5, NBQX, and PTX are the pharmacological agents for
suppressing NMDA-Rs, AMPA-Rs, and GABAA-Rs respectively. In the first three columns one single thin line
corresponds to firing rate in one electrode averaged over all bursts (aligned to NB onsets). In all columns, thick lines
correspond to overall average burst firing rates over all electrodes and over all bursts (bin width 30ms).
Figure 3. Network burst electrode fraction curves. Upper row corresponds to AMPA driven experiment and
lower row to NMDA driven experiment. Columns from left: Percentage of the electrodes attending to the NBs for
control, suppression of excitation, suppression of inhibition, and the average curves for all three cases. D-AP5,
NBQX, and PTX are the pharmacological agents for suppressing NMDA-Rs, AMPA-Rs, and GABAA-Rs
respectively. In the first three columns, one single thin line corresponds to fraction of attending electrodes in one
network burst. In all columns, thick lines correspond to average fraction of electrodes over all network bursts. In the
most right column, shaded areas show the standard error of mean (bin width 100ms).
Pharmacological
condition
AMPA driven
activity
NMDA driven
activity
Disinhibition
for AMPA
driven activity
14.1±2.3(n=3)
Disinhibition
for NMDA
driven activity
1.8±0.5 (n=3)
-
4.6±2.0 (n=3)
Increase in overall firing rate
Decrease in overall firing rate 9.8±2.9 (n=3)
Increase in burst frequency
Decrease in burst frequency
-
2.2±0.7 (n=3)
-
-
3.2±1.7 (n=3) -
-
21.3±9.2(n=3)
3.2 (n=1)
2.4±1.3 (n=2) -
Table1. Computed changes in overall firing rate (spikes/s) and burst frequency (bursts/min) compared to the
previous pharmacological condition (mean ± standard deviation). Disinhibition chronologically follows the
associated AMPA-R driven or NMDA-R driven activity.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.